[Senate Hearing 118-25]
[From the U.S. Government Publishing Office]




                                                         S. Hrg. 118-25
 
            ARTIFICIAL INTELLIGENCE: RISKS AND OPPORTUNITIES

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

                                HEARING

                               before the

                              COMMITTEE ON
               HOMELAND SECURITY AND GOVERNMENTAL AFFAIRS
                          UNITED STATES SENATE

                    ONE HUNDRED EIGHTEENTH CONGRESS


                             FIRST SESSION

                               __________

                             MARCH 8, 2023

                               __________

        Available via the World Wide Web: http://www.govinfo.gov
        
  
        
        

                       Printed for the use of the
        Committee on Homeland Security and Governmental Affairs
        
        
       
        
        
        
         [GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
         
         
         
                         ______

             U.S. GOVERNMENT PUBLISHING OFFICE 
 52-483 PDF          WASHINGTON : 2023
 
     
         
        
        

                   GARY C. PETERS, Michigan, Chairman
THOMAS R. CARPER, Delaware           RAND PAUL, Kentucky
MAGGIE HASSAN, New Hampshire         RON JOHNSON, Wisconsin
KYRSTEN SINEMA, Arizona              JAMES LANKFORD, Oklahoma
JACKY ROSEN, Nevada                  MITT ROMNEY, Utah
ALEX PADILLA, California             RICK SCOTT, Florida
JON OSSOFF, Georgia                  JOSH HAWLEY, Missouri
RICHARD BLUMENTHAL, Connecticut      ROGER MARSHALL, Kansas

                   David M. Weinberg, Staff Director
                    Zachary I. Schram, Chief Counsel
                  Michelle M. Benecke, Senior Counsel
                        Evan E. Freeman, Counsel
           William E. Henderson III, Minority Staff Director
              Christina N. Salazar, Minority Chief Counsel
                  Andrew J. Hopkins, Minority Counsel
                     Laura W. Kilbride, Chief Clerk
                   Ashley A. Gonzalez, Hearing Clerk

                            C O N T E N T S

                                 ------                                
Opening statements:
                                                                   Page
    Senator Peters...............................................     1
    Senator Johnson..............................................     3
    Senator Blumenthal...........................................    16
    Senator Hassan...............................................    17
    Senator Padilla..............................................    20
    Senator Sinema...............................................    22
    Senator Rosen................................................    24
Prepared statements:
    Senator Peters...............................................    33

                               WITNESSES
                        Wednesday, March 8, 2023

Alexandra Reeve Givens, President and Chief Executive Officer, 
  Center for Democracy and Technology............................     8
Suresh Venkatasubramanian, Ph.D., Professor of Computer Science 
  and Data Science, Brown University.............................    10
Jason Matheny, Ph.D., President and Chief Executive Officer, RAND 
  Corporation....................................................    12

                     Alphabetical List of Witnesses

Givens Alexandra Reeve:
    Testimony....................................................     8
    Prepared statement...........................................    35
Matheny, Jason Ph.D.:
    Testimony....................................................    12
    Prepared statement...........................................    60
Venkatasubramanian, Suresh Ph.D.:
    Testimony....................................................    10
    Prepared statement...........................................    53

                                APPENDIX

Peterson testimony submitted by Senator Johnson..................    65
Data and Society Statement for the Record........................    69
R Street Initiative Statement for the Record.....................    71


                        ARTIFICIAL INTELLIGENCE:



                        RISKS AND OPPORTUNITIES

                              ----------                              


                        Wednesday, March 8, 2023

                                     U.S. Senate,  
                           Committee on Homeland Security  
                                  and Governmental Affairs,
                                                    Washington, DC.
    The Committee met, pursuant to notice, at 10 a.m., in room 
SD-562, Dirksen Senate Office Building, Hon. Gary C. Peters, 
Chairman of the Committee, presiding.
    Present: Senators Peters [presiding], Hassan, Sinema, 
Rosen, Padilla, Blumenthal, Johnson, and Scott.

             OPENING STATEMENT OF SENATOR PETERS\1\

    Chairman Peters. The Committee will come to order.
---------------------------------------------------------------------------
    \1\ The prepared statement of Senator Peters appears in the 
Appendix on page 33.
---------------------------------------------------------------------------
    Today's hearing will discuss both the potential risks as 
well as the opportunities associated with artificial 
intelligence (AI), examining how artificial intelligence 
affects our nation's competitiveness on a global stage, and 
discuss ways to ensure that these technologies are used both 
safely and responsibly.
    The adoption of artificial intelligence in government, 
industry, and civil society has led to the rapid growth of 
advanced technology in virtually every sector, transforming 
millions of Americans' lives, millions of Americans all across 
our country.
    From the development of lifesaving drugs and advanced 
manufacturing to helping businesses and governments better 
serve the public, to self-driving vehicles that will improve 
mobility and make our roads safer, artificial intelligence 
certainly holds great promise.
    But this rapidly evolving technology also presents 
potential risks that could impact our safety, our privacy, and 
our economic and national security. We must ensure that the use 
of this technology becomes more widespread. We have to make 
sure that there are also the right safeguards in place to 
ensure it is being used appropriately.
    One of the greatest challenges presented by artificial 
intelligence is the lack of transparency and accountability in 
how algorithms reach their results. Often, not even the 
scientists and the engineers who design the AI models fully 
understand how they arrive at the outputs that they produce. 
This lack of visibility into how AI systems make decisions 
creates challenges for building public trust in their use. AI 
models can also produce biased results that can have 
unintended, but harmful consequences for the people interacting 
with those systems.
    Some AI models, whether because of the data sets they are 
trained on or the way in which the algorithm is applied, are at 
risk of generating outputs that discriminate on the basis of 
race, sex, age, or disability.
    Whether these systems are being used in criminal justice, 
college admissions, or even determining eligibility for a home 
loan, biased decisions and the lack of transparency surrounding 
them, can lead to adverse outcomes for people who may not be 
even aware that AI has played a role in the decisionmaking 
process. Building more transparency and accountability into 
these systems will help prevent any kind of bias that could 
undermine the utility of AI.
    While many government organizations and businesses are 
working to build AI systems that enhance our daily lives, we 
must be open-eyed about the risks presented by bad actors and 
adversaries who may use AI to intentionally cause harm, or 
undermine our national interests.
    Generative artificial intelligence like Chat Generative 
Pre-trained Transformer (ChatGPT) or deepfakes can be used to 
create convincing, but false information that can distort 
reality, undermine public trust, and even be used to cause 
widespread panic and fear in a worst-case scenario.
    The risks from this kind of improper use also extend beyond 
our borders. Adversaries like the Chinese government are racing 
to be the world leaders in these technologies and to harness 
the economic advantages that dominance in artificial 
intelligence will certainly create. The United States must be 
at the forefront of developing our own AI systems and training 
people how to use them appropriately, to protect our global 
economic competitiveness.
    If we do not, not only are we at risk of American entities 
having to purchase these mature technologies from an economic 
competitor like the Chinese government, there will be tools 
with little accountability that have been developed by an 
adversary that does not share our core American values, a 
serious national security risk.
    Finally, artificial intelligence will have a significant 
impact on the future of work. There is no question that AI 
systems have the potential to disrupt the workplace as we 
currently know it. That is why it is essential as the United 
States develops these technologies, we are also developing a 
workforce that is ready to work alongside them. We must address 
concerns that AI tools could replace human workers and instead 
focus on how they can assist humans and enhance the workplace.
    Our goal in today's hearing is to examine these types of 
risks and challenges and discuss what steps Congress should 
take to ensure that we are able to harness these benefits and 
opportunities with this technology. This includes ensuring that 
these technologies are used appropriately, and to protect the 
civil rights and civil liberties of all Americans.
    Last Congress, I passed bipartisan laws that took steps to 
ensure the appropriate use of artificial intelligence by 
government, including through procurement safeguards and by 
boosting the knowledge of the acquisition workforce, to ensure 
they are properly trained to understand the risks and 
capabilities of these technologies. I look forward to building 
on those efforts this Congress, and working alongside my 
colleagues on the Committee to support the development of AI 
technologies, and ensure that they are being used both 
appropriately and effectively.
    I hope that today's discussion will be the first of several 
on this important topic, and I am pleased to have our panel of 
witnesses with us today who are experts in the field of 
artificial intelligence and who can discuss the adoption of 
these systems and the broader impacts on industry, civil 
society, and government.
    With that I would like to now turn it over to our acting 
Ranking Member, Senator Johnson.

              OPENING STATEMENT OF SENATOR JOHNSON

    Senator Johnson. Thank you, Mr. Chairman. I would like to 
start by asking consent to enter Dr. Jordan B. Peterson's 
testimony into the record.\1\
---------------------------------------------------------------------------
    \1\ The statement submitted by Dr. Peterson appears in the Appendix 
on page 65.
---------------------------------------------------------------------------
    Chairman Peters. Without objection.
    Senator Johnson. Thank you. Now let me explain why I had to 
do that. Last Thursday, I was pretty late in the process, and I 
was asked by Ranking Member Paul to act as Ranking Member 
because he had a conflict with a pretty important hearing in 
Senate Foreign Relations Committee (SFRC). I was happy to do so 
because I have been very interested in the subject. Artificial 
intelligence has incredible impact, or will have incredible 
impact on our society and on individuals, and so I have been 
doing a fair amount of research on the topic. As a result I 
became aware of Jordan Peterson's interest in the topic as 
well.
    As a matter of fact, two weekends ago I watched about an 
hour-and-a-half-long video of him interviewing Jim Keller and 
Jonathan, I think it is Pageau--I apologize if I am 
mispronouncing his name--on this topic. Again, they were 
thinking deeply about this subject and its impact on society.
    First of all, who is Jordan B. Peterson? He is an author, a 
psychologist, an online educator, and professor emeritus at 
University of Toronto. For 20 years he taught some of the most 
highly regarded courses at Harvard and the University of 
Toronto, while publishing more than 100 well cited scientific 
papers and maintaining an active clinical and consulting 
practice. His international lecture tour has sold out more than 
400 venues, and his best-selling books include 12 Rules for 
Life: An Anecdote for Chaos and Beyond Order: 12 More Rules for 
Life.
    Unfortunately, the Chairman did not allow him to appear 
remotely, and we can talk about that a little bit later. But in 
lieu of an opening statement what I would like to do is read 
some of the key excerpts out of Dr. Peterson's testimony. We 
will see the insight and the thoughtfulness that we are missing 
by not having him here today.
    He starts his testimony talking about the large language 
models, for example, like ChatGPT. He says, ``Advanced large 
language models such as ChatGPT have burst onto the scene with 
a vengeance in the last six months. ChatGPT recently completed 
the standardized test (SAT) and scored 1020. A score of 1020 is 
equivalent to an intelligent quotient (IQ) of about 110, which 
make ChatGPT more intelligence than 75 percent of people.
    ``The significance of all this should not be 
underestimated. We now have AI systems capable of engaging in 
genuine conversation, able to write, able to produce computer 
code, able to 'think,' and they will be much smarter very 
soon.''
    He goes on to talk about the rights given to the extended 
digital self. He writes, ``For centuries we were also simple 
enough so that our name sufficed to identify us. Online, 
however, things are very different. Our digital identity is 
composed of the tools we use--the apps, programs, services, 
websites, et cetera--that we choose voluntarily to employ, as 
well as the records of our virtual behavior, our browsing 
patterns, our purchases, our records of travel, but the written 
communications and images we issue on platforms such as 
Instagram, Facebook, and more ominously, TikTok, which 
essentially operates under the control of the Chinese Communist 
Party (CCP). That extended digital self has very few rights, 
and our legal structure has not been able to adapt itself to 
the immense changes on the virtual front.
     ``The logical extension of such danger, and most likely 
outcome,'' in his estimation, ``is the duplication in the West 
of something approximating the utter catastrophe of a so-called 
social credit system in China. Everything is tracked and 
controlled. The government can, with the stroke of a pen, seize 
the economic resources of any given individual or group.'' In 
parentheses he says, ``Something that happened very ominously 
in Canada in the case of the truckers' convoy.''
    He goes on, ``Developing AI capabilities will radically 
extend the surveillance State. China has about 400 cameras 
watching every 1,000 people. We could well be entering an era 
of authoritarian AI-mediated social shunning. The use of 
cameras should be banned. Machines should never be given the 
authority to ticket, try, punish, or limit the economic or 
practical activities with human beings.'' He goes on to talk 
about additional dangers. ``In the next year, AI wizards will 
produce intelligence systems that will be able to produce 
representations of any person, doing anything that can be 
described, the so-called deepfakes. Imagine those being 
released on the eve of a critical election. Then imagine that 
happening everywhere, on every issue, thousands of times. 
Imagine being entirely unable to determine day-to-day what 
communication, from what person, photos, videos, auto 
recordings, writings is real and what is false. Then imagine 
that now, not in some distant future. That is where we are at. 
Steps must be taken on the legal front to make false digital 
representations of living persons not only illegal but 
seriously illegal.''
    He concludes, ``The development of AI systems as 
intelligent as we are''--and I would add probably even more 
so--``is not some future possibility but a current actuality. 
The melding of AI-mediated intelligence systems with our 
capacity for monitoring and surveillance prepares the way for a 
tyranny so comprehensive that we can barely imagine it.''
    Now again, these are just excerpts from his testimony, and 
I wish Dr. Peterson could have been here remotely to offer 
that. But for whatever reason, even though we have the 
technology here, the Chairman said he could not appear, we 
could not make it possible for him to appear remotely.
    Now behind the scenes over the weekend, there were other 
reasons supplied. Talk about some book. That was all a ruse. It 
was a pretext for not allowing Dr. Peterson to testify, and I 
really cannot guess why. Some kind of ideological reason.
    By the way, it was not unusual to get a witness pretty late 
in the process. As Chairman of this Committee over six years, 
it was very rare that I got testimony much more than the day 
before. Sometimes it could be hard to arrange witnesses. This 
was a somewhat unusual circumstance but not that unusual. So 
that should not be an excuse.
    So blocking Dr. Peterson because we supposedly could not 
accommodate a remote witness is simply not credible. For 
whatever reason, the Chairman and his staff did not want to 
allow our witness. This is an action that is beyond unfortunate 
and something we will not condone, which is why no Republicans 
will attend this hearing.
    I sincerely hope the Chairman will reconsider this partisan 
action and not repeat it in the future.
    Chairman Peters. Senator Johnson, if I could respond to 
that. I have been the Chair now, this is going into the third 
year. We have never blocked the minority from having a witness, 
and we are not blocking the minority from having this witness 
here now.
    Senator Johnson. Yes, you are. He is not here.
    Chairman Peters. Let me go through the process. We started 
putting this together a month ago, one month, we would hope, 
that staff, in one month's time, could come up with witnesses. 
We did. We have three eminent witnesses that were presented to 
the Ranking Member. We go through interviews. All three of you 
had interviews with staff from both the majority and minority. 
It is what we do with every witness. For every single hearing 
we do that. We do not want to change that policy. That is a 
very important policy, so we have an understanding of who the 
witnesses are. We have an opportunity to prepare, to make this 
a good hearing.
    A month ago we did that. We went through the process. We 
continually went to the Ranking Member and said, ``Please 
provide your minority witness. We would like to move forward. 
We are excited about this hearing.'' We did not hear anything. 
We had to actually put a deadline. Please, by this deadline, 
last week, on Thursday, please provide a witness. We did not 
hear.
    We finally got a witness, not from the Ranking Member but 
another Member, at 8 p.m. on Friday, with two business days 
prior to a hearing. There was a request for video. This is not 
a hybrid hearing. We have, for well over a year, everybody has 
appeared in person. I know maybe Senator Johnson likes----
    Senator Johnson.--On the technology here.
    Chairman Peters. The witnesses appear in person. They have 
always appeared in person, for a long period of time. Perhaps 
Senator Johnson likes Coronavirus Disease 2019 (COVID-19) 
protocols. I am not sure. But we have had personal folks here, 
because I think it is important to have witnesses in person. 
Each and every one of you arranged your schedule to be here in 
person. You could have done video but you knew that was the 
rule of the Committee. This is not a hybrid hearing. This is to 
be in person, and I think you have a much better hearing as a 
result of that.
    We said that with this new person that came in at the end 
that we would need them to appear in person, just like each and 
every one of you took the time and trouble to get here, they 
would have to do the same thing. Perhaps if they had more time, 
if we actually heard from the minority in a normal time, they 
would have been able to make those arrangements to be here in 
person.
    He was welcome to be here. If he wanted to sit here today 
we would have welcomed that. He would have had to go through 
the interview. It would have been short because we only had two 
business days to do this. We would have had to have an 
interview, like each and every one of you have done, and every 
single witness that comes before this Committee does it.
    All we are asking Senator Johnson, is let us have the same 
process. I told you, or I told the Ranking Member, that your 
witnesses, we are going to have more AI hearings, he is 
welcome. If you want him to be your witness at a future hearing 
we would welcome him. He will be the minority witness. It was a 
time constraint.
    Senator Johnson. OK. We will definitely take you up on that 
offer. But again, there were things happening behind the 
scenes, and again, I did not get brought into this process 
until late Thursday. We scrambled. We got him to agree to be a 
witness. We let you know it was going to be remote. The 
technology is obviously available.
    But again, as Chairman of this Committee, I did not take it 
upon myself to vet your witnesses, the minority witnesses. That 
is your job. If you end up with somebody with troubling 
circumstances around his testimony, that is on you, not on the 
Committee. Dr. Peterson is eminently qualified. He has been 
talking about this. He put a lot of work into his testimony and 
not able to provide it.
    Again, this situation, it is just not credible that we 
could not accommodate him remotely. It is not unusual that it 
is hard to sometimes find witnesses. I cannot speak for Senator 
Paul in terms of why he did not make the decision not to be 
Ranking Member, but I acted very expeditiously. I asked an 
eminently qualified individual to be a witness. He agreed. He 
put in the work. He provided insightful and thoughtful 
testimony. We should have allowed him to testify remotely, but 
we will take your offer for the next hearing and we will 
communicate that to Dr. Peterson.
    Chairman Peters. Senator, we want witnesses to be here in 
person. This is not a hybrid hearing. It was never noticed as a 
hybrid hearing.
    Senator Johnson. That is fine. We just do not want----
    Chairman Peters. I understand.
    Senator Johnson [continuing.] The majority blocking----
    Chairman Peters. We are not blocking.
    Senator Johnson [continuing.] Or even vetting our minority 
witnesses. That is honestly not your job. The minority has a 
right to have witnesses appear before the Committee on the 
topic at hand, and to have you have veto power over that is not 
proper.
    Chairman Peters. Again, Senator Johnson, we can provide. We 
sent the letter to the Ranking Member, your witness can 
testify. They have to be in person, and they have to have an 
interview like every other witness, and yet that did not 
happen, and the reason it did not happen was because it was 
such a short timeline. I get that. I know you were thrown this 
responsibility at the last moment.
    Senator Johnson. I acted expeditiously, and I came up with 
an excellent witness, and it would have been great to have him 
appear remotely.
    Chairman Peters. We would have welcomed him.
    Senator Johnson. Hopefully we will see him in person, as 
long as he is not too insulted by not being able to testify 
here today.
    Chairman Peters. Hopefully he is not insulted that he is 
being treated like everybody else. If he thinks that he should 
be treated differently than everybody else, well, in this 
Committee we treat everybody fairly. Everybody is treated the 
same way, and we believe that those rules should be followed.
    We would hope that in the future, when you have a month to 
prepare for a hearing that you actually do the work and prepare 
for a hearing, and do not expect that everybody is just going 
to drop everything and change all the rules and do something 
different. Do the work. This is an important Committee. We have 
always worked on a consensus basis. You and I worked on a 
consensus basis.
    Senator Johnson. That is right, but I never blocked any 
witnesses. But anyway, enough of this. Just get on with the 
hearing and we will attend the next one.
    Chairman Peters. Let us hope we can return to working in a 
bipartisan way and have folks do the work necessary so that 
these hearings go forward.
    With that, let us get to the important business at hand.
    It is the practice of the Homeland Security and 
Governmental Affairs Committee (HSGAC) to swear in witnesses, 
so if each of you will please stand and raise your right hand.
    Do you swear that the testimony that you will give before 
this Committee will be the truth, the whole truth, and nothing 
but the truth, so help you, God?
    Ms. Givens. I do.
    Mr. Venkatasubramanian. I do.
    Mr. Matheny. I do.
    Chairman Peters. Great. Thank you.
    Today's first witness is Alexandra Reeve Givens. Ms. Givens 
is the President and Chief Executive Officer (CEO) of the 
Center for Democracy and Technology (CDT), whose mission is to 
ensure democracy and individual rights are at the center of the 
digital revolution. Previously, Ms. Givens served as the 
founding Executive Director of the Institute for Technology Law 
and Policy at Georgetown Law, and as Chief Counsel for 
Intellectual Property and Antitrust on the Senate Judiciary 
Committee. Ms. Givens has also served as an adjunct professor 
at Columbia University School of Law.
    Ms. Givens, welcome to the Committee and thank you for 
appearing. You are recognized for your opening statement.

  TESTIMONY OF ALEXANDRA REEVE GIVENS,\1\ PRESIDENT AND CHIEF 
     EXECUTIVE OFFICER, CENTER FOR DEMOCRACY AND TECHNOLOGY

    Ms. Givens. Thank you very much, Senator Peters, and to 
Members of the Committee, thank you for inviting me to speak 
about the challenges and opportunities presented by AI.
---------------------------------------------------------------------------
    \1\ The prepared statement of Mr. Givens appears in the Appendix on 
page 35.
---------------------------------------------------------------------------
    The Center for Democracy and Technology, is a 28 year-old 
nonprofit, nonpartisan organization that works to protect civil 
rights, civil liberties, and democratic values in the digital 
age. CDT protects users' interests in areas ranging from 
commercial data practices to government surveillance to online 
content moderation to the use of technology in education and 
government services. AI is already transforming each one of 
these areas, so I am grateful for the Committee's focus on the 
topic today.
    While AI has the potential to generate new insights and 
make processes more efficient, it also poses risks, of being 
unreliable, biased, and hard to explain or hold accountable.
    My written testimony focuses on these risks in several 
areas that directly impact consumers. First, when AI or 
automated systems are used in decisions impacting people's 
access to economic opportunities, such as in employment, 
housing, and lending, and second, in the administration of 
government services, such as when AI or automated systems are 
used to detect fraud or determine benefits eligibility.
    When AI systems are used in these high-risk settings 
without responsible design and accountability, it can devastate 
people's lives. A person may be unfairly rejected from a job, 
be denied or unable to find housing, or be wrongly accused of 
fraud and stripped of the benefits they need to support their 
family. When this happens, the harm is felt not only by the 
people whose lives are upended by the decision but also by the 
businesses and government programs that are relying on these 
systems to work. Those businesses or government agencies are 
now bought into a system that is unfit for purpose, and may 
face legal, financial, and reputational consequences. That is 
why it benefits everyone to address the potential risks and 
limitations of AI.
    My written testimony details harms that have already arisen 
in these contexts. For example, hiring tools that 
systematically downgraded women's resumes or an automated video 
interview system where a reporter gave answers in German and 
yet was still found to be a 73 percent match for a company.
    In the government setting, the Michigan Integrated Data 
Automated System (MiDAS) in Michigan wrongfully classified up 
to 40,000 people's unemployment insurance applications as 
fraudulent based on design errors in the system. People who 
were already on the financial brink had their wages garnished, 
bank accounts levied, and were driven into bankruptcy. The 
State faced years of litigation and recently paid millions of 
dollars to victims.
    Government programs in Europe, the United Kingdom (UK), and 
Australia have had similar problems.
    When assessing these concerns, policymakers should consider 
several factors. First, poorly designed and governed AI systems 
can cause not just individual but systemic harm. In the hiring 
context, for example, an AI tool might replace the risk of a 
bad apple in human resource (HR), but it does so with a system 
that could be ineffective and discriminatory at scale. The 
resulting harms may impact an entire sector when a tool is used 
by multiple companies.
    Second, harms do not just impact the people who are the 
subject of the decision but the businesses and agencies relying 
on those tools. That is why we need robust, specific guidance 
to help people navigate these issues and to enforce existing 
laws to ensure that developers take their obligations 
seriously.
    Third, the subjects of AI decisionmaking often have no idea 
they are being assessed by an automated program, let alone how 
that tool may work, and neither do regulators.
    Without increased transparency about when AI systems are 
being used and how they have been designed and tested, society 
will be hamstrung in its efforts to identify and address harms.
    Fourth, AI systems need ongoing testing in their applied 
environment to make sure they are working as intended. But this 
is complicated because AI tools are often designed by one 
company and then deployed by many others in different settings. 
We need to work through the pathways of responsibility in this 
diffuse value chain.
    Given these challenges, we need a cross-society effort for 
the responsible design, deployment, use, and governance of AI. 
My written testimony outlines several ways in which the 
government can lead in this work.
    The first is to rapidly scale up guidance and resources to 
identify AI-related harms and mitigations. We need to help 
those non-expert businesses and agencies think about and 
address risk and when to say no to these tools altogether. The 
National Institute of Standards and Technology (NIST) AI Risk 
Management Framework (RMF) and the Blueprint for an AI Bill of 
Rights are good examples of this, but agencies across the 
Federal Government must lead in their respective sectors.
    Second is that we must increase transparency, which is 
where legislation like the Algorithmic Accountability Act or 
similar models can be useful. It is time to normalize the idea 
that companies designing and deploying AI tools in high-risk 
settings must first analyze and document how they work, 
accounting for potential risks and steps they have taken to 
address them.
    Third, as this Committee has well recognized, the Federal 
Government has an essential role to play in its own 
procurement, design, use, and funding of AI systems.
    Congress directed Office of Management and Budget (OMB) to 
issue guidance and principles for the Federal acquisition and 
use of AI, which was boosted by Executive Orders (EO) from both 
the Trump and Biden administrations. This work must continue 
without delay and we must continue to support agencies in this 
work, such as through the National AI Initiative Office (NAIIO) 
that this Committee and Congress created.
    I thank the Committee for its continued work in this and 
related areas, and I look forward to your questions.
    Thank you.
    Chairman Peters. Thank you, Ms. Givens.
    Our next witness is Dr. Venkatasubramanian who currently 
serves as Professor of Computer Science and Data Science at 
Brown University. His expertise includes data mining, machine 
learning (ML), algorithms, and computational geometry, 
specifically algorithmic fairness and their impacts on 
decisionmaking on society.
    Previously Professor Venkatasubramanian served as the 
Assistant Director for Science and Justice in the White House 
Office of Science and Technology Policy.
    Professor, welcome to the Committee, and thank you for 
appearing. You are recognized for your opening statement.

TESTIMONY OF SURESH VENKATASUBRAMANIAN,\1\ Ph.D., PROFESSOR OF 
      COMPUTER SCIENCE AND DATA SCIENCE, BROWN UNIVERSITY

    Mr. Venkatasubramanian. Thank you, Senator Peters and 
Members of the Homeland Security and Government Affairs 
Committee. I thank you for inviting me to testify at this 
important hearing on the risks and opportunities of AI. I am a 
professor of computer science and director of the Center for 
Technological Responsibility at Brown University.
---------------------------------------------------------------------------
    \1\ The prepared statement of Dr. Venkatassubramanian appears in 
the Appendix on page 53.
---------------------------------------------------------------------------
    I recently completed a stint as tech policy advisor in the 
White House and helped develop the Blueprint for an AI Bill of 
Rights.
    I have spent the last decade studying and researching the 
impact of automated systems, and AI, on people's rights, 
opportunities, and access to services. I have also spent time 
advising State and local governments on sound approaches to 
governing the use of technology that impacts people's lives.
    We are here today to talk about AI, a field of study trying 
to design systems that can sense, interact, reason, and behave 
in the way humans do, and in some cases even surpass us. People 
learn from the data we receive, and thus one sub-area of AI 
that is dominant right now, fueled by the collection of vast 
amounts of data, is machine learning, the design of systems 
that can incorporate historical data into the predictions they 
produce, and in some cases keep adapting as more data appears.
    Virtually every sector of society is now touched by machine 
learning, and the most consequential decisions and experiences 
in our lives are mediated by algorithms--where we go to school, 
how we learn, how we get jobs, whether we can buy a house, what 
kind of loan we get, whether we get credit to start a small 
business, whether we are surveilled by law enforcement or 
incarcerated before a trial, how long a sentence for a 
convicted individual is, and whether we can get paroled.
    The list goes on and on, and keeps expanding, with systems 
like GPT3, ChatGPT, and Bard, and many others that ingest 
extremely large amounts of data and huge compute power to 
create the plausibly realistic conversations that have caught 
our imagination over the past few months.
    All these systems have something in common. They are 
algorithms for making algorithms. The distinctive feature of a 
machine learning system is that the output of the system is 
itself an algorithm that purports to solve an underlying 
problem, whether it is predicting your loan worthiness, 
searching for a face in a video stream, or even having a 
conversation with an individual.
    As a consequence of the above, we do not actually know for 
sure whether and how these algorithms work and why they produce 
the output that they do. This might come as a surprise given 
how much we hear every day about the amazing and miraculous 
successes of AI. Yet AI systems fail.
    They fail when the algorithms draw incorrect conclusions 
from data. They fail when they make predictions based on faulty 
or biased data. They fail when the results of one AI system are 
fed into another, or even the same one, amplifying errors along 
the way. They fail when they are so opaque that errors in how 
they function cannot even be detected.
    The truth is AI systems are not magic. AI is technology, 
and like any other piece of technology that has benefited us--
drugs, cars, planes--AI needs guardrails so we can be protected 
from the worst failures while still benefiting from the 
progress AI offers.
    What should these guardrails look like? Any automated 
system that has meaningful impact on our rights, opportunities 
for advancement, and access to critical services should be 
tested so it works, and works well. It should not exhibit 
discriminatory behavior, be limited and careful in its use of 
our personal data, be transparent, and easily understandable, 
and be accompanied by human supervision for all the times that 
it fails. Moreover, all these protections should be documented 
and reported on clearly for independent scrutiny. Congress 
should enshrine these ideas in legislation, not just for 
government use of AI but for private sector uses of AI that 
have people-facing impact.
    I am a computer scientist--a card-carrying computer 
scientist, I like to say--and my work is to imagine 
technological futures. There is a future in which automated 
technology is an assistant. It enables human freedom, liberty, 
and flourishing, where the technology we build is inclusive and 
helps all of us achieve our dreams and maximize our potential.
    But there is another future in which we are at the mercy of 
technology, where the world is shaped by algorithms and we are 
forced to conform, in which those who have access to resources 
and power control the world and the rest of us are left behind. 
I know which future I want to imagine and work toward. I 
believe it is our job to lay down the rules of the road, the 
guardrails and the protections, so that we can achieve that 
future. I know we can do it if we try.
    Thank you for giving me this opportunity to speak.
    Chairman Peters. Thank you, Professor.
    Our next witness is Dr. Jason Matheny. Dr. Matheny 
currently serves as President of the RAND Corporation, a 
nonprofit institution that helps provide research and analysis 
to solve public policy challenges. Prior to his current role, 
Dr. Matheny led White House policy on technology and national 
security at the National Security Council (NSC) and the Office 
of Science and Technology Policy, and was the founding director 
of the Georgetown Center for Security and Emerging Technology. 
Dr. Matheny was congressionally appointed as a commissioner to 
the National Security Commission on Artificial Intelligence. 
Welcome to the Committee.
    You may proceed with your opening statement.

   TESTIMONY OF JASON MATHENY, Ph.D.,\1\ PRESIDENT AND CHIEF 
              EXECUTIVE OFFICER, RAND CORPORATION

    Mr. Matheny. Thank you, Chairman Peters and members of the 
Committee for the opportunity to testify today.
---------------------------------------------------------------------------
    \1\ The prepared statement of Mr. Matheny appears in the Appendix 
on page 60.
---------------------------------------------------------------------------
    For the past 75 years, RAND has conducted nonpartisan 
policy research, and we currently manage four federally funded 
research and development (R&D) centers for the Federal 
Government, including one for the Department of Homeland 
Security and three for the Department of Defense (DOD). Today I 
will focus my comments on how AI affects national security and 
U.S. competitiveness.
    Among a broad set of technologies, AI stands out both for 
its rate of progress and for its scope of applications.
    AI holds the potential to broadly transform entire 
industries, including ones that are critical to our future 
competitiveness, such as medicine, manufacturing, and energy. 
Applications of AI also pose grave security challenges for 
which we are currently unprepared, including the development of 
novel cyber weapons, large-scale disinformation attacks, and 
the design of advanced biological weapons.
    Threats from AI pose special challenges for national 
security for several reasons: the technologies are driven by 
commercial entities that are frequently outside our national 
security frameworks; the technologies are advancing quickly, 
typically outpacing policies and organizational reforms within 
government; assessments of the technologies require expertise 
that is concentrated in the private sector and that has rarely 
been used for national security; and the technologies lack 
conventional intelligence signatures that distinguish benign 
from malicious use, differentiate intentional from accidental 
misuse, or that permit attribution with confidence.
    By most measures, the United States is currently the global 
leader in AI. However, this may change as the People's Republic 
of China seeks to become the world's primary AI innovation 
center by 2030, an explicit goal of China's AI national 
strategy. In addition, both China and Russia are pursuing 
militarized AI technologies, intensifying the challenges that I 
just outlined. In response, I will highlight eight actions that 
national security organizations, including the Department of 
Homeland Security (DHS), could take.
    First, ensure that DHS cybersecurity strategies and cyber 
Red Team activities track developments in AI that are likely to 
affect cyber defense and cyber offense.
    Second, within the National Institute of Standards and 
Technology industry stakeholders and U.S. allies and partners 
ensure that international standards for AI prioritize safety, 
security, and privacy, so that the technologies are less prone 
to misuse by surveillance States.
    Third, consider creating a regulatory framework for AI that 
is informed by an evaluation of risks and benefits of AI to 
U.S. national security, civil liberties, and competitiveness.
    Fourth, identify the high-performance computing hardware 
that is used for AI as critical infrastructure that can be 
stolen or subverted, and consider requirements for tracking 
where high-performance computing hardware goes and what it is 
being used for.
    Fifth, work with the intelligence community (IC) to 
significantly expand the collection and analysis of information 
on key foreign public-and private-sector actors in adversary 
States involved in AI, and create new partnerships and 
information-sharing agreements among Federal, State, and local 
government agencies, the research community, and industry.
    Sixth, leverage AI expertise in the private sector through 
short-term and part-time Federal appointments and security 
clearances for leading academic and industry AI experts who can 
advise the government on key technology developments, with 
appropriate checks on conflicts of interest.
    Seventh, in Federal purchases and development of AI 
systems, include requirements for security, safety, and privacy 
measures that prevent AI systems from misbehaving due to 
accidents or adversaries, and require socially beneficial 
techniques, such as privacy-preserving machine learning and 
watermarking to detect generated text and deepfakes.
    Eighth and last, increase our investments in biosecurity 
and biodefense, given the potential applications of AI to 
design pathogens that are much more destructive than those 
found in nature.
    I thank the Committee for the opportunity to testify, and I 
look forward to questions.
    Chairman Peters. Thank you, Dr. Matheny.
    Professor Venkatasubramanian, this question is for you. In 
your statement you describe the so-called black box of the AI 
systems, where developers themselves do not fully understand 
exactly what happened in that black box as it is making those 
decisions. You mentioned in your opening comments and your 
written comments some of those risks, but for the Committee's 
benefit could you tell us more about the risks that are 
associated when you have non-transparent algorithms?
    Mr. Venkatasubramanian. Thank you, Senator, and you can 
call me Professor V. That is fine. My students do that too.
    Chairman Peters. Professor V?
    Mr. Venkatasubramanian. Professor V is just fine.
    Chairman Peters. Thank you.
    Mr. Venkatasubramanian. To your question, when we do not 
know how an algorithm works or why it works, we also do not 
know how it fails and under what circumstances it fails, and 
that is where the biggest problem is. We do not even know how 
to tell whether it is failing or not.
    If I use, for example, an algorithm to analyze a tissue 
scan, to determine whether a patient has cancer, such a failed 
algorithm could either falsely declare a patient free from 
cancer, which would be catastrophic, or falsely declare that 
they were positive for the test and therefore have to undergo 
harmful treatments that could be very harmful to them. We would 
not be able to tell the difference.
    That is why safety testing, investigation, and transparency 
are so critical, because of the way in which machine learning 
algorithms, and the fact that there are algorithms for 
generating algorithms, create these procedures that are very 
hard to understand. This comes up again with things like 
ChatGPT, where we do not know how they do what they do. They 
seem to be providing plausible answers, but as we have seen, it 
is very easy to get them to lie, or not lie but give answers 
that are false because we do not understand how they are 
working. That is where the lack of transparency is one of the 
biggest problems with understanding the effectiveness and 
whether these systems can work.
    Chairman Peters. I appreciate that.
    Ms. Givens, you have done a lot of work in this area as 
well. I would certainly love to have your thoughts on the black 
box and accountability.
    Ms. Givens. Thanks for the question. The thing that I think 
about is what is meaningful transparency, and the way to think 
about that is as somebody is deciding, as a small business, for 
example, whether to use one of these tools or even large and 
mid-sized businesses deciding right now whether they could 
integrate ChatGPT into some of their offerings.
    What are the resources that will help them make an informed 
decision? Right now there are many different tests and 
approaches to safety measurement, to mitigating and measuring 
bias, but we really need to fast-track that conversation to 
make sure that we are talking about well-established, robust 
approaches to identifying and addressing risks.
    We also need to think about a conversation of internal 
audits versus how we make that an external process that can 
have more accountability and visibility from the outside.
    Then, of course, how to make guidance and disclosures that 
are useful for users. All of those are areas where there is 
nascent work now, but we need to turbocharge those efforts to 
actually make transparency have value.
    Chairman Peters. Thank you. Dr. V, we are talking about 
bias in these systems. As a computer scientist you have 
considerable expertise in this area. Could you tell the 
Committee how does bias actually get into these AI systems? We 
should know how it gets in so we can figure out how to deal 
with it.
    Mr. Venkatasubramanian. Thanks for that. There is a phrase 
in computer science that is called ``garbage in, garbage out.'' 
It means that if you put bad data into an algorithm you will 
get a bad outcome. In machine learning, what we talk about now 
is ``bias in, bias out.'' A machine learning algorithm that 
takes data that has hidden biases in it will invariably, almost 
certainly, detect and amplify those biases in its output.
    We saw this happening when a company was training a system 
to predict who would be good people to hire. The system started 
picking up signals that the candidate was a woman, even if it 
was not explicitly mentioned--for example, a person whose 
curriculum vitae (CV) said that they went to Smith College--and 
then it started rejecting them. It turns out that in this case 
it was because the data being used to train the algorithm was 
itself biased. It was historical data on hiring from the 
company, and the company, as it turned out, had skewed and 
gender biased hiring practices.
    One very important example where bias gets into an AI 
system is when the underlying data used to train the algorithm 
has biases coming from historical context.
    Chairman Peters. Can you mitigate that by having larger 
datasets? Is that one way to do it, or you still have to, in 
some way, examine those sometimes very large datasets that are 
training AI.
    Mr. Venkatasubramanian. Unfortunately, merely having more 
data does not actually solve the problem because if that more 
data continues to have those kinds of biases then you will just 
make the problem even worse. What is required is a collection 
of procedures, among them procedures that examine the sources 
of data, examine the biases in the data, even if it is a large 
dataset, and try to understand how those biases might be 
affecting what the algorithm would do.
    Another set of procedures is to understand how the 
algorithm training is being done. There are certain best 
practices for how to train algorithms to try to mitigate these 
forms of bias, and they need to be put into place. When you do 
that you can mitigate a lot of these biases.
    Similarly looking at, in context, how the algorithm is used 
and deployed and how the results are showing up and whether 
biases are showing up in the output as well.
    In these three ways, if you have the appropriate practices 
put into place you could try to mitigate some of these biases. 
You may not remove all of them but you can definitely go a long 
way toward doing that.
    Chairman Peters. Thank you.
    Ms. Givens, you have told us how public conversation about 
responsible AI has been evolving. Could you help us understand, 
what would a truly responsible AI system actually look like?
    Ms. Givens. You have already started an important 
conversation around bias, but I think we also need to pull out 
the broader frame of are these systems working as intended. 
There is a functionality question to be had about are we 
actually able to rely on rational and predictable outcomes. Is 
the model structured in a way to actually allow people to have 
trust in the results that are being generated?
    When NIST produced its AI Risk Management Framework they 
identified a number of characteristics of what makes a 
trustworthy AI system, and I think it is actually a very useful 
way to think about these issues. For them, the factors are is 
it valid and reliable; safe, secure, and resilient; accountable 
and transparent; explainable and interpretable; privacy 
enhanced; and fair with harmful bias managed. Really each of 
those elements is its own inquiry. We need our own bodies of 
work as to how to make sure each of those are being maintained. 
But I think that is an incredibly useful way of breaking down 
these different elements of what it is to develop responsible 
AI.
    Then the final piece is we have to think about this through 
the entire lifecycle, so not just at the moment the tool is 
being designed in the first place but how and where it is being 
deployed, what that looks like in its contextual setting, and 
then because these tools, the whole way that they work is by 
learning over time, ongoing auditing and checks to make sure 
that they are still reliable, trustworthy, and have not brought 
in additional biases.
    That is the way we need to think about a holistic approach 
to these questions.
    Chairman Peters. Thank you.
    Senator Blumenthal, you are recognized for your questions.

            OPENING STATEMENT OF SENATOR BLUMENTHAL

    Senator Blumenthal. Thank you very much, Mr. Chairman, and 
I want to thank you for having this hearing, and the panel that 
we have which is, as you referred to it as truly eminent, 
informed, very helpful, and I welcome your willingness to have 
additional hearings, which I think most certainly we will want 
to do.
    Professor, I was interested in your reference to 
algorithms, quoting Princeton professor Narayanan, as ``snake 
oil.'' For me the danger of that snake oil is not only the 
mistakes that can be made, that is, the failings, and you all 
have identified some of those failings, but sometimes how they 
work all too well, the algorithms which are essentially, for 
most people in this world, black boxes, driving content to 
children. I want to thank the Chairman for his support in the 
efforts that we are making to protect children better than we 
have before. But these algorithms that work all too well will 
identify an interest that a child has and then continue driving 
content to that child. The idea that artificial intelligence is 
something way off in the future I think is a little bit 
misleading because right now Google and others are using these 
algorithms to drive that content.
    Could you describe whether they have control--and I will 
ask the other Members of the panel as well--whether they have 
and could exercise more control over what these algorithms do 
and whether they could make them more transparent.
    Mr. Venkatasubramanian. Thank you for the question, 
Senator. I should say up front I am not an expert on matters 
linked to children's safety online, but as an AI expert what I 
can tell you is that for all of these algorithms, and the kinds 
you mentioned, the things we have talked about today so far, 
the importance of governance, the importance of transparency of 
how these algorithms work, of having independent review and 
ongoing monitoring, are critically important to make sure that 
they do not have the consequences that we do not want them to 
have.
    That idea of governance in AI, it is an important part of 
the process of determining what is it we want out of these 
algorithms we are deploying. Oftentimes we do not ask that 
question, and algorithms are used for engagement or for selling 
ads, and we do not ask the question of what impact they are 
having.
    Having a broad framework, an overarching, comprehensive 
framework, where we can evaluate what these algorithms are, how 
they work, and what they are doing is a way, in general, that 
we can make sure that we can get the benefits of these systems 
and not get the harms.
    But to your specific point about child online safety I will 
defer to others on the panel who have more expertise.
    Senator Blumenthal. Ms. Givens.
    Ms. Givens. Senator, I know you have been a longtime leader 
on this issue, and we have worked for a long time with your 
staff on comprehensive privacy protections, not just for kids 
but for all consumers, frankly, engaging in these online 
platforms, where the hyper-targeting of content and of ads 
really can have harmful effects.
    I agree that this is a priority area. We have heard 
policymakers across the country and internationally focused on 
these issues and thinking about what responses can look like.
    Within my organization, one of the things we think about is 
how do you create the right incentives for companies to do well 
without creating adverse incentives that may end up, 
unfortunately, impacting kids, teens and their ability to 
access important information online. I think sometimes there 
can be questions about what are the right levers to push, how 
do we incentivize responsible design practices without creating 
a culture where, for example, it might be hard for teenagers 
online to access information about reproductive care or 
information that might be useful for them when they are 
exploring their gender identity or their family identity.
    There is a balance to be struck here, but on the overall, 
making sure that companies are being responsible in this space 
is incredibly important.
    Senator Blumenthal. Yes, I thank you for the work that you 
have done in this area, and particularly with my office, I know 
you have been very positive and constructive, so I thank you.
    I am hoping, to cut right to the chase, that we can move 
forward on the Kids Online Safety Act, which provides for more 
transparency and at the same time provides for tools and 
safeguards for children and parents to make judgments that give 
them, in effect, control back over their lives, which many feel 
now they are losing, and avoid the unintended consequences that 
you just referenced, unintended consequences that may involve 
constraints on free expression or other goals. I think there is 
a balance to be struck here. I think that is our goal. That is 
what the legislation has attempted to do.
    I do not know whether anyone, whether you have any comments 
on this question. Mr. Matheny.
    Mr. Matheny. Thanks, Senator. The one thing I would add is 
just that the potential for misuse is grounds for considering 
an appropriate regulatory framework, and I think reason to be 
especially cautious about open sourcing large language models 
that could be misused.
    Senator Blumenthal. One of the goals of the legislation is, 
in fact, greater transparency, and open sourcing certainly is a 
way of addressing that issue.
    Thank you, Mr. Chairman.
    Chairman Peters. Thank you. Senator Hassan, you are 
recognized for your questions.

              OPENING STATEMENT OF SENATOR HASSAN

    Senator Hassan. Thank you, Mr. Chair, and I want to thank 
the panel for being here and for your work. I want to start 
with a question to you, Dr. Metheny.
    I am Chair of the Subcommittee on Emerging Threats (ETSO), 
a Subcommittee of this Homeland Security Committee, and I focus 
there, among other things, on the risks that artificial 
intelligence could pose to the cybersecurity of critical 
infrastructure like electric grids and hospitals.
    You talked a little bit about some of the risks that AI 
poses, but can you expand a little bit, how does AI impact the 
cybersecurity threat landscape and are there opportunities to 
utilize AI to counter these threats?
    Mr. Matheny. Thanks, Senator, for the question. The 
application that has probably gained the most public attention 
of these large language models is generating language that we 
are familiar with, natural language, so creating an English 
poem or an English short story. What is getting less attention, 
but could be more impactful on security, is the application of 
these large language models to be used for software generation, 
code generation, and computer programming languages rather than 
a natural language. Some of these applications are already 
fairly sophisticated, and an increasing fraction of new 
software engineering is taking place with the use, or 
assistance, of large language models.
    If this trend continues, it is quite possible that the 
offense of cyber capabilities that today are accessible only to 
state-level actor offensive cyber programs could be accessible 
to a much larger number of actors, simply by having access to 
tools that are able to generate software at scale and requiring 
much less technical sophistication to do so. Those could pose 
risks then to critical infrastructure and other networks that 
are sensitive.
    Senator Hassan. Thank you for that. Are there capacities 
that could help counter that, that AI gives us?
    Mr. Matheny. The same tools can also be used to scale up 
cyber defense, and I think this will be a cat-and-mouse race to 
figure out, are the applications on the defensive side keeping 
up with the applications on the offensive side.
    I do not know the answer to that question. I think it will 
be a continuous competition between offense and defense.
    But we need to make sure that our cybersecurity 
organizations are keeping up with the trends in these large 
language models as they are applied.
    Senator Hassan. OK. Thank you. Another question for you, 
Dr. Matheny. AI capabilities will offer new opportunities for 
the intelligence community, theoretically at least, to improve 
national security. Are there ways you believe that AI can 
improve intelligence analysis?
    Mr. Matheny. Yes. I think that the application of AI 
systems, particularly in open source data, where the volumes of 
data exceed our ability to analyze using manual methods, is one 
of the most important areas for intelligence. We could be 
making use of a much broader range of open source imagery, open 
source text in order to understand what is happening in the 
world much faster, and be able to share it with the world much 
more quickly.
    As we are seeing from the war in Ukraine, when we are able 
to share open source information we are able to change the way 
the world understands what is happening in a part of the world 
that we do not have direct access to.
    Senator Hassan. Thank you. Then another question for you, 
Dr. Matheny. We know that government initiatives generally 
involve a number of different Federal agencies, and one of the 
things I am interested in is how can the Federal Government 
ensure that their agencies are coordinating with one another on 
AI research and deployment for potential joint projects or 
initiatives?
    Mr. Matheny. Thanks, Senator. One of the things that RAND 
has been working on over the years is how investments by one 
organization within the Federal Government, say one of our R&D 
organizations, can be more broadly shared across the government 
faster and how we can harmonize different efforts so that we 
are not duplicating efforts in one area of research, so that 
tool that are created by one agency can be leveraged by 
another, and so that standards that are used by one agency, say 
for AI being used for a particular application, can be 
harmonized with those in another agency.
    I think there are great gains in efficiency.
    One of the ways of harmonizing this would be through 
Federal procurement and ensure that we are using a consistent 
set of standards. Another would be through agencies like the 
National Institute of Standards and Technology, that have a key 
role to play in creating test frameworks and testbeds where we 
can robustly evaluate the performance of these AI systems.
    Senator Hassan. Thank you.
    Now a question for Professor V, as I will call you, and Ms. 
Givens. This is a question for both of you. There is growing 
concern among workers in many industries that AI could 
fundamentally change the nature of work in unpredictable ways. 
You have touched on this a little bit, but do you have 
recommendations for how the Federal Government should be 
addressing challenges that companies and employees face from 
the use of AI in the workplace? Dr. V, I will start with you.
    Mr. Venkatasubramanian. Thank you for the question, 
Senator. I think there are two parts to this, to helping 
workers deal with displacement due to AI. One, of course, is 
training and skilling, and the Federal Government can invest 
effort and research into helping workers train for our science, 
technology, engineering and mathematics (STEM)-enabled world. I 
think the Federal Government is doing that, and we can do 
definitely a lot more on that.
    I think it is even more important that we make sure that 
that training and that access to those skills is widely 
distributed and not just to those who have access to those 
already. That is one thing.
    I think another component of this is when we talk about 
worker displacement due to AI. I fundamentally believe it is 
because of overpromising on the part of AI systems, that tends 
to not play out when these systems are deployed.
    Systems are presented as being able to replace because of 
efficiencies, workers, but in fact they cause more problems 
than they deserve, and it is precisely because there is not 
governance, there is not the supervision, there is not the 
human supervision around these systems.
    I would argue that rather than thinking about workers 
displaced by AI, if we put proper governance and structures in 
place we will need more jobs for workers, in fact, to make sure 
that these systems, that are supposed to assist them, are not 
replacing them and doing it badly.
    Senator Hassan. Thank you. Ms. Givens.
    Ms. Givens. One of the questions is to think not just about 
displacement but if we are striving for a goal of workers 
working alongside AI systems, what does that interaction 
actually look like? We are seeing this play out now. You can 
think about fulfillment centers, for example, where workers are 
actually tasked with extreme specificity to every motion that 
they take, in the name of efficiency. There are business 
reasons for doing that, but there are also very real human 
impacts on the workers who are micromanaged at that level and 
live in a far more surveilled environment than they did before. 
Delivery van drivers, there are many other examples of this.
    There we need to think about things like workplace health 
and safety. The Occupational Safety and Health (OSHA) has a 
role to play. The Department of Labor (DOL) has a role to play. 
We need to think about enforcement, both of existing laws and 
how we create a movement for employers to understand what 
responsible practices look like and for workers to know and 
understand their rights.
    Senator Hassan. Thank you very much. Thank you, Mr. Chair.
    Chairman Peters. Thank you, Senator Hassan.
    Senator Padilla, you are recognized for your questions.

              OPENING STATEMENT OF SENATOR PADILLA

    Senator Padilla. Excited about the opportunities that 
advances in technology will offer to society. But as this 
conversation has already shown, with every disruptive 
technology there are risks that demand mitigation. For example, 
automated decisionmaking systems and tools risk actually 
exacerbating the many existing inequities in our society, and 
that actually leads me to my first question.
    It is clear that investments in AI research and education 
have not been distributed equally across the nation's 
researchers and innovators. Racial and gender diversity in AI 
and computer science programs are severely lacking. This lack 
of diversity among students gives rise to the corresponding 
lack of diversity in the workforce. A lack of diversity in the 
workforce then contributes to the development of AI tools and 
approaches that either do not account for or actively 
perpetuate systemic bias and limits the breadth of ideas 
incorporated into AI innovation.
    My question is for Dr. Venkatasubramanian. As an educator, 
how can we ensure our AI and computer science students and 
workforce reflect the diversity of our nation?
    Mr. Venkatasubramanian. Thank you, Senator, for that 
question. This is an issue that concerns me greatly, as you 
might imagine, as an educator. I see the students who come to 
me who are concerned about this, and more often than not the 
students who are most concerned about these issues are students 
who truly reflect the broad diversity in this country, which is 
in one way a very good thing, but it also shows where the gaps 
in our ability to deliver STEM education effectively to our 
population is.
    I would say that in my experience when students are able to 
see themselves in the work that they do, and the topics they 
study, they are more engaged with it and they feel like 
technology, in this case, can speak to them. Thestudents who 
come to speak to me about concerns around bias and responsible 
AI come to me because they have literally said, ``I finally see 
a place for myself in this tech ecosystem.''
    One of the reasons why I spend a lot of time talking about 
concerns about bias and inequities in technology isbecause it 
is only by speaking out loud about those issues and pointing to 
the ways in which we can use technology to mitigate those 
issues that we can actually bring in a population that feels 
like they are now being heard and that their concerns are being 
heard.
    I view these as part of the same story, that by spending 
time recognizing the inequities of AI, by spending time 
recognizing the need to govern areas to take account of these 
inequities we are actually telling people, ``We welcome you in 
this technology and in this technology-enabled world.''
    Senator Padilla. Thank you. Ms. Givens, I would be remiss 
if I did not take the opportunity to ask the former Chief 
Intellectual Property Counsel for the Senate Judiciary 
Committee a question about intellectual property. AI is 
introducing novel questions about the extent of a creator's 
intellectual property rights, most notably in the world of 
copyrights. Do you have any advice for those of us on the 
Judiciary Committee as we enter this new era of internet 
protocol (IP) complexity?
    Ms. Givens. I am afraid I do not have a solution for you on 
this incredibly complex issue, but I do think it is an area 
where much attention is needed. There are photographers and 
designers and artists out there who understandably are deeply 
worried about the erosion of their industry and the role that 
they can play with the creation of generative AI, and also that 
their work is being used to train those tools.
    On the other hand, we have had a very long tradition of 
fair use principles, and uses for transformative works in the 
creative space. There has to be a healthy conversation around 
how we appreciate some of those concerns of creators without 
inhibiting what is, in itself, an expressive act, the creation 
of new and diverse and transformative works through these 
tools.
    Senator Padilla. Thank you. To be continued. Dr. Matheny, 
large language models are rapidly improving and generative AI 
can have many important and positive applications. However, as 
a former elections administrator, I want to share a specific 
concern that I have about the ease with which this technology 
could facilitate election disinformation campaigns. Generative 
AI could radically reduce the cost and time while increasing 
the impact of misinformation and disinformation and propaganda. 
Not only could someone make it seem like one of us on the dais 
said something that we did not say or endorsing something that 
we do not endorse, but also the ability of foreign actors to 
supercharge their efforts to interfere in our elections is 
absolutely clear.
    Referencing back to the Judiciary Committee, we know from 
our law enforcement officials that it is actually domestic 
extremism and white supremacy that pose the largest national 
security threats to the United States.
    It is bad enough that Speaker McCarthy was willing to share 
with Tucker Carlson all the footage of January 6th, which is 
now being repackaged to make it seem like a whole different 
January 6, 2021, took place than what is reality.
    That is using actual footage. Imagine AI-generated video 
and the power that it can have in reshaping people's 
perspectives and attempts to redefine the truth.
    Doctor, in light of your testimony, how do you recommend 
that we prepare our elections infrastructure and political 
processes to address propaganda that is harder to detect?
    Mr. Matheny. Thanks for the question, Senator. For several 
years RAND has had a project on something we called ``truth 
decay,'' which is the vulnerability of democracies to 
disinformation attacks and other attacks against norms of 
evidence used in policy debates. One concern that we have had 
for several years is that the application of AI to 
disinformation campaigns could, as you point out, radically 
reduce the costs and increase the scale and speed of text, and 
speech potentially, that is used in disinformation, in ways 
that are very difficult to distinguish from human-generated 
forms of text and speech.
    I think one important area is in research on distinguishing 
generative model text and speech compared to ones that are 
authentic. First, how can we watermark the products of 
generative AI systems in ways that we can distinguish them, and 
second, for those systems that have not used watermarking, can 
we find other signatures that allow digital forensics to be 
able to distinguish that which is disinformation from that 
which is legitimate.
    Senator Padilla. Also to be continued. Thank you, Mr. 
Chairman.
    Chairman Peters. Thank you, Senator Padilla.
    Senator Sinema, you will be recognized for your questions. 
The vote has been called. I am going to run to vote. If you 
could take the gavel while I vote and then come back, and then 
Senator Rosen, I will be back shortly.

              OPENING STATEMENT OF SENATOR SINEMA

    Senator Sinema. [presiding.] Sure. Thank you, Mr. Chairman. 
Thank you to our witnesses for being here today. AI has the 
potential to revolutionize Arizonans' lives in countless ways, 
both good and bad. As we continue to integrate AI into our 
society we must ensure that this technology is developed and 
deployed in an ethical, transparent, and responsible manner 
that safeguards our values, preserves our privacy, and protects 
our national security.
    My first question is for Dr. Matheny. Generative AI is 
suddenly everywhere, including ChatGPT and deepfakes--those are 
the fake videos that make people appear to say or do things 
they did not actually say or do. I have some experience with 
that. The key to solving this challenge is transparency, and 
one of the most promising solutions is so called content 
provenance data. This allows digital creators to embed data in 
content that disclosures whether it is authentic, altered, or 
entirely synthetic.
    Do you believe that increasing transparency around what 
content is original versus what is AI-generated should be a 
policy priority for policymakers, and if so, is promoting 
content provenance efforts one of the most promising ways to 
create that transparency?
    Mr. Matheny. Some work that RAND has done over the past few 
years has identified watermarking and other ways of asserting 
provenance for digital media as being an important 
countermeasure against deepfakes, other forms of generated 
media that could be malicious.
    We also need to increase our ability to do forensics on 
media that may have been generated but does not leave as easy 
telltale signatures, either because the entities that generated 
that media have not participated in various kinds of regulatory 
efforts to introduce watermarking or provenance.
    I think what is likely to be required are investments in 
each of these categories, some way of asserting provenance, 
some way of watermarking, and investments and research on 
forensics for those that do not participate in the other two.
    Senator Sinema. Thank you.
    Ms. Givens, as Chair of this Committee's Government 
Operations Subcommittee I am committed to ensuring that the 
Federal Government serves as a role model for society when it 
comes to responsibly and ethically deploying AI. I also serve 
as the Chair of the Commerce Subcommittee that oversees NIST, 
which just released its first-ever AI Risk Management 
Framework.
    What is your assessment of the Federal Government's current 
AI practices, particularly with respect to transparency, bias, 
accuracy, and effectiveness, and how can government better 
manage these risks when it deploys AI?
    Ms. Givens. This Committee has taken some important steps 
to show the need for rigorous processes and how agencies think 
about their use, design, procurement of AI.
    I think there is still quite a lot of room for growth. The 
AI Risk Management Framework released by NIST is an excellent 
starting point, but we really need to operationalize it. We 
need to make sure that it is useful for people in the sectors 
of applicability where they are working. NIST needs to keep up 
its work on measurement strategies and ways to actually 
identify bias and assess whether or not interventions are 
working and are appropriate.
    Then one of the leading things that this Committee helped 
generate, and it is being bolstered by a number of Executive 
Orders, is the inventory of agency uses of AI and guidance 
coming from OMB, and those are still works in progress as far 
as I understand. One of the priorities, I think, needs to be 
expediting that work, for OMB to play its central coordinating 
role, helping guide acquisition and use principles, and then 
starting the cycle of agencies inventorying their uses and 
showing how they are going to comply with that guidance in a 
meaningful way.
    Senator Sinema. Thank you.
    My next question is for Dr. Venkatasubramanian--I practiced 
that one--and Dr. Metheny.
    I would like to continue on the topic of ethical AI but in 
the context of U.S.-China competition. As we compete against 
Beijing to win the AI race, America may lose if we focus solely 
on the size of our datasets, since, frankly, China's 
authoritarian system lends itself to vacuuming up vast volumes 
of data with few privacy protections. In contrast, America's 
competitive advantage may be our values, if we can translate 
these values into developing AI that is transparent, efficient, 
and fair.
    What advantages and disadvantages does our country face in 
the AI competition with China, and do you agree thatinstead of 
viewing our values as a liability in this competition America 
could and should view them as an asset?
    Mr. Venkatasubramanian. Thank you for the question, 
Senator. I completely agree with the idea that the United 
States has values that can be transmitted into the systems we 
build, and I would argue this is happening already, but 
unfortunately the United States is not leading on this. For 
example, in the European Union (EU), with the development of 
the AI Act and other legislation that is going to govern the 
use of technology, especially AI technology, there is an 
attempt to push forward on the kinds of responsible practices 
that I think have been, frankly, developed here in the United 
States but are now being used in Europe. I think the United 
States can take the innovative lead, on these practices and 
provide a model for, frankly, the rest of the world to follow 
in how we do AI that is innovative, as well as responsible, as 
well as ethical at the same time.
    I think we should push forward on that, we should emphasize 
that, and we should prioritize investments in those directions 
by prioritizing it within Congress and within the Federal 
Government.
    Mr. Matheny. I think that the United States has a couple of 
asymmetric advantages compared to China in AI. The first is 
that we are a much more attractive destination for the world's 
computer scientists and engineers. The United States has only 
four percent of the global population.
    China has only 18 percent. The other 78 percent is sort of 
up for grabs. The United States does a much better job of 
attracting scientists and engineers from overseas. Many of the 
scientists and engineers are attracted by our values, so I 
think those values are a deep part of our asymmetric advantage.
    A second advantage that we have is our ability to work with 
allies and partners. The United States and China each are 
responsible for about 25 percent of global research and 
development spending. When you add the United States and its 
allies, and add China and its allies, China's percentage does 
not increase because it does not have alliances with strong 
technological powers. The United States increases from 25 
percent to about 65 percent.
    Again, this is a place where having friends who are 
attracted to our values, who share our commitment to privacy, 
democratic governance, and the rule of law works to our 
advantage.
    Senator Sinema. Thank you. Senator Rosen.

               OPENING STATEMENT OF SENATOR ROSEN

     Senator Rosen. Thank you, Senator Sinema, and thank you to 
the witnesses for testifying today. I want to really speak a 
lot about skilled workforce because it is challenging across 
all platforms, as we see. Everyone who comes to talk to me is 
challenged with finding a skilled workforce, and our Federal 
agencies and the digital Workforce, no different.
    The National Security Commission on Artificial Intelligence 
does warn, and I am going to quote here, ``The human talent 
deficit is the government's most conspicuous AI deficit and the 
single greatest inhibitor to buying, building, and fielding AI-
enabled technologies for national security purposes.''
    The government, of course, we cannot compete with private 
sector salaries. We suffer from recruitment and retention 
issues, and the sustained AI talent shortage at government 
agencies, everyone would argue, could undermine our 
competitiveness.
    Dr. Matheny, what are the specific ways you think the 
Federal agencies can really work to improve and expand that AI 
talent pipeline, and how might academic partnerships and 
initiatives be leveraged right now, public-private sector, to 
fill some of these gaps perhaps?
    Mr. Matheny. Thanks so much for the question, Senator.
    I think that one of our most important levers or tools like 
the Intergovernmental Personnel Act, which allows the Federal 
Government to leverage expertise that is in academia and that 
is in other parts of the private sector, to bring in technical 
experts for short-term appointments, where they can serve as 
subject matter experts within Federal agencies.
    I think we also have roles, like special government 
experts, that allow those in the private sector to maintain 
their positions in the private sector while they still advise 
government.
    We certainly need to buildup the expertise within our own 
Federal workforce, but we also need to find more agile ways of 
leveraging the expertise that is distributed throughout the 
private sector, and those are two, I think, of our most 
important authorities to do so.
    Senator Rosen. Thank you. Professor V, I am going to turn 
to you because what kind of research and development 
investments should we be making to do just this, to uphill, 
reskill, or some might say right-skill the folks that are out 
there that do want to work, giving them an onramp to these jobs 
that can continue to grow?
    Mr. Venkatasubramanian. Thank you, Senator. As I mentioned 
earlier in response to Senator Padilla, one of the reasons that 
animates students from across the spectrum to work in 
technology, especially those who have not been seen by 
technology or are not being represented by technology earlier, 
is a desire to do something in the public good, to do something 
to improve the way all of us get the benefits of technology.
    I feel like the Federal Government is a place where a lot 
of these students, university students, come and say they want 
to work in the Federal Government. They do not want to work in 
the private sector because they want to do some good. I think 
that is where the Federal Government has a comparative 
advantage over the private sector, becausethe Federal 
Government can articulate a value of public good in working 
with technology. I think the Federal Government should 
advertise that, should focus on developing technology to help 
bring it to all in a responsible and ethical manner.
    I think Congress should continue its work to grow Federal 
expertise through training and skilling programs in the Federal 
Government. I think Congress should bring back the Office of 
Technology Assessment to help Members of Congress, the 
legislature, get more expertise on these topics as well.
    Senator Rosen. I could not agree more, as a former software 
developer, and so I am going to continue on this vein as we 
think about AI, the application that we use it for, 
cybersecurity, that can help us in these hunt forward 
operations, highlighting, or flagging, if you will, things for 
then humans to discern what seems right. So AI technology is 
rapidly evolving, and like I said, we really have to work on 
this. The National Cybersecurity Commission calls for more AI 
funding for AI-enabled cyber defenses.
    Again, Professor V, how do you think we can enable and use 
AI to detect malware, pattern recognition, the things that 
computers are really good at, on the defensive side, and how 
can we use that to harden our security against cyber threats?
    Mr. Venkatasubramanian. Senator Rosen, I think I will defer 
that question to Mr. Matheny here. He has much more expertise 
than I do on the national security side.
    Mr. Matheny. You are too kind. I am worried about the long 
run arms race between offense and defense on cyber. I think 
both sides are amplified in their abilities by applications of 
different kinds of AI approaches.
    On the defense side, as you mentioned, pattern recognition 
for looking for network activity that could suggest that there 
is an attack in progress. Most attacks are discovered weeks 
after. It would be nice if we detected them while they were 
happening so that we could do something about them. I do think 
that AI offers some applications in this area and there are 
active projects at the Intelligence Advanced Research Projects 
Activity (IARPA) and Defense Advanced Research Projects Agency 
(DARPA) to apply AI to cyber defense.
    On the offensive side, I think one concern is that we are 
going to see increased levels of sophistication among 
relatively moderately skilled programmers in developing code 
much more quickly that can be used offensively.
    I think the main thing here is for Federal agencies to be 
aware of how AI is being applied both offensively and 
defensively so that we are not surprised.
    Senator Rosen. Yes, I think you are right about that.
    I am going to continue in this vein about this national 
strategy because you spoke earlier about the EU publishing 
their coordinated plan on AI, and they are encouraging each of 
its member States to develop their own national strategies. Of 
course, last week the White House released our national 
cybersecurity strategy.
    What do you think would be the potential value for the U.S. 
national artificial intelligence strategy, more broadly, and 
how can interagency collaboration on AI be improved so we can 
detect and respond to threats more rapidly?
    Mr. Matheny. First I think all agencies would benefit from 
being able to draw in greater expertise, and that need not just 
mean full-time employees. It can mean advisors, consultants. 
Second is having a common frameworkfor AI standards that all 
Federal agencies can leverage.
    Here I think there is a key role for NIST to serve in 
developing uniform guidance for standards, ensuring that we 
also participate in international standards like ISO, SC 42.
    Then third, I think shared Federal procurement rules that 
allow agencies to be developing tools that are built toward 
common standards with a common test framework.
    Senator Rosen. Speaking as a former software developer, the 
word ``common framework'' is music to my ears, so I am just 
going to leave it at that.
    Mr. Chairman, I yield back.
    Chairman Peters [presiding.] Thank you. Thank you, Senator 
Rosen.
    Dr. Matheny, you have extensive experience investigating 
threats posed by AI and national security, which is why it is 
so wonderful to have you here today. You have also written in 
support of export bans on the Chinese government. Could you 
tell us more about the threats that AI poses in the hands of 
the Chinese government and its State-sponsored companies and 
why bans may be appropriate to look at?
    Mr. Matheny. Thanks, Mr. Chairman. One of the things that I 
worry about, and I am a bit of a Debbie Downer on this, is that 
AI can be used to accelerate the development of other 
technologies. We are seeing early forms of this, where tools 
like AlphaFold were used to solve a very hard problem in 
biochemistry, the protein folding problem.
    The upside potential of this is enormous. We can imagine 
this being applied to medicine in a variety of beneficial ways. 
It can also be used, though, to develop novel pathogens, and 
States that have historically not hadas many taboos as 
democracies around the use of technologies such as 
biotechnology for malicious use, I worry deeplyabout how AI 
will be used to supercharge different research and development 
efforts.
    The same goes for offensive cyber, and the same goes also 
for disinformation used both domestically within China's own 
population for human rights abuses, for surveillance 
applications in Xinjiang and elsewhere in China, and used to 
influence foreign populations.
    Chairman Peters. You talk about other uses, the dual use of 
this, and we know that AI has a great deal ofpotential to deal 
with diseases that we have been attempting to cure forever, 
diseases like cancer. But I am curious of your thoughts about 
AI systems being weaponized perhaps, to find biotoxins or 
chemical warfare agents. How concerned should we be about that?
    Mr. Matheny. Countries like China that have historically 
invested in biological weapons and that havedemonstrated an 
interest in ethnically targeted weapons greatly concern me. The 
use of AI for so-called genome-wideassociation studies to try 
to identify how one would ethnically target particular 
pathogens is one area of special concern. We know, from a 
variety of research efforts historically, that the most 
virulent or transmissible pathogens are not those that are 
found in nature but ones that can be constructed artificially. 
AI creates opportunity to enhance pathogens much more quickly 
and perhaps in ways that deliver effects to specific 
populations that are vulnerable.
    Chairman Peters. Ms. Givens, you have talked about AI and 
privacy and how our privacy is in danger, and this actually 
picks up a little bit on this question about creating 
pathogens. Would you talk a little bit about the privacy risk 
associated with using AI in the context of biometric data? We 
are providing more biometric data in databases. What are some 
of the concerns that you have associated with that?
    Ms. Givens. Absolutely. Biometric data is one of the most 
sensitive types of data we can have. If there is a data breach 
and my faceprint is taken--I am not changing my face any time 
soon and I do not have the capacity to do so--so this 
information is highly in need of protection.
    That makes it challenging when we think about the use of 
biometric identifiers, for example, in the delivery of 
government services. An increasing focus in fraud detection, 
for example, uses face recognition technology, one-to-one 
matching. Of course, there are law enforcement uses that are 
underway in the United States as well. We really need to think 
long and hard about the security vulnerabilities that can be 
created through this technology.
    In addition, there are real concerns about equity when 
these types of technologies are being used. When, for example, 
your ability to access government services is contingent on you 
being able to snap a good selfie on your phone, that can 
exclude a large number of people that do not have that 
technology on their phone. Government agencies need to think 
about responsible onramps, responsible transitions for others 
as well.
    But the cybersecurity and privacy vulnerabilities are real, 
and that is why it is so important to come back tothis language 
we have been talking about around real procurement standards, 
real safeguards, to make sure that when the government is 
considering using this technology there is a weighing of pros 
and cons, and then making sure that risks are mitigated.
    Chairman Peters. Thank you.
    Professor V, I have heard concerns about effective 
computing, which tries to discern someone's emotion from those 
facial expressions that Ms. Givens was just talking about. 
Could you tell the Committee more about effective computing and 
if you have concerns?
    Mr. Venkatasubramanian. Yes. Thank you for that.
    The premise, or the stated premise of effective computing 
is that we can infer information about people's internal 
States, their emotions, their cognitive States, their affect, 
from external features, external features like facial 
recognition, external features like how they walk, what kind of 
microtargeted expressions on their face, wrinkles, frowns, and 
so on.
    I have great concerns about this. The premise of effective 
computing is unfounded. It has no basis. AI systems cannot do 
this. They might claim they do but they cannot because there is 
no underlying science to back this up. There is no underlying 
science that says that you can, in fact, do this kind of 
inference of people's internal States from external features. 
It just does not work, and most claims are, pardon my 
expression, completely bogus.
    Chairman Peters. Professor V, based on your time at the 
Office of Science and Technology Policy and your contributions 
to the Blueprint for an AI Bill of Rights, could you paint a 
picture for us of what a truly accountable AI system would look 
like within a Federal agency?
    Mr. Venkatasubramanian. Yes. This Federal agency procures 
and wants to procure an AI system that would be used to impact 
people, it would start by consulting with advocates, community 
partners, and other stakeholders to ensure that any system it 
might want to procure truly benefits those being impacted, in 
an equitable manner.
    The agency will lay out strict guidelines and 
specifications to make sure that only the specific task is 
being sold, and that the system is not being repurposed for 
other tasks as well. It will make sure that the procurement 
process incorporates information about testing and validation 
for the specific task, that the system, in fact, works, and 
that as appropriate, disparity mitigation has been performed 
and results of these disparity mitigations are presented to the 
agency before procurement. It would not hand over people's data 
to the vendor, and if necessary would only share data with the 
vendor in a very controlled environment, for development 
purposes only.
    Any deployed algorithm, once the system is deployed, would 
be supervised by agency experts who have expertise in the 
domain of interest and can tell when the algorithm or the 
system might be generating inaccurate outputs. The system would 
be regularly re-evaluated on a standard, on a cadence, to make 
sure data shifts have not affected its behavior. The vendor 
would need to provide tools to explain the algorithm's 
behavior.
    I think an agency that is doing deployment of accountable 
AI well would be doing all of these things.
    Chairman Peters. Professor, if Congress were to requirement 
all the practices that you mentioned, what government body do 
you think would be best suited to hold agencies accountable?
    Mr. Venkatasubramanian. I think it is helpful to maybe 
distinguish between private sector use cases and government use 
cases. For private sector use cases, the FTC and its new Office 
of Technology would be perhaps the best place to do this, and 
should be given the resources to do this kind of work. For 
government uses, using the National AI Office that Congress had 
created, and OMB would probably be the best place to have high-
level guidance and supervision of these systems.
    Chairman Peters. Is there an example of an agency now that 
is using AI effectively and responsibly, in your opinion?
    Mr. Venkatasubramanian. The Department of Health and Human 
Services (HHS) has done an excellent job complying with 
congressional mandates around the inventory of AI, for example, 
and around executive orders around AI. They are being very 
careful, for example, in their Updates Rule 1557 and the 
development of guidelines together with the Food and Drug 
Administration (FDA) around the use of AI in diagnostics, and 
that is one agency I would definitely hold up as doing a good 
job in this space.
    Chairman Peters. Great. Dr. Metheny, this Committee has 
focused on laying some of the groundwork for responsible agency 
use and acquisition of AI. In our legislation we require 
standards and safeguards for acquiring and deploying these 
technologies and ensuring that the Federal workforce is up to 
the task to do that.
    Can you elaborate on what else we could be doing to make 
sure that government procures and uses AI effectively and 
responsibly?
    Mr. Matheny. Thank you, Mr. Chairman. I think the U.S. 
Government has a fair amount of purchasing power that it can 
leverage to require that procured technologies meet certain 
standards of safety, reliability, robustness, and those 
standards could be verified in compliance through a third-party 
audit. I think that is one important lever that the Federal 
Government has. It will still not be the primary purchaser, but 
the private sector, in order to comply with such standards, it 
would simply make business sense for them to ensure that their 
systems, on the whole, are compliant.
    A second key area, I think, is ensuring that democracies--
the United States, its allies, and partners ensure that the 
international standards for AI systems are ones that support 
democratic norms around privacy and self-determination. We have 
the opportunity, through the international standards processes 
such as SC 42 that I mentioned earlier, to make those standards 
be ones that are privacy preserving, that are compatible with 
encryption, for example, and I think that is an opportunity we 
should seize.
    Chairman Peters. Thank you. The last question before we 
wrap up this hearing I am going to pose to each of you.
    I will start with you, Ms. Givens, and then we will just 
work down the dais there.
    We have heard commentators and academics have warned about 
the risk of human-like artificial intelligence, or artificial 
general intelligence, and those tend to be a lot of 
apocalyptic, scary stories that people talk about. But my 
question to each of you is, what are the risks that artificial 
general intelligence pose, and realistically, how likely is 
that actually in the near future? What is your assessment of 
how fast this is going and when we may beconfronted with some 
of those even more challenging questions and issues?
    I will start with you, Ms. Givens.
    Ms. Givens. I will leave to some of my more technical 
colleagues to do the likelihood question. I never want to make 
a prediction on a congressional panel. But I will say that when 
we are talking about such sophisticated technology, it raises 
many of the issues that we are already facing now, but simply 
supercharged, which is why we have to get the fundamentals in 
front of us correct now. When we are thinking about, for 
example, rules-based systems and controls, we already have a 
hard enough time thinking about how to respond to machine 
learning models now. When we think about these advanced 
systems, the notion that those are going to evolve rapidly over 
time makes it even harder to contemplate.
    We have to address these questions of competency, of 
responsible design practices from the beginning, and we have to 
get our fundamentals right now, in the opportunity before us 
immediately, the ways in which AI is harming people in their 
daily lives right now, and the lack of ability for government 
agencies right now to be able to meaningfully respond to it, 
for us to even begin to think about how we tackle the next 
generation of issues.
    Chairman Peters. I think it is an important point.
    The technology we know, that we heard from the experts 
here, is advancing very rapidly. In the past we have tended to 
look at technology as it is developed and just be excited about 
the promise of it. It gets developed and then we start seeing 
some adverse consequences, and then we look at regulation or 
other types of ways of dealing with it.
    In this case this is moving so fast that I am concerned 
that if it gets way ahead of us that we cannot use the model of 
the past, where we see how things work out and then we address 
it. We really have to be thinking ahead, thinking a few steps 
ahead, which is why I am asking this question about the 
probability of even more powerful systems.
    Professor V, that is in your wheelhouse.
    Mr. Venkatasubramanian. Yes. People ask me what keeps me up 
at night. AGI does not keep me up at night.
    The reason why it does not is because, as Ms. Givens 
mentioned, the problems we are likely to face with the 
apocalyptic visions of AGI are the same problems we are already 
facing right now with the systems that are alreadyin play.
    I worry about people being sent to jail because of an error 
in an ML system. Whether you use some fancy AGI to do the same 
thing, it is the same problem, and we are seeing this problem 
right now.
    I think that the Committee's time is well spent pondering 
the harms that we are facing right now from these systems, and 
I would say, again, it is hard to predict. I am a computer 
scientist so maybe I should predict. But I would say that my 
bet is that the harms we are going to see as these more 
powerful systems come online, even with ChatGPT, are no 
different from the harms we are seeing right now. If we focus 
our efforts and our energies on governance and regulation and 
guardrails to address the harms we are seeing right now, they 
will be able to adjust as the technology improves. I am not 
worried that what we have put in place today will be out of 
date or out of sync with the new tech. The new tech is like the 
old tech, just supercharged.
    Chairman Peters. Thank you. Dr. Matheny, you will have the 
last word.
    Mr. Matheny. As is typically my last words, I do not know, 
and I think it is a really hard question. I think whether or 
not artificial general intelligence proves to be nearer than 
thought or farther than thought, I think there are things that 
we can do today that are important in either case, including 
regulatory frameworks that include standards with third-party 
tests or audits. The governance of our hardware supply chain so 
that we understand where large amounts of computing is going, 
and we prevent large amounts of computing from going to places 
that do not have the same ethical standards that we and other 
democracies have. Increasing the overall level of awareness and 
capability within the policy community, as you are doing today.
    Chairman Peters. Great. Thank you.
    I would like to thank our witnesses for joining us today, 
and certainly I am grateful for your contributions to this very 
important discussion. As you heard at the outset, this is not 
the end. We are going to have more hearings on this and 
continue to dig deeper into the subject matter and look forward 
to working with you on that.
    We know that today, as has been pretty clearly outlined, 
that AI systems can write like humans, they can assess business 
outlooks for companies, and they can even, hopefully, help us 
cure cancer at some point in the future.
    As we have heard, however, these new developments certainly 
bring new risks, and without responsible designs, the use of AI 
can be devastating and discriminatory. Biased AI systems can 
unfairly deny people job opportunities and open users to legal 
liability. AI can supercharge the privacy risks posed by 
biometric data collection.
    We also have heard that advancements in AI pose new 
challenges for our global competitiveness and national 
security. China is challenging the United States for leadership 
in AI innovation, and both China and Russia are developing 
military applications for AI as well. AI developments can 
create entirely new types of cyber and biological threats, and 
we must prepare for this new AI--enhanced world.
    As we have heard today, recent advancements in computing 
research and data collection and processing power means that 
now is the moment to act on artificial intelligence.
    As Chairman of the Committee I am going to work to ensure 
the United States continues to lead on AI, and we can be 
leaders in both AI research and production and in responsible 
AI design. They are not mutually exclusive. We can do all of 
the above, and we must. Your testimony here today will help 
inform the Committee's future legislative activities and 
oversight actions on that issue, and we look forward to being 
continually engaged with each and every one of you.
    The record for this hearing will remain open for 15 days, 
until 5 p.m. on March 23, 2023, for the submission of 
statements and questions for the record.
    This hearing is now adjourned.
    [Whereupon, at 11:30 a.m., the hearing was adjourned.]

                            A P P E N D I X

                              ----------                              

 [GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]