[Senate Hearing 115-649]
[From the U.S. Government Publishing Office]


                                                        S. Hrg. 115-649

                        DIGITAL DECISION-MAKING:
  THE BUILDING BLOCKS OF MACHINE LEARNING AND ARTIFICIAL INTELLIGENCE

=======================================================================

                                HEARING

                               BEFORE THE

    SUBCOMMITTEE ON COMMUNICATIONS, TECHNOLOGY, INNOVATION, AND THE 
                                INTERNET

                                 OF THE

                         COMMITTEE ON COMMERCE,
                      SCIENCE, AND TRANSPORTATION
                          UNITED STATES SENATE

                     ONE HUNDRED FIFTEENTH CONGRESS

                             FIRST SESSION
                               __________

                           DECEMBER 12, 2017
                               __________

    Printed for the use of the Committee on Commerce, Science, and 
                             Transportation

                  [GRAPHIC NOT AVAILABLE IN TIFF FORMAT]
                  
                Available online: http://www.govinfo.gov
                
                              ___________

                    U.S. GOVERNMENT PUBLISHING OFFICE
                    
37-295 PDF                  WASHINGTON : 2019                
       
       
       SENATE COMMITTEE ON COMMERCE, SCIENCE, AND TRANSPORTATION

                     ONE HUNDRED FIFTEENTH CONGRESS

                             FIRST SESSION

                   JOHN THUNE, South Dakota, Chairman
ROGER F. WICKER, Mississippi         BILL NELSON, Florida, Ranking
ROY BLUNT, Missouri                  MARIA CANTWELL, Washington
TED CRUZ, Texas                      AMY KLOBUCHAR, Minnesota
DEB FISCHER, Nebraska                RICHARD BLUMENTHAL, Connecticut
JERRY MORAN, Kansas                  BRIAN SCHATZ, Hawaii
DAN SULLIVAN, Alaska                 EDWARD MARKEY, Massachusetts
DEAN HELLER, Nevada                  CORY BOOKER, New Jersey
JAMES INHOFE, Oklahoma               TOM UDALL, New Mexico
MIKE LEE, Utah                       GARY PETERS, Michigan
RON JOHNSON, Wisconsin               TAMMY BALDWIN, Wisconsin
SHELLEY MOORE CAPITO, West Virginia  TAMMY DUCKWORTH, Illinois
CORY GARDNER, Colorado               MAGGIE HASSAN, New Hampshire
TODD YOUNG, Indiana                  CATHERINE CORTEZ MASTO, Nevada
                       Nick Rossi, Staff Director
                 Adrian Arnakis, Deputy Staff Director
                    Jason Van Beek, General Counsel
                 Kim Lipsky, Democratic Staff Director
              Chris Day, Democratic Deputy Staff Director
                      Renae Black, Senior Counsel
                                 ------                                

    SUBCOMMITTEE ON COMMUNICATIONS, TECHNOLOGY, INNOVATION, AND THE 
                                INTERNET

ROGER F. WICKER, Mississippi,        BRIAN SCHATZ, Hawaii, Ranking
    Chairman                         MARIA CANTWELL, Washington
ROY BLUNT, Missouri                  AMY KLOBUCHAR, Minnesota
TED CRUZ, Texas                      RICHARD BLUMENTHAL, Connecticut
DEB FISCHER, Nebraska                EDWARD MARKEY, Massachusetts
JERRY MORAN, Kansas                  CORY BOOKER, New Jersey
DAN SULLIVAN, Alaska                 TOM UDALL, New Mexico
DEAN HELLER, Nevada                  GARY PETERS, Michigan
JAMES INHOFE, Oklahoma               TAMMY BALDWIN, Wisconsin
MIKE LEE, Utah                       TAMMY DUCKWORTH, Illinois
RON JOHNSON, Wisconsin               MAGGIE HASSAN, New Hampshire
SHELLEY CAPITO, West Virginia        CATHERINE CORTEZ MASTO, Nevada
CORY GARDNER, Colorado
TODD YOUNG, Indiana
                            C O N T E N T S

                              ----------                              
                                                                   Page
Hearing held on December 12, 2017................................     1
Statement of Senator Wicker......................................     1
    Letter dated December 11, 2017 to Hon. Roger Wicker and Hon. 
      Brian Schatz from Dean Garfield, President and CEO, 
      Information Technology Industry Council (ITI)..............    83
    Letter dated December 12, 2017 to Hon. John Thune and Hon. 
      Bill Nelson from Marc Rotenberg, President, EPIC; Caitriona 
      Fitzgerald, Policy Director, EPIC; and Christine Bannan, 
      Policy Fellow, EPIC........................................    84
Statement of Senator Schatz......................................     2
Statement of Senator Moran.......................................    41
Statement of Senator Peters......................................    44
Statement of Senator Udall.......................................    46
Statement of Senator Young.......................................    48
Statement of Senator Cantwell....................................    50
Statement of Senator Markey......................................    55
Statement of Senator Cruz........................................    56
Statement of Senator Cortez Masto................................    58
Statement of Senator Blumenthal..................................    60

                               Witnesses

Dr. Cindy L. Bethel, Associate Professor, Department of Computer 
  Science and Engineering, Mississippi State University..........     4
    Prepared statement...........................................     5
Daniel Castro, Vice President, Information Technology and 
  Innovation Foundation (ITIF)...................................     8
    Prepared statement...........................................    10
Victoria Espinel, President and CEO, BSA  The Software 
  Alliance.......................................................    17
    Prepared statement...........................................    18
    Report entitled ``The $1 Trillion Economic Impact of Software    63
Dr. Dario Gil, Ph.D., Vice President, AI and IBM Q...............    26
    Prepared statement...........................................    27
Dr. Edward W. Felten, Ph.D., Robert E. Kahn Professor of Computer 
  Science and Public Affairs, Princeton University...............    32
    Prepared statement...........................................    34

                                Appendix

Response to written questions submitted to Dr. Cindy L. Bethel 
  by:
    2Hon. Amy Klobuchar..........................................    87
    Hon. Tom Udall...............................................    87
    Hon. Maggie Hassan...........................................    88
Response to written questions submitted to Daniel Castro by:
    Hon. Tom Udall...............................................    90
    Hon. Gary Peters.............................................    90
    Hon. Maggie Hassan...........................................    91
Response to written questions submitted to Victoria Espinel by:
    Hon. Gary Peters.............................................    92
    Hon. Maggie Hassan...........................................    94
Response to written questions submitted to Dr. Dario Gil, Ph.D. 
  by:
    Hon. Amy Klobuchar...........................................    96
    Hon. Tom Udall...............................................    97
    Hon. Gary Peters.............................................    98
    Hon. Maggie Hassan...........................................    99
Response to written questions submitted to Dr. Edward W. Felten, 
  Ph.D. by:
    Hon. Amy Klobuchar...........................................   102
    Hon. Tom Udall...............................................   103
    Hon. Gary Peters.............................................   103
    Hon. Maggie Hassan...........................................   104

 
                        DIGITAL DECISION-MAKING:
 THE BUILDING BLOCKS OF MACHINE LEARNING AND ARTIFICIAL INTELLIGENCE

                              ----------                              


                       TUESDAY, DECEMBER 12, 2017

                               U.S. Senate,
       Subcommittee on Communications, Technology, 
                      Innovation, and the Internet,
        Committee on Commerce, Science, and Transportation,
                                                    Washington, DC.
    The Subcommittee met, pursuant to notice, at 10 a.m. in 
room SR-253, Russell Senate Office Building, Hon. Roger Wicker, 
Chairman of the Subcommittee, presiding.
    Present: Senators Wicker [presiding], Schatz, Blunt, Cruz, 
Fischer, Moran, Sullivan, Heller, Inhofe, Capito, Young, 
Cantwell, Klobuchar, Blumenthal, Markey, Booker, Udall, Peters, 
Hassan, and Cortez Masto.

          OPENING STATEMENT OF HON. ROGER F. WICKER, 
                 U.S. SENATOR FROM MISSISSIPPI

    Senator Wicker. This hearing will come to order. Senator 
Schatz will be here in a few moments and has sent word that we 
should go ahead and proceed. Today's Subcommittee meets to 
examine the commercial applications of artificial intelligence 
and machine learning for the U.S. economy. We are also gathered 
to discuss how the responsible design and deployment of 
intelligent systems can foster innovation and investment, 
propelling the United States as a leader in artificial 
intelligence.
    I'm glad to convene this hearing, and as I mentioned, my 
colleague and friend Senator Schatz will be here in a moment.
    Artificial intelligence refers to technology that is 
capable of taking on human-like intelligence. Through data 
inputs and algorithms, AI systems have the potential to learn, 
to reason, to plan, perceive, process, make decisions, and even 
act for themselves.
    Although AI applications have been around for decades, 
recent advancements, particularly in machine learning, have 
accelerated in their capabilities because of the massive growth 
in data gathered from billions of connected devices and the 
digitization of everything. Developments in computer processing 
technologies and better algorithms are also enabling AI systems 
to become smarter and perform more unique tasks.
    Every day consumers use technology that employ some degree 
of AI, smartphone mapping apps that suggest faster driving 
routes, for example. Online search platforms are learning from 
past queries to generate increasingly customized results for 
users. We all know that when we click on a site on the 
Internet, news suggestions and advertisements on social media, 
and semi-autonomous vehicles are just a few examples of how 
machines and computer programs are taking on increasingly 
cognitive tasks.
    The excitement surrounding this technology is deserved. AI 
has the potential to transform our economy, and so let's talk 
about that today. AI's ability to process and sort through 
troves of data can greatly inform human decisionmaking and 
processes across industries, including agriculture, health 
care, and transportation. In turn, businesses can be more 
productive, profitable, and efficient in their operations.
    As AI systems mature and become more accurate in their 
descriptive, predictive, and prescriptive capabilities, there 
are issues that should be addressed to ensure the responsible 
development and use of this technology. Some of these issues 
include: understanding how data is gathered, what data is 
provided for an intelligent machine to analyze, and how 
algorithms are programmed by humans to make certain 
predictions. Moreover, understanding how the human end user 
interacts with or responds to the digital decision, and how 
humans interpret or explain decisions of the AI system over 
time will also need to be addressed.
    These are important considerations to ensure that the 
decisions made by AI systems are based on representative data 
that does not unintentionally harm vulnerable populations or 
act in an unsafe, anticompetitive, or biased way. So there's a 
lot to think about.
    In addition to these issues, other considerations, such as 
data privacy and cybersecurity, AI's impact on the workforce, 
and human control and oversight over intelligence systems 
should also be addressed as the technology develops. 
Fundamental to the success of machine learning and AI in 
enhancing U.S. productivity and empowering human decisionmaking 
is consumer confidence and trust in these systems. To build 
consumer confidence and trust, it is critical that the 
integration of AI into our commercial and government processes 
be done responsibly.
    To that end, I look forward to learning from today's 
witnesses about how AI is advancing in our economy and what 
best practices our industry and AI researchers are considering 
to achieve all the promised economic and societal benefits of 
this technology.
    Senator Schatz, do you have anything to add to my 
comprehensive opening statement?

                STATEMENT OF HON. BRIAN SCHATZ, 
                    U.S. SENATOR FROM HAWAII

    Senator Schatz. I think everything has been said, but not 
everybody has said it. So good morning. Thank you very much, 
Mr. Chairman.
    AI is advancing fast. Each year, processing power gets 
better, hardware gets cheaper, and algorithms are easier to 
train thanks to bigger and better datasets. These advances are, 
for companies and economies, great opportunities.
    Technologists, historians, and economists say that we're at 
the cusp of the next industrial revolution, but there are 
concerns. We've seen that AI can be a black box. It can make 
decisions and come to conclusions without showing its 
reasoning. There are also known cases of algorithms that 
discriminate against minority groups, and when you start to 
apply these systems to criminal justice, health care, or 
defense, the lack of transparency and accountability is 
worrisome.
    Given the many concerns in a field that's advancing so 
quickly and is so revolutionary, it's hard to believe that 
there is no AI policy at the Federal level, and that needs to 
change. To start, the government should not purchase or use AI 
systems if we can't explain what it does, especially if these 
systems are making decisions about our citizens' lives.
    There also needs to be more transparency and consumer 
control on data collection. Too many consumers still do not 
know what data is being collected, how it's being collected, or 
who owns it. Some of our current laws and regulations work, but 
some of them are too old and outdated to be used as a strong 
foundation for AI.
    For example, companies often use data scraping to build 
their AI models. This falls under the Computer Fraud and Abuse 
Act, a 1986 law that was written before the Web was really in 
operation. AI is now used to write news articles, edit 
photographs, artificially reconstruct movies, all actions that 
fall under the Digital Millennium Copyright Act, which was 
passed in 1998, 10 years before the iPhone.
    Our laws used to apply to actions in the physical world, 
but now they apply to software systems that ultimately do the 
same thing. I'm glad to see that industry and academia are 
being proactive by coming up with policy, principles, and 
professional ethics codes, but with this kind of patchwork, the 
system is only as strong as its weakest link.
    From the private sector to academia to government, everyone 
has to wrestle with the ethical and policy questions that AI 
raises. And that's why I intend to introduce a bill creating an 
independent Federal commission to ensure that AI is adopted in 
the best interest of the public. If created, the commission 
would serve as a resource and coordinating body to the 16-plus 
Federal agencies that govern the use of AI. Otherwise, we risk 
bureaucratic turf wars and outdated inconsistent rules.
    The commission would also be tasked with asking the tough 
ethical questions around AI. For instance, employers may not 
legally ask interviewees about their religion, marital status, 
or race, but they use software that mines social media data 
that may make the same inferences. Judges cannot base their 
sentencing decisions on whether the defendant has family 
members in jail, yet these facts contribute to a risk score 
calculated by machine learning algorithms.
    In these cases, it's not clear yet where we draw the line 
on what is legal and what is not. In some instances, existing 
statutes will suffice. A lot of our laws actually work just 
fine when it comes to AI, and great laws can survive the test 
of time. But there are a few things that need to be wrestled 
with today in Congress and with our agencies, and that's why 
this hearing is so important.
    Thank you very much. I look forward to hearing the 
testimony.
    Senator Wicker. Thank you, Senator Schatz.
    Our witnesses today are Dr. Cindy Bethel, Associate 
Professor, Department of Computer Science and Engineering at 
Mississippi State University; Mr. Daniel Castro, Vice 
President, Information Technology and Innovation Foundation; 
Ms. Victoria Espinel, Chief Executive Officer, BSA-The Software 
Alliance; Dr. Dario Gil, Vice President, IBM Research, AI, and 
IBM Q; and Dr. Edward Felten, Robert E. Kahn Professor of 
Computer Science and Public Affairs, Princeton University.
    Friends, we will begin on my left with Dr. Bethel and 
proceed with 5-minute opening statements down the table. Thank 
you very much.
    Dr. Bethel.

          STATEMENT OF DR. CINDY L. BETHEL, ASSOCIATE

           PROFESSOR, DEPARTMENT OF COMPUTER SCIENCE

         AND ENGINEERING, MISSISSIPPI STATE UNIVERSITY

    Dr. Bethel. Good morning, Chairman Wicker, Ranking Member 
Schatz, and members of the Committee. Thank you for the 
opportunity to appear before you today. I'm Dr. Cindy Bethel. 
I'm Associate Professor of Computer Science and Engineering at 
Mississippi State University, or MSU.
    It is an honor to speak with the Committee today about 
digital decisionmaking associated with artificial intelligence, 
known as AI, from the academic perspective, and about 
applications of AI being developed at MSU.
    Today I will address three primary areas: first, a brief 
introduction to AI; next, three AI projects being developed at 
MSU; and last I will discuss some key points associated with 
AI.
    A critical aspect associated with the advancement of 
science and technology is the development of algorithms with AI 
including machine learning for digital decisionmaking. In order 
for a system to make a decision, it must first acquire 
information, process and learn from that information, and then 
use that information to make decisions.
    The gold standard would be the ability for a system to make 
a decision in a manner similar to a human expert in that area. 
This is a relatively new field of science that will be--provide 
ongoing research opportunities, including efforts to enhance 
existing algorithms and develop new and more efficient and 
effective algorithms. This is critical to the overall 
advancement of science in various disciplines, such as 
robotics, medicine, economics, and others.
    There are countless application areas for which AI is 
beneficial and may make a significant impact on society. At 
MSU, we are on the forefront of AI developments with our 
research efforts, and today I will focus on three of those--
excuse me--projects.
    First is the integration of robots into high-risk law 
enforcement operations, such as SWAT teams, who I've trained 
with monthly for the last 6 years. This is an application for 
highly dynamic situations that require officers to make life-
critical decisions. The more information they have prior to 
entering these dangerous and unknown environments, the better 
decisions the officers can make.
    We are developing algorithms using robots into an 
environment prior to entry to provide critical audio and video 
intelligence to allow officers to make more informed decisions. 
This information continues the dynamics of how they make entry 
or process a scene. Our research through intelligent interfaces 
will help inform how this information is delivered to the 
officers and maximize safety and performance.
    Second, Mississippi State University is researching the 
development of autonomous cargo transport systems used in the 
fast-paced dynamic environment of a top 100 logistics company. 
This involves choosing the proper sensors to ensure that data 
is able to make--help make informed decisions on the vehicle's 
path.
    Further, we are researching a variety of human factors 
because controlled vehicles will transfer from autonomous to 
human driver control and back again. Our work shows that if a 
human is not actively engaged in driving, they may--there may 
be insufficient situational awareness to control--take control 
on short notice. So how the human driver will be notified of 
the transfer of control is another critical aspect of our work.
    To ensure maximum safety, researchers at MSU are exploring 
situations in which humans are operating in close proximity to 
the autonomous vehicles, such as a vehicle that's docking to 
deliver cargo, and a worker is unloading and loading this 
cargo.
    Finally, another one of MSU's AI research projects involves 
a robotic therapy support system, known as Therabot that is in 
the form of a stuffed robotic dog. Therabot is an alternative 
to animal-assisted therapy for people who may be allergic to 
animals or may not be able to care for a live animal. Therabot 
will be used to provide support during clinical therapy 
sessions and home therapy practice with children or adults who 
are dealing with post-traumatic stress disorders or other 
mental health concerns. The algorithms being developed modify 
the autonomous behaviors of the robot based on the interactions 
with the human to accommodate his or her preferences and in 
response to different levels of stress detected by the robot, 
and to provide improved comfort and support.
    AI will only be as good as the data the system receives and 
its ability to process that information to make a decision. The 
better the quality and quantity of information available to the 
system, the better the results will be from the machine 
learning process, which results in better final decisions from 
the system. Otherwise, decisionmaking capabilities can be 
limited or inaccurate.
    The potential applications of AI are almost limitless. The 
United States can and should remain at the forefront of AI 
research, application development, and innovation provided that 
government proceeds with a light regulatory touch that doesn't 
stifle this potential.
    Thank you so much for the opportunity to testify today on 
these important topics. I appreciate your time and attention to 
the advancement and impact of digital decisionmaking and AI.
    [The prepared statement of Dr. Bethel follows:]

   Prepared Statement of Dr. Cindy L. Bethel, Associate Professor of 
Computer Science and Engineering; Director of the Social, Therapeutic, 
 and Robotic Systems (STaRS) Lab; Billie J. Ball Endowed Professor in 
               Engineering; Mississippi State University
    Chairman Wicker, Ranking Member Schatz, and Members of the 
Committee, thank you for the opportunity to appear before you today. I 
am Dr. Cindy Bethel, an Associate Professor of Computer Science and 
Engineering, the Billie J. Ball Endowed Professor in Engineering, and 
the Director of the Social, Therapeutic, and Robotic Systems (STaRS) 
Lab at Mississippi State University. It is an honor to speak with the 
committee today about digital decision-making associated with 
artificial intelligence (AI) from the academic perspective and about 
applications of AI being developed at Mississippi State University.
    A critical aspect associated with the advancement of science and 
technology is the development of algorithms associated with AI 
including machine learning for digital decision-making. In order for a 
system to make a decision it must acquire information, have a means of 
processing and learning from that information, and have the ability to 
use that information to make an informed decision. The gold standard 
would be the ability for a machine or system to make a decision in a 
manner similar to that of a human who is an expert in that area. We 
have made considerable progress toward this goal since the inception of 
what is considered artificial intelligence, which began in 1943 with 
the research performed by McCulloch and Pitts. Today many machines and 
systems rely upon the use of artificial intelligence for digital 
decision-making. This is a relatively new field of science that will 
provide many lifetimes of research including efforts to enhance 
existing algorithms and develop new, more efficient, and effective 
algorithms. This is critical to the overall advancements of science in 
many disciplines, such as robotics, medicine, economics, and many 
others. There are often disagreements in the field as to what is 
considered AI and what algorithms and techniques used for learning are 
considered the best.
    There are many application areas for which artificial intelligence 
is beneficial and may make a significant impact on society. At 
Mississippi State University, we are actively conducting AI research 
with several research projects that use AI and machine learning 
techniques, but I will focus on three primary projects.
    The first project is the integration of robots into law enforcement 
operations, especially high risk, life critical incident responses such 
as those used with special weapons and tactics (SWAT) teams. I have 
been training monthly with rural regional SWAT team members since 2011. 
This is an example of an application in which high risk, dynamic 
situations are encountered, that require often officers to make life-
critical decisions. The more information or intelligence they have 
prior to entering these dangerous and unknown environments, the better 
decisions the officers can make. We are investigating and developing 
algorithms related to computer vision, sensor fusion, and scene 
understanding to send a robot in prior to entry to provide audio and 
video feedback to officers during a response, highlighting what is 
critical information for them to attend to so that they are not 
overwhelmed with information when under high stress. The algorithms 
identify what is important to the officers in the environment, such as 
children, weapons, and other possible threats. This information can 
change the dynamics of how they make entry or process the scene. We are 
also researching in what ways this information needs to be provided to 
the officers, so that they can use it to their advantage to keep them 
and others safer in the performance of their duties. For example if 
officers are conducting a slow and methodical search of a building for 
a suspect, and the environment is quiet, dark and threat is nearby, 
they would not want to receive the information in a openly visual 
manner such as a video stream on a mobile phone that would highlight 
them in the environment and make them more likely to be the target of 
harm. In this case, they may want a verbal description of the scene 
that comes across on their radio earpiece. If they are in a gunfight or 
if there is an alarm that is sounding in the environment, but they are 
in a relatively ``safe'' location, then they may want to receive this 
information in a visual form, because audio transmission would be 
difficult to hear. We are researching the development of intelligent 
interface switching in which the manner that information is delivered 
to the officers may change depending on what is happening in the 
environment they are operating in. The officers are excited and ready 
to start deploying some of these artificial intelligence and machine 
learning applications in real-world responses.
    A second project that we are working on at MSU is the development 
of autonomous cargo transport systems to be used in the fast-paced, 
dynamic environment of a top 100 logistics company. A primary factor 
that needs to be considered is what sensors need to be used to make 
informed decisions on the path the vehicles must travel. We also need 
to consider humans, who are sitting in the driver's seat of these 
vehicles, because control of the vehicles will change between fully 
autonomous to human driver operated. Research has shown that if the 
human is not actively involved in the activity of driving, that there 
may not be adequate situation awareness to be able to take back control 
of the vehicle with relatively short notice if needed. This has 
potential for life critical decision-making. We are investigating how 
the system needs to be able to alert the driver that control is being 
transferred, either to the vehicle or back to the human driver.
    It is also important to consider the types of notifications that 
need to occur to ensure safety and situational awareness of what is 
happening around the vehicle. Also there are situations in which humans 
are operating in close proximity to the autonomous vehicle. As human 
drivers, we observe consciously or unconsciously behaviors of other 
drivers to infer what nearby vehicles will do next, but if there are 
not those cues, then how do humans in the environment understand what 
the vehicle or system will do next? This is a major issue of concern. 
This occurs also when the vehicle is docking to deliver cargo and a 
human is involved in unloading and loading this cargo. We are exploring 
methods of notification to the person about what the vehicle will do 
next. This is of significant concern, and if the incorrect decision is 
made, the vehicle could cause harm to the human.
    The third project involves a robotic therapy support system, known 
as TherabotTM that is in the form of a stuffed robotic dog. 
TherabotTM is an alternative to animal-assisted therapy, for 
people who may be allergic to animals or may not be able to care for a 
live animal. TherabotTM will be used to provide support 
during clinical therapy sessions and for home therapy practice with 
children or adults who are dealing with post-traumatic stress disorders 
and other mental health concerns. The algorithms being developed for 
this project modify the autonomous behaviors and responses of the robot 
based on the interactions with the human to accommodate his or her 
preferences and in response to different levels of stress detected. 
Machine learning is being used to understand what behaviors the user 
prefers and to provide better support. It allows TherabotTM 
to be customizable to each individual user. It will learn and respond 
to each user as a dog would each person it encounters. Currently, 
TherabotTM can detect different types of touch, such as 
petting, patting, and hugging and will respond in different ways. If a 
person is under stress during the interaction he or she may squeeze the 
robot and the robot will adapt its behaviors to provide more comfort 
and support.
    Algorithms used in the research and development of systems that are 
capable of digital decision-making are being developed and enhanced by 
researchers all across the world. Mississippi State University is at 
the forefront of these research developments and is continually 
contributing through publications and sharing of knowledge, algorithms, 
and software developments.
    Research as a whole involves exploring what others have performed 
and then determining if there are modifications that can be made to 
improve upon those algorithms or the development of new algorithms to 
meet the needs of the application of use. For example, algorithms are 
being developed to learn a user's preference for how close a robot 
stands to them and still feel comfortable, or which friend they like 
most on a social media account. There are many methods for solving a 
problem. I typically tell my students who are interested in pursuing 
research that ``you need to be first or best at something and it is 
always best to be first.''
    Artificial intelligence will only be as good as the data the system 
receives and its ability to process that information to make a 
decision. This is a critical aspect to the advancement of this field 
that can impact almost any other discipline. The system or machine must 
have the ability to perceive information and that typically comes from 
sensors and other forms of data. There are many types of sensing 
systems such as a camera, streaming video, thermal images based on the 
heat signature of items in the environment, radar, and many others. 
There are also methods of gathering information from sources such as 
social media, mobile device location history, purchase history, product 
preferences, websites visited, etc., that can assist in the decision-
making process. The better the quality and quantity of information 
available to the system, the better the results will be from the 
machine learning process, which results in a better final decision from 
the system.
    Algorithms are programmed to receive data as an input, process that 
data, learn from large amounts of data, and then use that information 
to make a digital decision. If there is not sufficient amounts of data 
available to train the machine learning algorithms or enough diversity 
in the data to allow the learning algorithms to adapt to different 
aspects, then the decision-making capabilities can be limited or 
inaccurate.
    Another major issue of concern is the processing power necessary to 
handle large amounts of information received and come to a decision. 
This is especially a concern on smaller systems such as robots that 
have limited onboard processing capabilities. Many of the AI problems 
that are being addressed in the research community are performed on 
high powered computing resources and simulations are performed to 
validate the results. This is fine for many scientific applications, 
but in order for AI and machine learning to be beneficial in real-world 
applications, it will be necessary to perform the decision-making 
processes in real-time. The results need to be made available in an 
instant and not have to wait for processing time to provide a result. 
This is improving, but sensing and processing are currently significant 
limitations to the application and use of AI in digital decision-
making.
    The level of human engagement necessary for digital decision-making 
depends on the state of the AI system. There are different levels of 
autonomous decision making. There is full autonomy, where the system 
receives the input and then processes the information from that data 
and makes a decision with no input from a human. There is supervised 
autonomy, in which the system receives information, processes the data, 
and comes up with possible results, and the human may have the ability 
to override or make the final decision. The more common level of 
autonomy is supervised autonomy. The level of human engagement also 
needs to consider the ramifications of the decision-making process. If 
it is a life-critical decision, then most people are more comfortable 
with a human remaining involved in the process to ensure an ethical and 
``good'' decision is the final result. There are many ethical hurdles 
that will need to be decided at some point as to who is responsible if 
an AI system makes an incorrect decision, especially if the decision 
could result in harm to humans. There have been discussions in the 
field among researchers regarding who is responsible for this decision, 
such as the programmer, the company that made the system, or others. 
The current state often requires a human to be involved at some level 
of the final decision-making process unless it is low risk or well 
validated that the system will always make a ``right'' decision.
    The fields of artificial intelligence and machine learning are at 
such an early stage of scientific development that standards and best 
practices are discussed among researchers; however, there is not a 
single standard or set of best practices that I am aware of that all 
researchers and scientists follow at this point. The biggest concern is 
providing the best possible answers that will not result in harm and 
provides benefits to the users.
    The algorithms developed for machine learning and artificial 
intelligence can be used in almost any area of research, development, 
and discipline. It can be used to improve the decision-making process 
of humans. The processing of some information by a computer can be 
faster than what can be achieved by a human brain. Almost any aspect of 
society can benefit from the use of high quality artificial 
intelligence capabilities.
    A critical aspect in the development of artificial intelligence 
using machine learning and other techniques is the impact on the humans 
that are involved. A current limitation to the advancement of 
artificial intelligence is the quality and cost effectiveness of 
sensing capabilities to provide high quality information or data to the 
system to make those digital decisions. Another critical limitation to 
current artificial intelligence capabilities is onboard processing 
capabilities and the cost effectiveness of those systems. We have come 
a long way in the advancement of artificial intelligence; however we 
still have a long way to go! The potential applications of AI are 
almost limitless. The United States can and should remain at the 
forefront of AI research, application development, and innovation 
provided that the government proceeds with a light regulatory touch 
that doesn't stifle this potential.
    Thank you so much for the opportunity to testify today on these 
important topics. I appreciate your time and attention to the 
advancement and impacts of digital decision-making and artificial 
intelligence.

    Senator Wicker. Thank you, Dr. Bethel. Precisely 5 minutes.
    Dr. Bethel. Thank you.
    [Laughter.]
    Senator Wicker. Mr. Castro, we're delighted to have you.

          STATEMENT OF DANIEL CASTRO, VICE PRESIDENT,

    INFORMATION TECHNOLOGY AND INNOVATION FOUNDATION (ITIF)

    Mr. Castro. Thank you. Chairman Wicker, Ranking Member 
Schatz, and members of the Committee, I appreciate the 
invitation to be here today.
    AI has the potential to create a substantial and lasting 
impact on the economy by increasing the level of automation in 
virtually every sector, leading to more efficient processes and 
higher quality outputs, and boosting productivity in per capita 
incomes.
    In the coming years, AI is expected to generate trillions 
of dollars of economic value and help businesses make smarter 
decisions, develop innovative products and services, and boost 
productivity. For example, manufacturers are using AI to invent 
new metal alloys for 3D printing, pharmaceutical companies are 
using AI to discover new lifesaving drugs, and agricultural 
businesses are using AI to increase automation on farms.
    Companies that use AI will have an enormous advantage 
compared to their peers that do not; therefore, the United 
States should prioritize policy initiatives that promote AI 
adoption in its traded sectors where U.S. firms will face 
international competition.
    Many other countries already see the strategic importance 
of becoming lead adopters of AI, and they have begun 
implementing policies to pursue this goal. For example, this 
past March, Canada launched the Pan-Canadian AI Strategy, which 
is intended to help Canada become an international leader in AI 
research. The U.K.'s new budget, which was published last 
month, includes several provisions that have the goal of making 
the U.K. a world leader in AI, including establishing a new 
research center and funding about 500 Ph.D.'s. Japan has 
created an AI technology strategy designed to develop and 
commercialize AI in a number of fields, including 
transportation and health care. And China has declared its 
intent to be the world's premier AI innovation center by 2030.
    However, to date, the U.S. Government has not declared its 
intent to be a global leader in this field, nor has it begun 
the even harder task of developing a strategy to achieve that 
vision. Moreover, China, which has launched this ambitious 
program to dominate the field, has already surpassed the United 
States in terms of the total number of papers published and 
cited in some AI disciplines, such as deep learning.
    The U.S. should not cede its existing advantages in AI. 
Instead, it should pursue a multipronged national strategy to 
remain competitive in this field. First, the Federal Government 
should continue to expand its funding to support strategic 
areas of AI, especially in areas industry is unlikely to invest 
in, as well as better plan and coordinate Federal funding for 
AI R&D across different agencies.
    Second, the Federal Government should support educational 
efforts to ensure a strong pipeline of talent to create the 
next generation of AI researchers and developers, including the 
retraining and diversity programs as well as pursue integration 
policies that allow U.S. businesses to recruit and retain 
highly skilled computer scientists.
    Third, Federal and state regulators should conduct 
regulatory reviews to identify regulatory barriers to 
commercial use of AI in various industries, such as 
transportation, health care, education, and finance.
    Fourth, the Federal Government should continue to supply 
high-value datasets that enable advances in AI, such as 
providing open access to standardized reference datasets for 
text analysis and facial recognition. Federal agencies should 
also facilitate data-sharing between industry stakeholders just 
as the Department of Transportation has done on safety for 
autonomous vehicles.
    And, fifth, the Federal Government should assess what types 
of economic data it needs to gather from businesses to monitor 
and evaluate AI adoption, much like it tracked rural 
electrification or broadband connectivity as key economic 
indicators.
    Now, as with any technology, there will be some risks and 
challenges associated with AI that require government 
oversight, but the U.S. should not replicate the European 
approach to AI, where rules creating a right to explanation and 
a right to human review for automated decisions risk severely 
curtailing the uses of AI.
    Instead, the U.S. should create its own innovation-friendly 
approach to providing oversight of the emerging algorithmic 
economy just as it has for the Internet economy. Such an 
approach should prioritize sector-specific policies over 
comprehensive regulation, outcomes over transparency, and 
enforcement actions against firms that cause tangible harm over 
those that merely make missteps without injury.
    In many cases, regulators will not need to intervene 
because the private sector will address problems about AI, such 
as bias or discrimination, on its own. Moreover, given that 
U.S. companies are at the forefront of efforts to build AI that 
is safe and ethical, maintaining U.S. leadership in this field 
will be important to ensure these values remain embedded in 
this technology.
    AI is a transformational technology that has the potential 
to significantly increase efficiency and innovation across the 
U.S. economy, creating higher living standards and improve 
quality of life. But while the United States has an early 
advantage in AI, many other countries are trying to be number 
one--they're trying to be number one. We need more leadership 
on this issue. And I look forward to working with any member of 
the Committee on their proposed legislation and new ideas in 
this space. And I commend you all for holding this hearing.
    Thank you for the opportunity to be here today. And I look 
forward to the questions.
    [The prepared statement of Mr. Castro follows:]

         Prepared Statement of Daniel Castro, Vice President, 
        Information Technology and Innovation Foundation (ITIF)
Introduction
    Chairman Wicker, Ranking Member Schatz and members of the 
subcommittee, I appreciate the opportunity to appear before you to 
discuss the importance of artificial intelligence (AI) to the U.S. 
economy and how best to govern this important technology. My name is 
Daniel Castro, and I am vice president of the Information Technology 
and Innovation Foundation (ITIF), a non-profit, nonpartisan think tank 
whose mission is to formulate and promote public policies to advance 
technological innovation and productivity, and director of ITIF's 
Center for Data Innovation.
What is Artificial Intelligence?
    AI is a field of computer science devoted to creating computer 
systems that perform tasks much like a human would, particularly tasks 
involving learning and decision-making.\1\ AI has many functions, 
including, but not limited to:
---------------------------------------------------------------------------
    \1\ Daniel Castro and Joshua New, ``The Promise of Artificial 
Intelligence,'' Center for Data Innovation, October 2016, http://
www2.datainnovation.org/2016-promise-of-ai.pdf.

   Learning, which includes several approaches such as deep 
        learning (for perceptual tasks), transfer learning, 
---------------------------------------------------------------------------
        reinforcement learning, and combinations thereof;

   Understanding, or deep knowledge representation required for 
        domain-specific tasks, such as medicine, accounting, and law;

   Reasoning, which comes in several varieties, such as 
        deductive, inductive, temporal, probabilistic, and 
        quantitative; and

   Interacting, with people or other machines to 
        collaboratively perform tasks, and for interacting with the 
        environment.

    The cause of many misconceptions about AI, particularly its 
potential harms, is that some people conflate two very distinct types 
of AI: narrow AI and strong AI. Narrow AI describes computer systems 
adept at performing specific tasks, but only those specific types of 
tasks--somewhat like a technological savant.\2\ For example, Apple's 
Siri virtual assistant is capable of interpreting voice commands, but 
the algorithms that power Siri cannot drive a car, predict weather 
patterns, or analyze medical records. While other algorithms exist that 
can accomplish those tasks, they too are narrowly constrained--the AI 
used for an autonomous vehicle will not be able predict a hurricane's 
trajectory or help doctors diagnose a patient with cancer.
---------------------------------------------------------------------------
    \2\ Irving Wladawksy-Berger, `` `Soft' Artificial Intelligence Is 
Suddenly Everywhere,'' The Wall Street Journal, January 16, 2016, 
http://blogs.wsj.com/cio/2015/01/16/soft-artificial-intelligence-is-
suddenly-everywhere/.
---------------------------------------------------------------------------
    In contrast, strong AI, also referred to as artificial general 
intelligence (AGI), is a hypothetical type of AI that can meet or 
exceed human-level intelligence and apply this problem-solving ability 
to any type of problem, just as the human brain can easily learn how to 
drive a car, cook food, and write code.\3\ Many of the dystopian fears 
about AI--that it will eliminate most jobs or go out of control and 
wipe out humanity, for example--stem from the notion that AGI is 
feasible, imminent, and uncontrollable.\4\ However, at least for the 
foreseeable future, computer systems that can fully mimic the human 
brain are only going to be found in scripts in Hollywood, and not labs 
in Silicon Valley.
---------------------------------------------------------------------------
    \3\ Ibid.
    \4\ Robert D. Atkinson, ``'It's Going to Kill Us!' and Other Myths 
About the Future of Artificial Intelligence,'' (Information Technology 
and Innovation Foundation, June 2016), http://www2.itif.org/2016-myths-
machine-learning.pdf?_ga=1.201838291.334601971.1460947053.
---------------------------------------------------------------------------
    The application of AI has seen a surge in recent years because of 
the development of machine learning--a branch of AI that focuses on 
designing algorithms that can automatically and iteratively build 
analytical models from data without needing a human to explicitly 
program the solution. Before machine learning, computer scientists had 
to manually code a wide array of functions into a system for it to 
mimic intelligence. But now developers can achieve the same, or better, 
results more quickly and at a lower cost using machine learning 
techniques. For example, Google uses machine learning to automatically 
translate content into different languages based on translated 
documents found online, a technique that has proven to be much more 
effective than prior attempts at language translation.\5\
---------------------------------------------------------------------------
    \5\ Pedro Domingos, The Master Algorithm: How the Quest for the 
Ultimate Learning Machine Will Remake Our World (New York: Basic Books, 
2015).
---------------------------------------------------------------------------
What Are the Potential Benefits of AI?
    AI will have a substantial and lasting impact on the economy by 
increasing the level of automation in virtually every sector, leading 
to more efficient processes and higher-quality outputs, and boosting 
productivity and per-capita incomes. For example, the McKinsey Global 
Institute estimates that by 2025 automating knowledge work with AI will 
generate between $5.2 trillion and $6.7 trillion of global economic 
value, advanced robotics relying on AI will generate between $1.7 
trillion and $4.5 trillion, and autonomous and semi-autonomous vehicles 
will generate between $0.2 trillion and $1.9 trillion.\6\ Deloitte 
estimates that the Federal Government could save as much as $41.1 
billion annually by using AI to automate tasks.\7\ And Accenture 
predicts that by 2035, AI could increase the annual growth rate of the 
U.S. economy by 2 percentage points, the Japanese economy by 1.9, and 
the German economy by 1.6.\8\ The report also found that, for the 12 
countries surveyed, AI would boost labor productivity rates by 11 to 37 
percent.\9\
---------------------------------------------------------------------------
    \6\ James Manyika et al., Disruptive Technologies: Advances That 
Will Transform Life, Business, and the Global Economy,'' (McKinsey 
Global Institute, May 2013), http://www.mckin
sey.com/business-functions/business-technology/our-insights/disruptive-
technologies.
    \7\ Peter Viechnicki and William D. Eggers, ``How much time and 
money can AI save government?'' (Deloitte, April 26, 2017), https://
dupress.deloitte.com/dup-us-en/focus/cognitive-technologies/artificial-
intelligence-government-analysis.html.
    \8\ Mark Purdy and Paul Daugherty, ``Why Artificial Intelligence Is 
the Future of Growth,'' (Accenture, September 28, 2016), https://
www.accenture.com/us-en/_acnmedia/PDF-33/Accen
ture-Why-AI-is-the-Future-of-Growth.pdf.
    \9\ Ibid.
---------------------------------------------------------------------------
    There are a vast and diverse array of uses for AI, and many U.S. 
businesses are already using the technology today. Manufacturers are 
using AI to invent new metal alloys for 3D printing; pharmaceutical 
companies are using AI to discover new lifesaving drugs; mining 
companies are using AI to predict the location of mineral deposits; and 
agricultural businesses are using AI to increase automation on farms. 
The International Data Corporation (IDC) estimates that the market for 
AI technologies that analyze unstructured data will reach $40 billion 
by 2020.\10\ And AI startups have attracted significant investment, 
with U.S. investors putting $757 million in venture capital in AI 
start-ups in 2013, $2.18 billion in 2014, and $2.39 billion in 
2015.\11\
---------------------------------------------------------------------------
    \10\ ``Cognitive Systems Accelerate Competitive Advantage,'' IDC, 
accessed September 29, 2016, http://www.idc.com/promo/thirdplatform/
innovationaccelerators/cognitive.
    \11\ ``Artificial Intelligence Explodes: New Deal Activity Record 
for AI Startups,'' CB Insights, June 20, 2016, https://
www.cbinsights.com/blog/artificial-intelligence-funding-trends/.
---------------------------------------------------------------------------
    In some cases, the principle benefit of AI is that it automates 
work that would otherwise need to be performed by a human, thereby 
boosting efficiency. Sometimes AI can complete tasks that it is not 
always worth paying a human to do but still creates value, such as 
writing newspaper articles to summarize Little League games.\12\ In 
other cases, AI adds a layer of analytics that uncovers insights human 
workers would be incapable of providing on their own, thereby boosting 
quality. In some cases, it does both. For example, researchers at 
Stanford have used machine learning techniques to develop software that 
can analyze lung tissue biopsies with significantly more accuracy than 
a top human pathologist and at a much faster rate.\13\ By analyzing 
large volumes of data, researchers can train their computer models to 
reliably recognize known indicators of specific cancer types as well as 
discover new predictors.
---------------------------------------------------------------------------
    \12\ Steven Levy, ``Can an Algorithm Write a Better News Story Than 
a Human Reporter?'' Wired, April 24, 2012, https://www.wired.com/2012/
04/can-an-algorithm-write-a-better-news-story-than-a-human-reporter/.
    \13\ Kun-Hsing Yu et al., ``Predicting non-small cell lung cancer 
prognosis by fully automated microscopic pathology image features,'' 
Nature, August 16, 2017, https://www.nature.com/articles/ncomms12474.
---------------------------------------------------------------------------
    AI is also delivering valuable social benefits, such as by helping 
authorities rapidly analyze the deep web to crack down on human 
trafficking, fighting bullying and harassment online, helping 
development organizations better target impoverished areas, reducing 
the influence of gender bias in hiring decisions, and more.\14\ Just as 
AI can help businesses make smarter decisions, develop innovative new 
products and services, and boost productivity to drive economic value, 
it can achieve similar results for organizations generating social 
value, and many of these solutions have the potential to scale 
globally.
---------------------------------------------------------------------------
    \14\ Larry Greenemeier, ``Human Traffickers Caught on Hidden 
Internet,'' Scientific American, February 8, 2015 http://
www.scientificamerican.com/article/human-traffickers-caught-on-hidden-
internet/; Davey Alba, ``Weeding Out Online Bullying Is Tough, So Let 
Machines Do It,'' Wired, July 10, 2015, https://www.wired.com/2015/07/
weeding-online-bullying-tough-let-machines/; Michelle Horton, 
``Stanford Scientists Combine Satellite Data, Machine Learning to Map 
Poverty,'' Stanford News, August 18, 2016 http://news.stanford.edu/
2016/08/18/combining-satellite-data-machine-learning-to-map-poverty/; 
Sean Captain, ``How Artificial Intelligence is Finding Gender Bias at 
Work,'' Fast Company, October 10, 2015, https://www.fastcompany.com/
3052053/elasticity/how-artificial-intelligence-is-finding-gender-bias-
at-work.
---------------------------------------------------------------------------
    Finally, AI will be an increasingly important technology for 
defense and national security. AI can address many goals, such as 
improving logistics, detecting and responding to cybersecurity 
incidents, and analyzing the enormous volume of data produced on the 
battlefield. Moreover, AI will be a core enabler of the Pentagon's 
``Third Offset Strategy,'' a policy designed to keep the United States 
ahead of adversaries, especially ones capable of fielding numerically 
superior forces, through technological superiority.\15\ Indeed, one top 
Pentagon general has suggested that the Defense Department should never 
buy another weapons system that does not have AI built into it.\16\
---------------------------------------------------------------------------
    \15\ Sydney Freedberg, ``Faster Than Thought: DARPA, Artificial 
Intelligence, & The Third Offset Strategy,'' Breaking Defense, February 
11, 2016, https://breakingdefense.com/2016/02/faster-than-thought-
darpa-artificial-intelligence-the-third-offset-strategy/.
    \16\ Jack Corrigan, ``Three-Star General Wants Artificial 
Intelligence in Every New Weapon System,'' Nextgov, November 2, 2017, 
http://www.nextgov.com/cio-briefing/2017/11/three-star-general-wants-
artificial-intelligence-every-new-weapon-system/142225/.
---------------------------------------------------------------------------
How Should Policymakers Support the Adoption and Use of AI?
    Given the potential economic impact of AI in raising productivity, 
policymakers should develop a national strategy to support the 
development and adoption of AI in U.S. businesses. In particular, given 
the enormous advantage that AI-enabled firms will have compared to 
their non-AI-enabled peers, the United States should focus on AI 
adoption in its traded sectors where U.S. firms will face international 
competition. Many other countries see the strategic importance of 
becoming lead adopters of AI, and they have begun implementing policies 
to pursue this goal. These include:

   Canada: In March 2017, Canada launched the Pan-Canadian 
        Artificial Intelligence Strategy which sets a goal of 
        establishing Canada as an international leader in AI research. 
        The strategy has four goals, which include increasing the 
        number of AI researchers and graduates; establishing three 
        major AI research centers; developing global thought leadership 
        on the economic, ethical, policy and legal implications of 
        advances in AI; and supporting the national AI research 
        community.\17\
---------------------------------------------------------------------------
    \17\ ``Pan-Canadian Artificial Intelligence Strategy Overview,'' 
Canadian Institute for Advanced Research, March 3, 2017, https://
www.cifar.ca/assets/pan-canadian-artificial-intelligence-strategy-
overview/.

   China: China's State Council issued a development plan for 
        AI in July 2017 with the goal of making China a leader in the 
        field by 2030. China's goal is to be equal to countries 
        currently leading in AI by 2020. Then, over the subsequent five 
        years, China will focus on developing breakthroughs in areas of 
        AI that will be a ``a key impetus for economic 
        transformation.'' \18\ Finally, by 2030, China intends to be 
        the world's ``premier artificial intelligence innovation 
        center.'' \19\ China's plan also signals its intent to require 
        high school students to take classes in AI, which is one of the 
        most ambitious efforts to develop human capital for the AI 
        economy of any nation.
---------------------------------------------------------------------------
    \18\ Graham Webster et al., ``China's Plan to `Lead' in AI: 
Purpose, Prospects, and Problems,'' New America Foundation, August 1, 
2017, https://www.newamerica.org/cybersecurity-initiative/blog/chinas-
plan-lead-ai-purpose-prospects-and-problems/.
    \19\ Ibid.

   Japan: Prime Minister Abe launched the Artificial 
        Intelligence Technology Strategy Council in April 2016 to 
        develop a roadmap for the development and commercialization of 
        AI.\20\ Published in May 2017, the roadmap outlines priority 
        areas for research and development (R&D), focusing on the 
        themes of productivity, mobility, and health. The strategy also 
        encourages collaboration between industry, government, and 
        academia to advance AI research, as well as stresses the need 
        for Japan to develop the necessary human capital to work with 
        AI. Japan also launched its Japan Revitalization Strategy 2017, 
        which details how the government will work to support growth in 
        certain areas of the economy. The 2017 strategy includes a push 
        to promote the development of AI for telemedicine and self-
        driving vehicles to address the shortage of workers in Japan.
---------------------------------------------------------------------------
    \20\ Josh New, ``How Governments Are Preparing for Artificial 
Intelligence,'' August 8, 2017, https://www.datainnovation.org/2017/08/
how-governments-are-preparing-for-artificial-intelligence/.

   UK: The United Kingdom has taken several steps to promote 
        AI. The UK Digital Strategy, published in March 2017, 
        recognizes AI as a key field that can help grow the United 
        Kingdom's digital economy.\21\ The UK's new budget, published 
        in November 2017, includes several provisions that have the 
        goal of establishing the UK as a world leader in AI, such as by 
        establishing a ``Centre for Data Ethics and Innovation'' to 
        promote the growth of AI, facilitating data access for AI 
        through ``data trusts,'' and funding 450 PhD researchers 
        working on AI.\22\
---------------------------------------------------------------------------
    \21\ Department of Digital, Culture, Media, and Sport, UK Digital 
Strategy, (United Kingdom: Department for Digital, Culture, Media, and 
Sport, 2017), https://www.gov.uk/government/publications/uk-digital-
strategy.
    \22\ Her Majesty's Treasury (HM Treasury), Autumn Budget 2017 
(United Kingdom: HM Treasury, 2017), https://www.gov.uk/government/
publications/autumn-budget-2017-documents/autumn-budget-2017.

    While the U.S. Government has put significant funding behind AI 
R&D--approximately $1.1 billion in 2015--it has not done enough to 
maintain U.S. leadership.\23\ The most ambitious AI program comes from 
China, which as of 2014 surpassed the United States in terms of total 
number of papers published and cited in AI fields, such as deep 
learning.\24\ For both economic and national security reasons, the 
United States cannot afford to cede its existing advantages in AI, and 
should instead look to capitalize on its head start by developing a 
strategy to support AI development and adoption. Such a strategy should 
include policies to address the following:
---------------------------------------------------------------------------
    \23\ Ibid.
    \24\ ``National Artificial Intelligence Research and Development 
Strategic Plan,'' (National Science and Technology Council, October 
2016), https://www.nitrd.gov/PUBS/national_ai
_rd_strategic_plan.pdf.

   Funding: The government should continue to expand its 
        funding to support the ``National Artificial Intelligence 
        Research and Development Strategic Plan,'' a set of R&D 
        priorities identified by the Networking and Information 
        Technology Research and Development (NITRD) program that 
        addresses strategic areas of AI in which industry is unlikely 
        to invest, as well as better plan and coordinate Federal 
        funding for AI R&D across different agencies.\25\
---------------------------------------------------------------------------
    \25\ Ibid.

   Skills: The Federal Government should support educational 
        efforts to ensure a strong pipeline of talent to create the 
        next generation of AI researchers and developers, including 
        through retraining and diversity programs, as well as pursue 
        immigration policies that allow U.S. businesses to recruit and 
---------------------------------------------------------------------------
        retain highly skilled computer scientists.

   AI-Friendly Regulations: Federal and state regulators should 
        conduct regulatory reviews to identify regulatory barriers to 
        commercial use of AI in various industries, such as 
        transportation, health care, education, and finance.

   Data Sharing: Some advances in AI are made possible when 
        large volumes of accurate and representative data are made part 
        of a data commons. The government should continue to supply 
        high-value datasets that enable advances in AI, such as its 
        efforts to produce standardize reference datasets for text 
        analysis and facial recognition. Similarly, Federal agencies 
        should facilitate data sharing between industry stakeholders, 
        such as the Department of Transportation's draft ``Guiding 
        Principles on Data Exchanges to Accelerate Safe Deployment of 
        Automated Vehicles.'' \26\
---------------------------------------------------------------------------
    \26\ ``Draft U.S. DOT Guiding Principles on Voluntary Data 
Exchanges to Accelerate Safe Deployment of Automated Vehicles,'' (U.S. 
Department of Transportation, December 1, 2017) https://
www.transportation.gov/av/data.

   Economic Indicators: Understanding the degree to which U.S. 
        firms have automated processes using AI will be a key metric to 
        assessing the effectiveness of various policies. The Census 
        Bureau should assess what type of economic data it should 
        gather from businesses to monitor and evaluate AI adoption, 
        much like it has tracked rural electrification or broadband 
        connectivity as key economic indicators.
How Should Policymakers Address Concerns About Workforce Disruption?
    One of the most common fears about AI is that it will lead to 
significant disruptions in the workforce.\27\ This fear is not new--
concerns about technology-driven automation have been a perennial 
policy concern since at least the 1930s when Congress debated 
legislation that would direct the Secretary of Labor to make a list of 
all labor-saving devices and estimate how many people could be employed 
if these devices were eliminated.\28\ This concern has been exacerbated 
by a frequently-cited study by two Oxford academics which predicted 
that 47 percent of U.S. jobs could be eliminated over the next 20 
years.\29\
---------------------------------------------------------------------------
    \27\ For a thorough rebuttal of this concern, see Robert D. 
Atkinson, ``'It's Going to Kill Us!' And Other Myths of Artificial 
Intelligence,'' (Information Technology and Innovation Foundation, June 
2016), http://www2.itif.org/2016-myths-machine-learning.pdf.
    \28\ John Scoville, ``Technology and the Volume of Employment,'' 
Proceedings of the Academy of Political Science 18, no. 1 (May 1938): 
84-99.
    \29\ Carl B. Frey and Michael A. Osbourne, ``The Future of 
Employment: How Susceptible Are Jobs to Computerisation?'' (Oxford 
Martin School, University of Oxford, Oxford, September 17, 2013), 
http://www.oxfordmartin.ox.ac.uk/downloads/academic/
The_Future_of_Employment.
pdf.
---------------------------------------------------------------------------
    This study's predictions are misleading and unlikely for at least 
three reasons. First, the estimate includes a number of occupations 
that have little chance of automation, such as fashion models and 
barbers. Second, while this rate of productivity seems high and even 
threatening, it is only slightly higher than rates enjoyed in the mid-
1990s when U.S. job creation was robust and unemployment rates low. 
Third, it succumbs to what economists call the ``lump of labor'' 
fallacy which holds that once a job is gone, there are no other jobs to 
replace it. The reality is that AI-driven productivity enables 
organizations to either raise wages or reduce prices. These changes 
lead to increases in spending, which in turn creates more jobs. And 
given that consumers' wants are far from satisfied, there is no reason 
to believe that this dynamic will change anytime soon.
    But while predictions about massive AI-driven unemployment are 
vastly overstated--indeed, by historical standards occupational churn, 
the rate at which some jobs expand while others contract, is at its 
lowest levels in 165 years--there will still be some worker 
displacement as AI creates higher levels of productivity.\30\ So 
policymakers can and should do more to help workers make transitions 
between jobs and occupations, such as by providing strong social safety 
net programs, reforming unemployment insurance, and offering worker 
retraining. The failure to give workers training and assistance to move 
into new jobs or occupations not only contributes to higher structural 
unemployment, but also increases resistance to innovation and 
automation.\31\
---------------------------------------------------------------------------
    \30\ Robert D. Atkinson and John Wu, ``False Alarmism: 
Technological Disruption and the U.S. Labor Market, 1850-2015,'' 
(Information Technology and Innovation Foundation, May 2017), http://
www2.itif.org/2017-false-alarmism-technological-disruption.pdf.
    \31\ See forthcoming report: ``Technological Innovation, 
Employment, and Workforce Adjustment Policies,'' (Information 
Technology and Innovation Foundation, January 2018).
---------------------------------------------------------------------------
How Should Policymakers Provide Oversight of AI?
    When it comes to AI, the primary goal of the United States should 
be to accelerate the development and adoption of the technology. But as 
with any technology, there will be some risks and challenges that 
require government oversight. The presence of risk, however, does not 
mean that the United States should embrace the precautionary principle, 
which holds that new technology must first be proven safe before it can 
be used. Instead, policymakers should rely on the innovation principle, 
which says that policymakers should address risks as they arise, or 
allow market forces to address them, and not hold back progress because 
of speculative concerns. The innovation principle is especially useful 
when fears about a new technology exceed public awareness and 
understanding about how the technology works and how potential problems 
will be mitigated.\32\
---------------------------------------------------------------------------
    \32\ Daniel Castro and Alan McQuinn, ``The Privacy Panic Cycle: A 
Guide to Public Fears About New Technologies,'' (Information Technology 
and Innovation Foundation, September 2015), http://www2.itif.org/2015-
privacy-panic.pdf.
---------------------------------------------------------------------------
    To understand why this is important, consider the differences 
between the United States and the European Union in the Internet 
economy. Compared to Europe, the United States has had more success in 
the Internet economy, at least in part, because of its vastly more 
simplified data protection regulations. Yet even as the United States 
continues to produce the majority of the major global Internet 
companies, the European Union has decided to double down on its onerous 
data protection rules in the forthcoming General Data Protection 
Regulation (GDPR), a far-reaching set of policies that will 
substantially raise the costs, and in some cases, limit the feasibility 
of using AI in Europe. For example, the GDPR creates both a right to 
explanation and a right to human review for automated decisions, two 
requirements that will make it difficult for companies to construct 
business models that rely extensively on complex algorithms to automate 
consumer-facing decisions. The GDPR also requires organizations to only 
use data for the purposes for which they originally collected it, a 
rule that strictly limits the application of AI to existing data.\33\ 
If the United States wants to compete for global leadership in AI, it 
should be careful not to follow Europe down this path.
---------------------------------------------------------------------------
    \33\ Nick Wallace, ``UK Regulations Need an Update to Make Way for 
Medical AI,'' Center for Data Innovation, August 12, 2017, http://
datainnovation.org/2017/08/uk-regulations-need-an-update-to-make-way-
for-medical-ai/.
---------------------------------------------------------------------------
    While the United States should not replicate the European model, it 
should create its own innovation-friendly approach to providing 
oversight of the emerging algorithmic economy just as it has for the 
Internet economy. Such an approach should prioritize sector-specific 
policies over comprehensive regulations, outcomes over transparency, 
and enforcement actions against firms that cause tangible harm over 
those that merely make missteps without injury. For example, rather 
than industry-wide rules requiring ``algorithmic transparency'' or ``AI 
ethics''--proposals that focus on means, rather than ends--policymakers 
should look to address specific problems, such as ensuring financial 
regulators have the skills necessary to provide oversight of fintech 
companies relying heavily on AI to make lending decisions or provide 
automated financial advisors.
    In many cases, regulators will not need to intervene because the 
private sector will address problems about AI, such as bias or 
discrimination, on its own--even if to outsiders an algorithm appears 
to be a ``black box.'' After all, one company's hidden biases are 
another company's business opportunities. For example, if certain 
lenders were to use algorithms that consistently denied loans to ethnic 
or religious minorities who have good credit, then their competitors 
would have an incentive to target these individuals to gain new 
customers.
    Moreover, the private sector is actively seeking out solutions to 
eliminate problems like unintentional bias in AI that may skew its 
results.\34\ For example, a group of leading AI companies in the United 
States have formed an association to develop and share best practices 
to ensure that AI is fair, safe, and reliable, while another technology 
trade association has publicly committed itself to ensuring that the 
private sector designs and uses AI responsibly.\35\ Indeed, given that 
U.S. companies are at the forefront of efforts to build AI that is safe 
and ethical, maintaining U.S. leadership in this field will be 
important to ensure these values remain embedded in the technology.
---------------------------------------------------------------------------
    \34\ Cliff Kuang, ``Can A.I. Be Taught to Explain Itself?'' New 
York Times, November 21, 2017, https://www.nytimes.com/2017/11/21/
magazine/can-ai-be-taught-to-explain-itself.html.
    \35\ See ``Partnership on AI,'' https://www.partnershiponai.org/and 
``AI Policy Principles,'' Information Technology Industry Council, 
https://www.itic.org/resources/AI-Policy-Principles-FullReport2.pdf.
---------------------------------------------------------------------------
    But policymakers should be careful not to misclassify certain 
concerns as ``AI problems'' that would be best dealt with on a 
technology-neutral basis. For example, discrimination in areas such as 
access to financial services and housing are best addressed through 
existing legal mechanisms. No new laws and regulations are needed 
simply because a company uses AI, instead of human workers, to make 
certain decisions.\36\ Companies cannot use AI to circumvent laws 
outlawing discrimination.
---------------------------------------------------------------------------
    \36\ Travis Korte and Daniel Castro, ``Disparate Impact Analysis is 
Key to Ensuring Fairness in the Age of the Algorithm,'' Center for Data 
Innovation, January 20, 2015, http://datainnovation.org/2015/01/
disparate-impact-analysis-is-key-to-ensuring-fairness-in-the-age-of-
the-algorithm/.
---------------------------------------------------------------------------
    Finally, certain problems, such as sexism in hiring practices, are 
not necessarily made worse by AI. On the contrary, using AI can 
actually reduce human biases. For example, companies can use AI to 
police undesirable behaviors, like automatically flagging job 
advertisements that use gender-specific terminology, such as 
``waitress'' instead of ``wait staff,'' or stereotypical images, such 
as a female nurse.\37\ And unlike human processes, where it may take 
years or decades to change social norms and company culture, businesses 
can refine and tweak code over a period of days or weeks. For example, 
Google changes its search engine 500 to 600 times per year.\38\ Thus 
companies will likely have more success eliminating bias when it 
appears in AI, than when it appears elsewhere in society.
---------------------------------------------------------------------------
    \37\ Amber Laxton, ``Critics of `Sexist Algorithms' Mistake 
Symptoms for Illness,'' Center for Data Innovation, August 3, 2015, 
http://datainnovation.org/2015/08/critics-of-sexist-algorithms-mistake-
symptoms-for-illness/.
    \38\ Daniel Castro, ``Data Detractors Are Wrong: The Rise of 
Algorithms Is a Cause for Hope and Optimism,'' Center for Data 
Innovation, October 25, 2016, http://datainnovation.org/2016/10/data-
detractors-are-wrong-the-rise-of-algorithms-is-a-cause-for-hope-and-
optimism/.
---------------------------------------------------------------------------
Conclusion
    AI is a transformational technology that has the potential to 
significantly increase efficiency and innovation across the U.S. 
economy, creating higher living standards and improved quality of life. 
But while the United States has an early advantage in AI given its top 
talent in computer science and deep bench of companies, large and 
small, investing in the field, many other countries are actively vying 
to challenge U.S. leadership in this domain. In particular, China, with 
its highly skilled computer science workforce and significant funding 
for AI R&D, could easily catch and surpass the United States, leading 
to it gaining economic and military advantages.
    Unfortunately, U.S. policy debates about AI too often overemphasize 
the potential impact on worker displacement from automation or bias 
from algorithms and ignore the much more pressing concern about the 
potential loss of competitiveness and defense superiority if the United 
States falls behind in developing and adopting this key technology.
    Yet, when it comes to AI, successfully integrating this technology 
into U.S. industries should be the primary goal of policymakers, and 
given the rapid pace at which other countries are pursuing this goal, 
the United States cannot afford to rest on its laurels. To date, the 
U.S. Government has not declared its intent to remain globally dominant 
in this field, nor has it begun the even harder task of developing a 
strategy to achieve that vision. Some may think this is unnecessary, 
believing that the United States will automatically prevail in this 
technology simply because it has a unique culture of innovation and has 
prevailed on past technologies.\39\ Such views are naive and dangerous, 
and if followed, likely will lead to the United States being surpassed 
as the global leader in AI with significant negative consequences for 
the U.S. economy and society. However, it is not too late to begin to 
ensure continued U.S. leadership, and I commend you for holding this 
hearing so that we can have this conversation.
---------------------------------------------------------------------------
    \39\ Patrick Tucker, ``What the CIA's Tech Director Wants from 
AI,'' Defense One, September 6, 2017, http://www.defenseone.com/
technology/2017/09/cia-technology-director-artificial-intelligence/
140801/.

    Senator Wicker. Thank you, Mr. Castro.
    Ms. Espinel.

STATEMENT OF VICTORIA ESPINEL, PRESIDENT AND CEO, BSA  
                     THE SOFTWARE ALLIANCE

    Ms. Espinel. Good morning, Chairman Wicker, Ranking Member 
Schatz, and members of the Subcommittee. My name is Victoria 
Espinel, and I am the President and CEO of BSA  The 
Software Alliance.
    BSA is the advocate for the global software industry in the 
United States and around the world. Our members are at the 
forefront of developing artificial intelligence and related 
software services. I commend the Subcommittee for a hearing on 
this important topic, and I thank you for the opportunity to 
testify.
    At the outset, I think it's important to answer a key 
question: What is AI? So let me provide you with a brief 
anecdote. A 60-year-old woman was initially diagnosed with a 
conventional form of leukemia. She went through chemotherapy to 
treat the disease, but her recovery was unusually slow. 
Conventional tests failed to reveal a problem, but her doctor 
suspected that something was still wrong. After several 
frustrating months, they turned to an AI-powered, cloud-based 
system capable of cross-referencing the patient's genetic data 
with insights gleaned from tens of millions of studies from 
around the world. Within minutes, the doctors learned that the 
patient might be suffering from an extremely rare form of 
leukemia that required a unique course of treatment. The 
doctors were able to quickly update her treatment plan and 
watch her condition improve significantly.
    This is AI: it's innovative; it's powerful; it's 
lifesaving. AI is not the image that we see in science fiction 
movies of robots demolishing tall buildings; instead, the AI 
provided by BSA members today is a tool that uses data to help 
people solve complex problems, simplify our daily lives, 
improve business operations, and enhance government services.
    AI is powered by software, which is itself a major engine 
of economic growth. The software industry contributed more than 
$1.14 trillion to the U.S. GDP in 2016--a $70 billion increase 
in just 2 years. The software industry is a powerful job 
creator supporting over 10.5 million jobs with a significant 
positive impact on jobs and economic growth in every one of the 
50 states.
    For example, in Mississippi, software is contributing over 
$800 million to its GDP and over 7,000 jobs, a 25 percent 
increase in jobs in just 2 years. Over 4,000 miles away in 
Hawaii, software is contributing over $1 billion to its GDP and 
over 16,000 jobs. Across every single state in the country, the 
economic impact of software is on the rise.
    AI is helping all industry sectors. Whether it is securing 
networks, improving health, or helping American farmers save 
money, the impact of AI is already visible in every industry, 
in every state, and across the globe.
    We should also be prepared to address important issues that 
may arise as AI-enabled services are used. Let me focus on two. 
First, AI will change the skill sets needed for certain jobs, 
and while new AI-related jobs will be created, there will be 
shifts in the economy. BSA members are already helping 
launching groundbreaking initiatives to provide free training, 
including to youth and military veterans, to ensure that both 
the current workforce and the next generation are prepared for 
the future. We are dedicated to this work, and we look forward 
to collaborating with all of you on this effort.
    Second, we are mindful of the need to ensure that AI is 
both trained and used fairly and responsibly. At the same time, 
we recognize the potential of AI to make human decisions more 
accurate and less biased and the need to push toward that 
outcome.
    As our companies seek to ensure responsible AI deployment, 
there are several steps that Congress and the administration 
can take. First, as I highlighted earlier, AI depends on data, 
so we urge Congress to pass the Open Government Data Act, which 
would make non-sensitive government data more open, more 
available, and more usable for the general public.
    Ranking Member Schatz, thank you for your great work as 
sponsor of the Open Government Data Act. We hope that Congress 
will act soon to send it to the President's desk. We also 
encourage Congress and the administration to be leaders on 
digital trade to encourage global data flows.
    Second, we encourage increased investment in government 
research, including on how AI can contribute to both positive 
economic and social outcomes and policies that incentivize 
private sector research and development.
    And, third, we need to prioritize education and workforce 
development so that our young people and our current workforce 
are prepared for the future.
    As part of all of this, we need to have a meaningful 
dialogue with all stakeholders about how to address any 
challenges that lie ahead. The legislation introduced by 
Senators Cantwell, Young, and Markey is a good step, and we 
thank you for that. Thanks to Senator Schatz as well for the 
legislation that you are currently working on.
    In closing, we look forward to working with all of you 
towards a clear understanding of AI and to address the 
challenges and embrace the opportunities ahead. BSA members are 
part of the solution to these challenges, and we are eager to 
work with you as we chart a responsible path forward.
    Thank you, and I look forward to your questions.
    [The prepared statement of Ms. Espinel follows:]

      Prepared Statement of Victoria Espinel, President and CEO, 
                       BSA--The Software Alliance
    Good morning Chairman Wicker, Ranking Member Schatz, and members of 
the Subcommittee. My name is Victoria Espinel, and I am the President 
and CEO of BSA  The Software Alliance.
    BSA is the leading advocate for the global software industry in the 
United States and around the world.\1\ Our members are at the forefront 
of developing cutting-edge artificial intelligence (AI) and related 
software-enabled technologies and services that are having a 
significant impact on the U.S. and global economy. I commend the 
Subcommittee for holding a hearing on this important topic, and I thank 
you for the opportunity to testify on behalf of BSA.
---------------------------------------------------------------------------
    \1\ BSA's members include: Adobe, ANSYS, Apple, Autodesk, Bentley 
Systems, CA Technologies, CNC/Mastercam, DataStax, DocuSign, IBM, 
Microsoft, Oracle, salesforce.com, SAS Institute, Siemens PLM Software, 
Splunk, Symantec, Trimble Solutions Corporation, The MathWorks, Trend 
Micro and Workday.
---------------------------------------------------------------------------
I. AI: Defining the Landscape
    The term ``artificial intelligence'' often conjures images of all-
knowing robots with physical and cognitive abilities far superior to 
those of their human creators. The actual AI services that are in the 
market today--and that BSA members provide--bear no resemblance to the 
sinister images of the future that consumers often see in the movies, 
with robots taking over big cities and small towns.
    Instead, they increasingly are becoming a foundational technology 
that drives many products and services that people use every day. 
Whether it is a personal digital assistant that helps consumers locate 
the nearest restaurant, a fraud detection monitoring service that 
prevents criminals from placing charges on credit cards, or a tool that 
helps teachers identify students with additional needs and develop 
personalized lesson plans, we increasingly rely on a diverse range of 
AI-enabled services every day.
But what is ``AI''?
    Although definitions of AI vary, one common description of AI is 
that it refers to machines that act intelligently in pursuit of human-
defined objectives. At its core, AI is simply a tool. It includes a 
broad range of technologies, but the AI systems that BSA members 
largely provide assist in the analysis of enormous volumes of data to 
find connections that improve the quality and accuracy of human 
decision-making. Although some AI systems have a limited degree of 
autonomy, such as submarines that map the ocean bed and measure ocean 
currents, and others are minutely supervised, such as robot surgical 
tools assisting doctors with hip replacement surgeries, the vast 
majority provide advice and recommendations to humans rather than 
acting independently. AI makes possible important tasks that would 
otherwise be economically or physically infeasible, such as inspecting 
wind turbine blades or the interior of oil pipelines.
    AI systems, like other software systems, use sophisticated 
algorithms. An algorithm is a set of instructions that processes 
various inputs and provides an output in a systematized way. The 
algorithms used in AI are particularly well-suited to analyzing massive 
volumes of data from many different sources, and in identifying 
patterns across the enormous number of variables in such data that may 
interact in complex and unexpected ways. Through this analysis, AI 
systems can enhance perception, learning, reasoning, and decision-
making, and improve the ability of people to solve complex and 
challenging problems.
    The use of systems, including software, to help people solve 
complex problems is not new. Research into AI dates back many decades, 
but we have witnessed tremendous advances in AI capabilities over the 
past five to ten years. These advances have been fueled by a number of 
related developments, including the proliferation of technologies that 
generate vast amounts of data, the affordability of data storage, and 
ever-growing data processing capabilities.
    BSA members have made significant investments in enhancing these 
data-driven technologies to develop innovative AI solutions for use 
across a broad range of applications in a wide variety of contexts.
II. AI Services Provide Substantial Benefits
    Advances in AI and software-enabled data analytics are fueling job 
and economic growth in the United States and around the world, 
improving how businesses in every sector operate, and producing real 
societal gains. We must recognize that AI will change the skill sets 
needed for certain jobs. And while new, AI-related jobs will be 
created, there will be shifts in the labor market. And although we 
should be mindful of the need to ensure that AI is deployed fairly and 
responsibly, we should also recognize the potential of AI to make human 
decisions more accurate and less biased, and thereby to promote 
fairness and inclusiveness across all segments of society.
A. AI and Related Software Services Are Creating Jobs and Economic 
        Growth
    In high-tech and low-tech industries alike, the analysis of data 
has made businesses more agile, responsive, and competitive, boosting 
the underlying productivity of many key pillars of our economy.
    The economic implications of the data revolution--and AI and 
related software solutions that leverage that data--are enormous. 
Economists predict that making better use of data could lead to a 
``data dividend'' of $1.6 trillion in the next four years, and that 
data-enabled efficiency gains could add almost $15 trillion to global 
GDP by 2030.\2\ In addition, experts predict that applications of AI 
technologies could grow the global economy by $7.1 to $13.17 trillion 
over the next eight years.\3\
---------------------------------------------------------------------------
    \2\ See BSA, What's the Big Deal With Data? 14 (Oct. 2015), 
available at http://data.bsa.org/wp-content/uploads/2015/12/
bsadatastudy_en.pdf. The potential of digital data to improve the 
healthcare system is substantial: some estimates predict that if the 
healthcare sector were to use data more effectively to drive efficiency 
and quality, the sector could save more than $300 billion every year. 
See James Manyika et al., Big Data: The Next Frontier for Innovation, 
Competition, and Productivity, McKinsey Global Institute (May 2011), 
available at http://www.mckinsey.com/insights/business_technology/
big_data_the_next_frontier_for_innovation.
    \3\ See Disruptive technologies: Advances that will transform life, 
business, and the global economy, McKinsey Global Institute (May 2013), 
available at http://www.mckinsey.com/business-functions/digital-
mckinsey/our-insights/disruptive-technologies.
---------------------------------------------------------------------------
    AI systems are powered by software, which itself is a major engine 
of economic growth. In September, Software.org: the BSA Foundation 
released a study with data from the Economist Intelligence Unit (EIU) 
showing that the software industry alone contributed more than $1.14 
trillion to U.S. GDP in 2016--a $70 billion increase in just two 
years.\4\ The study also showed that the software industry is a 
powerful job creator, supporting over 10.5 million jobs, with a 
significant impact on job and economic growth in each of the 50 
states.\5\
---------------------------------------------------------------------------
    \4\ Software.org: The BSA Foundation, The Growing $1 Trillion 
Economic Impact of Software 5 (Sept. 2017), available at https://
software.org/wp-content/uploads/2017_Software_Economic
_Impact_Report.pdf.
    \5\ Id.
---------------------------------------------------------------------------
B. AI and Related Software Services Are Improving Every Industry
    The benefits of AI are not limited to the software sector. In fact, 
AI innovation is stimulating growth across all industry sectors as 
businesses, big and small, use AI and related software services to 
improve supply chains, secure their networks, and evaluate how to 
improve their products and services. There are numerous examples of 
this positive impact across a wide swath of industries, for instance:

   Cybersecurity. AI tools are revolutionizing how we monitor 
        network security, helping analysts parse through hundreds of 
        thousands of security incidents per day to weed out false 
        positives and identify threats that warrant further attention 
        by network administrators. By automating responses to routine 
        incidents and enabling security professionals to focus on truly 
        significant threats, AI-enabled cyber tools are helping 
        enterprises stay ahead of their malicious adversaries.\6\
---------------------------------------------------------------------------
    \6\ For example, IBM's Watson for Cyber Security is a cybersecurity 
tool that can analyze 15,000 security documents per day--a rate 
essentially impossible for any individual to achieve. Watson's data 
processing capabilities enable analysts to more quickly identify 
incidents that require human attention. See IBM, IBM Delivers Watson 
for Cyber Security to Power Cognitive Security Operations Centers (Feb. 
13, 2017), https://www-03.ibm.com/press/us/en/press
release/51577.wss; Jason Corbin, Bringing the Power of Watson and 
Cognitive Computing to the Security Operations Center, Security 
Intelligence (Feb. 13, 2017), https://securityintelligence
.com/bringing-the-power-of-watson-and-cognitive-into-the-security-
operations-center/?cm_mc_uid
=70595459933115020631816&cm_mc_sid_50200000=1503364089&cm_mc_sid_5264000
0=150336
5578. Splunk uses a similar model, with machine-learning algorithms 
conducting real-time analysis and processing of massive volumes of data 
from all sensors on a network to identify anomalies, feeding 
visualization tools that help network administrators efficiently triage 
security incidents. See David Braue, Machine learning key to building a 
proactive security response: Splunk, CSO Online (Aug. 20, 2015), 
https://www.cso.com.au/article/582483/machine-learning-key-building-
proactive-security-response-splunk/. Microsoft's Windows 10 Anniversary 
Edition introduced AI-driven capabilities for automatically isolating 
suspicious network traffic pending adjudication by network 
administrators. See Chris Hallum, Defense Windows clients from modern 
threats and attacks with Windows 10, Channel 9 video content (Oct. 6, 
2016), available at https://channel9.msdn.com/events/Ignite/2016/
BRK2135-TS); ``Intelligent Security: Using Machine Learning to Help 
Detect Advanced Cyber Attacks,'' https://www.microsoft.com/en-us/
security/intelligence.

   Financial Services. AI is improving fraud detection by 
        providing companies with real-time information that helps them 
        identify and investigate different types of fraud, reducing the 
        losses attributed to fraudsters by billions of dollars. In a 
        matter of seconds, machine learning algorithms can generate a 
        risk score for a transaction by parsing through large volumes 
        of data about the vendor and the purchaser to determine the 
        likelihood of fraud.\7\ These tools are protecting consumers 
        from the risk of fraudulent charges and from the frustration 
        associated with ``false declines.''
---------------------------------------------------------------------------
    \7\ See generally Pablo Hernandez, CA Technologies Uses AI Tech to 
Combat Online Fraud, eSecurityPlanet, May 4, 2017, available at https:/
/www.esecurityplanet.com/network-security/ca-technologies-uses-ai-tech-
to-combat-online-fraud.html.

   Agriculture. AI is helping farmers tackle some of the 
        biggest issues they face, including declining crop yields and 
        changing weather patterns, through precision farming, better 
        data analysis, and improved operational efficiency. For 
        instance, tools like computer vision and deep-learning 
        algorithms are enabling farmers more effectively to process 
        data for purposes of monitoring crop and soil health.\8\
---------------------------------------------------------------------------
    \8\ See Kumba Sennaar, AI in Agriculture--Present Applications and 
Impact, techemergence (Nov. 17, 2017), https://www.techemergence.com/
ai-agriculture-present-applications-impact/.

   Manufacturing. AI-enabled tools are also helping factory 
        owners streamline their manufacturing processes and resolve 
        problems common to most factories, such as inaccurate demand 
        forecasting and capacity planning, unexpected equipment 
        failures and downtimes, and supply chain bottlenecks. 
        Predictive maintenance, for instance, allows manufacturers to 
        achieve 60 percent or more reduction in unscheduled system 
        downtime. Cameras powered by computer vision algorithms can fix 
        product defects immediately and identify root causes of 
        failure. AI thus enables manufacturers to reduce waste, shorten 
        production periods, increase yields on production inputs, and 
        improve both revenue and workplace safety.\9\
---------------------------------------------------------------------------
    \9\ See Mariya Yao, Factories Of The Future Need AI To Survive And 
Compete, Forbes.com (Aug. 8, 2017), https://www.forbes.com/sites/
mariyayao/2017/08/08/industrial-ai-factories-of-future/#2d7ab2fd128e.

   Healthcare. AI technologies are already providing solutions 
        that help save lives. A 2016 Frost & Sullivan report predicts 
        that AI has the potential to improve health outcomes by 30 to 
        40 percent.\10\ AI is helping fuel these improved health 
        outcomes not by replacing the decision-making of healthcare 
        professionals, but by giving these professionals new insights 
        and new ways of analyzing and understanding the health data to 
        which they have access. For example, AI tools are powering 
        machine-assisted diagnosis and surgical applications are being 
        used to improve treatment options and outcomes. Image 
        recognition algorithms are helping pathologists more 
        effectively interpret patient data, thereby helping physicians 
        form a better picture of patients' prognosis.\11\ The ability 
        of AI to process and find patterns in vast amounts of data from 
        disparate sources is also driving important progress in 
        biomedical and epidemiological research.\12\
---------------------------------------------------------------------------
    \10\ See From $600 M to $6 Billion, Artificial Intelligence Systems 
Poised for Dramatic Market Expansion in Healthcare, Frost & Sullivan 
(Jan. 5, 2016), https://ww2.frost.com/news/press-releases/600-m-6-
billion-artificial-intelligence-systems-poised-dramatic-market-
expansion-healthcare.
    \11\ See e.g., Meg Tirrell, From coding to cancer: How AI is 
changing medicine, cnbc.com (May 11, 2017), https://www.cnbc.com/2017/
05/11/from-coding-to-cancer-how-ai-is-changing-medicine.html.
    \12\ For instance, AI is helping biologists who are aiming to treat 
100 molecular genetic diseases by 2025. See Splunk, Machine Learning 
Helps Recursion Pharmaceuticals Treat Genetic Diseases (Nov. 7, 2017), 
https://www.splunk.com/en_us/newsroom/press-releases/2017/splunk-
machine-learning-helps-recursion-pharmaceuticals-treat-genetic-
diseases.html. In another example, Microsoft researchers are also using 
AI and related technologies to better understand the behavior of cells 
and their interaction, which could ultimately help ``debug'' an 
individual's specific form of cancer and allow doctors to provide 
personalized cancer treatment. See generally, Microsoft, Biological 
Computation, https://www.microsoft.com/en-us/research/group/biological-
computation/.

   Education. AI technologies offer tools for students, 
        teachers, and administrators to help students learn more 
        effectively both within and outside of the classroom. AI 
        programs can, for example, analyze a student's performance in a 
        particular skill across subjects over the course of a year and 
        automatically provide new content or specified learning 
        parameters, offering students continual, individualized 
        practice and feedback. They can also help teachers better 
        understand student performance, quickly identify students that 
        need particular attention, and develop lesson plans that 
        customize instruction, content, pace, and testing to individual 
        students' strengths and interests.\13\ AI solutions also are 
        helping administrators track attendance patterns and gain 
        insights on student performance more broadly.\14\
---------------------------------------------------------------------------
    \13\ See Software.org: The BSA Foundation, The Growing $1 Trillion 
Economic Impact of Software, supra note 4, at 7; see also Daniel 
Faggella, Examples of Artificial Intelligence in Education, 
TechEmergence (Mar. 7, 2017), https://www.techemergence.com/examples-
of-artificial-intelligence-in-education/.
    \14\ Benjamin Herold, Are schools ready for the power and problems 
of big data?, Education Week (Jan. 11, 2016), available at http://
www.edweek.org/ew/articles/2016/01/13/the-future-of-big-data-and-
analytics.html.
---------------------------------------------------------------------------
C. AI Services Provide Tremendous Societal Benefits
    The range of potential societal benefits from the use of AI 
services is equally vast. For example, AI solutions are at the heart of 
new devices and applications that improve the lives of people with 
disabilities, including helping people with vision-related impairments 
interpret and understand photos and other visual content.\15\ This 
technology opens new possibilities for people with vision impairments 
to navigate their physical surroundings, giving them increased 
independence and greater ability to engage with their communities.
---------------------------------------------------------------------------
    \15\ For instance, Microsoft recently released an intelligent 
camera app that uses a smartphone's built-in camera functionality to 
describe to low-vision individuals the objects that are around them. 
See Microsoft, Seeing AI, https://www.microsoft.com/en-us/seeing-ai/.
---------------------------------------------------------------------------
    AI is also helping governments improve constituent services in ways 
that save time, money, and lives. For example, cities are optimizing 
medical emergency response processes using AI-based systems, enabling 
them to more strategically position personnel and reduce both response 
times and the overall number of emergency trips.\16\ AI is also helping 
to leverage data to improve disaster response and relief efforts, 
including after the 2015 earthquake in Nepal.\17\
---------------------------------------------------------------------------
    \16\ See Kevin C. Desouza, Rashmi Krishnamurthy, and Gregory S. 
Dawson, Learning from public sector experimentation with artificial 
intelligence, Brookings Institution (June 23, 2017), https://
www.brookings.edu/blog/techtank/2017/06/23/learning-from-public-sector-
experimentation-with-artificial-intelligence/.
    \17\ See Patrick Meier, Virtual Aid to Nepal: Using Artificial 
Intelligence in Disaster Relief, Foreign Affairs (June 1, 2015), 
available at https://www.foreignaffairs.com/articles/nepal/2015-06-01/
virtual-aid-nepal.
---------------------------------------------------------------------------
                             *  *  *  *  *
    Whether it is detecting financial fraud, improving health outcomes, 
making American farmers more competitive, or enhancing government and 
emergency services, the impact of AI and related software services is 
already visible in every industry, in every state, and across the 
globe.
III. Fostering Consumer Trust in AI
    Even as society gains from the substantial benefits that AI offers, 
we also recognize that there may be legitimate concerns about how AI 
systems are deployed in practice, which may also affect trust and 
confidence in AI. In particular, as people increasingly apply AI 
services in new contexts, questions may arise about how they operate, 
whether they treat people fairly and are free from improper bias, and 
their impact on jobs. Like many technologies, AI has an almost infinite 
range of beneficial uses, but we should also take appropriate steps to 
ensure that it is deployed responsibly. We recognize that responsible 
deployment of AI should instill consumer confidence that these 
important issues will be appropriately addressed.
A. Enhancing Understanding of AI Systems
    Building trust and confidence in AI-enabled systems is an important 
priority. In some instances, the complexity of these technologies, 
which are designed to identify patterns and connections that humans 
could not easily identify on their own, can make it challenging to 
explain how certain aspects of AI systems work. BSA members understand 
that, in order to promote trust, companies that build and deploy AI 
systems will need to provide meaningful information to enhance 
understanding of how these systems operate.
    Indeed, ensuring that AI systems operate as intended and treat 
people fairly is an important priority. We are eager to participate in 
meaningful dialogues with other stakeholders about how best to 
accomplish that goal, and we welcome opportunities such as this one to 
help advance that dialogue. Currently, relevant technical tools and 
operational processes that could improve understanding and confidence 
in AI systems are still being developed, and it is an area of robust 
research. Although more work needs to be done, it is already clear that 
expectations are highly context-specific--and demands will vary based 
on this context. As we seek to address these important issues, we will 
aim to ensure that we remain sufficiently flexible to respond to 
concerns, and to adapt to the changing landscape as these emerging 
technologies, and potential solutions to new challenges, continue to 
evolve.
B. Preparing the Workforce for the Jobs of the Future
    As AI services improve every industry, they will likely have a 
multi-dimensional impact on employment. The deployment of AI in the 
workplace will enable employees to focus on tasks that are best suited 
to uniquely human skillsets, such as creativity, empathy, foresight, 
judgment, and other social skills. Although there appears to be no 
consensus on the precise impact AI will have on employment, there is 
broad recognition that widespread deployment of these technologies will 
create demand for new types of jobs, and that these jobs often will 
require skills that many workers today do not yet have.
    Current estimates indicate the United States will not have enough 
workers to meet the predicted high demand for computer science-related 
jobs. For example, by 2020, the U.S. Bureau of Labor Statistics 
predicts that there will be 1.4 million computing jobs, but just 
400,000 computer science students with the skills necessary to fill 
those jobs.\18\ It is imperative that the United States takes steps now 
to ensure that we have a sufficient pipeline of workers with the skills 
needed to perform these new, high-quality jobs.
---------------------------------------------------------------------------
    \18\ See Allie Bidwell, Tech Companies Work to Combat Computer 
Science Education Gap, U.S. News & world report, Dec. 27, 2013, 
available at https://www.usnews.com/news/articles/2013/12/27/tech-
companies-work-to-combat-computer-science-education-gap.
---------------------------------------------------------------------------
    Yet even these estimates do not take into account the extent to 
which the use of AI may require new skills. Because AI services will 
likely be integrated across all sectors of the economy, the new jobs AI 
creates, and the new skills that will be needed, will reach beyond the 
tech sector, and will also likely extend to workers in both urban and 
rural areas. Indeed, many of these jobs will ``look nothing like those 
that exist today,'' and will include ``entire categories of new, 
uniquely human jobs'' that will require ``skills and training that have 
no precedents.'' \19\ As a result, one key challenge that lies ahead is 
determining how to ensure that the U.S. workforce has the skills 
necessary for the future.
---------------------------------------------------------------------------
    \19\ H. James Wilson, Paul R. Daugherty, Nicola Morini-Bianzino, 
The Jobs that Artificial Intelligence will Create, MIT Sloan Management 
Review (Mar. 23, 2017), available at https://sloanreview.mit.edu/
article/will-ai-create-as-many-jobs-as-it-eliminates/.
---------------------------------------------------------------------------
    BSA members are working hard to help address this challenge. BSA 
recognizes that this will require a multi-faceted solution, including 
cooperation with public and private stakeholders. We seek to identify 
opportunities and partnerships that focus on retraining the workforce 
with new skills, creating a pipeline of workers with skills to fill the 
next generation of jobs, increasing access to those jobs for skilled 
workers, and increasing deployment of cloud services, which facilitate 
employment and collaboration in different geographic regions.
    Notably, BSA members already have begun helping workers and youth 
acquire new skills that will enable them to leverage AI systems.\20\ 
BSA members offer several high-tech and business training programs, 
including at the high school level. Some programs target populations 
not traditionally associated with tech jobs, such as military 
veterans.\21\ These initiatives illustrate just some of the ways in 
which AI-based employment concerns can be meaningfully addressed.
---------------------------------------------------------------------------
    \20\ See, e.g., Allen Blue, How LinkedIn is Helping Create Economic 
Opportunity in Colorado and Phoenix (Mar. 17, 2016), https://
blog.linkedin.com/2016/03/17/how-linkedin-is-helping-create-economic-
opportunity-in-colorado-and-phoenix; Markel Foundation, Why Microsoft 
and the Markle Foundation are Working Together to Connect Workers with 
New Opportunities in the Digital Economy, https://www.markle.org/
microsoft. IBM, for instance, has established Pathways in Technology 
Early College High Schools (P-TECH Schools). P-TECH schools are 
innovative public schools that offer students the opportunity to earn a 
no-cost associates degree within six years in fields such as applied 
science and engineering--and to acquire the skills and knowledge 
necessary to pursue further educational opportunities or to step easily 
into well paying, high-potential informational technology jobs. IBM 
designed the P-TECH model to be both widely replicable and sustainable 
as part of an effort to reform career and technical education. See IBM, 
IBM and P-TECH, https://www-03.ibm.com/press/us/en/presskit/42300.wss. 
Likewise, Salesforce offers free high-tech and business skills training 
through Trailhead, its online learning platform, with the goal of 
preparing them for the estimated 3.3 million jobs created by the 
Salesforce economy worldwide from 2016 to 2022, nearly 1 million of 
which are forecasted to be in the United States. See International Data 
Corporation, The Salesforce Economy Forecast: 3.3 Million New Jobs and 
$859 Billion New Business Revenue to Be Created from 2016 to 2022 (Oct. 
2017), available at http://www.salesforce.com/assets/pdf/misc/idc-
study-salesforce-economy.pdf; see also Gavin Mee, How the Salesforce 
Economy is Driving Growth and Creating Jobs, Oct. 24, 2017, available 
at https://www.salesforce.com/uk/blog/2017/10/idc-how-the-salesforce-
economy-is-driving-growth-and-creating-jo; Gavin Mee, Guest Blog: Gavin 
Mee, Salesforce--Evolving tech means change in digital skills, TechUK 
(Apr. 26, 2017), at https://www.techuk
.org/insights/opinions/item/10695-guest-blog-gavin-mee-salesforce-
evolving-tech-means-change
-in-digital-skills.
    \21\ For example, the Splunk4Good initiative, which partners with 
non-profits, is helping military veterans and their families, along 
with youth, train for careers in technology, providing free access to 
Splunk licenses and its extensive education resources to help them 
attain marketable skillsets. See Splunk, Splunk Trains Workforce of 
Tomorrow With Amazon Web Services, NPower, Wounded Warrior Project and 
Year Up, (Sept. 26, 2017) https://www.splunk.com/en_us/newsroom/press-
releases/2017/splunk-trains-workforce-of-tomorrow-with-amazon-web-
services-npower-wounded-warrior-project-and-year-up.html.
---------------------------------------------------------------------------
IV. Opportunities for Congress and the Administration to Facilitate AI 
        Innovation
    As innovation in AI and related software services increasingly 
fuels growth in the global economy, countries around the world are 
taking steps to invest in education, research, and technological 
development to become a hub for AI innovation. For example, the UK 
government recently released an Industrial Strategy, which identifies 
putting the UK at the forefront of the AI and data revolution as one of 
four key strategies that will secure its economic future.\22\ In the 
EU, the European Parliament recently issued a report on civil law rules 
regarding robotics, which highlights the opportunities robotics and AI 
offer and encourages investment in such technology so Europe can 
maintain leadership in this space.\23\ Likewise, in Japan, the 
government recently issued a new strategy designed to strengthen 
collaboration between industry, the government, and academia on matters 
related to robotics, and also issued a report offering the first 
systematic review of AI networking issues in Japan.\24\ In China, the 
government has issued a ``Next Generation Artificial Intelligence 
Development Plan,'' which lays out objectives for AI development in 
China for the next 13 years and calls on China to become a global AI 
innovation center by 2030.\25\
---------------------------------------------------------------------------
    \22\ See UK Secretary of State for Business, Energy and Industrial 
Strategy, Industrial Strategy Building a Britain fit for the future 
(Nov. 2017), available at https://www.gov.uk/government/uploads/system/
uploads/attachment_data/file/662541/industrial-strategy-white-paper-
print-version.pdf. 
    \23\ See European Parliament 2014-2019, Resolution of 16 February 
2017 with recommendations to the Commission on Civil Law Rules on 
Robotics, Eur. Parl. Doc. P8_TA (2017)0051, http://
www.europarl.europa.eu/sides/getDoc.do?pubRef=-//EP//NONSGML+TA+P8-TA-
2017-0051+0
+DOC+PDF+V0//EN.
    \24\ See Fumio Shimpo, Japan's Role in Establishing Standards for 
Artificial Intelligence Development, Carnegie Endowment for 
International Peace (Jan 12, 2017), http://carnegie
endowment.org/2017/01/12/japan-s-role-in-establishing-standards-for-
artificial-intelligence-development-pub-68311.
    \25\ See Elsa Kania, China's Artificial Intelligence Revolution, 
The Diplomat (Jul. 27, 2017), available at https://thediplomat.com/
2017/07/chinas-artificial-intelligence-revolution/.
---------------------------------------------------------------------------
    In the United States, a flexible policy framework that facilitates 
responsible AI deployment and increased investment will be key to 
preserving U.S. global economic competitiveness. An essential part of 
that effort will be ensuring the ability to access data, and to 
transfer that data seamlessly across borders, which are vital for AI to 
flourish. It also will be important to support investment in AI-related 
education, workforce development, and research. To that end, there are 
several steps that Congress and the Administration could take to spur 
AI innovation and continued economic growth.
A. Pass OPEN Government Data Act
    First, Congress should pass the OPEN Government Data Act. This 
legislation, which the House recently passed as Title II of the 
Foundations for Evidence-Based Policymaking Act, recognizes that 
government-generated data is a national resource that can serve as a 
powerful engine for creating new jobs and a catalyst for economic 
growth. To that end, the OPEN Government Data Act would require 
agencies to make non-sensitive government data more open, available, 
and usable for the general public. Making such data more readily 
available will improve government transparency, promote government 
efficiency, and foster innovation of data-driven technologies such as 
artificial intelligence.
    We would like to thank Ranking Member Schatz for his tireless work 
as an original sponsor of the OPEN Government Data Act. We are hopeful 
that the Senate will act soon to secure its final passage into law.
B. Support Efforts to Promote Digital Trade and Facilitate Data Flows
    We also urge Congress and the Administration to continue supporting 
efforts to expand digital trade. Indeed, the new digital data economy, 
which increasingly relies on AI and related software services, will 
benefit from a globally recognized system for digital trade that 
facilitates cross-border data flows and establishes clear rules, 
rights, and protections. There are several opportunities for Congress 
and the Administration to lead in this area.
    First, the ongoing NAFTA discussions provide an important 
opportunity to modernize the trade agreement, which was initially 
negotiated when digital services were in their infancy. We are 
encouraged that the Administration has made it an objective to seek to 
prohibit market access barriers to digital trade, including 
restrictions on data transfers, data localization mandates, and 
technology transfer requirements.
    Second, another key priority is ensuring that transatlantic trade 
continues to thrive. In particular, we appreciate Congress's and the 
Administration's leadership on issues relating to the EU-U.S. Privacy 
Shield, which both protects privacy and facilitates data transfers 
between the EU and United States. We encourage your continued support 
as the Administration proceeds with its ongoing successful 
implementation of the framework.
    Third, as other countries seek to modernize their trade policies, 
the Administration should engage key global partners to ensure that new 
trade initiatives facilitate data-driven innovation and protect against 
market access barriers for e-commerce and digital trade.
C. Invest in AI research, education, and workforce development
    Unlocking the full promise of AI technologies also requires a long-
term strategy of investing in education, workforce development, and 
research. Because human beings ultimately drive the success of AI, 
supporting education, training, and research is essential to extracting 
the maximum level of benefit that AI technologies offer.
    As an initial matter, Congress and the Administration should ensure 
that education programs are developing human talent more effectively. 
Broadly speaking, this means that Congress and the Administration 
should support science, technology, engineering, and mathematics (STEM) 
education at all levels. It also means creating and supporting programs 
that help educate researchers and engineers with expertise in AI, as 
well as specialists who apply AI methods for specific applications and 
users who operate those applications in specific settings.\26\ For 
researchers and engineers, these programs should include training in 
computer science, statistics, mathematical logic, and information 
theory, and for specialists, they should focus on software engineering 
and related applications.\27\
---------------------------------------------------------------------------
    \26\ See U.S. Executive Office of the President, Preparing for the 
Future of Artificial Intelligence, National Science and Technology 
Council Committee on Technology 26 (Oct. 2016), available at https://
obamawhitehouse.archives.gov/sites/default/files/whitehouse_files/
microsites/ostp/NSTC/preparing_for_the_future_of_ai.pdf.
    \27\ See id.
---------------------------------------------------------------------------
    Congress and the Administration should also support the development 
of new and innovative ways to ensure the U.S. workforce is prepared for 
the jobs of the future. Because AI will generate new jobs in categories 
both known and unforeseen, we need to develop thoughtful and effective 
approaches to equip the U.S. workforce with the skills necessary to 
seize the opportunities these new technologies create and to optimize 
the role of AI in modern life.
    Continued scientific research is essential to fully tapping the 
potential of AI technology. Congress and the Administration should 
therefore also promote both public and private sector research to help 
ensure that the United States remains a leader in this space. The U.S. 
Government should invest in the types of ``long-term, high-risk 
research initiatives'' in which the commercial sector may be reluctant 
to invest. In the past, such R&D investments have led to 
``revolutionary technological advances. . .[such as] the Internet, GPS, 
smartphone speech recognition, heart monitors, solar panels, advanced 
batteries, cancer therapies, and much, much more.'' \28\ Congress and 
the Administration should also adopt policies that incentivize private-
sector R&D, including by expanding access to financing.
---------------------------------------------------------------------------
    \28\ See The National Artificial Intelligence Research and 
Development Strategic Plan (Oct. 2016), available at https://
www.nitrd.gov/PUBS/national_ai_rd_strategic_plan.pdf.
---------------------------------------------------------------------------
    Passing the OPEN Government Data Act, supporting efforts to promote 
digital trade and facilitate cross-border data flows, and investing in 
AI research, education, and workforce development will be critical to 
maximizing the opportunities AI presents and helping to ensure that the 
United States maintains leadership in AI innovation and deployment, 
even as other nations increase their own efforts to take advantage of 
the possibilities that AI offers.
                                 * * *
    We appreciate Congress's leadership on the important issue of 
facilitating AI innovation and its responsible deployment. Thank you 
and I look forward to your questions.

    Senator Wicker. Thank you very much.
    Dr. Gil.

      STATEMENT OF DR. DARIO GIL, Ph.D., VICE PRESIDENT, 
                          AI AND IBM Q

    Dr. Gil. Chairman Wicker, Ranking Member Schatz, members of 
the Subcommittee, thank you for inviting me here today. My name 
is Dario Gil, and I am the Vice President of AI and quantum 
computing at IBM.
    The idea of creating a thinking machine is not new, and 
precedes modern computing. Calculating machines were built in 
antiquity and improved throughout history by many 
mathematicians. The term ``artificial intelligence'' was first 
introduced 61 years ago in 1956, and AI, as an academic 
discipline, took off. Three years later, IBM scientist Arthur 
Samuel coined the term ``machine learning'' to refer to 
computer algorithms that learn from and make predictions on 
data by building a model from sample inputs without following a 
set of static instructions.
    One type of machine learning and AI algorithm that has 
gained tremendous attention over the past several years is an 
artificial neural network, notably, deep learning. These 
networks are inspired by the architecture of the human brain, 
with neurons organized as layers, and different layers may 
perform different kinds of operations on their inputs. When 
presented with sample data, neural net can be trained to 
perform a specific task, such as recognizing speech or images. 
Over the last decade, the explosion of digital data and the 
growth in processing speed and power have made it possible to 
use neural nets in real-world solutions.
    While many tend to focus on the automation features of AI, 
we believe its true impact will be felt in humans carrying out 
complex tasks which they cannot do on their own. My prepared 
testimony provides detailed examples on the many ways in which 
IBM's AI platform for enterprise business, Watson, is being 
used to augment human abilities across many industries, from 
strengthening cybersecurity to enhancing the customer 
experience to improving agriculture and optimizing supply 
chains. AI is playing a bigger and bigger role in all realms of 
commerce, and its uses will only grow.
    Now, there is no question the advent of AI will impact the 
nature of jobs, yet history suggests that even in the face of 
technological transformation, employment continues to growth 
with economic expansion and the creation of entirely new jobs 
despite the disappearance of some occupations.
    Jobs are made out of tasks. Those tasks that cannot be 
automated, that cannot be automated by AI, are those in which 
workers will provide the greatest value, commanding higher 
wages and incomes as a result.
    Now, the creation of AI itself will require new job 
categories associated with how we design and train them, how we 
secure them, and verify that they work as planned, and how we 
integrate them into our workflows. The application of AI will 
change our professions, opening up new categories of work and 
increasing demand for some existing professions as we combine 
the capabilities of these AI systems with our own expertise. 
For example, many more cybersecurity professionals will be 
needed to engage with AI systems and act decisively upon the 
threat to information they provide.
    We must address the problem of shortage of workers with the 
skills needed for these and many other roles. A useful example 
in this regard is software programming, which is taught as a 
critical skill in many high schools and colleges. We should 
promote a similar movement for AI techniques, such as machine 
learning.
    To enjoy the full benefits of AI, we'll also need to have 
confidence in the recommendations, judgments, and uses of AI 
systems. In some cases, users will need to justify why an AI 
system produced its recommendations. For example, doctors and 
clinicians using AI systems to support medical decisionmaking 
may require to provide specific explanations for a diagnosis or 
course of treatment, both for regulatory and liability reasons.
    IBM is actively innovating in this field. We're deriving 
best practices for how, when, and where AI algorithms should be 
used. We're creating new AI algorithms that can be trusted, are 
explainable, and are more accurate.
    Of course, no single company can guarantee the safe and 
responsible use of such a pervasive technology. For this 
reason, IBM is a founding member of the Partnership on AI, a 
collaboration with other key industry leaders and many 
scientific and nonprofit organizations. It's focused on the 
formulation of best practices, on AI technologies, and 
advancing the public's understanding of AI.
    In addition, a recently created MIT-IBM Watson AI Lab has 
one of its research pillars advancing shared prosperity with 
AI, exploring how AI can deliver economic and societal benefits 
to a broader range of people, nations, and enterprises.
    In a similar way, we look forward to working closely with 
the Members of Congress to ensure the responsible, ethical, and 
equitable use of AI as this technology continues to evolve.
    [The prepared statement of Dr. Gil follows:]

     Prepared Statement of Dario Gil, Vice President, AI and IBM Q
Introduction
    Chairman Wicker, Ranking Member Schatz, members of the 
Subcommittee. Thank you for inviting me here today. My name is Dario 
Gil and I am Vice President, AI and quantum computing at IBM.
    We have arrived at a remarkable moment in the history of 
information technology. An explosion of data and computation has given 
us access to massive amounts of digitized knowledge. With it, we have 
enormous intelligence and power to see patterns and solve problems we 
never could have before.
    Increasingly, the engine we use to tap into this knowledge is 
artificial intelligence. We can now train algorithms with emerging 
Artificial Intelligence (AI) technologies to learn directly from data 
by example. Moreover, we can do this at scale cost through cloud 
networks to create machines that help humans think. For these reasons, 
AI is the most important technology in the world today.
    The rise of AI has been accompanied by both boundless enthusiasm 
about its ability to transform our lives, and fears it could 
potentially harm or displace us. At IBM, we take a different approach. 
We are guided by the use of artificial intelligence to augment human 
intelligence. We focus on building practical AI applications that 
assist people with well-defined tasks. We believe people working 
collaboratively with these learning systems is the future of expertise.
    In my testimony, I'll provide an overview of AI and describe how 
work in this field has evolved. Then I'll offer some examples of the 
rapidly growing commercial applications of IBM Watson, the best-known 
artificial intelligence platform for enterprise business today. I'll 
also look at how we're beginning to combine AI with other emerging 
technologies, such as blockchain, to optimize business and provide 
trust in transactions. I'll describe how AI will impact the nature of 
work leading to many new and improved job opportunities. Finally, I'll 
examine IBM's position on the responsible and ethical use of AI.
The Evolution of AI
    The idea of creating a `thinking' machine is not new and precedes 
modern computing. The study of formal reasoning dates to ancient 
philosophers such as Aristotle and Euclid. Calculating machines were 
built in antiquity and were improved throughout history by many 
mathematicians. In the 17th century Leibniz, Hobbes and Descartes 
explored the possibility that all rational thought could be made as 
systematic as algebra or geometry.
    In 1950, Alan Turing, in his seminal paper Computing Machinery and 
Intelligence, laid out several criteria to assess whether a machine 
could be deemed intelligent. They have since become known as the 
``Turing test.'' The term ``artificial intelligence'' was first 
introduced in 1956, sixty-one years ago, and AI as an academic 
discipline took off. Three years later, in 1959, IBM scientist Arthur 
Samuel coined the term ``machine learning'' to refer to computer 
algorithms that learn from and make predictions on data by building a 
model from sample inputs, without following a set of static 
instructions.
    An algorithm is simply a set of rules to be followed in 
calculations or other problem-solving operations. It can be as basic as 
the steps involved in solving an addition problem or as complex as 
instructing a computer how to perform a specific task. One type of 
machine learning and AI algorithm that has gained tremendous attention 
over the past several years is an artificial neural network. It has 
been essential to the explosive growth of AI systems today.
    Artificial neural networks are inspired by the architecture of the 
human brain. They contain many interconnected processing units, called 
artificial neurons, which are analogous to biological neurons in the 
brain. Typically, neurons are organized in layers. Different layers may 
perform different kinds of operations on their inputs. When presented 
with sample data, an artificial neural network can be trained to 
perform a specific task, such as recognize speech or images. For 
example, an algorithm can learn to identify images that contain cars by 
analyzing numerous images that have been manually labeled as ``car'' or 
``no car.'' It can then use those results to identify cars in images 
that it has not seen before.
    Even though neural networks and other machine learning algorithms 
were actively researched more than six decades ago, their practical use 
was hindered by the lack of digitized data from which to learn from and 
insufficient computational power. At the time, most data were in analog 
form and not easily available to a computer. Training of the neural 
network algorithm was and remains a computationally intensive process. 
Due to the limitations of processors, it could not be implemented 
effectively.
    Over the last decade, the explosion of digital data, the growth in 
processing speed and power, and the availability of specialized 
processing devices such as graphical processing units (GPUs) have made 
it possible to use artificial neural networks in real-world solutions. 
Today, computation is carried out not only in the cloud and in data 
centers, but also at the edge of the network, in sensors, wearable 
devices, smart phones, embedded electronics, factory machines, home 
devices, or components in a vehicle.
    These conditions have also allowed researchers and engineers to 
create incredibly complex neural networks, called deep learning 
networks. They perform in ways comparable to humans in many tasks. For 
certain tasks, such as speech and image recognition, game playing, and 
medical image classification, these networks can outperform people. 
Today, neural networks are used in a variety of applications, including 
computer vision, speech recognition, machine translation, social 
network analysis, playing board and video games, home assistants, 
conversational devices and chatbots, medical diagnostics, self-driving 
cars, and operating robots.
    In addition to machine learning, AI systems deploy a variety of 
other algorithms and technologies that include knowledge 
representation, machine reasoning, planning and scheduling, machine 
perception (speech and vision), and natural language processing and 
understanding. At IBM, we are actively researching and advancing these 
and other technologies so that we can continue to enhance AI systems.
    We are also envisioning and developing the next-generation 
infrastructure required for increasingly complex AI tasks and 
workloads. This is the physical hardware required to run AI algorithms: 
the processors, servers, databases, storage, data centers and cloud 
infrastructure. When all these pieces are aligned in a way that allows 
algorithms to analyze data with maximum efficiency, we refer to it as 
the ``full stack.''
    By successfully engineering the full stack, we can build AI-powered 
solutions that we can apply to a broad array of societal and industry 
challenges. While many tend to focus on the benefit of automation, we 
believe that AI's true impact will be felt in assisting people's daily 
lives, and by helping us carry out extremely complex tasks we cannot do 
on our own. That includes everything from forecasting the weather, to 
predicting how traffic will flow, to understanding where crops will 
grow the best. AI will also help us research the best combinations of 
compounds for drug development, repurpose chemical structures for new 
medicines, and optimize vastly intricate supply chains. I'd like to 
illustrate this further with a look at how IBM Watson is already being 
used across a range of different industries.
AI applications to industries
    My first example illustrates how AI can assist humans in reacting 
to a problem when there is very little time to react. The security of 
data and on-line transactions is fundamental to the growth of commerce. 
But the simple fact is that most organizations can't keep up with the 
threats. Security experts are too few and overstretched. Sophisticated 
attacks, including those using AI tools, are coming at a rate that 
makes them extremely difficult to stop. Entire networks are compromised 
in the blink of an eye. Watson for cybersecurity allows us to turn the 
tables. It sifts through the insights contained within vast amounts of 
unstructured data, whether it's documented software vulnerabilities or 
the more than 70,000 security research papers and blogs published each 
year. It instantly alerts security experts to relevant information, 
scaling and magnifying human cognition. It also learns from each 
interaction that has an alert, and works proactively to stop continued 
intrusion. Security analysts, armed with this collective knowledge, can 
respond to threats with greater confidence and speed.
    The second example shows how AI is enhancing customer experience. 
Tax preparation is an area ripe for AI solutions. H&R Block is using 
Watson to understand context, interpret intent and draw connections 
between clients' statements and relevant areas of their tax return. 
Watson is working alongside H&R Block Tax Pros as they take clients 
through the tax return process, suggesting key areas where they may 
qualify for deductions and credits. Clients can follow along and 
understand how their taxes are being computed and impacted by numerous 
aspects of the tax code. They can also see the many paths to file a 
return with the IRS, pinpointing the route that delivers a maximum 
refund.
    A third example demonstrates AI's ability to personalize the client 
experience. 1-800-Flowers launched an AI-powered gift concierge powered 
by Watson Conversation. It interacts with online customers using 
natural language. The service can interpret questions, then ask 
qualifying questions about the occasion, sentiment and who the gift is 
for to ensure that suggestions are appropriate and tailored to each 
customer. In this way, the customer can get the right flower for the 
right occasion.
    The next example highlights AI's role in enhancing agricultural 
productivity. A program led by our Research division called IBM PAIRS 
is bringing the power of AI to improve crop yields. It works by 
processing and analyzing vast amounts of geospatial data to generate 
vastly improved long term weather and irrigation forecasts. Using these 
methods, IBM Research and Gallo Winery co-developed a precision 
irrigation method and prototype system that allows Gallo to use 20 
percent less water for each pound of grapes it produces.
    A final example shows AI's ability to optimize supply chains. 
Traditional brick and mortar retailers are under tremendous pressure 
from e-commerce. They must find new, cost-effective and efficient ways 
to deliver goods to buyers in order to stay in business. That means 
offering customers a range of delivery options--pick up in store, ship 
from the nearest store, or move goods seamlessly between store and e-
commerce. Our client, a major American retailer, had to coordinate this 
effort across a thousand stores in their fulfillment chain. Our 
predictive models enabled them to determine optimal distribution across 
their entire chain, factoring in dozens of different variables. Over an 
eight-day period including Black Friday and Cyber Monday, they 
processed more than 4 million orders--a company record--at a savings of 
19 percent per order compared to the prior year. This led to an overall 
savings of $7.5 million dollars.
    From cybersecurity, to customer experience, to personalization, to 
productivity and optimization, AI is playing a bigger and bigger role 
in all realms of commerce. And its uses will only grow.
AI and blockchain
    The potential for AI becomes even greater when combined with other 
emerging technologies such as blockchain. Blockchain stores data in 
shared, secure, distributed ledgers that allow every participant 
appropriate access to the entire history of a transaction using a 
``permissioned'' network--one that is highly secure and can determine 
who can see what. Blockchain holds the promise to be the way we may do 
all transactions in the future.
    A typical AI process could include data collection, use of the data 
to train a neural network and then deployment of the pre-trained model 
to power the application. Blockchain supports the AI process by 
reducing the risk of data tampering and provides data in a form that 
can be used and audited. There's an old saying in the computer industry 
``garbage in, garbage out,'' and that applies to data and how you use 
it. The integrity of the data used as input to the AI model is a 
necessary criterion in ensuring the value and usefulness of the model.
    Because it can process and analyze massive quantities of data, AI 
can use blockchain data to gain valuable insights and detect patterns 
in how supply chains work and processes behave. Over time, this will 
generate a valuable source of clean, trusted transactional data that 
cuts across industries to give us new insights. That includes both 
structured and unstructured data--everything from Internet of Things 
(IoT) information to compliance and geospatial data that's stored on a 
blockchain. AI can use this information to generate valuable insights 
and detect patterns in near-real time, driving new efficiencies across 
business operations.
    For example, IBM Research is working with Everledger, a company 
that tracks and protects diamonds and other valuables. We're using AI 
to analyze digital information on one million diamonds Everledger has 
stored on a blockchain. We can cross-check that data against UN 
regulations to prevent the sale of conflict diamonds. We can verify 
time and date stamps. We can certify laser inscriptions in the girdle 
of the stone. We can perform these analytics directly on the 
blockchain, without the need to extract the data first. This minimizes 
opportunities for data tampering and fraud. While this is a specialized 
application, it shows some of the kinds of data we can collect and 
analyze at huge scale.
    We have also partnered with Walmart to use blockchain and AI 
techniques to ensure food safety. Today's food supply chains are highly 
complex and involve multiple components, stakeholders, and activities. 
This complexity makes it difficult to identify sources of 
contamination, counterfeit substitutions, loss of refrigeration, or 
food transportation safety issues as products move from their sources 
to their consumption by consumers. Blockchain supports traceability by 
tracking the food products from origin to destination and by allowing 
certification of respective transactions and events along the way. AI-
powered technologies are used to analyze this information to help 
ensure food that can be eaten safely.
AI and the Future of Work
    Artificial intelligence will alter the way people work. This has 
been true of many new technologies that have benefited human 
populations over time because they dramatically improved industrial 
output. They have led to fewer grueling jobs. In the process, new types 
of jobs have emerged. However, such disruptive improvements have always 
called for a period of training and adjustment.
    We need to openly understand and recognize this fact, so that we 
can create the right conditions to make this transition as successful 
as possible. As a nation, we need to be prepared to offer the 
appropriate education and support to manage this change well. There's 
no question the advent of artificial intelligence will impact jobs. 
Despite the fear, anxiety, and prediction of massive job loss, history 
suggests that, even in the face of technological transformation, 
employment continues to grow and very few occupations disappear.
    Rather, it is the transformation of occupations that is very likely 
to be widespread that will impact most workers. Occupations are made up 
of tasks. It is the tasks that are automated and reorganized where the 
transformation occurs. Workers will need new skills for the new 
transformed tasks and occupations. But, it is the tasks that cannot or 
will not be automated where workers provide the greatest value, 
commanding higher wages and incomes as a result.
    Some ``new collar jobs'' will emerge--jobs that require advanced 
technical skills but do not necessarily require a full undergraduate 
education. A study by Accenture of more than 1,000 large companies that 
are already using or testing AI and machine-learning systems identified 
the emergence of entire categories of new, uniquely human jobs with no 
precedents.
    For example, ``trainers'' will be required to teach AI systems how 
they should perform. They may write scripts for chatbots, helping them 
to improve their understanding of human communication, or help provide 
labeled data needed to train the algorithm. They may teach AI systems 
how to personalize their recommendations, or show compassion when 
responding to certain problems that require it. Another category of 
``explainers'' will be needed to help convey how AI systems have 
arrived at a decision or a recommendation. They'll monitor the 
operations of AI systems or perform forensic analyses on algorithms and 
make corrections to them if they generate an incorrect result. Earlier, 
I referenced the shortage of qualified cybersecurity professionals. In 
the future, we'll need far more of them to engage with AI systems, 
review the recommendations these systems offer and act decisively upon 
threats.
    There are actions we must take now to ensure the workforce is 
prepared to embrace the era of AI and the ways it will augment our 
economy. To begin, we must address the shortage of workers with the 
skills needed to make advances in AI, create new solutions and work in 
partnership with AI systems. We need to match skills education and 
training with the actual skills that will be required in the emerging 
age of AI.
    At IBM, we have an educational model called P-TECH to train new 
collar workers for a job in technology. P-TECH combines the best of 
high-school, community college, hands-on-skills training and 
professional mentoring, and provides public high school students in 
grades 9-14 a path to post-graduation opportunities in fields aligned 
with the skills American employers are looking for.
    Our goal must be to create multiple pathways like this for more 
people to acquire the skills that will be in demand, as AI use becomes 
more commonplace. We can use the example of the adoption of software 
programming as a critical skill that is taught in many high school and 
colleges. Some colleges require that all students learn how to code 
since they consider it a necessary skill for success. Students becoming 
proficient in programming have a wider range of job opportunities.
    In the future, we may promote and see a similar trend with students 
gaining understanding of and proficiency in AI techniques such as 
machine learning. Preparing more U.S. students and workers for success 
in these well-paying new collar jobs is essential if we want a 
workforce that is ready to capitalize fully on AI's economic promise.
    Let me also say that as well-intentioned as it may seem to some, 
taxing automation will not serve the cause of fostering employment in 
the new AI economy. It will only penalize technological progress. We 
should not adopt measures like this one that will harm America's 
competitiveness.
    Inevitably, people adapt best by finding higher value in new 
skills. Technologies that are easiest to integrate and integrate with 
will be those that improve human productivity. But they should not 
replace human judgment. IBM Watson was designed from the beginning to 
work in concert with human expertise. It will only be successful as 
long as there are people with the right skills to engage with it.
Building trust in AI
    To enjoy the full benefits of AI, we will also need to have 
confidence in the recommendations, judgments and uses of AI systems. 
IBM is deeply committed to the responsible and ethical development of 
AI. Last year, we published one of the first corporate white papers on 
this subject. The paper, which was intended to help launch a global 
conversation, centered around the need for safe, ethical, and socially 
beneficial management of AI systems.
    Trust in automated systems is not a new concept. We drive cars 
trusting the brakes will work when the pedal is pressed. We perform 
laser eye surgery trusting the system to make the right decisions. We 
have automated systems fly our airplanes trusting they will navigate 
the air correctly. In these cases, trust comes from confidence that the 
system will not make a mistake, leveraging system training, exhaustive 
testing, and experience. We will require similar levels of trust for AI 
systems, applying these methodologies.
    In some cases, users of AI systems will need to justify why an AI 
system produced its recommendation. For example, doctors and clinicians 
using AI systems to support medical decision-making may be required to 
provide specific explanations for a diagnosis or course of treatment, 
both for regulatory and liability reasons. Thus, in these cases, the 
system will need to provide the reasoning and motivations behind the 
recommendation, in line with existing regulatory requirements specific 
to that industry. In the European Union, this will be a requirement for 
all automated decision-making AI systems as of May 2018.
    These safeguards can also help to manage the potential for bias in 
the decision-making process, another important concern with AI. Bias 
can be introduced both in the datasets that are used to train an AI 
system and by the algorithms that process that data, and how people 
interpret and communicate the discerned insights. Our belief is that 
the data and algorithmic aspects can not only be managed, but also that 
AI systems themselves can help eliminate many of the biases that 
already exist in human decision-making models today.
    At the beginning of this year, IBM issued principles for 
transparency and trust to guide our development and use of AI systems. 
In summary, they state the following:

   We believe AI's purpose is to augment human intelligence

   We will be transparent about when and where AI is being 
        applied, and about the data and training that went into its 
        recommendations.

   We believe our clients' data and insights are theirs.

   We are committed to helping students, workers, and citizens 
        acquire the skills to engage safely, securely, and effectively 
        with cognitive systems, and to do the new kinds of work that 
        will emerge in an AI economy.

    In the same way that we are committed to the responsible use of AI 
systems, we are committed to the responsible stewardship of the data 
they collect. We also believe that government data policies should be 
fair and equitable and prioritize openness.
    IBM is actively innovating in this field. We are deriving best 
practices for how, when, and where AI algorithms should be used. We are 
creating new AI algorithms that are more explainable and more accurate. 
We are working on the algorithmic underpinnings of bias and AI, such as 
creating technologies that can identify and cleanse illegal biases from 
training datasets.
    Of course, no single company can guarantee the safe and responsible 
use of such a pervasive technology. For this reason, IBM is a founding 
member of the Partnership on AI, a collaboration with other key 
industry leaders and many scientific and nonprofit organizations. Its 
goal is to share best practices on AI technologies, advance the 
public's understanding, and serve as an open platform for discussion 
and engagement about AI and its influences on people and society.
    AI has enormous transformative power. Much has been said about its 
potential to transform sectors and industries. However, AI is also 
giving us a technological toolkit to address many societal challenges. 
At IBM we are committed to pioneering new solutions, and showcasing and 
promoting the opportunities to use AI in social good applications. 
Three years ago, we launched the AI for Social Good program and have 
executed a number of AI for Good projects, from using AI to understand 
patterns of opioid addiction, to prototyping recommendation systems 
that would aid low-income individuals and help them stay out of 
poverty, to applying machine learning to understand transmission 
mechanisms of the Zika virus.
    Earlier this year, we announced the MIT-IBM Watson AI Lab, a 
partnership with Massachusetts Institute of Technology (MIT) to carry 
out fundamental AI research. One of the research areas for the lab is 
focused on advancing shared prosperity through AI--exploring how AI can 
deliver economic and societal benefits to a broader range of people, 
nations and enterprises.
    Lastly, no discussion of the future of AI would be complete without 
acknowledging the critical role of government. Public investment and 
policy support have been the twin pillars of American global 
technological leadership for the past half-century. We hope and expect 
the same will be true in the coming age of AI. For this reason, we 
enthusiastically welcome the interest and support of the United States 
Senate as this technology continues to evolve. Together, we can ensure 
that AI serves people at every level of society and advances the common 
good.

    Senator Wicker. Thank you very much.
    Dr. Felten.

           STATEMENT OF DR. EDWARD W. FELTEN, Ph.D.,

          ROBERT E. KAHN PROFESSOR OF COMPUTER SCIENCE

            AND PUBLIC AFFAIRS, PRINCETON UNIVERSITY

    Dr. Felten. Chairman Wicker, Ranking Member Schatz, and 
members of the Subcommittee, thank you for the opportunity to 
testify today.
    Progress in AI has accelerated over the last decade. 
Machines have met and surpassed human performance on many 
cognitive tasks, and some longstanding grand challenge problems 
in AI have been conquered.
    Recent experience during this time teach us some useful 
lessons for thinking about AI as a developing technology.
    First, AI is not a single thing; it's different solutions 
for different tasks. Success has come in ``narrow AI,'' which 
applies a toolbox of specific technical approaches to craft 
solutions for specific applications. There has been a lot less 
progress on general AI, which tries to create a single, all-
purpose artificial brain, like we see in the movies.
    Second, successful AI doesn't think like a human. If it is 
an intelligence, it is sort of an alien intelligence. AI and 
people have different strengths and weaknesses, so teaming up 
with AI is promising if we can figure out how to work with an 
intelligence different from our own.
    And, third, more engineering effort or more data translates 
into better AI performance. Progress requires a lot of hard 
work by experts, and that's why our AI workforce is so 
important.
    The strategic importance of AI to the United States goes 
beyond its economic impact to include cybersecurity, 
intelligence analysis, and military affairs as well.
    The U.S. is currently the world leader in AI research and 
applications, but our lead is not insurmountable. Countries 
around the world are investing heavily in AI, so our industry 
researchers and workforce need support in their efforts to 
maintain American leadership in this area. American companies 
recognized the potential of AI early on and have been investing 
and moving aggressively to hire top talent.
    Our lead in research and development is less secure. 
Federal funding for AI research has been relatively flat. 
Aggressive hiring by industry has thinned the ranks of the 
academics who train the next generation of researchers. 
Industry does a lot of valuable research, but the public 
research community also plays an important role in basic 
research and in training young researchers, so investments in 
policies to support and grow the public research community are 
important.
    Policies to enhance access to high-quality education for 
all American children, especially in computing, lay the 
foundation for our future workforce. And America has always 
been a magnet for talent from around the world, and that has to 
continue if we are to retain our leadership.
    The many benefits of AI are tempered by some challenges. AI 
systems may pose safety risks, they may introduce inadvertent 
bias into decisions, and they may have unforeseen consequences. 
Much of the criticism of AI has centered on the risk of 
inadvertent bias, and real-world examples of biased AI are well 
documented.
    The good news is that there are technical ways to eliminate 
bias. Developers can improve datasets to be more representative 
of the population, they can use algorithms that are more 
resistant to bias. Promising results on debiasing both data and 
algorithms are emerging from the research community, and that 
research should continue to be supported because it points a 
way to deploying AI more widely with less concern about bias.
    In considering the risks of AI, it's important to remember 
that the alternative to relying on AI is to rely on people, and 
people are also at risk of error and bias. In the long run, AI 
systems will devise complex data-driven strategies to pursue 
goals, but people will continue to decide which goals the 
system should pursue. To better hold AI systems accountable, we 
need new technologies and new practices to connect AI with the 
human institutions that will govern it.
    Regarding regulation, there is no need to create special 
regulations just for AI at this time. In sectors that are 
already regulated, the existing regulations are already 
protecting the public, and regulators need only consider 
whether and how to adjust the existing regulations to account 
for changes in practices due to AI. For example, the Department 
of Transportation, in the previous administration and this one, 
has been adapting vehicle safety regulation to enable safe 
deployment of self-driving cars.
    Government agencies have important roles to play beyond 
regulation. More expertise, advice, and coordination is needed 
across the government to help agencies decide how to adapt 
regulations and use AI in their operations. New structures and 
new policies to strengthen this expertise would be very 
beneficial.
    With good policy choices and the continued hard work and 
investment of American companies, researchers, and workers, AI 
can improve the health and welfare of Americans, boost 
productivity and economic growth, and make us more secure.
    Americans currently lead the world in AI. We should not 
step on the brakes; instead, we should reach for the 
accelerator and the steering wheel.
    Thank you for the opportunity to testify today.
    [The prepared statement of Dr. Felten follows:]

  Prepared Statement of Edward W. Felten Robert E. Kahn Professor of 
       Computer Science and Public Affairs, Princeton University
    Chairman Wicker, Ranking Member Schatz, and members of the 
Committee, thank you for inviting me to speak today about how best to 
realize the benefits of artificial intelligence.
Artificial Intelligence (AI) and Machine Learning (ML)
    Artificial intelligence (AI) and machine learning (ML) have been 
studied since at least 1950. There has been an unexpected acceleration 
in technical progress over the last decade, due to three mutually 
reinforcing factors: the availability of big data sets, which are 
analyzed by more powerful algorithms, enabled by faster computers. In 
recent years, machines have met and surpassed human performance on many 
cognitive tasks, and some longstanding grand challenge problems in AI 
have been conquered.
    Industry has recognized the rise of AI as a technical shift as 
important as the arrival of the Internet or mobile computing. Companies 
around the world have invested heavily in AI research and development, 
and leaders of major companies have described adoption of machine 
learning as a bet-the-company opportunity.
    The strategic importance of AI/ML to the United States goes beyond 
its economic impact. These technologies will also profoundly affect the 
future of security issues such as cybersecurity, intelligence analysis, 
and military affairs.
    Fortunately, the United States is currently the world leader in AI/
ML research, development, and applications, in both the corporate and 
academic spheres. Our national lead is not insurmountable, however. 
Countries around the world are investing heavily in AI/ML, so our 
scientists, engineers, and companies need support in their efforts to 
maintain American leadership.
The Nature of AI/ML Today
    The history of AI teaches some important lessons that are useful in 
considering policy choices.
    AI is not a single thing--it is different solutions for different 
tasks. The greatest progress has been in ``narrow AI,'' which applies a 
toolbox of specific technical approaches to craft a solution specific 
to one application or a narrow range of applications. There has been 
less progress on ``general AI,'' which strives to create a single, all-
purpose artificial brain that could address any cognitive challenge and 
would be as adaptive and flexible as human intelligence. Indeed, there 
is no clear technical path for achieving general AI, so it appears that 
for at least the next decade the policy focus should be on the 
implications of narrow AI.
    In a world of narrow AI, there will not be a single moment at which 
machines surpass human intelligence. Instead, machines may surpass 
human performance at different times for different cognitive tasks; and 
humans might retain an advantage on some cognitive tasks for a long 
time. Even if machines surpass humans in the lab for some task, 
additional time and effort would need to be invested to translate that 
advance into practical deployment in the economy.
    Successful AI does not think like a human--if it is an 
intelligence, it is an alien intelligence. Because AI solutions are 
task-specific and do not directly mimic the human brain, AI systems 
tend to ``think'' differently than people. Even when successful, AI 
systems tend to exhibit a different problem-solving style than humans 
do. An AI system might handle some extremely complex situations well 
while failing on cases that seem easy to us. The profound difference 
between human thinking and AI operation could make human-AI teaming 
valuable, if the strengths of people and machines can complement each 
other. At the same time, these differences create challenges in human-
AI teaming because the teammates can have trouble understanding each 
other and predicting their teammates' behavior.
    On many cognitive tasks, more engineering effort or more data 
translates into better AI performance. Many AI systems learn from data. 
Such systems can be improved by re-engineering them to learn more from 
the available data or by increasing the amount of data available for 
training. Either way, devoting more effort to engineering and operating 
an AI system can improve its performance. Machines are generally worse 
than humans at learning from experience, but a machine with a very 
large data set has much more ``experience'' from which to learn. Using 
the narrow AI approaches that have been successful so far, expert AI 
developers must invest significant effort in applying AI to each 
specific task.
Benefits of AI/ML
    AI is already creating huge benefits, and its potential will only 
grow as the technology advances further.
    For example, AI is a key enabler of precision medicine. AI systems 
can learn from data about a great many patients, their treatments, and 
outcomes to enable better choices about how to personalize treatment 
for the particular needs, history, and genetic makeup of each future 
patient.
    AI is also enabling self-driving cars, which will eventually be 
much safer than human drivers, saving thousands of American lives every 
year. Self-driving vehicles will improve mobility for elderly and 
disabled people who cannot drive and will lower the cost and increase 
the convenience of transporting people and goods.
    Given the tremendous benefits of AI in these and other areas and 
the likelihood that the technology will be developed elsewhere even if 
the United States does not lead in AI, it would be counterproductive to 
try to stop or substantially slow the development and use of AI. We 
should not ask the industry and researchers to slam on the brakes. 
Instead, we should ask them to use the steering wheel to guide the 
direction of AI development in ways that protect safety, fairness, and 
accountability.
Policies to Support AI Progress
    America's leadership in AI has been driven by three factors: our 
companies, our researchers, and our talented workforce.
    American companies recognized the potential of AI early on and have 
been investing heavily in AI and moving aggressively to hire top 
talent. This is the area in which our national leadership in AI seems 
safest, at least in the short run. In the longer run, however, industry 
must be able to work with world-leading American researchers and 
workforce to sustain its advantage.
    Our lead in research and development is less secure. Federal 
funding for AI research and development has been relatively flat, even 
as the importance of the field has dramatically increased. Aggressive 
hiring by industry has thinned the ranks of the academic researchers 
and teachers who are needed to train the next generation of leaders. 
Although industry has carried out and supported a great deal of 
research, it cannot and does not cover the full spectrum. The public 
research community plays an important role in basic research, in 
research areas such as safety and accountability, and in training young 
researchers, so investments and policies to support and grow that 
community are a key enabler of continued American leadership.
    The foundations of the future workforce are laid in our K-12 
schools. Policies to enhance access to high-quality education for all 
American children, especially in computing, can grow the number of 
students who enter higher education eager and able to pursue studies in 
technical fields such as AI.
    The American AI workforce has also been boosted immeasurably over 
the years by the attractiveness of our universities and industry to the 
most talented people from around the world. America has been a magnet 
for talent in AI and other technical fields, and that must continue if 
we are to retain our leadership. Policies to ensure that America 
remains an attractive place for foreign-born experts to live, study, 
work, and start companies are among the most important steps for the 
future health of our AI enterprise.
Risks and Challenges of AI/ML
    The benefits of AI are tempered by some risks and challenges: AI 
systems may pose safety risks; they may introduce inadvertent bias into 
decisions; and they may suffer from the kinds of unforeseen 
consequences brought on by any novel, complex technology. These are 
very serious issues that require attention from policymakers, AI 
developers, and researchers.
    Much of the criticism of AI/ML systems centers on the risk that 
adoption of AI/ML will lead inadvertently to biased decisions. There 
are several ways this could happen. If a system is trained to mimic 
past human decisions, and those decisions were biased, the system is 
likely to replicate that bias. If the data used to train a system is 
derived from one group of people more than another, the result may 
serve the overrepresented group to the detriment of the 
underrepresented group. Even with ideal data, statistical artifacts can 
advantage larger groups to the detriment of smaller ones. Real-world 
examples of these sorts of biases are well-documented.
    The solution is not to stop pursuing AI, but rather to take steps 
to prevent and mitigate bias. Practitioners should work to improve 
their data, to ensure that datasets are representative of the 
population and do not rely on past biased decisions. They should also 
improve their algorithms by developing and using AI systems that are 
more resistant to bias, so that even if flaws remain in the data, the 
system can produce results that are more fair. In both areas, data 
improvement and algorithm improvement, the research community is 
producing promising early results that will improve the anti-bias 
toolkit available to practitioners. A robust national AI research 
effort should include studies of algorithmic bias and how to mitigate 
it.
    In considering the risks of bias and accountability in AI, it is 
important to remember that in most cases the alternative to relying on 
AI is to rely on human decisions, which are themselves at risk of 
error, bias, and lack of accountability. In the long run, we will 
likely rely much more on algorithms to guide decisions, while retaining 
the human role of determining which goals and criteria should guide 
each decision.
Accountability, Transparency, and Explainability
    The importance of the decisions now made or assisted by AI/ML 
systems requires that the systems and their operators are accountable 
to managers, overseers, regulators, and the public. Yet accountability 
has proven difficult at times due to the complexity of AI systems and 
current limitations in the theory underlying AI. Improving practical 
accountability should be an important goal for the AI community.
    Transparency is one approach to improve accountability. Disclosing 
details of a system's code and data can enable outside analysts to 
study the system and evaluate its behavior and how well the system 
meets the goals and criteria it is supposed to achieve. Full 
transparency is often not possible, however. For example, a system's 
code might include valuable trade secrets that justify withholding 
aspects of its design, or the data might contain private information 
about customers or employees that cannot be disclosed.
    Even where transparency is possible, it is far from perfect as an 
accountability mechanism. Outside analysts may have limited practical 
ability to understand or test a system that is highly complex and meant 
to operate at very large scale. Indeed, even the designers of a system 
may struggle to understand the nuances of its operation. Computer 
science theory says that examining a system beforehand cannot hope to 
reveal everything the system will do when it is exposed to real-world 
inputs. So transparency, though useful, is far from a complete solution 
to the accountability problem.
    Another approach to accountability is inspired by the field of 
safety engineering. The approach is to state clearly which safety, 
fairness, or compliance properties a system is designed to provide, as 
well as the operating conditions under which the system is designed to 
provide those properties. This is backed up with detailed evidence that 
the system will have the claimed properties, based on a combination of 
design reviews, laboratory testing, automated analysis tools, and 
safety monitoring facilities in place during operation. Rather than 
revealing everything about how the system works, this approach focuses 
on specific safety, fairness, or compliance requirements and allows 
system developers to use the full range of technical tools that exist 
for ensuring reliable behavior, including the tools that the system 
developers will already be using internally for quality control.
    Much needs to be done to make this approach feasible for routine 
use. Research can develop and test different approaches to proving 
behavioral properties of systems. Professionals can convene to develop 
and pilot best practices and standards. The overarching challenge is to 
understand how to relate the technical process of engineering for 
reliable operation to the administrative processes of management, 
oversight, and compliance.
Regulation and the Role of Government Agencies
    There is no need to create special regulations for AI. Where AI is 
used in sectors or activities that are already regulated, the existing 
regulations are already protecting the public and regulators need only 
consider whether and how to adjust the existing regulations to account 
for changes in practices due to AI.
    For example, the Department of Transportation (DOT) and National 
Highway Traffic Safety Administration (NHTSA) have taken useful steps, 
under the previous and current Administrations, to clarify how existing 
safety regulations apply to self-driving vehicles and how Federal 
safety regulations relate to state vehicle laws. These changes will 
serve to smooth the adoption of self-driving vehicles which, once they 
are mature and widely adopted, will save many thousands of lives.
    Similarly, the Federal Aviation Administration (FAA) has been 
striving to adapt aviation regulations to enable safe, commercial use 
of unmanned aerial systems (UAS, or ``drones''), which have benefits in 
many sectors, such as agriculture. The FAA has taken some steps to 
increase the flexibility to use UAS commercially, but the interagency 
process on UAS has been moving slowly. Agencies should be urged to work 
with the FAA to advance this important process.
    Government agencies have important roles to play beyond regulation. 
For example, the National Institute of Standards and Technology (NIST) 
and the Department of Commerce can contribute by setting technical 
standards, codifying best practices in consultation with the private 
sector, and convening multi-stakeholder discussions, much as they have 
done in the area of cybersecurity.
    All agencies should consider how they might use AI to better 
accomplish their missions and serve the American people. AI can reduce 
costs, increase efficiency, and help agencies better target their use 
of taxpayer dollars and other limited resources. The National Science 
and Technology Council's subcommittee on Machine Learning and AI can 
serve as a focal point for interagency coordination and sharing of 
ideas and best practices.
    With good policy choices and the continued hard work and investment 
of American companies, researchers, and workers, AI can improve the 
health and welfare of Americans, boost productivity and economic 
growth, and make us more secure. Americans currently lead the world in 
AI. We should not step on the brakes. Instead, we should reach for the 
accelerator and the steering wheel.
    Thank you for the opportunity to testify. I look forward to 
answering any questions.

    Senator Wicker. Great. Thank you so much.
    Let me start with Dr. Gil. Machine learning and artificial 
intelligence capabilities are accelerating, and I think 
Americans listening today or maybe insomniacs listening at 2 in 
the morning 2 weeks from now on C-SPAN, know that we use AI for 
social media and online search queries and smartphone apps. Ms. 
Espinel mentioned health diagnosis, and I think she grabbed our 
attention there.
    What other industries, Dr. Gil, stand to benefit the most 
from what we're talking about today?
    Dr. Gil. I believe actually artificial intelligence is 
going to touch every profession and every industry, but just to 
give some concrete examples.
    Senator Wicker. OK, good.
    Dr. Gil. From security and cybersecurity, there's a class 
of category of problems that have to do with responding with 
low latency. So the nature of the problem that one has to 
address is too complex, there are too many dimensions of it, 
and AI can assist a professional detect a threat and be able to 
assess the proper response. So sometimes it has to do with how 
much time one has to make a decision, and can you be assisted?
    In other areas in which, for example, in the health care 
profession, even though we may have more time in some occasions 
to perform a diagnosis or select a treatment, just the sheer 
complexity of the number of documents, or in this case, it 
could be genomic information that comes into play, goes beyond 
the expertise that any given person can have. So in that 
instance, it can assist, for example, in the process of, you 
know, medical diagnosis, as an example.
    In agriculture, here we're talking about being able to 
integrate sensory measurements from soil and weather data, 
satellite data, to be able to predict what kind of irrigation 
one may have to deploy to improve the productivity of it, to 
give another example.
    So I think that, you know, in every situation where we're 
trying to integrate knowledge that comes from sometimes 
measurements from the physical world and also bodies of 
evidence that we've accumulated through our expertise, combined 
with our own expertise, we facilitate every worker and 
professional to be able to make those better decisions.
    Senator Wicker. OK. Now, Ms. Espinel and to all members of 
the panel, Ms. Espinel says that our government needs to talk 
about three changes to policy: open data, more government 
research, and prioritizing education and workforce development. 
Also, Dr. Felten says there's no need to create special 
regulations for AI.
    Who wants to comment on this? Does anybody want to take 
issue that we have or is everybody in agreement with all four 
of these statements? Is there some room for nuances and 
disagreements? Does anybody want to respond?
    Yes, Mr. Castro.
    Mr. Castro. Thanks. So I think one of the most important 
things to look at in this space is it's about technology 
adoption. When we're looking at AI, the big question for the 
United States in terms of competitiveness is, are we going to 
be the lead in terms of adopting this technology rapidly in our 
industries before other countries do it in their industries? 
Because that's going to determine U.S. competitiveness long 
term.
    So when we're talking about policies in this space, there 
are two types. There are the policies that help accelerate 
deployment and adoption of the technology and our R&D in this 
space, and then there are the regulations that might slow down 
adoption or, you know, kind of skew or realign how we do this 
adoption.
    So when we're comparing ourselves to Europe, for example, 
which is also pursuing this, we have to ask the question: Why 
aren't we doing what Europe is doing to accelerate adoption? 
And, two, are we having smart regulations that allow us to 
apply it in our industries kind of better, smarter, and faster 
than they're doing?
    Senator Wicker. Do you sort of agree with Dr. Felten, that 
regulations may actually impede our development of AI?
    Mr. Castro. I think in most cases, regulation that's 
focused on AI specifically is probably misguided. If there's a 
problem there, we need to look broader and say, Why is this 
problem happening? And is it caused in human situations as 
well?
    Senator Wicker. Dr. Felten, have I mischaracterized 
anything you've said?
    Dr. Felten. No, you have not. I would agree with Ms. 
Espinel's points. With regard to regulation, in addition to 
sectoral regulation, there's an important role for agencies 
sometimes to create new regulatory structures to allow more 
activity, as the FAA has been working to do with drones. The 
FAA has been working to create new rules which allow broader 
commercial use of drones in the United States. And so although 
that is a change to regulation, it's one that enables more 
activity by the commercial sector.
    Senator Wicker. Thank you very much.
    Senator Schatz.
    Senator Schatz. Thank you, Mr. Chairman.
    Thank you for all of your testimony.
    The way I see this is there's a competitiveness side of the 
ledger, which I think is not easy to do, but relatively 
straightforward. And Ms. Espinel's testimony spells out some of 
the steps that we can take on a bipartisan basis to make sure 
that we win that race internationally. That part is, again, not 
easy, but relatively straightforward morally and as a matter of 
policy.
    Where I think it does get difficult is that I think I 
quibble with you, Mr. Castro, in the sense that I don't think 
we're--that the endgame here is just that we race as fast as we 
can in all sectors without regard to consequences and view any 
regulatory effort as contradictory to our national goals. I 
think part of what we have to do is recognize that in areas 
like health care and agriculture, it's a pretty much unalloyed 
good to have more data, to save lives, to make agriculture more 
productive, for instance. In defense, it's a little tricky. In 
criminal justice and policing, it's extremely tricky.
    And so I don't think anybody in the Senate is talking about 
European style regulation. I think what we are saying is that 
this is a complex area that's going to revolutionize society 
and the way we interact with each other, the way machines 
interact with each other, and with ourselves, and if we're not 
careful, we could enshrine some of our thorniest societal 
problems in datasets. And I'm not as persuaded, Dr. Felten, as 
maybe you are, that we can program our way out of that.
    And one of the challenges that I would just maybe ask, if 
we can start with Dr. Felten and go down the line, I'm worried 
about diversity in the industry. I think that to the extent 
that you have software engineers and decisionmakers both at the 
line level writing the code, but all the way up to project 
management and the people who are wrestling with some of these 
moral questions, are mostly white men, and I think that's not a 
trivial thing because they're not thinking about biases in 
policing. They may be thinking differently about autonomous 
weapons.
    And so I'm wondering how you view--and I don't think this 
is a place for regulation, but I do think this is a place for 
us, as a society, to grapple with if this is going to be 
transformational and change everything, is it fair, is it 
rational, to have only, or I should say predominantly, white 
men in charge of setting up these algorithms that most of the 
rest of society can't even access because it's all proprietary?
    Dr. Felten.
    Dr. Felten. Sure. With respect to the question of whether--
of the role of technology in addressing these issues of bias, I 
think technology has an important role to play, but it can't be 
the entire solution, as you suggested, Senator. What we need is 
a combination of institutional oversight and governance with 
technology. Technology can provide the levers, but we still 
need institutions that are able to pull those levers to make 
sure that the results are conducive to what our institutions 
and our society wants.
    With respect to the question about diversity in the 
workforce, this is certainly an issue. The AI workforce is even 
less diverse than the tech workforce generally. And it's 
important to take efforts to improve that so we can put our 
whole team on the field as a nation. And I commend groups like 
AI for All that are working on that to you.
    Senator Schatz. Dr. Gil and then Ms. Espinel, we have a 
minute and 20 seconds left.
    Dr. Gil. OK. Perhaps I could address the topic of bias 
associated with these models because bias can be introduced at 
the level of the dataset, as you properly pointed out, if the 
data that has been collected, you know, is not representative 
of the whole population, in this case, it's to make the right 
assessments. You can introduce then a bias in the----
    Senator Schatz. Well, just to be clear, it can be 
empirically valid, right? I mean, the simplest example is 
here's where crimes have been committed in the past, right? It 
turns out Ferguson, right? Load that into the dataset, it's a 
predictive algorithm, works every time. You go over there, you 
find more and more crime, and it spins and spins. And then on 
top of it, enshrining that bias in an algorithm, I'm not sure 
that--it gets more permanent than it would be if it were up to 
the individual judgment of sheriffs, VAs, people.
    Dr. Gil. Yes. So this is a very active field of research in 
which we're very active that actually has also to do with in 
the cases in which you may have high degrees of prediction, but 
you're incorporating protected variables or protected 
individuals in the case of the assessment that you cannot use 
because of law, that is a variable that you cannot incorporate.
    So there are ways actually to perform the data science to 
be able to achieve, you know, a prediction. I'm not----
    Senator Schatz. But if a police department like deals with 
a vendor, and they say, ``We've got a predictive algorithm, and 
we can't show you how this predictive algorithm works,'' but in 
back of that, you know, inside of that black box, they've got 
census block stuff, they've got all kinds of stuff that you 
would be not allowed to use in policing, how do we even know?
    Ms. Espinel, how worried should I be about this?
    Ms. Espinel. So we talk about bias, but I would actually--I 
would love to address that as well because it's an issue I'm 
really passionate about and focused on, and I know we're 
running low on time here, but you also have sort of led into 
explainability and accountability, and that's really important, 
too. So I will try to briefly touch on both and I'm happy to 
continue this conversation.
    Senator Schatz. Yes.
    Ms. Espinel. So in terms of bias, I think there are really 
two parts of it. So part of it is how the AI systems are 
trained, how they're built, in essence. And, obviously, as you 
point out, data can be inaccurate or it can be incomplete, or 
it can have its own biases that are going to skew the outcomes. 
And there are a number of things that can be done right now to 
try to help with that.
    So part is making sure that data scientists that are 
building them have the tools and the training to try to counter 
that. Second, and you already raised this, I think this is 
another area, or you have another reason why diversity in tech 
is so important, and I think the more experience and background 
you have at the table as AI systems are being trained, the more 
helpful it's going to be to try to avoid that.
    I think, third, as has been mentioned, there is a lot of 
research going on in this area, so continuing to support and 
invest in research that will help lessen the chances of bias in 
AI is very important.
    And last I would say, you know, to the extent bias is 
discovered, obviously companies should be working immediately 
to try to address that.
    So I think there are a number of things that can and should 
be happening now to try to counter that.
    I think there's another part of this discussion, which 
we've heard less about, which I think is really important, 
which is how AI can be used, not trained and built, but how it 
can be used to try to counter bias and to try to broaden 
inclusion. And there are a number of really interesting 
examples here both in terms of hiring practices and in terms of 
broadening inclusion for people with diseases like--or people 
with conditions like autism or people that are visually 
impaired where AI can dramatically transform their ability to 
interact with society and in workplaces. And so I think having 
more discussion about how AI can and should be used to try to 
lessen bias and to try to broaden inclusion is very important.
    I could talk more. I know we're running low on time.
    Senator Schatz. I think my time is up. I'll let the other 
members ask their questions.
    Senator Wicker. Senator Schatz's time is close to expiring.
    [Laughter.]
    Senator Wicker. But we'll take another round.
    Senator Moran.

                STATEMENT OF HON. JERRY MORAN, 
                    U.S. SENATOR FROM KANSAS

    Senator Moran. I have four questions, and I'm glad to know 
that the standard has now increased beyond the 5 minutes.
    [Laughter.]
    Senator Moran. Let me quickly try to ask these four 
questions. First about research and development in the Federal 
role. A number of us on this Committee are members of the 
Appropriations Committee. We would think of opportunities to be 
supportive of this endeavor by funding of NSF, NIH, DoD. What 
am I missing? Is there something out there that we ought to be 
paying attention to from an appropriations process that 
supports Federal research in this regard to AI? Don't pause 
very long here.
    Dr. Gil. No, well, in addition to the agencies you listed, 
I think the DOE also has an important element of it in the 
intersection of high-performance computing and artificial 
intelligence and how the computational platform to support 
historical approaches to be able to do things like modeling of 
chemical processes and sophisticated approaches to do that to 
combine it with the more statistical approaches that are being 
enabled now with artificial intelligence and machine learning. 
So I think being able to combine those two disciplines in the 
context of the DOE would be very helpful as well.
    Senator Moran. Thank you.
    The significance--the difference between private research, 
business research, and government research, where do we see the 
focus? Let me say, where are the most dollars being spent? Is 
the private sector more engaged than the Federal Government?
    Dr. Gil. I would say at this point it would be fair to say 
that in the private sector, certainly in the technology world, 
AI is the single most important technology in the world today. 
So--and the levels of investment that we're all making around 
that is commensurate with that statement.
    Senator Moran. Yes.
    Dr. Felten. The private sector investment in AI research 
and development is much larger than the Federal investment 
currently. I would agree with the list of agencies that you 
listed, Senator, and that Dr. Gil listed. And I'd also commend 
to you the ``National AI Research and Development Strategic 
Plan'' that was published last October that was put together by 
the research agencies.
    Senator Moran. Thank you very much.
    As we attempt to promote STEM education, in that broad 
phrase of ``STEM,'' is that sufficient to describe the kind of 
intellectual and academic excellence that we need in order to 
develop AI? Is it something more than just promoting science, 
mathematics, engineering, and research, the traditional kinds 
of STEM things, as we pursue support of education?
    Ms. Espinel.
    Ms. Espinel. So that is very important, and I think trying 
to ensure that every child in every state, if they want to go 
into tech, they have the skills to do that, and that's a real 
viable, realistic career opportunity for them. I think that's 
critical.
    I think there are other things that could be very helpful 
as well. So one of the things that we've been thinking about is 
trying to modernize vocational training. So for, you know, not 
just for--important for very young children learning as well, 
but as young adults are coming out of school, then thinking 
about where their career path could take them, I think there's 
a lot that could be done to try to improve those programs as 
well.
    Senator Moran. If anyone has suggestions in addition to the 
short time-frame that I have, that Senator Schatz didn't have, 
please let me know. We'd like to figure out how we focus our 
educational support in a way that adds this new dimension or 
additional dimension to what kids in grade school, middle 
school, are learning. They're the future.
    Ms. Espinel. That's fantastic. Thank you.
    Senator Moran. As a Kansan, I need to ask a question about 
agriculture. This could be Dr. Gil, Ms. Espinel, Dr. Castro, 
or, well, really any of you, Dr. Bethel, where is the research 
taking--who are the leaders in research when it comes to 
agriculture? Is it the universities or is it the private sector 
again who is focused on large data and what it can mean to 
increase productivity and efficiency and better on-the-farm 
income? Who should I be talking to that's fully engaged in this 
world?
    Dr. Gil. Yes. I think there is wonderful work going on 
across a number of universities, and we can--I can give you 
some more details of some specific programs that are, you know, 
very well tailored to this. But certainly in the private sector 
there is a lot of activity that has had to do with focusing in 
the instrumentation aspect and the measurement of fields, 
particularly with satellite data, being able to combine also 
very unique datasets.
    As we've aggregated datasets in terms of I alluded before 
in terms of like soil characteristics, you know, 
evapotranspiration models that we have, weather data, and be 
able to combine all these layers of data together to be able to 
have accurate forecasts and prediction, and to the degree that 
we have more autonomy also in the agricultural fields to be 
able to, you know, irrigate with more precision--right?--or be 
able to use fertilizers with more precision as well, the 
combination of all of those factors is what is increasing 
productivity and what it's enabled to do that.
    Senator Moran. Dr. Felten, my final question. You indicate 
in your testimony there is no clear technical path for 
achieving general AI, you have narrow and general. Is what 
you're telling me is that that's more science fiction, like 
more of a James Bond movie than where we are today?
    Dr. Felten. That is the case today. We can't rule out the 
possibility that general AI may come along far in the future, 
but from a policymaking standpoint, narrow AI is what we have, 
and it, I think, should be setting the agenda. We should be 
alert for the possibility that sometime down the road general 
AI may come, but it's not close.
    Senator Moran. Thank you all very much.
    Thank you, Mr. Chairman.
    Ms. Espinel. Senator Moran, if I may, you mentioned 
advances in farming technology, and since you are from Kansas, 
I just wanted to let you know that we put out a study earlier 
this year looking at the impact of software across the United 
States, and Kansas is one of the states we saw the biggest jump 
in jobs. So over 30 percent growth in software jobs in Kansas, 
and you're up to nearly 40,000, and part of that is farming 
technology, but it is other types of software services as well. 
So Kansas is doing great.
    Senator Moran. I wouldn't want to forget the aviation world 
that we live in, too, in Kansas. Thank you very much.
    Senator Wicker. Senator Peters.
    Senator Peters. Thank you, Mr. Chairman.
    Senator Wicker. How's Michigan doing?
    Senator Peters. Yes, how's Michigan doing?
    Ms. Espinel. Michigan is doing really, really well.
    Senator Wicker. But not as good as Kansas, I'm sure.
    [Laughter.]
    Ms. Espinel. Well, Kansas did see a huge jump, but 
Michigan, $13 billion in GDP from software into Michigan. And 
Michigan is definitely doing better in jobs overall in terms of 
numbers, maybe not quite as big a jump year to year, but 
Michigan is a really--is a really strong state, and our----
    Senator Wicker. And I guess New Mexico really isn't even in 
the game.
    [Laughter.]
    Ms. Espinel. We actually had an event in Detroit last week 
through our foundation talking about software and tech, and 
very focused on the educational system in Michigan and the 
great things that Michigan is doing to try to advance software 
and technology in the state, so thank you.
    Senator Wicker. Senator Peters.

                STATEMENT OF HON. GARY PETERS, 
                   U.S. SENATOR FROM MICHIGAN

    Senator Peters. Thank you, Chairman Wicker, for bringing up 
that question so I didn't have to use my time for that. That 
was very well done.
    [Laughter.]
    Senator Peters. And you're right, Michigan is moving very 
aggressively in this area, and it's primarily driven by self-
driving cars and what's happening in that space, which is very 
exciting, something that I've been intimately involved in over 
the last few years and months. And we now have some significant 
legislation moving forward for that. In fact, it's been 
described to me by folks in the AI space that having self-
driving vehicles may be the Moonshot for artificial 
intelligence, that when AI can pilot a car through a complex 
city environment with all sorts of things happening all around 
it, that means AI has developed to the point where it's going 
to be transformative in every single industry. It's going to be 
a key barometer of where we are going forward. So we are 
pleased that that's happening in Michigan.
    In fact, we had General Mattis, who mentioned the four 
places for technology in the country, and Michigan was one of 
those four. So a little different vision than a lot of people 
here in Washington may have for my great state, so I appreciate 
the opportunity for that to come up.
    The question I have--and I'm a believer in all of the 
wonderful things that you're talking about. I believe this is 
the most transformative technology in a long, long time, just 
as it is in the auto industry. It's probably as big as when the 
first car came off of the assembly line. We know what happened 
after that, in creating the American middle class, changing 
everything about our economy. We think the same thing will 
happen with AI.
    And so there's incredible promise for it, but I think we 
also have to be very open to the potential downside to this. 
And I know some of you have addressed the employment issue, and 
I want to just talk about that because my experience has been 
the folks who are big proponents of the technology downplay the 
employment aspects. Folks who are scared probably overplay the 
employment aspects. And the truth is going to be somewhere in 
the middle.
    And I think one thing it will have an impact on that 
employment growth and what we've been seeing in the economy 
recently is we have further concentration of industry and fewer 
and fewer companies that have larger shares. That has actually 
suppressed wage growth, it has created a less dynamic 
environment when it comes to new business formation. I mean, we 
can go through the economic arguments associated with that.
    And so, Mr. Castro, you mentioned that whoever comes up--
whatever companies embrace AI will have a significant 
technological advantage when they do that. We see that in the 
auto industry. It's why the auto industry is racing to be 
first, or at least very early, knowing that there is probably 
going to be fewer car companies once AI is fully implemented as 
well.
    So my question to you and other panelists, what sort of 
implications will AI have for the concentration of business in 
those companies or those industries, or I should say those 
individual companies within those industries, that have the 
resources to be able to utilize this? And will it be more 
difficult for small businesses?
    Mr. Castro. Yes, so I think there are two effects. I mean, 
if you talk to a company like Amazon, it's a--you know, they're 
using AI more than anyone, and they're growing faster than 
anyone.
    And so in some cases, we're going to see--especially when 
we're talking about global competitiveness, U.S. companies 
growing because of the technology, and that that growth will 
outpace the jobs offset. Of course, that won't happen 
everywhere, and in many cases, what we want to see is more 
productivity, which means fewer workers in a given space per 
output.
    In those cases, and what we've seen historically in this 
phase, is that the new jobs are not necessarily AI jobs. It's 
not that, you know, everyone is going to now be, you know, 
building self-driving cars or designing them, it's that we see 
more people in other professions, more doctors, you know, 
classrooms with higher teacher-to-student ratios. The kinds of 
changes we often say we want to see and can't pay for right 
now, we can get in the future.
    We did a really interesting study earlier this year that 
was looking at occupational change over the last 165 years, and 
it's actually at the lowest change, so the least disruption, 
that we've seen in this entire time period. And the reason for 
that is we get a lot of misperceptions when we see ATMs on the 
corner, and you think there are fewer of these jobs, when, in 
reality, you have more banks, you have more prosperity. And so 
the job losses are usually more visible than the job creation, 
which is why we have the skewed perception.
    Senator Peters. But I think we go beyond job losses and 
actually look at wage differentials and income inequality, and 
that's probably what I was alluding to, is when you have 
increased concentration, you have less dynamism in the economy, 
which I think is consistent with what you just said about the 
job churn has gone down, it's becoming less dynamic. Many 
economists believe that's a big reason why we have growing 
inequality in this country as well.
    Certainly folks who are able to get these software jobs are 
going to do extremely well, and God bless them for doing that, 
but based on trends that we already see of stagnant wages, even 
though there are increases in productivity, doesn't necessarily 
translate into everyday wages for everyday folks, that that all 
could accelerate at a very quick pace that we should be 
thinking about, and I think it's important for us to be 
conscious of that impact, it's not just jobs, it's income 
inequality.
    Now, Ms. Espinel, if you want to comment on that, please 
do.
    Ms. Espinel. I was just going to say briefly that, one, I 
think you're right, it's something we need to be thinking 
about. You said, or maybe you were alluding to what Mr. Castro 
said, in terms of businesses using AI. I guess I would say, you 
know, we don't think that big businesses or concentrated 
businesses are the ones that should be using AI. I think our 
hope is that AI and the technology behind it will be 
democratized sufficiently so that small businesses will have 
the ability to use it as well and help them in whatever their 
business objectives are.
    So I guess I would say, yes, I would agree. It would be a 
concern if we saw AI being used primarily by just some large 
companies, but I personally don't think that will be the future 
of its deployment, and I know, certainly speaking for our 
members, they would like for small businesses, for the public 
sector, for any type of organization that is trying to make 
decisions and trying themselves to make decisions in a way that 
is better informed to be able to have the use of the benefits 
of AI.
    Senator Peters. Well, I would agree that that's the goal, 
and we hope we could democratize it, but it hasn't necessarily 
played out that way. It does take significant amount of 
capital.
    And as Mr. Castro mentioned, Amazon is using AI to the 
greatest extent of any other company, it's growing the fastest, 
and that's why we have brick-and-mortar retailers that are 
going out of business all over the country as well. It's great 
for Amazon, great for AI, great for productivity for Amazon, 
but it may not be so great for the mom-and-pop store that's 
there. And I would argue that mom-and-pop store thinks they 
can't have an AI system like Amazon, that's just simply not 
realistic for them.
    And so we need to be thinking. I don't have the answers, 
but I think we need to be thinking about that or we're going to 
be facing some significant societal challenges in the future. 
But thank you for your comments.
    Senator Wicker. Senator Udall.

                 STATEMENT OF HON. TOM UDALL, 
                  U.S. SENATOR FROM NEW MEXICO

    Senator Udall. Thank you, Mr. Chairman.
    And thank you to all the witnesses here today. I think this 
has been very excellent testimony. And obviously there are some 
very positive aspects to AI. But I wanted to ask about the bots 
and the software used to hurt consumers. The New York Times 
recently reported on so-called ``Grinch bots,'' software used 
to predict the web links of the in-demand toys and other 
merchandise and to purchase the goods before the general public 
has access to these items. This practice has caused many of 
this holiday season's most popular items, such as Fingerlings, 
Super Nintendo Classic Edition, Barbie Hello Dreamhouse, to be 
sold out online within seconds. However, one can go to eBay and 
easily find these same products available for increased costs, 
sometimes dramatically increased.
    Have either of your organizations worked with retailers to 
prevent software bots from taking advantage of Internet deals 
to jack up the prices of goods? Can you identify other ways to 
prevent these software bots from using machine learning to 
become more and more sophisticated?
    Ms. Espinel. I'm happy to take this as the parent of two 
young boys that are busy filling out their Christmas lists. 
So--and I actually think this is an example of an area where AI 
can be really helpful. So Dr. Gil and others have talked about 
the use of AI and cybersecurity because AI is really good at 
looking at large amounts of data, detecting patterns, and then 
predicting where threats might lie.
    The Grinch bots I think is an area--AI is also great at 
fraud detection, and there are a lot of the characteristics of 
these Grinch bots that are similar. So part of it is, you know, 
if you see--if you see unusual patterns like huge, large, 
amounts of purchases that are happening very, very quickly, and 
these Grinch bots work so they can process transactions in just 
a few seconds, that is a pattern, and an unusual pattern.
    And AI is really good at looking at patterns like this, and 
then alerting the retailers so they can shut down those 
purchases in much the same way that AI is being used now for 
credit card fraud detection and can help detect unusual 
patterns and then, you know, give your bank or you, as the 
credit card user, the ability to say, ``I didn't approve those, 
so you need to block my card.'' That--this type of activity I 
think is actually a great example of how AI could be deployed 
to try to detect those unusual patterns and then tell the 
retailers to not process those transactions and shut it down.
    Senator Udall. Yes, we hope that that happens more because 
you can see the consumer damage, I think every day, and in a 
lot of the business coverage.
    Dr. Bethel, in your testimony, you spoke of the need to 
access more and more data, including social media and mobile 
device location history to help refine the uses of artificial 
data. I'm concerned about the privacy implications that this 
kind of sharing could have. Could you speak to the ways to 
obtain relevant data while still protecting consumers' 
sensitive information? And do others on the panel share my 
concerns on privacy?
    Please, Dr. Bethel.
    Dr. Bethel. I mean, privacy, anytime you're dealing with 
data, is going to be an issue of concern. And I think we have 
to be responsible in how we use that data and how we obtain the 
data. So there does probably need to be some consideration 
given to how that data is obtained, what that information is 
used for, and trying to ensure the privacy of the individuals 
that that information is coming from. It's critical to the 
success, I think, of AI, and being responsible with that.
    So there may need to be some regulation in that to protect 
the general public from harm basically is where regulation I 
feel is needed. But we need to be aware of how that is 
happening and try to put in place measures to ensure the safety 
of that data.
    Senator Udall. Great. Thank you.
    Please.
    Mr. Castro. Thank you. On the question of privacy, there is 
actually some really interesting where we're seeing, you know, 
if you look at kind of trends, increased use of AI can actually 
increase privacy for individuals because it takes personal data 
out of the hands of humans, which is what generally people are 
concerned about. There are general privacy concerns when you 
actually start talking to people, is that, ``I don't want you 
personally to see my medical record. I don't want this person 
over here to see my financial data or use it in a harmful 
way.'' But when they say, you know, ``Are you OK with a 
computer seeing it?'' suddenly everything is OK. And you see 
this in a lot of sensitive areas. This is one reason why people 
like to shop online for certain personal items, they don't want 
to see a clerk at checkout, but they're absolutely fine with 
Amazon having that information.
    So in many cases, what we want to see is how we can have 
policies that actually shift it so companies can guarantee the 
data isn't ever touched by a human, it's only touched by a 
computer, and that is a privacy guarantee. But then we need to 
make sure we can actually hold companies accountable if they 
ever allow that data to escape in certain ways.
    Senator Udall. Yes.
    Dr. Felten. AI raises the stakes in the privacy debate 
because it boosts both the uses of your data that you might 
like as well as the ones that you might not like. And so the 
importance of protecting data privacy continues and maybe is 
higher in a world with AI.
    Dr. Gil. Yes. And I would just say from the perspective of 
IBM, we take the strong view that our clients' data is their 
own, and the insights that derive from the application of AI 
when we help them to do that are their own, too. So we take a 
very strong view of not using data for other purposes that are 
not the intended ones.
    Senator Udall. Great.
    Ms. Espinel. I'll just say very briefly, I think it's 
important to distinguish between different types of data and 
different types of AI. Our BSA companies are at the forefront 
of protecting privacy, and much of the AI that we built is not 
built using personal data. So I think it's important to bear in 
mind.
    Senator Udall. Thank you very much.
    I yield back, Mr. Chairman.
    Senator Wicker. Senator Young.

                 STATEMENT OF HON. TODD YOUNG, 
                   U.S. SENATOR FROM INDIANA

    Senator Young. I thank our panelists for being here today. 
I want to build on some previous line of questioning about our 
labor markets and their preparation or lack thereof, as we 
start to move into an AI-driven economy. What sort of questions 
should policymakers be asking right now to ensure that we 
optimize the skills our workforce has to the extent possible to 
prepare for this, in many ways, exciting and promising new 
technology?
    Dr. Felten. Well, I think--I think there's one set of 
questions with regard to education to make sure--especially in 
the K-12 system, that kids are getting some basic education in 
computer science. Things are more in hand, I think, at the 
university level; it's more a matter of resources there. The 
more difficult questions relate to adult works and displaced 
adult workers and what is available in terms of retraining or 
apprenticeship programs to help them to make sure that those 
are available and to make sure that those are backed by good 
data on the effectiveness of programs is important.
    Ms. Espinel. So I would say I think there are three things 
to think about. One is, you know, if you look at the situation 
today, the Department of Labor BLS is saying that by 2020 
there's going to be 1.4 million jobs that need a computer 
science degree, and yet we only have about 400,000 graduates in 
the pipeline. So we have a gap in our labor pipeline today that 
definitely needs to be addressed if the United States is going 
to stay competitive in this area. So that's one thing we need 
to focus on.
    Second is in terms of education. I think we need to be 
rethinking our educational system for young people to ensure 
that they have access to those skills and the opportunity to 
acquire them so that if they decide that they want to go into 
tech--and not every tech job requires a 4-year college degree, 
and I think that's also something we need to be better about 
explaining--that they have a realistic ability to do so. But 
that is going to require some rethink of the educational system 
that we have now. And I think we need to be modernizing the 
vocational training programs we have for young people that are 
coming out of school.
    And then the third area, which Dr. Felten referred to, in 
terms of people that are in the workforce now, I think we need 
to not just be investing in reskilling and retraining programs, 
but, again, thinking differently about them. I think we need to 
do a better job of matching the skills that people have with 
the employment needs that are out there across the country, 
which is an area where I think there's a lot of work to be 
done, but a lot of potential.
    Senator Young. Yes. With respect to this issue of the labor 
markets and AI, what sort of assumptions are you hearing or 
reading about that you think either overstate some of the 
forces that will be unleashed as AI continues to develop and be 
adopted, perhaps strike you as a bit alarmist, or understate 
these forces?
    Dr. Gil. I think----
    Senator Young. Yes, sir.
    Dr. Gil. Oh, sorry. Dr. Felten was, you know, very helpful 
in providing this journey from the narrow form of artificial 
intelligence that we have today to a general form of artificial 
intelligence that ultimately could perform, you know, 
arbitrary, you know, tasks and domains. We're far away from 
that. So very often the discussion gets framed in terms of 
either the policy implications or the implications to the labor 
market associated with that future form of artificial general 
intelligence that frankly is decades away at best.
    So in that sense, when the conversation is framed in that 
lens, it does come across as alarmist--right?--you know. When 
we talk about the more narrow form of artificial intelligence 
that exists today, it's more of a focus of, what domains can it 
have an impact? Where are proven examples that we can do better 
on that? And at best, you will perform some narrow specific 
tasks that can be complemented with human labor.
    Senator Young. Mr. Castro, do you have some thoughts?
    Mr. Castro. Yes. I mean, so the number one study that 
everyone thinks about in terms of jobs is the study that came 
out of Oxford through research of Frey and Osborne that said 47 
percent of U.S. jobs would be eliminated by I think it was 2025 
or 2030. That study was--first of all, it's not peer reviewed, 
and when you look at the data, they use a very kind of, you 
know, flawed methodology to come up with this estimate.
    So they list in the back in their Appendix all the 
different professions that they say will be automated with AI, 
and it's things like fashion models and barbers, which they 
tried to walk a robot down the runway in Japan once, and they 
haven't done it since.
    So, I mean, you know, kind of realistically, those numbers 
are very much inflated, and they're not actually tied to what 
we're seeing in the market today. So, you know, the main thing 
is I think when you see these studies, and there have been a 
number of studies that have used their data, it doesn't reflect 
reality.
    Senator Young. I'll just close here and indicate, you know, 
the reason we're focused on this I think is we see some 
incredible potential here. There are some studies that indicate 
AI has the potential to increase the rate of economic growth in 
the U.S. from 2.6 percent to 4.6 percent by 2035. I mean, 
that's just amazing. That would benefit all Americans. There 
are some serious policy things we need to wrestle with.
    I've been partnering with Senator Cantwell, who has a lot 
of expertise and professional background in the area of 
technology, and we've developed legislation that would 
establish a Federal advisory committee to help us better 
understand some of these issues. So if you have thoughts moving 
forward about things that a Federal advisory committee should 
look at as we consider the policy implications and broader 
market implications of AI, I'd certainly welcome those, and I 
suspect Senator Cantwell would as well.
    So thank you, Mr. Chairman.
    Senator Wicker. What a nice segue to Senator Cantwell.
    Senator Young. Yes.
    Senator Wicker Senator Cantwell is now recognized.

               STATEMENT OF HON. MARIA CANTWELL, 
                  U.S. SENATOR FROM WASHINGTON

    Senator Cantwell. Thank you, Mr. Chairman. And I do look 
forward to working with Senator Young on this issue and the 
various aspects of investigation.
    I wanted to bring up applications because one of the things 
that I think we should be thinking about is our role as an 
actual user. And one of the things I'm most interested in is 
AI's application for cybersecurity. One of our biggest threats 
obviously we face now is the threat of cybersecurity in all 
sorts of ways. I've seen some applications by MIT and I think 
University of Louisville, several entities that are both 
finding fault in code, basically doing a better job--why wait 
to find out some guy forgot to do the Apache patch, forgot to 
do the patch? You know? I mean, one employee at that company 
cost everybody a lot of money because he didn't put in a patch. 
AI could help us find errors in code or actually in some of 
these areas I think predict cyber attacks.
    So I don't know who on the panel could speak to that, but 
to me, one of the applications that I hope that we will look at 
is the government's use of this as it relates to combating 
cybersecurity.
    Dr. Felten. There's a huge opportunity there, and it's 
rapidly becoming a necessity, as the bad guys are adopting AI 
and automation in their cyber attacks. Government and other 
institutions need to be using more AI in defense in order to, 
as you said, Senator, find vulnerabilities before they're 
exploited in order to be able to react at machine speed when 
things start to go wrong, and in order to better understand 
what are the possible implications of the way that systems are 
set up so we don't get surprised in the way that institutions 
nowadays too often are by both that a breach occurs and by how 
bad the consequences are.
    Dr. Gil. Yes, there are two dimensions are essential in 
this topic. One of it actually has to do with securing AI 
itself. The very models that we create to enable these 
predictions are actually vulnerable to attack, and that in 
itself, there are many steps that can be taken to be able to 
secure those models. In fact, you can even extract data from a 
model that has been created.
    And the second I mention is the one you alluded to, which 
is AI itself has to be an integral component to protect against 
other AI-powered attacks. So in a way, we are going to have AI 
against AI in the realm of cybersecurity because, you know, 
some of the bad guys are already using these techniques to be 
able to attack our networks with the speed and accuracy and 
adaptability that without the presence of AI, and Ms. Espinel 
alluded to this already, it will be impossible to defend 
against.
    Senator Cantwell. Did you want to comment on that as well?
    Ms. Espinel. Just I'll just say briefly that IBM is among 
the BSA members that are focused on this. We have a number of 
companies that are very focused on cybersecurity and are using 
AI to try to detect patterns and then predict threats before 
they happen. So I agree with you, that I think it's an area 
where AI is already being used, and there is potential for it 
to be doing even more.
    Senator Cantwell. Well, I like the notion just in the fact 
that because there are things called human error, that you can, 
you know, use this to look for faults in code and weak points 
and do--because that's what someone else is doing, right? And 
so the sooner that you can find that yourselves and detect that 
and create another security layer, the better.
    Is there any other government application that you think we 
should be investigating on? Some people say on like health 
statistics and things of that nature, but I don't know if you--
anybody else on the panel has----
    Mr. Castro. We did look last week into government use of 
AI, and one of the biggest challenges is that we're just not 
measuring what we're doing within government, that there are 
lots of opportunities, but there's not a good place where if 
you're an agency, not even a CIO, but just a team manager 
basically, who wants to start using, you know, automated 
calendaring, you know, how do I actually go out and do that 
quickly? You know, can I quickly procure this? Is there an 
approved list of, you know, best practices from other agencies? 
Is there good information sharing? We're starting to do that 
through GSA, but we're not there yet.
    So one of the things that this Committee could hopefully 
help do is to really push agencies to ask--bring the CIOs in 
here and ask them, ``What are you doing around AI? Have you 
identified the top three opportunities and are you pursuing 
them?'' Just like, you know, many agencies were directed to 
find the high-value datasets, pick three, and get them out 
there as open data, we can do something similar around AI, pick 
the top three opportunity targets and pursue that over the next 
12 months. Similarly----
    Ms. Espinel. I would agree with that, but I think city 
planning is another area where AI could be really helpful. So 
like traffic congestion, you know, one of the things that city 
planners are trying to do is optimize traffic patterns and the 
changing of traffic lights to try to improve traffic flow, and 
that's something that AI is really good at. You know, there are 
a lot of variables, so traffic congestion can seem like a 
relatively simple question, but if you think of all the 
variables and traffic patterns, it's actually quite complicated 
in taking all that data and then giving city planners 
recommendations for how to optimize traffic flow is something 
AI can do really well.
    Senator Cantwell. Thank you.
    Thank you, Mr. Chairman.
    Senator Wicker. Dr. Bethel, you are doing practical 
applications on high-risk law enforcement situations, 
autonomous cargo transfer, and animal-assisted therapy. Do you 
receive Federal research dollars for that? And to what extent?
    Dr. Bethel. The law enforcement application we have 
submitted to NSF. We have currently one grant under review. The 
Therabot was funded under NSF funding. And some other projects 
that we're doing are also funded through the NSF or Department 
of Defense.
    Senator Wicker. OK. Now, Senator Schatz mentioned a real 
concern with regard to law enforcement. What you all are doing, 
though, at Mississippi State, is scenarios where there is 
already a threat, and law enforcement needs information about 
how to respond, how to get inside the building. Is there a 
child there? Is there something explosive there? Has anyone 
raised the concerns that he raised, any of your law enforcement 
people or victims or defendants, raised concerns about data 
bias?
    Dr. Bethel. Law enforcement has not mentioned any data bias 
because when we are using these algorithms, we are looking at 
more like objects in the scene----
    Senator Wicker. Something is already occurring.
    Dr. Bethel. Something is occurring. We're not trying to use 
it to target people, we're using it because the scene--some 
kind of event has occurred, and law enforcement is responding. 
And so we are trying to provide them as much information as we 
can prior to them going in so that hopefully they can make 
better, more safer, decisions when they enter into the 
environment. For instance, their protocol changes completely if 
there's a child inside the home that they're going into. They 
can't use a flash bang, they can't do things they would 
normally do.
    So we can, by sending a robot in to be able to see what's 
happening prior, they can make decisions, so they're probably 
going to end up saving lives, both civilian and law enforcement 
lives, because--and they will have better performance because 
of it. So bias hasn't so far come into our discussions when 
we've been looking at law enforcement applications, but we are 
not using it, I think, in the manner in which Senator Schatz 
has indicated.
    Senator Wicker. Now, with regard--first of all, on the law 
enforcement application, how long has Mississippi State been 
doing this?
    Dr. Bethel. Six years. I started training with them in 
2011. We train monthly.
    Senator Wicker. OK. AI, according to your testimony, is 
only as good as the data the system receives. How has your data 
improved over the 6 years, and could you elaborate on that?
    Dr. Bethel. So each training we do, we do video recordings, 
we do sensor recordings, we do different manners of data 
collection. The more examples of data that we can get and 
obtain, the better the system is at classifying information, 
doing the sensor fusion, and incorporating that information to 
make more informed decisions. So as time has gone on, we've 
been able to obtain more and more samples of data to be able to 
use that for the systems.
    Senator Wicker. OK. Now, most of the panel, if not all the 
panel, seems to think we lead the world, it's just a question 
of whether we're going to continue leading the world. Senator 
Schatz said he doesn't think with regard to national policy 
we'll take a European approach. Rather than ask him what he 
means by that, let me ask the panel, are there mistakes that 
our international competitors, whether in Europe or Asia or 
Africa or wherever, are making that we need to avoid? Are there 
overreaches in terms of regulation that we need to avoid?
    Everybody is trying to get ahead. The testimony is that 
China is making an important effort, the U.K., Japan, our 
neighbor to the north, but in terms of things we can avoid, 
mistakes that other countries have made, does anybody have a 
suggestion that we need to be mindful of as we work on Senator 
Young and Senator Cantwell's and Senator Schatz's legislation?
    Yes, Mr. Castro.
    Mr. Castro. Yes, so, I mean, the two big things I think 
some countries are thinking about is the right to explanation 
that's across the board or right to opt out of automated 
decisions. So the problem with those two policies is that it 
minimizes the ability or limits the ability of companies to 
deploy AI and many commercial applications.
    So, for example, if you want to be a company, you know, 
just like before you had companies suddenly become online 
lenders, if you want to be a company that's an AI lender, 
streamline your entire business process through AI, you 
basically can't do that because anyone who got rejected for a 
loan application, for example, could ask for a human appeal, 
and you would have to have human support to do that.
    And the problem with that kind of framework is it basically 
limits those types of business models, and it's especially 
problematic when you talk about global competition because it 
doesn't limit necessarily a foreign competitor from offering a 
similar service. Now, in certain industries, of course, you 
have to be licensed in the United States, so there are limits 
there, but overall, that would limit these kind of 
opportunities.
    Ms. Espinel. Can I just say briefly? I think most other 
governments are still considering what their AI policy 
environment is going to look like. So many governments are 
having this discussion right now. I do think that there are 
governments that are--seem to be skewing toward considering a 
more regulatory approach, a more heavy-handed regulatory 
approach, and that raises concerns.
    And I think one issue I'd raise in particular I think for 
us, I think any regulatory approach that would compel the 
disclosure of an algorithm, that would compel companies to hand 
over source code or algorithms, to a government agency is one 
that raises a lot of concerns with us. We don't think it's 
actually going to be an effective way to address policy 
outcomes, and it raises a lot of competitive concerns to be 
handing over algorithms or source codes to governments.
    Senator Wicker. Dr. Gil, on cyber attacks from hostile 
forces internationally and our defense against that, do I 
understand you to--well, no, let me rephrase it. Is it 
conceivable that artificial intelligence will be empowered by a 
hostile force to make a decision to go forward with a cyber 
attack without a human being at that end making the decision to 
pull the trigger?
    Dr. Gil. Humans will definitely design AI-powered attacks 
to attack, you know, infrastructure or, you know, an industrial 
setting or another nation-state----
    Senator Wicker. But whether that attack is made today or 
Valentine's Day next year, an artificial intelligence----
    Dr. Gil. System.
    Senator Wicker.--system might make that decision.
    Dr. Gil. That's correct. That is possible----
    Senator Wicker. How close is that to reality?
    Dr. Gil.--because, I mean, think about it in the past for--
you could design it with an explicit program model to be able 
to carry out such attacks, and we would have to, you know, 
create defense mechanisms against those kinds of attacks. But 
because the program was well stipulated, you could imagine 
somebody defending against the attack to be able to interpret 
what those rules of attack were.
    Now, the moment that you're employing a more machine-
learning-based approach where the type of attack could morph 
depending on the environment it's detecting, now being able to 
detect what is the form of the attack that is taking place 
requires another pattern detection mechanism. So that's why I 
was referring to--and Ms. Espinel was talking before--about you 
need AI to defend against AI-powered attacks because it's the 
only way to make it really adaptive to an adaptive attack.
    Senator Wicker. OK. Senator Markey, do you mind if we--if 
we let the two witnesses respond here? And then I'll be 
generous with your time.
    Dr. Felten?
    Dr. Felten. Thank you. What you're talking about in terms 
of automated cyber attack is something we see already with 
computer viruses. A virus is software which autonomously 
spreads itself from place to place, and then does whatever it's 
programmed to do to cause harm at the places that it infects. 
So to bring AI into this would just be to have a more 
sophisticated, more adaptive form of virus. It's not 
fundamentally a new thing, it is--it's a path that we are 
already on, or a path that the bad guys are already on.
    Senator Wicker. And quickly, Ms. Espinel.
    Ms. Espinel. I still think it's more akin to making a 
recommendation than making a decision. Of course, you, the 
person, can decide that whatever recommendation the system 
gives you, you're going to have it automatically act on, but 
you still made that decision up front, and I don't think we 
should be--as a general matter, I don't think we should be 
abrogating decisionmaking authority, but I think that what 
you're talking about really is AI making a recommendation, and 
then the person who designed it deciding whether or not they're 
going to accept that recommendation automatically or not.
    Senator Wicker. We're really talking about what's coming at 
us.
    Senator Markey.

               STATEMENT OF HON. EDWARD MARKEY, 
                U.S. SENATOR FROM MASSACHUSETTS

    Senator Markey. Thank you, Mr. Chairman, very much.
    I thank you all for being here.
    The digital revolution's newest innovations--augmented 
reality, autonomous vehicles, drones--all of these industries 
use artificial intelligence, which rely heavily on a free and 
open Internet. Regrettably, these disrupting technologies may 
be dealt a major blow in their infancy, and that's because in 
just 2 days, the FCC will vote on a proposal to eliminate net 
neutrality.
    Without enforceable net neutrality rules in place, 
broadband providers, like Comcast, Verizon, AT&T, and Spectrum, 
could block or slow down the content of innovators and 
businesses using AI all in an effort to leverage their 
gatekeeper role to favor their own content and generate 
additional profits. And what will replace these robust net 
neutrality protections? Nothing, absolutely nothing is going to 
replace those rules.
    Dr. Felten, how would eliminating net neutrality 
protections impact the deployment and development of innovative 
technologies that use AI?
    Dr. Felten. A lot of AI technologies operate within an 
institution or within a company's data center, and so those 
technologies would not be much affected by changes in network 
policy. To the extent that changes in network policy affect the 
ability of companies to deliver their products to consumers, 
that would obviously be a policy concern, but in my mind, it is 
important but also somewhat separate from the issue of 
development of AI.
    Senator Markey. So you don't think that the additional cost 
to the developer, to the innovator, won't have an inhibiting 
impact upon their ability to go to the capital markets, raise 
the dough, in order to produce their innovation knowing that 
there's no guarantee that they can reach the 320 million 
Americans without paid prioritization or without a threat of 
throttling or blocking?
    Dr. Felten. I would say that all pro-innovation policies 
and all policies that are designed to help small companies 
enter and compete provide value and are pro-innovation and 
valuable. As I said, I don't think that net neutrality plays a 
special role with respect to AI.
    Senator Markey. But does it----
    Dr. Felten. There are other areas of innovation.
    Senator Markey. Will it play a role in your opinion?
    Dr. Felten. Yes. I think those decisions do play a role in 
almost any area of innovation.
    Senator Markey. OK. And on the question of child privacy, 
the Child Online Privacy Protection Act of 1998 is still the 
communications constitution for safeguarding children online, 
and that's a law I was able to get passed back in 1998.
    As emerging technologies like AI are deployed, it's 
important that they honor core American values, including 
privacy. Dr. Felten, could AI technologies pose a threat to 
children's privacy? And is there a threat that AI technologies 
could produce inappropriate content for children?
    Dr. Felten. This is an issue to pay attention to. AI does 
raise the stakes on policy--I'm sorry--AI raises the stakes on 
privacy discussions generally, and that's true with respect to 
children and others. And, of course, parents are very concerned 
about what their kids see and what happens, and that's one of 
the reasons why COPA, for example, requires parental consent 
before certain uses of data are allowed.
    Senator Markey. So could relying on AI in children's toys 
negatively impact kids' ability to develop empathy if we 
substitute real people with computers that cannot fully 
understand emotion as humans do, Dr. Felten?
    Dr. Felten. Well, I think kids are more interested in 
playing with other kids or using toys as a vehicle for playing 
with other kids. I'm less worried about kids bonding with 
something that is not human-like or not companionable. I think 
kids will reject those on their own.
    Senator Markey. Yes. Earlier this year I wrote to Mattel 
with serious concerns about their plan to bring the first all-
in-one voice-controlled smart baby monitor to the market. 
Mattel had planned for the device, Aristotle, to use artificial 
intelligence to help teach children and respond to their needs. 
After an outcry of questions, Mattel canceled that product.
    Dr. Felten, what does that experience expose as to 
potential negatives of using artificial intelligence with 
children's devices?
    Dr. Felten. Well, I think stories like this illustrate that 
it's important to understand the implications of the 
technologies that are being deployed. As I would imagine, that 
parents would be happy to, say, be notified if there is 
indication that their child is in distress, and AI may help to 
do that more effectively. But these issues of unintended 
consequences and safety are paramount, and that's one of the 
important aspects of clearing the road for responsible 
deployment of AI, is to make sure these issues are taken care 
of.
    Senator Markey. Thank you.
    Senator Wicker. Senator Cruz.

                  STATEMENT OF HON. TED CRUZ, 
                    U.S. SENATOR FROM TEXAS

    Senator Cruz. Thank you, Mr. Chairman.
    Thank you to each of the witnesses for coming here this 
morning to testify.
    A little over a year ago, the Subcommittee on Science and 
Space, which I chair, held the first congressional hearing on 
artificial intelligence. And then, as now, we heard testimony 
about the extraordinary transformative process that we are 
engaged in right now and how AI in time can be expected to 
touch virtually every area of human endeavor, and indeed that 
this transformation may be of comparable import to the 
transformation we engaged in, in a prior era in the Industrial 
Age.
    Anytime we're seeing dramatic transformations in our 
economy and our workforce and how we interact with each other, 
that poses the risk of dislocations, but it also poses policy 
and government and regulatory challenges for how to interact 
with the new terrain. In your judgment, what are the biggest 
barriers right now to developing and expanding AI and its 
positive impacts on our society and our economy?
    Ms. Espinel. I'll head off another--so I think one of the 
biggest barriers is a lack of understanding, a lack of 
understanding about what AI is, what the actual technology is, 
and then what it does and what the--you know, what both the 
intended and unintended consequences are. And so I think, you 
know, this hearing, the legislation that Senator Schatz is 
working on, I think trying to increase our collective 
understanding is critical, it's fundamental.
    I think there are a number of specific policy issues that 
would be helpful in terms of eliminating barriers. So, you 
know, one of those is AI is all about data, and so good data 
policy in various ways I think is very important. I think 
investing in research, which we've talked about already, 
investing both in government research and incentivizing private 
sector research, is very important. And then I think thinking 
about jobs and workforce development, both the jobs today, but 
what will happen tomorrow? And rethinking our educational 
system and our training and reskilling programs are vitally 
important. So those are the three specific areas, but I think a 
greater understanding needs to be part of all of those 
discussions.
    Mr. Castro. So I agree that skills are very important. 
Especially much of this technology is new, and we need people 
that can basically rapidly be credentialed in how to deploy it. 
But when we look at the biggest opportunities for AI, it's 
really in some of the regulated industries. And so I think 
that's where we head up on challenges because there are two 
challenges: one, regulators aren't necessarily prepared with 
how to deal with this; and, two, they don't necessarily have 
the skill set or capabilities internally within the regulatory 
system to handle it. And so I think that's something we need to 
be very focused on, is asking questions----
    Senator Cruz. What agencies in the industries in 
particular?
    Mr. Castro. So financial regulation, for example. Education 
is another example, especially when we're talking about using 
AI in primary education, there are a lot of questions about 
privacy that get raised. And the financial system it's the 
questions of, you know, do we have regulation basically to 
understand the technology? And we'll do things more than just 
ask, ``Can I see the algorithm?'' but be able to say, ``Can I 
look at outcomes? Can I actually measure outcomes?'' and then 
ask questions about, ``OK, is this fair? Is this different than 
what we have now? Is this moving in the right direction?'' and 
also have some of that regulatory flexibility. So we need kind 
of, you know, fewer cops and more sandboxes.
    Senator Cruz. So one area that has generated fears and 
concern is general AI, and scientists and innovators ranging 
from Stephen Hawking to Bill Gates to Elon Musk have raised 
concerns. Stephen Hawking stated, quote, ``Once humans develop 
artificial intelligence, it would take off on its own and 
redesign itself at an ever-increasing rate. Humans, who are 
limited by slow biological evolution, couldn't compete and 
would be superseded.'' Elon Musk has referred to it as, quote, 
``Summoning the demon.''
    How concerned should we be about the prospect of general 
AI? Or to ask the question differently, in a nod to Terminator, 
when does Skynet go online?
    [Laughter.]
    Dr. Felten. Hopefully never.
    Senator Cruz. That's the right answer.
    Dr. Felten. I think there's a lot of debate within the 
technical community about how likely these sorts of scenarios 
might be. I think virtually everyone agrees that it would be 
far in the future. And generally the people who are most 
involved in AI research and development tend to be the most 
skeptical about the Skynet or existential risk type of 
scenarios.
    In any case, the sorts of risks and concerns that exist now 
about AI are really baby versions of the ones that we would 
face with a more sophisticated general AI. And so the tactics 
we--the policies that make sense now to deal with the issues we 
face now are the same ones we would use to warm up for a 
general AI future if it comes. And so from a policy choice 
standpoint, the possibility of distant general AI seems less 
important to me.
    Dr. Gil. Yes, I would completely agree. I think if you ask 
practitioners in the field when they would envision the 
possibility of that, I think everybody would say 20-plus years 
out. And whenever scientists says 20-plus years out is our code 
word for saying we just don't know. Right? We're nowhere near 
close to be able to do that.
    So I do think that while it's an area of very important 
study, and there are many universities who are now creating a 
rate of monitoring of what is the progress of AI and the 
implications that it would have if we eventually reach general 
artificial intelligence, I do think it would be a mistake to 
guide our policy decisions at present based on the sort of like 
long-term hypothetical that we don't even have, even as 
practitioners, even a credible path to get there.
    Senator Cruz. Thank you.
    Senator Wicker. Senator Cortez Masto.

           STATEMENT OF HON. CATHERINE CORTEZ MASTO, 
                    U.S. SENATOR FROM NEVADA

    Senator Cortez Masto. Thank you, Mr. Chairman.
    And thank you to the panel members. I'm very excited about 
the conversation today as well.
    Obviously, we are standing at the edge of a technological 
revolution that we must ensure will, and there has been 
discussion, take our labor force with it. So this is a timely 
conversation.
    Workforce development has been a focus of mine since I've 
entered the Senate, and so has innovation. I've proudly been 
leading a bipartisan legislation on drone expansion, the use of 
smart technology in transportation, and in trying to spur the 
next generation of women to be equal at the forefront of STEM, 
specifically computer programming, through the introduction of 
the Code Like a Girl Act. I've worked on this because I've seen 
the future of these developments through my state in Nevada.
    And just last week I was visited by the leadership at the 
Truckee Meadows Community College in Reno, who, in partnership 
with Panasonic, has developed a curriculum to provide 
individuals to get the specific training that local jobs are 
hiring for. These are conversations that constantly are going 
on in my state.
    So let me start here, Mr. Felten and Ms. Espinel. The 
skills gap has been discussed extensively today, and something 
that obviously is on all of our minds and how we address it. 
Obviously, we have China investing dramatically in the area of 
AI, so it begs the question of whether there are investments we 
can be making at the Federal level to help close any potential 
skills gaps. I'm curious your thoughts on that.
    Ms. Espinel. So in terms of the skills gap specifically, we 
also talked a little bit about Federal support for research 
funding. But I would say a few, and then others may add to 
them.
    I think so one is trying to improve access to computer 
science education at all states and at very early stages of 
education. So I think that's one area where the Federal 
Government could be helpful.
    I think the second is in rethinking how our vocational 
training programs work. There are vocational training programs 
in place, but they could be streamlined and they could be 
better adapted to the world that we live in today. So I think 
that's an area where there's a lot that could done at the 
Federal level.
    And then the third I would say is that I think there is 
now, and there will maybe be only an increasing sort of 
information gap between the skills that people have and 
employers knowing about those, and that seems like an area 
where there's a real deficiency now, so therefore real 
opportunity to try to create programs that not only create--
either create pathways directly into employment or do a better 
job of matching up skills that people have with the jobs that 
employers have to offer.
    Senator Cortez Masto. Thank you.
    Dr. Felten. There are opportunities to widen the pipeline 
at all stages starting with K-12, making sure that basic 
computer science education is available to every child at that 
level, and more advanced education at the high school level for 
those who are ready for it.
    There are opportunities to improve--increase the number of 
teachers and trainee researchers at the university level, and 
that's education and research funding specifically in areas of 
AI and computer science at the university level. And then 
vocational and adult training and apprenticeship programs also 
are very important to get people on other career paths, give 
them an on-ramp into this area.
    Senator Cortez Masto. Go ahead.
    Ms. Espinel. Can I just say one more thing?
    Senator Cortez Masto. Yes.
    Ms. Espinel. There's a lot that BSA companies are doing in 
this area, IBM among them. And so I think another area is 
working with the Federal Government both to scale up programs 
that are effective now and be more collaborative in this area I 
think is----
    Dr. Gil. Just to touch on just one example of that. On the 
initiative, on a program, we started a number of years called 
P-TECH, which is a 9 through 14 educational program that 
combines both education, hands-on program on vocational, you 
know, schools. And when they graduate through this--right?--we 
are talking about the creations of new collar professions that 
have enough skills to be able to practice and benefit from the 
advances that are happening in AI without having to go through 
a full college education, and now that's touching, you know, 
more than 10,000 students.
    Senator Cortez Masto. Thank you. No, this is something 
that's happening in Nevada now. It seems like it's such common 
sense, but we just don't do it, which is that collaboration and 
partnership between whether it's a government, private sector, 
and then our education system or our skills vocational trade, 
to really say, ``What are the jobs of the future? What are they 
going to look like?'' work with those employers to develop the 
curriculum that you're going to need for that skilled workforce 
so that we can really start educating them now. And that starts 
I think at a very young age and working with our education 
system, but also the vocation trades that are out there as 
well. Not every child is going to go on to get a higher degree, 
but there are some that are absolutely and rightfully are going 
to go get that vocation or that trade and that skill that's 
necessary.
    So I appreciate the comments today. I am running out of 
time, so I will submit the rest of my questions to this 
incredible panel for the record. And I appreciate you being 
here. Thank you.
    Senator Wicker. Thank you very much.
    Senator Blumenthal.

             STATEMENT OF HON. RICHARD BLUMENTHAL, 
                 U.S. SENATOR FROM CONNECTICUT

    Senator Blumenthal. Thanks, Mr. Chairman. I apologize that 
I missed some of your testimony so far, but I know that this 
panel has been very important and enlightening, and I thank you 
for being here.
    As you well know, the success of autonomous vehicles is 
closely linked with the success of AI. An autonomous vehicle 
can only perform as well as the input it receives. A lot of us 
know the statistic that 94 percent of the 37,000 deaths on the 
road each year are attributable to human error--97 percent. The 
hope is that autonomous vehicles will eliminate that human 
error.
    What we really lack is information on the extent, to what 
extent that is caused by human error will be possibly replaced 
by computer error. And we know computers are not infallible.
    I wonder, Dr. Bethel, whether you could talk about some of 
the tasks in which humans still perform better than computers 
or an AI system in the context of autonomous vehicles, if 
you're able.
    Dr. Bethel. In the context of autonomous vehicles, humans 
are able to adjust to very rapidly changing, unpredictable 
environments, things happening in the environment. Our sensor 
systems and our onboard processing in autonomous vehicles is 
just not to a point where it can make those kind of adjustments 
that rapidly currently with current technology to be able to 
adjust to that. So in those cases where you have an erratic 
driver who is doing unpredictable behaviors, it's really hard 
sometimes for the system to be able to detect that and react 
accordingly. In those cases, a human currently is making better 
decisions on autonomy--over an autonomous vehicle in that kind 
of environment.
    Senator Blumenthal. So if I can just extrapolate from your 
answer, a computer trying to deal with a drunk driver, either 
ahead or next to that computer, would have trouble because the 
drunk driver, by definition, is acting in not only 
unpredictable, but irrational and sometimes actually self-
destructive ways.
    Dr. Bethel. Right. So it would be much more difficult for a 
computer to predict that kind of behavior than it would be for 
a human to slow down and react. I mean, it would react, but it 
probably is not going to be as effective as a human driver is 
at this point in the stage of where AI is.
    Senator Blumenthal. Are there ways to program a computer or 
create software that deals with those unpredictable situations?
    Dr. Bethel. To some extent, but I think there's a long way 
to go on that.
    Senator Blumenthal. In your testimony, you say, and I'm 
quoting, ``A current limitation to the advancement of 
artificial intelligence is the quality and cost effectiveness 
of sensing capabilities to provide high-quality information or 
data to the system to make those digital decisions. We have a 
long way to go in the advancement of artificial intelligence--
,'' ``We have come a long way in the advancement of artificial 
intelligence,'' excuse me, ``however, we still have a long way 
to go.''
    Sensing, perceiving, interpreting, surroundings are 
essential to driving a vehicle. Can you describe some of the 
limitations in the current computer vision and sensing 
technologies?
    Dr. Bethel. I'll probably need to get back to you on some 
of that because it's not exactly my area of expertise related 
to computer vision, but there are limitations in the sensing 
capabilities we currently have. Every sensor has its own 
limitations, so there's no perfect sensor out there.
    There are also differences in processing power, so trying 
to be able to handle large amounts of data coming in from these 
sensors, and processing that in a timely manner can be a 
challenge. So that's another area. And computer vision has just 
mega amounts of data coming in that has to be processed, and so 
another limitation is the actual processing power to handle 
that, especially in a timely real-time manner onboard a system.
    Senator Blumenthal. At what point do you think it would be 
appropriate to completely remove the human from an autonomous 
vehicle?
    Dr. Bethel. Thank you for the question. I--depending on the 
application, I think there are applications currently where 
autonomy could be--fully autonomous systems are capable. I 
think it's not realistic anytime in the near future, and 
especially in autonomous cars, to be able to say that a fully 
autonomous car is going to be possible all the time.
    Senator Blumenthal. Thank you.
    Thank you very much to the panel.
    Thank you, Mr. Chairman.
    Senator Wicker. Senator Schatz.
    Senator Schatz. Dr. Felten, I just wanted to follow up on 
the question around autonomous weapons. It seems to me that 
this is an area that is different than a lot of these other 
ethical, societal, micro, macro, economic questions; this is 
about how we engage in warfighting. And so to the degree and 
extent that some of these algorithms are hackable, to the 
degree and extent that we have a body of history around how 
we're supposed to engage our military in an ethical manner, 
that this is an area where the Federal Government has to make 
policy.
    And I'm wondering if you can give us any insight into, in 
the absence of policymaking, who's making these decisions? Is 
it defense contractors? Is it individual procurement officers? 
How is all of this getting decided?
    Dr. Felten. Well, there does need to be a policy which 
deals with the very serious issues that you mentioned, both 
from the standpoint of what our military should be willing to 
do and what safeguards are needed on the systems that they're 
using consistent with the need for them to be as effective as 
possible in battle, and also how we deal with adversaries, who 
might not be as scrupulous in following the international 
humanitarian law.
    And this is a national policy issue that ought to be a 
matter of policy discussion at the highest levels. And if it's 
done in a decentralized way, if each contractor, each 
contracting officer, does one thing, if the State Department 
goes to international arms control discussions and does their 
own thing, we get uncoordinated policy and we get a result that 
doesn't serve the American people well.
    Senator Schatz. Thank you.
    Senator Wicker. Well, I would like to thank everyone who 
has participated. And I think Senator Cortez Masto said this 
has been an incredible panel, and I would have to agree.
    Ms. Espinel, you brought information about software and the 
way a number of states have benefited. Did you have information 
for all 50 states or just for the states represented by this 
panel?
    Ms. Espinel. We have it for all 50 states. Our foundation 
software data put out a study in September for each of the 50 
states.
    Senator Wicker. OK. Well, then if you would, please enter 
that into the record.
    Ms. Espinel. I'd be delighted to.
    [The information referred to follows:]

    [GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
    

    Senator Wicker. I would like to enter into the record a 
letter from the Information Technology Industry Council and a 
letter from the Electronic Privacy Information Center without 
objection. And that's so ordered.
    [The information referred to follows:]

                    Information Technology Industry Council
                                  Washington, DC, December 11, 2017

Hon. Roger Wicker, Chairman,
Hon. Brian Schatz, Ranking Member,
Subcommittee on Communications, Technology, Innovation, and the 
            Internet,
U.S. Senate Committee on Commerce, Science, and Transportation,
Washington, DC.

Dear Chairman Wicker and Ranking Member Schatz:

    In advance of your hearing on ``Digital Decision-Making: The 
Building Blocks of Machine Learning and Artificial Intelligence,'' I am 
writing to thank you for interest in and attention to the exciting 
innovation that is Artificial Intelligence (AI). ITI represents more 
than 60 of the world's leading information and communications 
technology (ICT) companies. Our companies are the most dynamic and 
innovative companies from all corners of the ICT sector, including the 
development and deployment of AI. I submit this letter on behalf of ITI 
and its members, and respectfully request that you enter it into the 
hearing record.
    Artificial intelligence (AI) technology is an integral part of our 
daily lives, work, and existence. It's already made an important mark 
on much of our society and economy, and the exciting part is that we're 
just seeing the beginning of its benefits.
    Go to any hospital and medical research center and you will see how 
doctors and medical providers use AI to save lives. For example, the 
company Berg uses AI to analyze large amounts of oncological data to 
create a model of how pancreatic cancer functions, enabling us to 
develop chemotherapy to which cancer cells are more responsive.
    Educators across the country use AI to enhance the potential for 
future generations to grow and learn. Thanks to IBM's Teacher Advisor, 
a new tool based on its Watson cognitive computing platform, third-
grade math teachers can develop personalized lesson plans. Teacher 
Advisor analyzes education standards, sets targets for skills 
development, and uses student data to help teachers tailor 
instructional material for students with varying skill levels.
    AI also makes day-to-day of life easier--for everyone. Many of our 
everyday tasks, like making shopping lists and ordering groceries, are 
streamlined through devices like Alexa. And, through AI technology, 
researchers at the International Islamic University of Malaysia have 
developed the Automatic Sign Language Translator (ASLT) that uses 
machine learning to interpret sign language and convert it into text, 
easing communications for many.
    We know AI will revolutionize the way we do business and our 
overall economy. It's projected AI will add between $7.1 trillion and 
$13.17 trillion to the global economy by 2025.\1\
---------------------------------------------------------------------------
    \1\ https://www.itic.org/resources/AI-Policy-Principles-
FullReport2.pdf
---------------------------------------------------------------------------
    There's no question AI will continue to transform our lives, 
society, and economy for the better. We understand, however, that there 
are many questions about this technology and that with transformative 
innovation, there are going to be points of tension. The tech industry 
is committed to working with all stakeholders to identify and resolve 
challenges.
    In conjunction with the global leaders in AI innovation, ITI 
recently published AI Policy Principles. These principles are designed 
to be guidelines for the responsible development and deployment of AI 
as we develop partnerships with governments, academia, and the public.
    Our Policy Principles are a conversation catalyst, encouraging all 
stakeholders, public and private, to collaborate to create smart 
policies that allow this emerging technology to flourish while 
addressing the complex issues that arise out of its growth and 
deployment. Given the reach of AI, we think this kind of partnership 
and engagement is critical to advance the benefits and responsible 
growth of AI while also endeavoring to answer the public's questions 
about the use of this nascent technology.
    We look forward to working with this Committee, other members of 
Congress, academia, industry partners, and the public to advance AI 
responsibly. Thank you, again, for holding the timely and important 
hearing.

                                             Dean Garfield,
                                                 President and CEO,
                          Information Technology Industry Council (ITI)
                                 ______
                                 
                      Electronic Privacy Information Center
                                  Washington, DC, December 12, 2017

Hon. John Thune, Chairman,
Hon. Bill Nelson, Ranking Member,
U.S. Senate Committee on Commerce, Science, and Transportation,
Washington, DC.

Dear Chairman Thune and Ranking Member Nelson:

    We write to you regarding the ``Digital Decision-Making: The 
Building Blocks of Machine Learning and Artificial Intelligence'' 
hearing.\1\ EPIC is a public interest research center established in 
1994 to focus public attention on emerging privacy and civil liberties 
issues.\2\ EPIC has promoted ``Algorithmic Transparency'' for many 
years.\3\
---------------------------------------------------------------------------
    \1\ Digital Decision-Making: The Building Blocks of Machine 
Learning and Artificial Intelligence, 115th Cong. (2017), S. Comm. on 
Commerce, Science, and Transportation, https://www.commerce.senate.gov/
public/index.cfm/hearings?ID=7097E2B0-4A6B-4D92-85C3-D48E100
8C8FD (Dec. 12, 2017).
    \2\ EPIC, About EPIC, https://epic.org/epic/about.html.
    \3\ EPIC, Algorithmic Transparency, https://epic.org/algorithmic-
transparency/.
---------------------------------------------------------------------------
    Democratic governance is built on principles of procedural fairness 
and transparency. And accountability is key to decision making. We must 
know the basis of decisions, whether right or wrong. But as decisions 
are automated, and organizations increasingly delegate decisionmaking 
to techniques they do not fully understand, processes become more 
opaque and less accountable. It is therefore imperative that 
algorithmic process be open, provable, and accountable. Arguments that 
algorithmic transparency is impossible or ``too complex'' are not 
reassuring.
    It is becoming increasingly clear that Congress must regulate AI to 
ensure accountability and transparency:

   Algorithms are often used to make adverse decisions about 
        people. Algorithms deny people educational opportunities, 
        employment, housing, insurance, and credit.\4\ Many of these 
        decisions are entirely opaque, leaving individuals to wonder 
        whether the decisions were accurate, fair, or even about them.
---------------------------------------------------------------------------
    \4\ Danielle Keats Citron & Frank Pasquale, The Scored Society: Due 
Process for Automated Predictions, 89 Wash. L. Rev. 1 (2014).

   Secret algorithms are deployed in the criminal justice 
        system to assess forensic evidence, determine sentences, to 
        even decide guilt or innocence.\5\ Several states use 
        proprietary commercial systems, not subject to open government 
        laws, to determine guilt or innocence. The Model Penal Code 
        recommends the implementation of recidivism-based actuarial 
        instruments in sentencing guidelines.\6\ But these systems, 
        which defendants have no way to challenge are racially biased, 
        unaccountable, and unreliable for forecasting violent crime.\7\
---------------------------------------------------------------------------
    \5\ EPIC v. DOJ (Criminal Justice Algorithms), EPIC, https://
epic.org/foia/doj/criminal-justice-algorithms/; Algorithms in the 
Criminal Justice System, EPIC, https://epic.org/algorithmic-
transparency/crim-justice/.
    \6\ Model Penal Code: Sentencing Sec. 6B.09 (Am. Law. Inst., 
Tentative Draft No. 2, 2011).
    \7\ Julia Angwin et al., Machine Bias, ProPublica (May 23, 2016), 
https://www.propublica.org/article/machine-bias-risk-assessments-in-
criminal-sentencing.

   Algorithms are used for social control. China's Communist 
        Party is deploying a ``social credit'' system that assigns to 
        each person government-determined favorability rating. 
        ``Infractions such as fare cheating, jaywalking, and violating 
        family-planning rules'' would affect a person's rating.\8\ Low 
        ratings are also assigned to those who frequent disfavored 
        websites or socialize with others who have low ratings. 
        Citizens with low ratings will have trouble getting loans or 
        government services. Citizens with high rating, assigned by the 
        government, receive preferential treatment across a wide range 
        of programs and activities.
---------------------------------------------------------------------------
    \8\ Josh Chin & Gillian Wong, China's New Tool for Social Control: 
A Credit Rating for Everything, Wall Street J., Nov. 28, 2016, http://
www.wsj.com/articles/chinas-new-tool-for-social-control-a-credit-
rating-for-everything-1480351590

   In the United States, U.S. Customs and Border Protection has 
        used secret analytic tools to assign ``risk assessments'' to 
        U.S. travelers.\9\ These risk assessments, assigned by the U.S. 
        Government to U.S. citizens, raise fundamental questions about 
        government accountability, due process, and fairness. They may 
        also be taking us closer to the Chinese system of social 
        control through AI.
---------------------------------------------------------------------------
    \9\ EPIC v. CBP (Analytical Framework for Intelligence), EPIC, 
https://epic.org/foia/dhs/cbp/afi/.

    In a recent consumer complaint to the Federal Trade Commission, 
EPIC challenged the secret scoring of young athletes.\10\ As EPIC's 
complaint regarding the Universal Tennis Rating system makes clear, the 
``UTR score defines the status of young athletes in all tennis related 
activity; impacts opportunities for scholarship, education and 
employment; and may in the future provide the basis for `social 
scoring' and government rating of citizens.'' \11\ As we explained to 
the FTC, ``EPIC seeks to ensure that all rating systems concerning 
individuals are open, transparent and accountable.'' \12\
---------------------------------------------------------------------------
    \10\ EPIC, EPIC Asks FTC to Stop System for Secret Scoring of Young 
Athletes (May 17, 2017), https://epic.org/2017/05/epic-asks-ftc-to-
stop-system-f.html; See also Shanya Possess, Privacy Group Challenges 
Secret Tennis Scoring System, Law360, May 17, 2017, https://www
.law360.com/articles/925379; Lexology, EPIC Takes a Swing at Youth 
Tennis Ratings, June 1, 2017, https://www.lexology.com/library/
detail.aspx?g=604e3321-dfc8-4f46-9afc-abd47c5a5179
    \11\ EPIC Complaint to Federal Trade Commission, In re Universal 
Tennis at 1 (May 17, 2017).
    \12\ Id.
---------------------------------------------------------------------------
    In re Universal Tennis, EPIC urged the FTC to (1) Initiate an 
investigation of the collection, use, and disclosure of children's 
personal information by Universal Tennis; (2) Halt Universal Tennis's 
scoring of children without parental consent; (3) Require that 
Universal Tennis make public the algorithm and other techniques that 
produce the UTR; (4) Require that Universal Tennis establish formal 
procedures for rectification of inaccurate, incomplete, and outdated 
scoring procedures; and (5) Provide such other relief as the Commission 
finds necessary and appropriate.\13\
---------------------------------------------------------------------------
    \13\ Id. at 13.
---------------------------------------------------------------------------
    ``Algorithmic Transparency'' must be a fundamental principle for 
consumer protection The phrase has both literal and figurative 
dimensions. In the literal sense, it is often necessary to determine 
the precise factors that contribute to a decision. If, for example, a 
government agency or private company considers a factor such as race, 
gender, or religion to produce an adverse decision, then the decision-
making process should be subject to scrutiny and the relevant factors 
identified.
    On October 12, 2016, The White House announced two reports on the 
impact of Artificial Intelligence on the U.S. economy and related 
policy concerns. Preparing for the Future of Artificial Intelligence 
concluded that ``practitioners must ensure that AI-enabled systems are 
governable; that they are open, transparent, and understandable; that 
they can work effectively with people; and that their operation will 
remain consistent with human values and aspirations.'' \14\
---------------------------------------------------------------------------
    \14\ Preparing for the Future of Artificial Intelligence, (Oct 
2016), Executive Office of the President, National Science and 
Technology Council, Comm. on Technology, https://obama
whitehouse.archives.gov/sites/default/files/whitehouse_files/
microsites/ostp/NSTC/preparing_
for_the_future_of_ai.pdf.
---------------------------------------------------------------------------
    Some have argued that algorithmic transparency is simply 
impossible, given the complexity and fluidity of modern processes. But 
if that is true, there must be some way to recapture the purpose of 
transparency without simply relying on testing inputs and outputs. We 
have seen recently that it is almost trivial to design programs that 
evade testing.\15\ And central to the science and innovation is the 
provability of results.
---------------------------------------------------------------------------
    \15\ Jack Ewing, In '06 Slide Show, a Lesson in How VW Could Cheat, 
N.Y. Times, Apr. 27, 2016, at A1.
---------------------------------------------------------------------------
    Europeans have long had a right to access ``the logic of the 
processing'' concerning their personal information.\16\ That principle 
is reflected in the U.S. in the publication of the FICO score, which 
for many years remained a black box for consumers, establishing credit 
worthiness without providing any information about the basis of 
score.\17\
---------------------------------------------------------------------------
    \16\ Directive 95/46/EC--The Data Protection Directive, art 15 (1), 
1995, http://www.data
protection.ie/docs/EU-Directive-95-46-EC--Chapter-2/93.htm.
    \17\ Hadley Malcom, Banks Compete on Free Credit Score Offers, USA 
Today, Jan. 25, 2015, http://www.usatoday.com/story/money/2015/01/25/
banks-free-credit-scores/22011803/.
---------------------------------------------------------------------------
    The continued deployment of AI-based systems raises profound issues 
for democratic countries. As Professor Frank Pasquale has said:

        Black box services are often wondrous to behold, but our black 
        box society has become dangerously unstable, unfair, and 
        unproductive. Neither New York quants nor California engineers 
        can deliver a sound economy or a secure society. Those are the 
        tasks of a citizenry, which can perform its job only as well as 
        it understands the stakes.\18\
---------------------------------------------------------------------------
    \18\ Frank Pasquale, The Black Box Society: The Secret Algorithms 
that Control Money and Information 218 (Harvard University Press 2015).

    We ask that this Statement from EPIC be entered in the hearing 
record. We look forward to working with you on these issues of vital 
importance to the American public.
            Sincerely,
                                            Marc Rotenberg,
                                                         President,
                                                                  EPIC.
                                      Caitriona Fitzgerald,
                                                   Policy Director,
                                                                  EPIC.
                                          Christine Bannan,
                                                     Policy Fellow.
                                                                  EPIC.

    Senator Wicker. The hearing record will remain open for 2 
weeks. During this time, Senators are asked to submit any 
questions for the record. Upon receipt, the witnesses are 
requested to submit their written answers to the Committee as 
soon as possible.
    Thank you very much. The hearing is now adjourned.
    [Whereupon, at 12 p.m., the hearing was adjourned.]

                            A P P E N D I X

    Response to Written Question Submitted by Hon. Amy Klobuchar to 
                          Dr. Cindy L. Bethel
    Question. Political ads on the Internet are more popular now than 
ever. In 2016, more than $1.4 billion was spent on digital 
advertisements and experts project that number will continue to 
increase. In October, I introduced the Honest Ads Act with Senators 
Warner and McCain, to help prevent foreign interference in future 
elections and improve the transparency of online political 
advertisements. We know that 90 percent of the ads that Russia 
purchased were issue ads meant to mislead and divide Americans. 
Increasing transparency and accountability online will benefit 
consumers and help safeguard future elections.
    Dr. Bethel, could making more data about political advertisements 
publicly available help improve the performance of algorithms designed 
to prevent foreign interference?
    Answer. More data that is high quality and has varied examples of 
content may be helpful in improving AI algorithms in general. The data 
though needs to have features that are detectable to learn from as part 
of the machine learning process. It is not clear from the political 
advertisements that have been promoted, as in the previous election, it 
was not evident what features could be used to learn from to detect 
involvement by foreign parties of these political advertisements. More 
data does not always equate to better results. There needs to be 
sufficient variations in key features that can be detected to be able 
to develop and test algorithms that will be effective and will have 
beneficial and meaningful results.
                                 ______
                                 
      Response to Written Question Submitted by Hon. Tom Udall to 
                          Dr. Cindy L. Bethel
    Question. Can you speak of some of the ways that government funded 
artificial intelligence development is now being used in the private 
sector?
    Answer. AI research has been funded for many years through the 
National Science Foundation, National Institutes of Health, the 
Department of Defense, USDA, among other agencies. It is possible that 
there has been government-funded research into artificial intelligence 
that has moved directly into the private sector for inclusion in 
product development. Generally, concepts, algorithms, and advancements 
developed as part of artificial intelligence research, has been 
integrated into product developments in the private sector but it is 
typically not a direct transfer from research into the private sector.
    Currently, the research I am performing with Therabot 
TM, the robotic therapeutic support companion, funded by the 
National Science Foundation, developed algorithms associated with 
artificial intelligence and machine learning that are being used in 
this application. This is a project that is planned for 
commercialization and making it available to the public. Mississippi 
State University has received government funding for projects that 
include artificial intelligence and have leveraged and transitioned 
some of the that knowledge into industry-based projects that benefit 
the private sector.
    The private sector has been active in funding their own 
advancements of artificial intelligence and machine learning. Many of 
the developments and advancements in applications of AI have been 
developed from private sector research and development groups. There 
are also cases, where a researcher receives government-funded grants 
for the development of AI and machine learning, and then may later 
transition into a private sector position and takes that knowledge with 
him or her to advance product development that benefits the private 
sector.
    There are numerous research developments that have been government 
funded under SBIR/STTR mechanisms that are joint funding for industry 
and academic researchers. NIH and USDA have been very active in 
transitioning the technology developments using their funds into the 
private sector and commercially available products. NSF and DoD also 
has programs such as Innovation Corps (I-Corps) that transition 
developments into commercialized and publicly available products. These 
have been successful programs available to researchers who have been 
funded under government grants and contracts.
    I am not sure of all of the government-funded research to date in 
artificial intelligence, so I am not sure exactly which projects have 
benefited or ended up being applied in the private sector. That would 
be a research project in itself that may be a worthwhile effort.
                                 ______
                                 
   Response to Written Questions Submitted by Hon. Maggie Hassan to 
                          Dr. Cindy L. Bethel
    Question 1. Artificial intelligence, or AI, holds tremendous 
promise for individuals who experience disabilities. For example, 
Google and Microsoft have technologies to process language and speech 
and translate it into a text format to assist individuals who are deaf 
and hard of hearing. Other technologies will go even further to improve 
the lives of people with disabilities and I would like to learn more 
from the panel about what we can expect. What other specific 
technologies are you aware of in the AI space that will help people who 
experience disabilities?
    Answer. There are numerous technologies that use artificial 
intelligence that are being developed to assist people with different 
types of disabilities to improve their quality of life. There is the 
development and use of brain-machine interfaces for operating 
wheelchairs and other devices through detecting signals sent with 
intention from the brain. There are exoskeletons that are being 
developed and prosthetics that learn and detect the signals and 
impulses from the nervous system to be able to customize how these 
devices work and enhance the capabilities of the end users. There are 
in-home assistive robots that are being developed to assist disabled 
and elderly people in their homes with reminders to take medications, 
to fetch different items, and to remotely monitor users so they can 
remain in their homes longer. The Therabot TM robot is being 
developed as a home therapy support tool to provide comfort and support 
for people who have experience post-traumatic stress or other types of 
mental health disorders that can be used to detect and alert clinicians 
or others when problems are detected. There are continual enhancements 
to technologies for the hearing and visually impaired users. These are 
just some of the many examples that use artificial intelligence to 
improve quality of life for people with different types of 
disabilities.

    Question 2. How will manufacturers and developers work to perfect 
this technology so that it can truly be a reliable tool for these 
individuals?
    Answer. These technologies will need to go through extensive 
testing and user studies/clinical trials to ensure the safety of the 
developments before they are sold to the public or used by the public. 
Edge cases need to be tested for events that may not occur frequently 
but have the possibility of happening. Once these types of technologies 
are developed and tested, then standards need to be established to 
ensure ongoing quality of the products for safe use as would be the 
case of any product used in medical applications or for consumer use.

    Question 3. What more can Congress to do assist with these efforts?
    Answer. While research and development is occurring it is important 
to not establish highly restrictive legislative policies or it will 
stifle the creativity and development by researchers. Once something is 
established and has been tested, then it may be necessary to legislate 
standards of practice for the protection and safety of the public using 
these items. This would be later in the process. Providing funding to 
support and an environment supportive of development of these items 
would allow the U.S. to stay on the top of research developments that 
use artificial intelligence and machine learning. Legislation should be 
limited to restrictions and standards for consumer and user safety.

    Question 4. As we see machine learning and AI increasingly embedded 
in products and services that we rely on, there are numerous cases of 
these algorithms falling short of consumer expectations. For example, 
Google and Facebook both promoted fraudulent news stories in the 
immediate wake of the Las Vegas Shooting because of their 
algorithms.\1\
---------------------------------------------------------------------------
    \1\ NYT: After Las Vegas Shooting, Fake News Regains Its Megaphone, 
Kevin Rose, 10/02/2017 https://www.nytimes.com/2017/10/02/business/las-
vegas-shooting-fake-news.html
---------------------------------------------------------------------------
    YouTube Kids is a service designed for children, and marketed as 
containing videos that are suitable for very young children. In 
November, YouTube Kids promoted inappropriate content due to 
algorithms.\2\ While the use of machine learning and AI holds limitless 
positive potential, at the current point, it faces challenges where we 
should not risk getting it wrong.
---------------------------------------------------------------------------
    \2\ NYT: On YouTube Kids, Startling Videos Slip Past Filters, Sapna 
Maheshwari, 11/04/2017 https://www.nytimes.com/2017/11/04/business/
media/youtube-kids-paw-patrol.html
---------------------------------------------------------------------------
    Should there be any formal or informal guidelines in place for what 
tasks are suitable to be done by algorithms, and which are still too 
important or sensitive to turn over; and what more can be done to 
ensure better and more accurate algorithms are used as you work to 
better develop this technology?
    Answer. In cases, where the algorithms are related to safety or 
life critical decisions then it may be necessary to have a human in the 
loop for sanity checks to ensure the best possible decision is made. 
When it comes to children, this would be case when the system needs to 
be thoroughly tested with a human involved to ensure the system is 
working well and there needs to be testing and validation that occurs 
to ensure those ``edge'' cases or rarer situations are also tested for 
no matter how unlikely it is for it to occur. Validation and testing 
should be performed extensively with adults prior to using the system 
with children and then tested with children with adult supervision. 
Mistakes can happen, but everything possible needs to be done to 
prevent issues that could potentially cause harm. In the case of 
decisions to weaponize autonomous systems there should be a human in 
the loop when it comes to decisions that impact human lives. There does 
need to be established standards and benchmarks to assist developers in 
testing to ensure the safety of a product before it is put in the hands 
of the public.

    Question 5. Machine learning and AI hold great promise for 
assisting us in preventing cybersecurity attacks. According to an IBM 
survey of Federal IT managers, 90 percent believe that artificial 
intelligence could help the Federal Government defend against real-
world cyber-attacks. 87 percent think AI will improve the efficiency of 
their cybersecurity workforce.\3\
---------------------------------------------------------------------------
    \3\ INFORMATION MANAGEMENT: AI seen as key tool in government's 
cybersecurity defense, Bob Violino, 11/30/2017 https://www.information-
management.com/news/artificial-intelligence-seen-as-key-tool-in-
governments-cybersecurity-defense
---------------------------------------------------------------------------
    While this is promising, the Federal Government currently faces a 
shortage of qualified cybersecurity employees, and to make matters 
worse, the pipeline of students studying these topics is not sufficient 
to meet our needs. A recent GAO report found that Federal agencies have 
trouble identifying skills gaps, recruiting and retaining qualified 
staff, and lose out on candidate due to Federal hiring processes.
    The George Washington University Center for Cyber & Homeland 
Security recently released a report titled ``Trends in Technology and 
Digital Security'' which stated:

        ``Traditional security operations centers are mostly staffed 
        with tier one analysts staring at screens, looking for unusual 
        events or detections of malicious activity. This activity is 
        similar to physical security personnel monitoring video cameras 
        for intruders. It is tedious for humans, but it is a problem 
        really well-suited to machine learning.'' \4\
---------------------------------------------------------------------------
    \4\ https://cchs.gwu.edu/sites/cchs.gwu.edu/files/downloads/
Fall%202017%20DT%20symposi
um20compendium.pdf 

    What effect will machine learning and AI will have on 
cybersecurity; and how do you think the Federal Government can best 
leverage the benefits offered by machine learning and AI to address our 
cybersecurity workforce shortage?
    Answer. The use of artificial intelligence and machine learning can 
definitely help with the tedious task of identifying potential threats 
to security. The initial evaluation could be performed using computer 
systems and they are adept at detecting anomalies and maybe even ones 
that a human may not readily detect. In cases where it is a threat to 
life or safety of humans involved, then it may need to verified by a 
cybersecurity trained specialist. Overall, the use of good algorithms 
and machine learning systems could help fill the gap that has occurred 
with the lack of trained cybersecurity personnel. If the initial 
detection work can be performed by the computer systems, then it would 
require less personnel to verify those findings. Overall, the use of 
well-developed AI and machine learning systems could be leveraged to 
address some of the workforce shortage issues associated with 
cybersecurity professionals. It is also important to better recruit for 
these programs and to show the benefits of being involved in this type 
of career. There are limitations to the government hiring practices and 
payscales, but these can be overcome. It may require changes though to 
these practices to entice students entering and choosing their career 
fields to consider careers in these areas.
                                 ______
                                 
      Response to Written Question Submitted by Hon. Tom Udall to 
                             Daniel Castro
    Question. In your testimony, you discussed how regulators will not 
need to intervene because the private sector will address artificial 
intelligence's problems--such as bias and discrimination. However, 
there have been studies that show implicit bias even when artificial 
intelligence is deployed. For example, a study \1\ about using AI to 
evaluate resumes found that candidates with names associated with being 
European American were 50 percent more likely to be offered an 
interview than candidates with names associated with being African-
American. What role should the Federal Government play where there is 
implicit bias and discrimination--particularly when companies are 
required to be ``Equal Opportunity Employers''?
---------------------------------------------------------------------------
    \1\ http://www.sciencemag.org/news/2017/04/even-artificial-
intelligence-can-acquire-biases-against-race-and-gender
---------------------------------------------------------------------------
    Answer. This is an important question. To clarify, regulators will 
need to continue to intervene to address specific policy goals, such as 
ensuring non-discrimination in hiring practices. However, policymakers 
do not necessarily need to create new laws and regulations only for AI 
to achieve those goals. Existing laws that make these practices illegal 
still apply, regardless of whether or not a company uses AI to 
discriminate against a protected class. For example, a company cannot 
circumvent its obligations in Title VII of the Civil Rights Act to 
discriminate against a particular race in its hiring practices simply 
by using an algorithm to review job applicants.
    There are additional steps policymakers can take to reduce bias. 
One way to assess bias, whether it be in analog processes or digital 
algorithms, is to have businesses conduct disparate impact analyses. 
For example, if a company is using an AI system to screen job 
applicants, and it has concerns about potential racial bias, it should 
test this system to assess its accuracy. If government agencies are 
early adopters of such AI-driven services, they can help identify 
potential areas of concerns. However, disparate impact analysis is only 
possible if organizations have data available to them. Moreover, 
regulators can also identify practices that are known to have a 
disparate impact, such as using certain criteria for making a credit or 
housing decision, and discouraging businesses from using those 
methods.\2\
---------------------------------------------------------------------------
    \2\ Travis Korte and Daniel Castro, ``Disparate Impact Analysis is 
Key to Ensuring Fairness in the Age of the Algorithm,'' Center for Data 
Innovation (2015), http://datainnovation.org/2015/01/disparate-impact-
analysis-is-key-to-ensuring-fairness-in-the-age-of-the-algorithm/.
---------------------------------------------------------------------------
    In addition, there are likely areas where additional policy is 
needed to protect workers. For example, Congress should consider 
passing laws such as the Employment Non-Discrimination Act (ENDA) to 
ensure that data about sexual orientation and gender identity cannot be 
used to unfairly harm workers.\3\ These types of laws address specific 
concerns of vulnerable populations, but they do not apply only to AI.
---------------------------------------------------------------------------
    \3\ For more on this topic as it relates to data, see: Joshua New 
and Daniel Castro, ``Accelerating Data Innovation: A Legislative Agenda 
for Congress,'' Center for Data Innovation (2015), http://
datainnovation.org/2015/05/accelerating-data-innovation-a-legislative-
agenda-for-congress/.
---------------------------------------------------------------------------
    Finally, policymakers should recognize that using AI can often 
reduce discrimination by limiting the potential for both implicit and 
explicit human bias. For example, a company that uses AI to screen 
applicants has the potential to reduce implicit bias of human managers 
in hiring practices. And while AI systems may not be perfect at the 
outset, as people identify problems, they will be able to more quickly 
resolve these issues. The same is not true for strictly human 
processes, where eliminating bias, is much more difficult.
                                 ______
                                 
     Response to Written Question Submitted by Hon. Gary Peters to 
                             Daniel Castro
    Question. Many have predicted that AI will have a profound effect 
on the labor market. Most predict that low-wage, routine-based jobs 
will be under the most pressure for replacement by AI. Meanwhile, 
recent advancements in technology has led to job creation that will 
mostly require highly-skilled, highly-educated workers. What evidence 
have you seen regarding businesses incorporating this labor shift into 
their business plans?
    Answer. Some of these predictions are not based on sound analysis. 
Bureau of Labor Statistics (BLS) projections show that the fastest 
growing jobs are not in high-skilled occupations. For example, the 
industry that BLS projects will have the most job growth between 2016-
2026 is the ``food services and drinking places'' industry.\4\ These 
are not high-wage jobs. Increased use of AI can yield higher rates of 
automation and hopefully fewer of these low-wages jobs. The way to 
achieve higher-wage jobs is by increasing productivity. In particular, 
increasing productivity in low-skill jobs will grow wages in these 
occupations.
---------------------------------------------------------------------------
    \4\ See ``Projections of industry employment, 2016-2026,'' Bureau 
of Labor Statistics, https://www.bls.gov/careeroutlook/2017/article/
projections-industry.htm.
---------------------------------------------------------------------------
    Some companies have taken steps to address disruption in the 
workforce. For example, Google and Facebook have made substantial 
commitments to funding for job retraining programs.\5\ However, 
overall, U.S. companies are investing less in training now than they 
were 15 years ago. A knowledge tax credit, where corporates receive a 
credit for qualified expenditures on worker training, would help 
address this problem.\6\
---------------------------------------------------------------------------
    \5\ ``Google pledges $1 billion to prepare workers for 
automation,'' Engadget, October 13, 2017, https://www.engadget.com/
2017/10/13/grow-with-google/.
    \6\ See Rob Atkinson, ``How a knowledge tax credit could stop 
decline in corporate training,'' The Hill, http://thehill.com/blogs/
pundits-blog/finance/235018-how-a-knowledge-tax-credit-could-stop-
decline-in-corporate.
---------------------------------------------------------------------------
                                 ______
                                 
   Response to Written Questions Submitted by Hon. Maggie Hassan to 
                             Daniel Castro
    Question 1. Artificial intelligence, or AI, holds tremendous 
promise for individuals who experience disabilities. For example, 
Google and Microsoft have technologies to process language and speech 
and translate it into a text format to assist individuals who are deaf 
and hard of hearing.
    Other technologies will go even further to improve the lives of 
people with disabilities and I would like to learn more from the panel 
about what we can expect.
    What other specific technologies are you aware of in the AI space 
that will help people who experience disabilities?
    Answer. AI will have widespread benefits for people with 
disabilities. As noted in the question, one of the major areas of 
impact AI will have is by allowing more people to interface with 
computer systems using their voice, instead of a keyboard. In 
particular, the combination of AI with the Internet of Things, will 
give people with many types of disabilities a better quality of life as 
they will now be able to control more of the world around them. These 
functions will allow more people with disabilities to participate in 
the workforce, go to school, and be more active in their communities. 
In addition, AI can be used to create smart agents that automate 
specific tasks, such as scheduling meetings, setting a thermostat, or 
re-ordering groceries. While these types of actions are conveniences 
for some people, for people with significant disabilities, they can be 
empowering and allow individuals significantly more autonomy and 
independence.

    Question 2. How will manufacturers and developers work to perfect 
this technology so that it can truly be a reliable tool for these 
individuals?
    Answer. One way to improve is by having industry work more closely 
with different populations of people with disabilities throughout the 
design and testing of new products. Working closely with different 
groups helps developers better anticipate user needs and pursue 
universal design.

    Question 3. What more can Congress to do assist with these efforts?
    Answer. One significant challenge is that the need to design for 
accessibility for people with disabilities is still underappreciated 
among technologists. One way to change this is to address this problem 
at the colleges and universities training the next generation of 
computer scientists and engineers. For example, Congress could 
establish NSF-funded Centers of Excellence for Accessible Design to 
prioritize this skillset and develop more curriculum. In addition, 
Congress should explore ways to encourage and support more people with 
disabilities to pursue careers in technology-related fields so they can 
be involved from the outset in the design and creation of more 
technologies. Finally, Congress should work to increase access to 
technology for people with disabilities, including by ensuring that 
programs designed to close the digital divide, such as PC or Internet 
access, are updated for newer technologies.

    Question 4. As we see machine learning and AI increasingly embedded 
in products and services that we rely on, there are numerous cases of 
these algorithms falling short of consumer expectations. For example, 
Google and Facebook both promoted fraudulent news stories in the 
immediate wake of the Las Vegas Shooting because of their 
algorithms.\7\ YouTube Kids is a service designed for children, and 
marketed as containing videos that are suitable for very young 
children. In November, YouTube Kids promoted inappropriate content due 
to algorithms.\8\ While the use of machine learning and AI holds 
limitless positive potential, at the current point, it faces challenges 
where we should not risk getting it wrong. Should there be any formal 
or informal guidelines in place for what tasks are suitable to be done 
by algorithms, and which are still too important or sensitive to turn 
over; and what more can be done to ensure better and more accurate 
algorithms are used as you work to better develop this technology?
---------------------------------------------------------------------------
    \7\ NYT: After Las Vegas Shooting, Fake News Regains Its Megaphone, 
Kevin Rose, 10/02/2017 https://www.nytimes.com/2017/10/02/business/las-
vegas-shooting-fake-news.html
    \8\ NYT: On YouTube Kids, Startling Videos Slip Past Filters, Sapna 
Maheshwari, 11/04/2017 https://www.nytimes.com/2017/11/04/business/
media/youtube-kids-paw-patrol.html
---------------------------------------------------------------------------
    Answer. AI, much like humans, is fallible. There should always be 
some oversight of AI, just as there should always be some oversight of 
humans. It is not a problem if AI systems make mistakes, unless these 
mistakes go undetected. So the key objective, whether a decision is 
being made by a computer of a human, is whether there is sufficient 
oversight appropriate to the level of risk for the individuals 
involved. This will likely be context dependent. This is one reason why 
it is inappropriate to talk about industry-wide regulation of AI, and 
much more appropriate to talk about industry-specific regulations of 
AI. For example, the Department of Transportation may have specific 
requirements for the types of oversight it wants for autonomous 
vehicles that looks very different than the type of oversight the 
Securities and Exchange Commission needs for AI-driven stock trading.

    Question 5. Machine learning and AI hold great promise for 
assisting us in preventing cybersecurity attacks. According to an IBM 
survey of Federal IT managers, 90 percent believe that artificial 
intelligence could help the Federal Government defend against real-
world cyber-attacks. 87 percent think AI will improve the efficiency of 
their cybersecurity workforce.\9\ While this is promising, the Federal 
Government currently faces a shortage of qualified cybersecurity 
employees, and to make matters worse, the pipeline of students studying 
these topics is not sufficient to meet our needs. A recent GAO report 
found that Federal agencies have trouble identifying skills gaps, 
recruiting and retaining qualified staff, and lose out on candidate due 
to Federal hiring processes. The George Washington University Center 
for Cyber & Homeland Security recently released a report titled 
``Trends in Technology and Digital Security'' which stated:
---------------------------------------------------------------------------
    \9\ INFORMATION MANAGEMENT: AI seen as key tool in government's 
cybersecurity defense, Bob Violino, 11/30/2017 https://www.information-
management.com/news/artificial-intelligence-seen-as-key-tool-in-
governments-cybersecurity-defense

        ``Traditional security operations centers are mostly staffed 
        with tier one analysts staring at screens, looking for unusual 
        events or detections of malicious activity. This activity is 
        similar to physical security personnel monitoring video cameras 
        for intruders. It is tedious for humans, but it is a problem 
        really well-suited to machine learning.'' \10\
---------------------------------------------------------------------------
    \10\ https://cchs.gwu.edu/sites/cchs.gwu.edu/files/downloads/
Fall%202017%20DT%20symposi
um%20compendium.pdf

    What effect will machine learning and AI will have on 
cybersecurity; and how do you think the Federal Government can best 
leverage the benefits offered by machine learning and AI to address our 
cybersecurity workforce shortage?
    Answer. AI is very good at specific tasks, such as pattern 
recognition and anomaly detection. This means that it will be useful 
for identifying attacks in real time, and it will be an especially 
important line of defense against zero-day attacks (i.e., attacks that 
use a previously undisclosed vulnerability). AI might also help 
developers eliminate certain types of vulnerabilities which may be 
identifiable at the outset, much like a spell-checker or grammar-
checker can review documents. However, AI will not be a panacea, as 
many cybersecurity risks are the result of poor implementation and a 
lack of adherence to best practices.
                                 ______
                                 
    Response to Written Questions Submitted by Hon. Gary Peters to 
                            Victoria Espinel
    Question 1. I am concerned by recent reports in Nature, The 
Economist, and Wall Street Journal about large tech firms monopolizing 
the talent in AI and machine learning. This concentration of talent can 
lead to several negative outcomes including long-term wage stagnation 
and income inequality.
    In your opinion, what steps or incentives might mitigate this 
concentration, encourage AI-experts to work at small and medium 
enterprises, or launch their own start-up with the goal of growing a 
business (rather than having a goal of being bought out by one of the 
tech giants)? Similarly, what incentives might encourage AI experts to 
become educators and trainers to help develop the next generation of AI 
experts?
    How can the Federal Government compete with the tech giants to 
attract experts needed to develop and implement AI systems for defense 
and civil applications?
    Answer. Artificial intelligence (AI) is a burgeoning field, with 
market dynamics that are quickly evolving. While the competition for AI 
expertise is certainly fierce, it is important to remember that the 
economic benefits of AI will be spread throughout the economy. By 
helping people make better data-driven decisions, AI is stimulating 
growth in every industry sector. It is helping to optimize 
manufacturing, improve supply chains, secure networks, and enhance 
products and services.
    The history of the technology industry suggests that innovation 
will continue to emerge from enterprises of all sizes. Indeed, BSA's 
membership is a testament to just how fiercely competitive the 
enterprise technology landscape is. Datastax, DocuSign, Salesforce, 
Splunk and Workday are just a few of the young companies that have 
disrupted the industry over the past 10 years and contributed to a wave 
of innovation that has made the U.S. software industry the envy of the 
world. Moreover, despite intense competition for AI expertise, small 
and medium-sized firms continue to play an incredibly important role in 
driving AI innovation. In fact, a recent study found that there are 
currently more than 2,000 AI startups that have raised almost $30 
billion in funding.\1\
---------------------------------------------------------------------------
    \1\ See Vala Afshar, AI Is Tranformational Technology and Major 
Sector Disruptor, Huffington Post (Dec. 5, 2017), https://
www.huffingtonpost.com/entry/ai-is-transformational-technology-and-
major-sector_us_5a259dbfe4b05072e8b56b6e.
---------------------------------------------------------------------------
    Because AI will be a huge driver of the global economy in the years 
ahead, it is vital that we examine the important issues that you have 
raised to ensure that the United States remains the global hub for AI 
innovation. There are three specific ways in which the government can 
increase the talent pool and attract that talent to the government for 
defense and civil applications. First, the government should increase 
its commitment to STEM education so that the United States prepares 
more men and women for AI-related careers. As part of this effort, 
government agencies can explore partnerships with academic institutions 
to provide internships to students studying in this field, which could 
pique their interests in pursuing careers in public service, and 
develop opportunities for academic researchers to share their technical 
expertise with government agencies. Second, the government should 
increase its funding for AI research, providing targeted investments 
into the ``high risk, high reward'' areas of basic research that are 
typically underfunded by the private sector.\2\ Third, the government 
should be ambitious in its goals. The greater the vision for how AI 
will improve government services and capabilities, the better it will 
do in attracting talent.
---------------------------------------------------------------------------
    \2\ See Jason Furman, Is this Time Different? The Opportunities and 
Challenges of Artificial Intelligence, AI Now: The Social and Economic 
Implications of Artificial Intelligence in the Near Term (July 7, 
2016), available at https://goo.gl/pzFDYw (``In 2015, American 
businesses devoted almost 1.8 percent of GDP to research and 
development, the highest share on record. But government investments in 
R&D have fallen steadily as a share of the economy since the 1960s. 
While business investment is critical, it is not sufficient. Basic 
research discoveries often have great social value because of their 
broad applicability, but there tends to be underinvestment in basic 
research by private firms because it is difficult for a private firm to 
appropriate the gains from such research. In fact, while the private 
sector accounts for roughly two-thirds of all spending on R&D, it is 
important to keep in mind that it largely invests in applied research 
while the Federal Government provides 60 percent of the funding for 
basic research.'').

    Question 2. Many have predicted that AI will have a profound effect 
on the labor market. Most predict that low-wage, routine-based jobs 
will be under the most pressure for replacement by AI. Meanwhile, 
recent advancements in technology has led to job creation that will 
mostly require highly-skilled, highly-educated workers. What evidence 
have you seen regarding businesses incorporating this labor shift into 
their business plans?
    Answer. The benefits of AI will be widespread, likely enhancing 
operations in every industry. As a result, AI also will likely create 
shifts in the labor market across the economy. The precise impact of AI 
on employment is uncertain. However, it is clear that AI will create 
new opportunities within existing jobs and new roles that require 
skills that the current workforce does not yet have. As a result, many 
BSA companies have launched initiatives to train employees, youth, and 
military veterans to help meet the demands of the future labor market. 
BSA would like to work with Congress to ensure we have the right 
programs and resources in place for the jobs of the future. We would be 
happy to come in and discuss with you the initiatives of the software 
industry that address this important issue.
                                 ______
                                 
   Response to Written Questions Submitted by Hon. Maggie Hassan to 
                            Victoria Espinel
    Question 1. Artificial intelligence, or AI, holds tremendous 
promise for individuals who experience disabilities. For example, 
Google and Microsoft have technologies to process language and speech 
and translate it into a text format to assist individuals who are deaf 
and hard of hearing. Other technologies will go even further to improve 
the lives of people with disabilities and I would like to learn more 
from the panel about what we can expect. What other specific 
technologies are you aware of in the AI space that will help people who 
experience disabilities?
    Answer. There are numerous ways in which AI is being used to 
improve the lives of people who experience disabilities. Below, I 
highlight a few examples.

   Visual impairment--Microsoft recently released an 
        intelligent camera app that uses a smartphone's built-in camera 
        functionality to describe to low-vision individuals the objects 
        that are round them. See Microsoft, Seeing AI, https://
        www.microsoft.com/en-us/seeing-ai/. The app opens up new 
        possibilities for the visually impaired to navigate the world 
        with more independence.

   Autism--IBM researchers are using AI to develop tools that 
        will help people with cognitive and intellectual disabilities, 
        such as autism, by breaking down complex sentences and phrases 
        to help them better understand normal speech and communicate 
        more effectively. See https://www-03.ibm.com/able/content-
        clarifier.html.

   Accessible public transportation--As part of a public-
        private partnership, an innovative project is underway that 
        aims to help disabled people and those with special needs 
        access public transportation by providing real-time information 
        through an Internet of Things system that helps them find the 
        right track, platform, train, and place to board, and alerts 
        them when to disembark. See David Louie, Artificial 
        Intelligence Research Is Helping the Disabled Use Public 
        Transportation, (July 12, 2017), http://abc7news.com/
        technology/ai-being-used-to-help-disabled-using-public-
        transportation/2210112/.

   Mobility Impairments--Microsoft's Windows 10 operating 
        system introduced Eye Control, a built-in eye tracking feature 
        that enables people with motor neurone disease and other 
        mobility impairments to navigate their computers. See Tas 
        Bindi, Microsoft Using AI to Empower Living With Disabilities, 
        Zdnet (Nov. 15, 2017), http://www.zdnet.com/article/microsoft-
        using-ai-to-empower-people-living-with-disabilities/.

   Alzheimer's--Researchers in Italy and Canada have developed 
        machine-learning algorithms to help identify patients that are 
        at risk of developing Alzheimer's. In early tests, the 
        technology has identified changes in the brain that lead to 
        Alzheimer's almost a decade before clinical symptoms would 
        appear. See Daisy Yuhas, Doctors Have Trouble Diagnosing 
        Alzheimer's. AI Doesn't, NBC News (Oct. 30, 2017), https://
        www.nbcnews.com/mach/science/doctors-have-trouble-diagnosing-
        alzheimer-s-ai-doesn-t-ncna815561.

    As researchers continue to apply AI to new settings, the myriad 
ways in which AI is used to enhance the lives of people with 
disabilities will only increase.

    Question 2. How will manufacturers and developers work to perfect 
this technology so that it can truly be a reliable tool for these 
individuals?
    Answer. BSA members that design and offer AI products and services 
have strong incentives to ensure that the technology is reliable, as 
they understand that building trust and confidence in AI systems is 
integral to successful and widespread deployment of AI services.
    There are a number of strategies that companies already employ to 
accomplish this objective. For example, one key priority is ensuring 
access to vast, robust, and representative data sets. Because AI 
technologies process and learn from data inputs, ensuring sufficient 
quantity and quality of data used to train AI systems is very important 
to enhancing the reliability of these services.
    In addition, another key step companies take to enhance reliability 
is testing their AI systems to ensure that they operate as intended, 
and making appropriate adjustments where they identify errors.
    Companies also recognize the need to protect AI systems from 
cyberattacks and are investing heavily in the development of advanced 
security tools.
    As companies continue to seek to expand the capabilities of AI 
technologies, investment in research and development will continue to 
be important to unleash the full potential of innovation, strengthen 
cybersecurity, and enhance the overall reliability of AI systems.

    Question 3. What more can Congress to do assist with these efforts?
    Answer. Congress can play a very important role in facilitating the 
deployment of AI services that help people with disabilities by 
ensuring, more broadly, that the U.S. maintains a flexible policy 
framework that spurs innovation in AI.
    Specifically, as I highlighted in my testimony, I think Congress 
can assist with these efforts in three key ways. First, Congress should 
pass the OPEN Government Data Act, which recognizes that government-
generated data is a national resource that can serve as a powerful 
engine for creating new jobs and a catalyst for economic growth, and 
that it is incredibly valuable in fostering innovation in AI and other 
data-driven services.
    Second, Congress should support efforts to promote digital trade 
and facilitate data flows. In a global economy, real-time access to 
data around the world has become increasingly critical for AI and other 
digital services to function. As a result, Congress should support 
modernizing trade initiatives, such as NAFTA, that seek to facilitate 
digital trade and limit inappropriate restrictions on cross-border data 
transfers.
    Third, Congress should promote U.S. investment in AI research, 
education, and workforce development to ensure that the U.S. remains 
globally competitive. Strategic investment in education and workforce 
development can help ensure that the next generation and our current 
workforce are prepared for the jobs of the future. In addition, 
promoting public sector and incentivizing private sector research will 
be essential to unlocking additional capabilities that AI can provide.

    Question 4. As we see machine learning and AI increasingly embedded 
in products and services that we rely on, there are numerous cases of 
these algorithms falling short of consumer expectations. For example, 
Google and Facebook both promoted fraudulent news stories in the 
immediate wake of the Las Vegas Shooting because of their 
algorithms.\3\ YouTube Kids is a service designed for children, and 
marketed as containing videos that are suitable for very young 
children. In November, YouTube Kids promoted inappropriate content due 
to algorithms.\4\ While the use of machine learning and AI holds 
limitless positive potential, at the current point, it faces challenges 
where we should not risk getting it wrong. Should there be any formal 
or informal guidelines in place for what tasks are suitable to be done 
by algorithms, and which are still too important or sensitive to turn 
over; and what more can be done to ensure better and more accurate 
algorithms are used as you work to better develop this technology?
---------------------------------------------------------------------------
    \3\ NYT: After Las Vegas Shooting, Fake News Regains Its Megaphone, 
Kevin Rose, 10/02/2017 https://www.nytimes.com/2017/10/02/business/las-
vegas-shooting-fake-news.html
    \4\ NYT: On YouTube Kids, Startling Videos Slip Past Filters, Sapna 
Maheshwari, 11/04/2017 https://www.nytimes.com/2017/11/04/business/
media/youtube-kids-paw-patrol.html
---------------------------------------------------------------------------
    Answer. Because AI is ultimately a technology that is intended to 
help people and organizations make better uses of data, I would be 
reluctant to prescribe any bright line rules about when its use may or 
may not be appropriate. However, it is important for companies that 
develop AI systems, and their customers, to consider the unique risks 
and potential unintended consequences that can arise when AI is 
deployed in particular settings. While AI is an invaluable tool for 
making sense of large quantities of data, there are settings where the 
intuition of subject matter experts will remain important. For 
instance, while AI systems certainly have an important role to play in 
helping to diagnose patients, they are a resource for a medical 
professional to consider in making a diagnosis or prescribing 
treatment, they should not be replacing a doctor's judgment.

    Question 5. Machine learning and AI hold great promise for 
assisting us in preventing cybersecurity attacks. According to an IBM 
survey of Federal IT managers, 90 percent believe that artificial 
intelligence could help the Federal Government defend against real-
world cyber-attacks. 87 percent think AI will improve the efficiency of 
their cybersecurity workforce.\5\
---------------------------------------------------------------------------
    \5\ INFORMATION MANAGEMENT: AI seen as key tool in government's 
cybersecurity defense, Bob Violino, 11/30/2017 https://www.information-
management.com/news/artificial-intelligence-seen-as-key-tool-in-
governments-cybersecurity-defense
---------------------------------------------------------------------------
    While this is promising, the Federal Government currently faces a 
shortage of qualified cybersecurity employees, and to make matters 
worse, the pipeline of students studying these topics is not sufficient 
to meet our needs. A recent GAO report found that Federal agencies have 
trouble identifying skills gaps, recruiting and retaining qualified 
staff, and lose out on candidate due to Federal hiring processes.
    The George Washington University Center for Cyber & Homeland 
Security recently released a report titled ``Trends in Technology and 
Digital Security'' which stated:

        ``Traditional security operations centers are mostly staffed 
        with tier one analysts staring at screens, looking for unusual 
        events or detections of malicious activity. This activity is 
        similar to physical security personnel monitoring video cameras 
        for intruders. It is tedious for humans, but it is a problem 
        really well-suited to machine learning.'' \6\
---------------------------------------------------------------------------
    \6\ https://cchs.gwu.edu/sites/cchs.gwu.edu/files/downloads/
Fall%202017%20DT%20symposi
um%20compendium.pdf

    What effect will machine learning and AI will have on 
cybersecurity; and how do you think the Federal Government can best 
leverage the benefits offered by machine learning and AI to address our 
cybersecurity workforce shortage?
    Answer. AI tools are revolutionizing network security, helping 
analysts parse through hundreds of thousands of security incidents per 
day to weed out false positives and identify threats that warrant 
further attention by network administrators. By automating responses to 
routine incidents and enabling security professionals to focus on truly 
significant threats, AI-enabled cyber tools are helping enterprises 
stay ahead of their malicious adversaries. For instance, AI has helped 
an enterprise security operations center ``reduce the time to remediate 
spearphishing attacks from there hours per incident to less than two 
minutes per incident.'' \7\ Importantly, AI is also helping to train 
the next generation of security analysts, teaching them to more quickly 
identify threats that need to be escalated through the chain of 
command.\8\ Greater deployment of AI is therefore a critical factor for 
addressing the cyber workforce shortage, which experts now estimate 
will climb 1.8 million positions by 2022.
---------------------------------------------------------------------------
    \7\ See Robert Lemos, AI Is Changing SecOps: What Security Analysts 
Need to Know, TechBeacon (Dec. 19, 2017), https://techbeacon.com/ai-
changing-secops-what-security-analysts-need-know.
    \8\ Id.
---------------------------------------------------------------------------
    However, AI alone will not solve the cyber workforce shortage. It 
is therefore incumbent on governments and industry to work 
collaboratively to grow the pipeline of cyber talent. To that end, BSA 
recently launched a new cybersecurity agenda \9\ that highlights four 
pathways for developing a 21st century cybersecurity workforce:
---------------------------------------------------------------------------
    \9\ BSA/The Software Alliance, A Cybersecurity Agenda for the 
Connected Age, available at www.bsa.org//media/Files/Policy/
BSA_2017CybersecurityAgenda.pdf.

   Increase access to computer science education: Expand 
        cybersecurity education for K-12 as well as in undergraduate 
        computer science programs, increase scholarships, and 
---------------------------------------------------------------------------
        incentivize minority students.

   Promote alternative paths to cybersecurity careers: Launch 
        careers through apprenticeship programs, community colleges, 
        cybersecurity ``boot camps,'' and government or military 
        service.

   Modernize training for mid-career professionals: Reform 
        Trade Adjustment Assistance, and update other mid-career re-
        training programs, to provide American workers with high-demand 
        cybersecurity and IT skills as digitalization transforms the 
        global economy.

   Improve the exchange of cybersecurity professionals between 
        the government and private sector: Enable private sector 
        experts to join the government for periodic or short-term 
        assignments.
                                 ______
                                 
    Response to Written Question Submitted by Hon. Amy Klobuchar to 
                          Dr. Dario Gil, Ph.D.
    Question. While I was at the hearing there was significant 
discussion about the future security applications for machine learning 
and artificial intelligence. As the Ranking Member on the Rules 
Committee, I am working with Senators Lankford, Harris and Graham on a 
bill to upgrade our election equipment to protect against cyber-
attacks. The Department of Homeland Security recently confirmed that 
hackers targeted 21 states' election systems in the run-up to the 2016 
election. As we prepare for 2018 and beyond, we must ensure that our 
election systems are secure, both from a hardware and a software 
perspective because election security is national security. Dr. Gil, 
can artificial intelligence and machine learning be used to identify 
and prevent cyber-attacks?
    Answer. The power of AI, like most machine learning techniques, 
lies in identifying broader trends, building models of what normal and 
expected behavior is and flagging anomalies. There has been tremendous 
value shown by deploying AI in the field of security, and we use it in 
a plethora of use cases: in, applications to flag systems and networks, 
to monitor for anomalies and raise alerts when these behaviors change. 
Such anomalous behavior may indicate an attack (or a benign error). AI 
has also been leveraged for generalizations to make it easier to 
identify new instances of known attacks. It can be used to learn about 
malware and exploits of vulnerabilities and can use that to detect new 
infections and intrusions better than rule-based systems. AI and AI-
based techniques can also help in hardening security protections to 
make it more difficult for attackers to successfully exploit a system. 
For example, automating tests (fuzzing) to probe for vulnerabilities 
using reinforcement learning can be more efficient than an exhaustive 
scan. It should be borne in mind that while there is no panacea for 
security, AI is a very powerful tool that can be employed to increase 
security when combined with a standard suite of best practices.
                                 ______
                                 
     Response to Written Questions Submitted by Hon. Tom Udall to 
                          Dr. Dario Gil, Ph.D.
    Question 1. As you are aware, New Mexico is home to two national 
laboratories--Sandia and Los Alamos. Can you speak of any partnership 
you have with the national laboratories?
    Answer. We have had, and continue to have, a number of partnerships 
with Los Alamos and Sandia National Labs. At this time, we cannot 
comment on individual projects, but we value the joint work we have 
with the labs and all of our external partners.

    Question 2. Can you speak of some of the ways that government 
funded artificial intelligence development is now being used in the 
private sector?
    Answer. Government-funded AI development is currently being used by 
IBM in the following ways:

   U.S. Army Research Labs funded the development of 
        technologies for rule-based systems and policy management. The 
        technology developed is being used now by IBM to improve and 
        support the ability of employees in a company to configure 
        computer software to efficiently execute high-volume, highly 
        transactional process functions, boosting capabilities and 
        saving time and money.

   U.S. Army Research Labs funded the development of AI enabled 
        algorithms for analyzing and understanding properties of large 
        numbers of moving objects. The technology developed has been 
        used to create commercial cloud-based services such as The 
        Weather Company LongShip service to predict the influence of 
        weather impacts on traffic. It also has been incorporated into 
        commercial software products such as DB/2.

    Looking more broadly, there are a variety of ways in which 
government funded research can be used by the private sector. These 
include:

   Government funded technology has been used to create dual-
        purpose technologies--those which serve the needs of the 
        private sector as well as the needs of the government agency 
        which sponsored the work. One expression of this is when the 
        private sector produces a COTS (Custom Off the Shelf) product 
        which allows the government to meet their specific requirement 
        at a lower overall cost.

   Government funded research (primarily in basic sciences) has 
        a research horizon which is typically longer than the private 
        sector. As a result, government funded research has been used 
        to create technology at the earlier stage. There are many 
        instances where government support has been used for 
        technologies at the TRL (Technology Readiness Level) of 1-4, 
        and the successful ones from this level of exploration have led 
        to commercialized products at TRL level 5 and above.

   Government funded technology has been used to produce open 
        source software, which private sector companies develop further 
        and use to create new offerings. In many cases, the open source 
        software is used by academics and the public at large for 
        knowledge creation.

   Government funded alliances have driven collaboration 
        between private sector researchers, academics and government 
        researchers. This cross-fertilization of researchers working 
        towards a common goal has been beneficial to all parties--
        including employment opportunities for students, infusion of 
        new ideas in industry activities, improvements in government, 
        and formation of lasting collaborations.
                                 ______
                                 
     Response to Written Question Submitted by Hon. Gary Peters to 
                          Dr. Dario Gil, Ph.D.
    Question. A major challenge AI and machine learning developers need 
to address is the ability to ensure prolonged safety, security, and 
fairness of the systems. This is especially true of systems designed to 
work in complex environments that may be difficult to replicate in 
training and testing, or systems that are designed for significant 
learning after deployment. Dr. Gil, you testified that IBM is looking 
to build trust in AI by following a set of principles to guide your 
development and use of AI systems. Would you please provide more detail 
about how these principles are being implemented? How will these 
principles prevent a system designed to learn after deployment from 
developing unacceptable behavior over time?
    Answer. The currently available AI products, such as factory 
robots, personal digital assistants, and healthcare decision support 
systems, are designed to perform one narrow task, such as assemble a 
product, provide a weather forecast or make a purchase order, or help a 
radiologist interpret an X-ray. When these technologies learn after 
deployment, they do so in the context of that narrow task, and do not 
have the ability to learn other tasks on their own, even similar ones. 
The kind of AI systems that can acquire new skills and perform new 
cognitive and reasoning tasks autonomously are actively being 
researched. This effort does not include only the development of 
underlying learning and reasoning capabilities; the AI research 
community is actively pursuing the capabilities that would ensure the 
safety, security and fairness of these systems.
    To start with, the principles of safe design are applied to a wide 
variety of engineered systems, such as trains, safety breaks, 
industrial plants, flight autopilot systems, and robotic laser surgery. 
Some of these principles apply directly to the design of AI systems, 
some will be adapted, and new ones will have to be defined. For 
example, it is possible to constrain the space of outcomes or actions a 
robot can perform, to ensure that it does not accidentally come into 
contact with human workers and cause injury. Similarly, robots in 
complex environments that encounter completely new situations could be 
designed to require human intervention. Another direction is to embed 
principles of safe and ethical behavior in the AI reasoning mechanisms, 
so that they can distinguish between right and wrong actions.
    With respect to the fairness of the AI systems, we are currently 
pursuing a range of efforts aimed at developing and embedding in our 
services and offerings techniques for bias detection, certification, 
and mitigation. For example, we have developed algorithms that can de-
bias training data so that any AI system that learns from such data 
does not discriminate against protected groups (e.g., those defined by 
race or gender). We also are working on using blockchain to ensure the 
integrity of an AI system by making sure that it is secure, auditable 
and used as intended. We also are developing capabilities to enhance 
the explainability and interpretability of AI systems, so that 
unacceptable behaviors can be easily discovered and removed.
    IBM has established the following principles for the artificial 
intelligence/cognitive era:

    Purpose: The purpose of AI and cognitive systems developed and 
applied by the IBM company is to augment human intelligence. Our 
technology, products, services and policies will be designed to enhance 
and extend human capability, expertise and potential. Our position is 
based not only on principle but also on science. Cognitive systems will 
not realistically attain consciousness or independent agency. Rather, 
they will increasingly be embedded in the processes, systems, products 
and services by which business and society function--all of which will 
and should remain within human control.

    Transparency: For cognitive systems to fulfill their world-changing 
potential, it is vital that people have confidence in their 
recommendations, judgments and uses. Therefore, the IBM company will 
make clear:

   When and for what purposes AI is being applied in the 
        cognitive solutions we develop and deploy.

   The expertise that informs the insights of cognitive 
        solutions, as well as the methods used to train those systems 
        and solutions.

   The principle that clients own their own business models and 
        intellectual property and that they can use AI and cognitive 
        systems to enhance the advantages they have built. We will work 
        with our clients to protect their data and insights, and will 
        encourage our clients, partners and industry colleagues to 
        adopt similar practices.

    Skills: The economic and societal benefits of this new era will not 
be realized if the human side of the equation is not supported. This is 
uniquely important with cognitive technology, which augments human 
intelligence and expertise and works collaboratively with humans. 
Therefore, the IBM company will work to help students, workers and 
citizens acquire the skills and knowledge to engage safely, securely 
and effectively in a relationship with cognitive systems, and to 
perform the new kinds of work and jobs that will emerge in a cognitive 
economy.

    Data: Since AI is heavily based on data, IBM has developed a 
framework of best practices for data stewardship \1\ that ensures great 
care and responsibility in data ownership, storage, security, and 
privacy. IBM abides by these practices and, as a result, serves as a 
data steward providing transparent and secure services. For example, we 
write client agreements with full transparency and will not use client 
data unless they agree to such use. We will limit that use to the 
specific purposes clearly described in the agreement. IBM does not put 
`backdoors' in its products for any government agency, nor do we 
provide source code or encryption keys to any government agency for the 
purpose of accessing client data.
---------------------------------------------------------------------------
    \1\ IBM: Data Responsibility@IBM, Khttps://www.ibm.com/blogs/
policy/wp-content/uploads/2017/10/IBM_DataResponsibility-A4_WEB.pdf
---------------------------------------------------------------------------
    We are working on a range of efforts aimed at developing and 
embedding in our services and offerings techniques for bias detection, 
certification, and mitigation. For example, we are working on improving 
the accuracy of directly interpretable decision-support algorithms, 
such as decision trees and rule sets, as well as enhancing the 
interpretability of deep learning neural net models.
    Moreover, as we develop innovative AI systems, we are guided by the 
principles of safety engineering. Some of these principles could be 
directly applied to the design of AI systems, some will be adapted, and 
new ones will have to be defined. For example, robots in complex 
environments that encounter completely new situations could be designed 
to require human intervention. Another direction is to embed principles 
of safe and ethical behavior in their reasoning mechanisms, so they can 
distinguish between right and wrong actions.
    Finally, we are working to develop AI systems that act according to 
human values that are relevant for the scenarios and communities in 
which such systems will be deployed. This means constraining the 
learning, reasoning and optimization machinery inside AI systems to 
behavioral constraints. These constraints will ensure that the actions 
of the AI system comply with values and guidelines that humans define 
as appropriate for the specific use case and application. Such 
behavioral constraints should be learned offline  (i.e., by training 
the system with data or via simulation), modified only by humans, and 
given a higher priority compared to online policies (outcomes that an 
AI system learns post-deployment, which are based on reinforcement 
learning or other machine learning approaches aimed at reward 
maximization and optimization).
                                 ______
                                 
   Response to Written Questions Submitted by Hon. Maggie Hassan to 
                          Dr. Dario Gil, Ph.D.
    Question 1. Artificial intelligence, or AI, holds tremendous 
promise for individuals who experience disabilities. For example, 
Google and Microsoft have technologies to process language and speech 
and translate it into a text format to assist individuals who are deaf 
and hard of hearing. Other technologies will go even further to improve 
the lives of people with disabilities and I would like to learn more 
from the panel about what we can expect. What other specific 
technologies are you aware of in the AI space that will help people who 
experience disabilities?
    Answer. AI technologies are enabling many exciting new assistive 
functions by enhancing machines' ability to see, hear, interpret 
complex signals, and operate in the real world through action and 
dialog. Essential building blocks for these new capabilities are 
machine vision, speech to text, text to speech, natural language 
understanding and generation, emotion recognition, and machine learning 
to interpret sensor data.
    For example, with AI vision, it is now becoming possible to 
describe an image, a local environment, or a video to a person with 
visual impairment. Further, these technologies will soon support 
wearable assistants that can recognize people, objects, landmarks and 
obstacles in the environment, and guide a person safely to an 
unfamiliar destination.
    AI speech to text capabilities, coupled with natural language 
understanding, enable a quadriplegic individual to control their 
environment through speech commands, providing a new level of autonomy. 
Machine learning techniques can translate brain and nerve signals into 
commands for prosthetic limbs and convey a sense of touch.
    AI natural language understanding and generation enable 
communication of knowledge in the form most easily understood by an 
individual, whether that means generating a description of a graph for 
a blind person, reading text aloud for a person with dyslexia, 
simplifying a complex document for a person with an intellectual 
disability, or, one day, translating between spoken languages and sign 
languages used by people who are deaf and hard of hearing.
    AI embedded in autonomous vehicles, intelligent wheelchairs and 
interactive assistance robots will provide physical independence and 
assistance for many. For example, IBM, the CTA (Consumer Technology 
Association) Foundation, and Local Motors are exploring applications of 
Watson technologies to developing the world's most accessible self-
driving vehicle, able to adapt its communication and personalize the 
overall experience to suit each passenger's unique needs.
    Machine learning on sensor data from instrumented environments can 
support an older adult in living independently and safely at home by 
learning their normal patterns of behavior and providing assistance and 
alerts.
    Just as importantly, AI technologies will benefit people with 
disabilities by analyzing websites and applications, finding and fixing 
accessibility problems in a more automated way than was previously 
possible.

    Question 2. How will manufacturers and developers work to perfect 
this technology so that it can truly be a reliable tool for these 
individuals?
    Answer. Perfecting AI technologies will take significant 
experimentation and depends on the availability of data. For these core 
technologies to be reliable, applications for people with disabilities 
should be explored as early as possible and used to drive requirements. 
This includes consulting with, and testing by, people with disabilities 
in realistic environments.
    At IBM, we are exploring the potential of AI technologies to 
support people with disabilities, and older adults through several 
initiatives and collaborations. IBM researchers are exploring how 
Watson's language-processing software could help people with cognitive 
disabilities by simplifying text, how older adults' patterns of 
activity can be learned, how a blind person can navigate and find 
objects effectively using machine vision, and how AI can enable our 
accessibility test tools to move from pointing out problems to actively 
suggesting solutions.
    Secondly, it is essential that people with disabilities are 
represented adequately in training data, to prevent new forms of 
discrimination from emerging. For example, an authentication system 
based on voice recognition should be able to recognize people with 
dysarthric speech. Manufacturers and developers applying AI 
technologies should incorporate mechanisms to recognize and gracefully 
handle exceptions, falling back on human judgment for cases that are 
outside their training.

    Question 3. What more can Congress to do assist with these efforts?
    Answer. For AI to deliver the promised economic and societal 
benefits to a broader range of people, including people with 
disabilities, both policy support and public investment from the U.S. 
Government are critical.
    Given the great diversity of human abilities, it is a challenge for 
manufacturers and developers to ensure diversity in training data. For 
example, speech recognition training data should ideally include people 
who stutter. Government investment in initiatives to make diverse data 
broadly available would accelerate our ability to make AI technology 
more inclusive, and to apply AI techniques to new accessibility 
problems. Government support for controlled studies with people with 
disabilities will also accelerate the inclusion of people with 
disabilities.
    Access to other forms of data is also critical. An indoor-outdoor 
navigation system for blind people relies on public outdoor maps, but 
indoor maps are privately owned. A centralized mechanism to share such 
maps for accessibility purposes would remove a practical barrier to the 
widespread use of such systems. AI vision techniques depend on the use 
of images or video to describe people and objects to people with visual 
impairment. Government leadership is needed to address privacy concerns 
of individuals and copyright concerns of organizations over the use of 
images of their faces or products for accessibility purposes. There is 
a copyright exception for converting books into braille, and a similar 
solution could be effective here.
    Secondly, policy support can help to counter the danger of new 
forms of discrimination. For example, reinstating the Department of 
Justice rulemaking on accessibility guidelines for public websites 
would emphasize the importance of accessibility, and spur efforts by 
industry to include people with disabilities in development.
    Most of the examples in question 1 describe ways that AI 
technologies can assist people with sensory or physical impairments. 
There is a need to foster standards and policies to address the needs 
of people with cognitive disabilities, which will encourage application 
of AI technologies to these challenges.

    Question 4. As we see machine learning and AI increasingly embedded 
in products and services that we rely on, there are numerous cases of 
these algorithms falling short of consumer expectations. For example, 
Google and Facebook both promoted fraudulent news stories in the 
immediate wake of the Las Vegas Shooting because of their 
algorithms.\2\ YouTube Kids is a service designed for children, and 
marketed as containing videos that are suitable for very young 
children. In November, YouTube Kids promoted inappropriate content due 
to algorithms.\3\ While the use of machine learning and AI holds 
limitless positive potential, at the current point, it faces challenges 
where we should not risk getting it wrong. Should there be any formal 
or informal guidelines in place for what tasks are suitable to be done 
by algorithms, and which are still too important or sensitive to turn 
over; and what more can be done to ensure better and more accurate 
algorithms are used as you work to better develop this technology?
---------------------------------------------------------------------------
    \2\ NYT: After Las Vegas Shooting, Fake News Regains Its Megaphone, 
Kevin Rose, 10/02/2017 https://www.nytimes.com/2017/10/02/business/las-
vegas-shooting-fake-news.html
    \3\ NYT: On YouTube Kids, Startling Videos Slip Past Filters, Sapna 
Maheshwari, 11/04/2017 https://www.nytimes.com/2017/11/04/business/
media/youtube-kids-paw-patrol.html
---------------------------------------------------------------------------
    Answer. AI is already more capable than humans in narrow domains, 
some of which involve delicate decision making. Humanity is not 
threatened by them, but many people could be affected by their 
decisions. Examples are autonomous online trading agents, media and 
news services, and soon autonomous cars. Even though AI algorithms are 
usually evaluated based on their accuracy, that is, their ability to 
produce correct results, this is only one component of a bigger 
picture. We need to be able to assess the impact of their decisions in 
the narrow domains where they will function.
    To understand the suitability of an AI system with respect to 
performing a specific task, one must consider not only their accuracy, 
but also the context, the possible errors, and the consequences on the 
impacted communities. Furthermore, the assessment of risk should be 
carried out with respect to both the risk of ``doing it'' and the risk 
of ``not doing it'', as in many fields we already know the consequences 
of wrong decisions made by humans. For example, melanoma detection from 
skin images is a task that AI algorithms can perform at high levels of 
accuracy. Even though there is still a possibility of error, it is 
beneficial to deploy such systems in healthcare decision support, in a 
way that would augment human decision-making process. On the other 
hand, let us consider automated trading systems. A bad decision in 
these systems may be (and has been) a financial disaster for many 
people. That will also be the case for self-driving cars. Some of their 
decisions will be critical and possibly affect lives. Because sectors 
like finance and transportation can carry large risks, protections have 
always been in place through existing regulations. These existing 
protections are properly designed to provide consumer protection even 
with the advent of new technologies like AI.
    Finally, we believe that in many applications, rather than 
considering only fully autonomous AI solutions, the most effective 
approach is to build AI systems that support humans and work with them 
in performing a task. For example, in a breast cancer detection study, 
it has been shown that doctors and AI working together achieve a higher 
degree of accuracy than just doctors or AI separately.

    Question 5. Machine learning and AI hold great promise for 
assisting us in preventing cybersecurity attacks. According to an IBM 
survey of Federal IT managers, 90 percent believe that artificial 
intelligence could help the Federal Government defend against real-
world cyber-attacks. 87 percent think AI will improve the efficiency of 
their cybersecurity workforce.\4\
---------------------------------------------------------------------------
    \4\ INFORMATION MANAGEMENT: AI seen as key tool in government's 
cybersecurity defense, Bob Violino, 11/30/2017 https://www.information-
management.com/news/artificial-intelligence-seen-as-key-tool-in-
governments-cybersecurity-defense
---------------------------------------------------------------------------
    While this is promising, the Federal Government currently faces a 
shortage of qualified cybersecurity employees, and to make matters 
worse, the pipeline of students studying these topics is not sufficient 
to meet our needs. A recent GAO report found that Federal agencies have 
trouble identifying skills gaps, recruiting and retaining qualified 
staff, and lose out on candidate due to Federal hiring processes.
    The George Washington University Center for Cyber & Homeland 
Security recently released a report titled ``Trends in Technology and 
Digital Security'' which stated:

        ``Traditional security operations centers are mostly staffed 
        with tier one analysts staring at screens, looking for unusual 
        events or detections of malicious activity. This activity is 
        similar to physical security personnel monitoring video cameras 
        for intruders. It is tedious for humans, but it is a problem 
        really well-suited to machine learning.'' \5\
---------------------------------------------------------------------------
    \5\ https://cchs.gwu.edu/sites/cchs.gwu.edu/files/downloads/
Fall%202017%20DT%20symposi
um%20compendium.pdf

    What effect will machine learning and AI will have on 
cybersecurity; and how do you think the Federal Government can best 
leverage the benefits offered by machine learning and AI to address our 
cybersecurity workforce shortage?
    Answer. AI and machine learning will be a disruptive force in the 
field of cybersecurity by providing the potential for aiding both in 
the defense and protection of critical infrastructure, leveling the 
playing field between large nation states and smaller niche players.
    From a defensive standpoint, AI has shown promise in automating 
defenses, such as probing systems for weaknesses, including software 
vulnerabilities and configuration errors. Penetration testing and bug 
finding tools have benefited tremendously from AI techniques in 
improving their efficiency to more quickly evaluate systems for 
weaknesses and increase coverage of the evaluated space. Security 
monitoring tools have also benefited greatly from AI and will continue 
to do so as AI systems improve. Automation can be leveraged to process 
suspicious alerts and events that warrant investigation, performing 
many of the rote tasks typically performed by low-level analysts. These 
automated tools will provide an analyst a more complete picture of the 
events unfolding, highlight meaningful information and context, triage, 
and allow the analyst to provide a higher-level response. This can 
allow security analysts to investigate far more alerts than are 
currently possible, and hopefully make fewer errors in how those alerts 
are processed. Security operations can be conducted at machine-scale as 
opposed to human-scale.
    The Federal Government can learn from the experience in research, 
industry and academia in leveraging AI to develop and deploy the next 
generation of AI-powered defenses that will be necessary to protect the 
Nation's critical infrastructure. This requires significant leadership 
and outreach on behalf of the Government to industry and academia on 
the following fronts:

   Declare AI Leadership in Cyber Security as a national 
        research and development priority.

   Evolve and develop the Nation's cybersecurity strategy to 
        address the AI-powered threats to critical infrastructure with 
        AI-powered defenses.

   Initiate U.S. Government programs, through various policy 
        and funding agencies (e.g., OSTP, DARPA, IARPA, NSF, NIST etc.) 
        to fund and sponsor leading edge research in areas of 
        intersection between AI and security

   Set policies and standards for procurement of next 
        generation security controls by the U.S. Government.
                                 ______
                                 
    Response to Written Question Submitted by Hon. Amy Klobuchar to 
                      Dr. Edward W. Felten, Ph.D.
    Question. Political ads on the Internet are more popular now than 
ever. In 2016, more than $1.4 billion was spent on digital 
advertisements and experts project that number will continue to 
increase. In October, I introduced the Honest Ads Act with Senators 
Warner and McCain, to help prevent foreign interference in future 
elections and improve the transparency of online political 
advertisements. We know that 90 percent of the ads that Russia 
purchased were issue ads meant to mislead and divide Americans. 
Increasing transparency and accountability online will benefit 
consumers and help safeguard future elections. Dr. Felten, can machine 
learning be used to help identify issue ads and stop misinformation 
from spreading online?
    Answer. Yes, machine learning can be useful in several ways. 
Machine learning can help to classify the nature or topic of ads, to 
distinguish issue ads from others and to characterize the issue being 
addressed by an ad. Machine learning can be useful in in determining 
the source of an ad, including in identifying when a single source is 
trying to disguise itself as a set of separate, independent sources. 
More broadly, machine learning can be helpful in identifying 
misinformation and disinformation campaigns, and in targeting 
countermeasures to maximize the impact on a harmful campaign while 
minimizing collateral damage.
    Three caveats are in order, however. First, more research will be 
necessary to take full advantage of these opportunities. That research 
is best done using realistic datasets derived from platforms' 
experience with past disinformation campaigns. Second, machine learning 
methods will necessarily be less than perfectly accurate. Not only will 
they fail to spot some disinformation campaigns, they will also 
sometimes misclassify content or a user as malicious when they are in 
fact benign. Appropriate use of machine learning in this setting will 
require both a careful technical evaluation of the likelihood of 
errors, and a policy approach that recognizes the harm that might be 
done by errors. Finally, machine learning systems for detecting 
anomalies depend on a variety of data sources and signals, and the 
success of machine learning depends on the characteristics of those 
sources and signals. Real-world data is sometimes erroneous and often 
incomplete, in ways that could frustrate the use of machine learning 
for this application or render it less accurate. Where data signals 
derive from the votes or clicks of users, the resulting system may be 
subject to gaming or manipulation, so such signals should be used with 
caution, especially in systems that aim to limit disinformation.
                                 ______
                                 
      Response to Written Question Submitted by Hon. Tom Udall to 
                      Dr. Edward W. Felten, Ph.D.
    Question. In your testimony, you discussed how adoption of 
artificial intelligence can inadvertently lead to biased decisions. 
What specific steps should the Federal Government and other users take 
to improve the data and ensure that datasets minimize societal bias--
especially with regard to vulnerable populations?
    Answer. The results of a machine learning system can only be as 
accurate as the dataset on which the system was trained. If a community 
is underrepresented in the dataset, relative to its representation in 
the population, then that community is likely to be poorly served by 
the system, as the system will not put enough weight on the 
characteristics of the underrepresented community.
    Practitioners should take care to ensure that datasets are 
representative of the population, to the extent possible. Where this is 
not possible, deficiencies in the dataset should be noted carefully, 
and steps should be taken to mitigate the deficiencies. For example, it 
is sometimes possible to correct for a group's underrepresentation in a 
data analysis or machine learning procedure by putting greater weight 
on data points that represent that group. Additional statistical 
methods exist that can counteract the effect of non-representative 
datasets.
    Another common source of error or bias in machine learning occurs 
when a system is tasked with learning from examples of past decisions 
made by people. If those past decisions were biased, the machine is 
likely to learn to replicate that bias. Whenever a system is trained 
based on past human decisions, care should be taken to consider the 
social and historical context of those past decisions and to look for 
indications of bias in the system's output, and anti-bias techniques 
should be used in designing or training the system if possible.
    In addition to technical measures in the design and use of AI 
systems, the possibilities of bias--whether that AI will introduce bias 
or that AI will open opportunities to measure and counteract human 
bias--should be taken into account in making policy decisions. 
Consulting with technical experts, and including technical expertise in 
the policymaking conversation, are important steps toward good policy 
in this area.
                                 ______
                                 
     Response to Written Question Submitted by Hon. Gary Peters to 
                      Dr. Edward W. Felten, Ph.D.
    Question. I am concerned by recent reports in Nature, The 
Economist, and Wall Street Journal about large tech firms monopolizing 
the talent in AI and machine learning. This concentration of talent can 
lead to several negative outcomes including long-term wage stagnation 
and income inequality.
    In your opinion, what steps or incentives might mitigate this 
concentration, encourage AI-experts to work at small and medium 
enterprises, or launch their own start-up with the goal of growing a 
business (rather than having a goal of being bought out by one of the 
tech giants)? Similarly, what incentives might encourage AI experts to 
become educators and trainers to help develop the next generation of AI 
experts?
    How can the Federal Government compete with the tech giants to 
attract experts needed to develop and implement AI systems for defense 
and civil applications?
    Answer. One approach to making smaller companies attractive to top 
AI talent is to adopt pro-competition policies generally. Because AI 
often relies on large datasets, and those datasets are more likely to 
be held by large companies, there may be a natural tendency toward 
concentration in AI-focused industry sectors. Public policy can help to 
ensure that smaller companies can be viable competitors. The Federal 
Government can also provide some large, high quality datasets that may 
be useful to individuals and companies of all sizes.
    At present, the demand for highly skilled AI experts exceeds the 
supply, leading to a scarcity of those experts in all but the best-
funded environments. In the long run, steps to increase the education 
and training of AI professionals are the most important means to 
strengthen our national talent base and broaden the availability of 
expertise.
    The talent pipeline can be widened at every stage. At the K-12 
level, access to a good computer science course should be available to 
every student. A bipartisan coalition of states and nonprofit actors is 
working toward this goal. At the university and graduate level, access 
to education is limited by the number of trained faculty available to 
teach advanced AI and machine learning courses.
    It is difficult to underestimate the importance of supporting a 
large and robust public research community. This ensures that access to 
the latest knowledge and techniques in AI is available to the public 
and not limited to a few companies' researchers. It widens the talent 
pipeline because AI research funding enables faculty hiring in AI, 
which increases the national capacity to train AI leaders. Federally-
funded research projects serve as the main training ground for the next 
generation of research leaders.
                                 ______
                                 
   Response to Written Questions Submitted by Hon. Maggie Hassan to 
                      Dr. Edward W. Felten, Ph.D.
    Question 1. Artificial intelligence, or AI, holds tremendous 
promise for individuals who experience disabilities. For example, 
Google and Microsoft have technologies to process language and speech 
and translate it into a text format to assist individuals who are deaf 
and hard of hearing. Other technologies will go even further to improve 
the lives of people with disabilities and I would like to learn more 
from the panel about what we can expect. What other specific 
technologies are you aware of in the AI space that will help people who 
experience disabilities?
    Answer. There are many examples, of which I will highlight three 
here.
    First, self-driving vehicles will improve mobility and lower the 
cost of transportation for people who are unable to drive. These 
vehicles will have major safety benefits in the long run, and they are 
already starting to benefit people with disabilities. Maintaining 
policies to encourage safety-conscious testing and deployment of self-
driving vehicles will benefit all Americans, and especially those with 
disabilities.
    Second, computer vision and image interpretation systems have the 
potential to help those with visual disabilities process information 
about their surroundings. These systems are demonstrating an increasing 
capacity to identify specific objects and people in complex scenes, and 
to model and predict what might happen next, such as warning of 
potential dangers.
    Third, AI can help to identify barriers to accessibility. For 
example, Project Sidewalk at the University of Maryland combines 
crowdsourced data collection with AI techniques to build a database of 
curb, ramp, and sidewalk locations, and analyze it to identify 
accessibility problems. This can help city planners and property owners 
recognize accessibility failures.

    Question 2. How will manufacturers and developers work to perfect 
this technology so that it can truly be a reliable tool for these 
individuals?
    Answer. As with any new, complex technology, careful testing is 
needed to understand the implications of using a system. Such testing 
must be done in a realistic environment and must involve the community 
of potential users.
    The best design practices are user-centered, meaning that the 
potential user community for a product is involved throughout the 
design process, from initial concept exploration through final testing. 
This is especially important if the designer might experience the world 
differently than the user community.

    Question 3. What more can Congress to do assist with these efforts?
    Answer. Three significant things that Congress can do are (1) 
provide funding for research on applications of AI for use by people 
with disabilities; (2) work with agencies to ensure they are giving 
proper attention to these issues and the interests of people with 
disabilities; and (3) highlight the need for work in this area and 
highlight the successes of those already working in the area.

    Question 4. [citations omitted] As we see machine learning and AI 
increasingly embedded in products and services that we rely on, there 
are numerous cases of these algorithms falling short of consumer 
expectations. For example, Google and Facebook both promoted fraudulent 
news stories in the immediate wake of the Las Vegas Shooting because of 
their algorithms. YouTube Kids is a service designed for children, and 
marketed as containing videos that are suitable for very young 
children. In November, YouTube Kids promoted inappropriate content due 
to algorithms. While the use of machine learning and AI holds limitless 
positive potential, at the current point, it faces challenges where we 
should not risk getting it wrong. Should there be any formal or 
informal guidelines in place for what tasks are suitable to be done by 
algorithms, and which are still too important or sensitive to turn 
over; and what more can be done to ensure better and more accurate 
algorithms are used as you work to better develop this technology?
    Answer. In considering a switch from human to AI based decision 
making, we should not demand perfection of the AI system. The 
alternative to AI is often to rely on human judgment, which is also 
prone to bias and mistakes. Instead of demanding perfection of the AI 
system, an organization needs to understand the potential consequences 
of adopting the AI system well enough to conclude with justified 
confidence that switching to an automated system is an improvement and 
on balance the effects are benign.
    There are also considerations of scale. AI can operate at larger 
scale (that is, on larger number amounts of data or more decisions per 
second) than any human organization can hope to achieve. As a result, 
in some cases the choice in designing a function is not between using 
AI and using a human, but rather between using AI and not providing the 
function at all. The tremendous value provided by many of today's 
information, communication, and publishing tools relies at least in 
part on the use of AI.
    That said, the potential risks of AI-based systems must be 
considered and addressed. It is too early to establish formal 
guidelines, because not enough is known about how best to address these 
problems. Informal guidelines are needed, and the industry and other 
stakeholders should be encouraged to develop them collaboratively. 
Multi-stakeholder groups such as the Partnership on AI may be useful 
venues for these discussions. The guidelines, best practices, and 
technical tools to address these problems will evolve with time.
    Switching a process based on human decision-making to one based on 
AI can have unpredictable consequences, so experimentation is needed in 
a safe environment to adequately understand the implications of such a 
change before it is made. Organizations can be more transparent by 
publishing information about what kinds of testing and analysis were 
done in preparation for the introduction of AI into an existing 
process, thereby enabling stakeholders to understand why a change was 
made and query the organization if concerns remain.
    In many cases, an organization will have valid reasons, such as 
trade secrets or user privacy, to refrain from publishing the full 
details of how a system works. This does not preclude the organization 
from publishing information about how it tested the system and 
evaluated the pros and cons of adopting it.

    Question 5. [citations omitted] Machine learning and AI hold great 
promise for assisting us in preventing cybersecurity attacks. According 
to an IBM survey of Federal IT managers, 90 percent believe that 
artificial intelligence could help the Federal Government defend 
against real-world cyber-attacks. 87 percent think AI will improve the 
efficiency of their cybersecurity workforce. While this is promising, 
the Federal Government currently faces a shortage of qualified 
cybersecurity employees, and to make matters worse, the pipeline of 
students studying these topics is not sufficient to meet our needs. A 
recent GAO report found that Federal agencies have trouble identifying 
skills gaps, recruiting and retaining qualified staff, and lose out on 
candidate due to Federal hiring processes. The George Washington 
University Center for Cyber & Homeland Security recently released a 
report titled ``Trends in Technology and Digital Security'' which 
stated:

        ``Traditional security operations centers are mostly staffed 
        with tier one analysts staring at screens, looking for unusual 
        events or detections of malicious activity. This activity is 
        similar to physical security personnel monitoring video cameras 
        for intruders. It is tedious for humans, but it is a problem 
        really well-suited to machine learning.''

    What effect will machine learning and AI will have on 
cybersecurity; and how do you think the Federal Government can best 
leverage the benefits offered by machine learning and AI to address our 
cybersecurity workforce shortage?
    Answer. As the GWU report suggests, cybersecurity tasks of a 
routine nature can be automated, thereby reducing the need for human 
operators do to lower-level work. This can free up workers to 
concentrate on higher-level tasks requiring more skill and judgment, 
and can help to mitigate the Federal Government's cybersecurity 
personnel shortage.
    Notwithstanding these opportunities, the Federal Government will 
continue to face challenges in recruiting and retaining the best 
technology talent. Many proposals exist to address these challenges by 
improving hiring authorities, pay scales, and working conditions for 
Federal technology workers, and by instituting or expanding training 
and scholarship-for-service programs.

    Question 6. Mr. Felten, as we heard you recently served as the 
deputy U.S. Chief Technology Officer at the White Office of Science and 
Technology Policy. One of the projects you worked on in that office was 
a major report on Artificial Intelligence. That report was one of the 
many important projects taken on by the Office of Science and 
Technology Policy in recent years. And it's extremely disappointing 
that President Trump has failed to nominate leaders for that office, 
now more than ten months into his presidency. That's the longest a 
president has gone without a science advisor since the Office of 
Science and Technology Policy was established in law in 1976. I've led 
two letters to President Trump urging him to nominate well-qualified 
experts to lead this office, but so far we have seen nothing from this 
administration. As a former leader at the Office of Science and 
Technology Policy, could you please explain why this office is 
important, and what kinds of qualities you look for in good nominees 
for this office?
    Answer. OSTP's importance derives from the central role of science 
and technology in nearly every area of policy. In major policy areas 
such as national security and defense, transportation, education, and 
the economy, technology is critical to the most important challenges 
and opportunities. Making the best decisions in these areas requires 
input from and dialog with the technical community. Congress assigned 
that role within the White House to OSTP.
    AI is just one of the areas of interest for OSTP, but it connects 
to many important policy questions. What should DoD's policy be on 
autonomous weapons systems, and what position should the United States 
take in international talks about such weapons? How will AI-driven 
automation affect the job market, and how can American schoolchildren 
and adults be educated and trained for the future workplace? What needs 
to be done to improve highway safety as automated vehicles become 
practical? How can American farmers, journalists, and businesses be 
freed to use drones, while strengthening our defenses against potential 
terrorist uses of the same technology? How will changes in information 
technology affect the mission of the Intelligence Community, and what 
kinds of people and capabilities will the IC need in the future? How 
will cybersecurity concerns affect all of these goals? Each of these 
questions can be better answered with the help of technical advisors 
who have deep domain knowledge, connections to the relevant technical 
communities, and a seat at the policy table.
    A successful OSTP Director will be a trusted advisor to the 
President and the President's senior advisors, a liaison to departments 
and agencies on science and technology issues, and an ambassador to 
scientific and technical communities in the United States and around 
the world.
    A candidate for OSTP Director should be a highly respected member 
of the scientific/technical community, with a reputation for technical 
knowledge and policy judgment. The candidate should be able to work 
successfully across disciplines, acquiring knowledge and providing 
advice across many subject areas with appropriate staff support. They 
should be able to work successfully within the unusual administrative 
and legal environment of the White House, and they should be able to 
recruit, motivate, and lead a team of highly-skilled domain experts and 
policy advisors.
    Because the subject matter of science and technology is so 
extensive, and the United States is blessed with leading experts in so 
many specialties, no one person can hope to have the knowledge, 
experience, and connections needed to provide advice in all technical 
areas. A successful OSTP Director will recruit a team of topic-area 
advisors who can provide context and guidance in specific areas and can 
expand OSTP's ``surface area'' in coordinating with agencies, outside 
experts, and the public.

                                  [all]