[Senate Hearing 118-600]
[From the U.S. Government Publishing Office]
S. Hrg. 118-600
THE NEED FOR TRANSPARENCY
IN ARTIFICIAL INTELLIGENCE
=======================================================================
HEARING
BEFORE THE
SUBCOMMITTEE ON CONSUMER PROTECTION,
PRODUCT SAFETY, AND DATA SECURITY
OF THE
COMMITTEE ON COMMERCE,
SCIENCE, AND TRANSPORTATION
UNITED STATES SENATE
ONE HUNDRED EIGHTEENTH CONGRESS
FIRST SESSION
__________
SEPTEMBER 12, 2023
__________
Printed for the use of the Committee on Commerce, Science, and
Transportation
[GRAPHIC NOT AVAILABLE IN TIFF FORMAT]
Available online: http://www.govinfo.gov
__________
U.S. GOVERNMENT PUBLISHING OFFICE
59-704 PDF WASHINGTON : 2025
-----------------------------------------------------------------------------------
SENATE COMMITTEE ON COMMERCE, SCIENCE, AND TRANSPORTATION
ONE HUNDRED EIGHTEENTH CONGRESS
FIRST SESSION
MARIA CANTWELL, Washington, Chair
AMY KLOBUCHAR, Minnesota TED CRUZ, Texas, Ranking
BRIAN SCHATZ, Hawaii JOHN THUNE, South Dakota
EDWARD MARKEY, Massachusetts ROGER WICKER, Mississippi
GARY PETERS, Michigan DEB FISCHER, Nebraska
TAMMY BALDWIN, Wisconsin JERRY MORAN, Kansas
TAMMY DUCKWORTH, Illinois DAN SULLIVAN, Alaska
JON TESTER, Montana MARSHA BLACKBURN, Tennessee
KYRSTEN SINEMA, Arizona TODD YOUNG, Indiana
JACKY ROSEN, Nevada TED BUDD, North Carolina
BEN RAY LUJAN, New Mexico ERIC SCHMITT, Missouri
JOHN HICKENLOOPER, Colorado J. D. VANCE, Ohio
RAPHAEL WARNOCK, Georgia SHELLEY MOORE CAPITO, West
PETER WELCH, Vermont Virginia
CYNTHIA LUMMIS, Wyoming
Lila Harper Helms, Staff Director
Melissa Porter, Deputy Staff Director
Jonathan Hale, General Counsel
Brad Grantz, Republican Staff Director
Nicole Christus, Republican Deputy Staff Director
Liam McKenna, General Counsel
------
SUBCOMMITTEE ON CONSUMER PROTECTION, PRODUCT SAFETY,
AND DATA SECURITY
JOHN HICKENLOOPER, Colorado, Chair MARSHA BLACKBURN, Tennessee,
AMY KLOBUCHAR, Minnesota Ranking
BRIAN SCHATZ, Hawaii DEB FISCHER, Nebraska
EDWARD MARKEY, Massachusetts JERRY MORAN, Kansas
TAMMY BALDWIN, Wisconsin DAN SULLIVAN, Alaska
TAMMY DUCKWORTH, Illinois TODD YOUNG, Indiana
BEN RAY LUJAN, New Mexico TED BUDD, North Carolina
PETER WELCH, Vermont CYNTHIA LUMMIS, Wyoming
C O N T E N T S
----------
Page
Hearing held on September 12, 2023............................... 1
Statement of Senator Hickenlooper................................ 1
Statement of Senator Blackburn................................... 3
Statement of Senator Cantwell.................................... 4
Statement of Senator Moran....................................... 43
Statement of Senator Klobuchar................................... 45
Statement of Senator Young....................................... 47
Statement of Senator Lujan....................................... 61
Witnesses
Victoria Espinel, Chief Executive Officer, BSA | The Software
Alliance....................................................... 5
Prepared statement........................................... 7
Dr. Ramayya Krishnan, W. W. Cooper and Ruth F. Cooper Professor
of Management Science and Information Systems; Dean, Heinz
College of Information Systems and Public Policy; Founding
Faculty Director, The Block Center for Technology and Society,
Carnegie Mellon University..................................... 17
Prepared statement........................................... 19
Sam Gregory, Executive Director, WITNESS......................... 22
Prepared statement........................................... 23
Rob Strayer, Executive Vice President of Policy, Information
Technology Industry Council (ITI).............................. 32
Prepared statement........................................... 33
Appendix
Prepared Statement of Hon. Ted Cruz, U.S. Senator from Texas..... 65
Letter dated September 12, 2023 to Hon. John Hickenlooper and
Hon. Marsha Blackburn from Adam Thierer, Senior Fellow, R
Street Institute............................................... 66
Hodan Omaar, Senior Policy Analyst, Information Technology and
Innovation Foundation (ITIF), prepared statement............... 68
Jennifer Huddleston, Research Fellow, Cato Institute, prepared
statement...................................................... 71
Response to written questions submitted to Victoria Espinel by:
Hon. Brian Schatz............................................ 72
Hon. Ben Ray Lujan........................................... 73
Hon. John Hickenlooper....................................... 73
Hon. Peter Welch............................................. 74
Hon. Ted Cruz................................................ 75
Response to written questions to Dr. Ramayya Krishnan submitted
by:
Hon. Amy Klobuchar........................................... 76
Hon. Brian Schatz............................................ 78
Hon. Ben Ray Lujan........................................... 79
Hon. John Hickenlooper....................................... 79
Hon. Peter Welch............................................. 80
Hon. Ted Cruz................................................ 81
Response to written questions to Sam Gregory submitted by:
Hon. Brian Schatz............................................ 81
Hon. Ben Ray Lujan........................................... 84
Hon. John Hickenlooper....................................... 85
Hon. Peter Welch............................................. 87
Response to written questions to Rob Strayer submitted by:
Hon. Amy Klobuchar........................................... 88
Hon. Brian Schatz............................................ 89
Hon. Ben Ray Lujan........................................... 89
Hon. John Hickenlooper....................................... 90
Hon. Peter Welch............................................. 91
Hon. Ted Cruz................................................ 92
THE NEED FOR TRANSPARENCY
IN ARTIFICIAL INTELLIGENCE
----------
TUESDAY, SEPTEMBER 12, 2023
U.S. Senate,
Subcommittee on Consumer Protection, Product
Safety, and Data Security,
Committee on Commerce, Science, and Transportation,
Washington, D.C.
The Committee met, pursuant to notice, at 2:33 p.m., in
room SR-253, Russell Senate Office Building, Hon. John
Hickenlooper, Chairman of the Subcommittee, presiding.
Present: Senators Hickenlooper [presiding], Cantwell,
Klobuchar, Baldwin, Lujan, Welch, Blackburn, Fischer, Moran,
and Young.
OPENING STATEMENT OF HON. JOHN HICKENLOOPER,
U.S. SENATOR FROM COLORADO
Senator Hickenlooper. Good afternoon. Welcome to this
hearing of the Subcommittee on Consumer Protection, Product
Safety and Data Security will now come to order.
While Artificial Intelligence has been part of our lives
for years and years, its newer forms have now captured, it is
fair to say, the world's attention. We are now far beyond the
era of asking Alexa to play a song, or Siri to dial our spouse.
These are the examples of Narrow-AI.
ChatGPT, the new generative AI systems can now plan a
custom travel itinerary, create artwork, remix a song, and help
you write computer code. So it is obvious that AI is a powerful
technology that will revolutionize our economy.Just like the
first car, or personal computer, AI is a transformative
technology that has just both benefits and risks for consumers.
That means we have to proceed with intention and care.
Our goal today is to identify how we do that. Specifically
we need to begin to help Americans understand AI's capabilities
and limitations, to reduce AI's potential risks, relative to
consumers, and to increase the public's trust in AI systems
through transparency.
The fact that we need to be careful with AI doesn't negate
how important it is, or the massive potential it has to
transform our lives. From helping with your tedious daily
tasks, to helping doctors properly diagnose and find the right
treatments for an illness, the possibilities go beyond what we
can imagine today, far beyond.
But we must also confront the fact that AI can be misused
by bad actors, AI can be used to make scams, fraud, and cyber
attacks more harmful, and more effective. Companies developing
and deploying AI, we believe, have a role to build a safe,
secure, and reliable system that, over time, will earn the
trust of the public.
Congress will play a role by setting reasonable rules of
the road to inform and protect consumers. The federal
government, academia, and private sector will all need to work
together to establish thoughtful AI Policy.
In April, Senator Blackburn and I sent a letter to tech
companies asking how they are adopting the NIST AI Risk
Management Framework. The responses showed how the framework is
helping companies build accountability, transparency, and
fairness into their products.
Today, Senator Thune and I sent a letter to the Office of
Science and Technology Policy to stress the importance of
developing Federal standards to help consumers understand and
identify AI-generated content.
This is going to be more critical for building trust as AI
expands into larger and larger aspects of our lives.Several
other federal AI initiatives are currently underway. To name a
few: The White House has convened leading tech companies,
bringing people together to build a shared understanding and a
voluntary commitment to build trustworthy AI systems.
NIST formed a group, a Public Generative AI Working Group
to build on its AI Risk Management Framework. Also the National
AI Initiative Office is coordinating a whole-of-government
effort to develop AI safety and transparency with guidelines--
with input from experts in civil society, in academia, and the
private sector.
And we are fortunate to have two N-A-I-A-C, or NAIAC, two
NAIAC members as witnesses here today. These are all
encouraging steps, but it doesn't mean we have done--that we
are done when it comes to making sure we have created a
framework in which AI will be safe and transparent for
consumers.
The AI-powered future comes with many challenges that we
can already see, building a talented STEM-trained workforce,
providing efficient computing power, ensuring that we protect
consumer data privacy.
We know that AI trains on publicly available data, and this
data can be collected from everyday consumers everywhere, in
all parts of their lives. There are too many open questions
about what rights people have to their own data and how it is
used, which is why Congress needs to pass comprehensive data
privacy protections. This will empower consumers, creators, and
help us grow our modern AI-enabled economy.
This issue is complicated and it is going to require
bipartisanship to acquire results. Committees across Congress
are examining AI's impact on society through different lenses,
each hearing is an invitation for policymakers and families at
dinner tables across America to think about how AI will impact
their everyday lives.
Today's discussion is that next step as we work towards
building what ultimately will become, hopefully, necessarily a
global consensus. This committee is well positioned to examine
all of these important issues with the goal of promoting
transparency, and the goal of creating an AI system that
consumers will have confidence in.
I would like to welcome each of our witnesses who are
joining us today: Ms. Victoria Espinel, CEO of Business
Software Alliance, BSA; Dr. Ramayya Krishnan, Dean of the
College of Information Systems, Carnegie Mellon University; Mr.
Sam Gregory, Executive Director, WITNESS, I guess I should say,
W-I-T-N-E-S-S; and Mr. Rob Strayer, Executive Vice President
for Policy, Information Technology Industry Council, ITI.
I would now like to recognize Ranking Member Blackburn for
her opening remarks.
STATEMENT OF HON. MARSHA BLACKBURN,
U.S. SENATOR FROM TENNESSEE
Senator Blackburn. And thank you, Mr. Chairman. I certainly
appreciate that we are having this hearing. This is kind of AI
Week on the Hill, and we have a Judiciary Committee hearing
going on this afternoon. And of course, we have our Member
Forum, that is going to take place tomorrow and Thursday. So we
are pleased to be putting attention on this.
You know, AI as you said, has been around for years,
whether it is autocorrect, or auto fill, or voice assist, or
facial recognition, things that people have become accustomed
to using. But with ChatGPT in November, it is like people said:
Wait a minute, what is this? And of course, generative AI is
something that people have turned their attention to and
saying: How is this happening? How is it taking place?
And Tennessee, my state, is really quite a leader in this
field of AI, we have several automobile companies that are
investing and innovating with AVs, we have farmers that are
really leading the way in smart agriculture, and it is so
interesting to hear some of their concepts. We have got thought
leaders, people at the University of Tennessee, and also Oak
Ridge National Lab, who are pushing boundaries on AI every
single day.
With innovators like these, the future of AI can sometimes
border on the unimaginable, especially as the technology
continues advancing at a pace that is more rapid than any other
in our recent memory.
This swift rate of advancement, however, has caused many
concerns. Many of the discussions that I have heard around AI
has focused on the doomsday scenarios. While it is important to
prevent catastrophic events, we must also not lose sight of the
transformational benefits of AI.
For instance, in addition to the examples that I have
previously mentioned, AI has profound implications for the
financial and health care industries, two industries that are
critical to our state. That is why any regulatory action from
Congress, or from Federal agencies, must balance safety with
the preservation of an innovative economy.
Our adversaries, like China, are not slowing down on AI,
and we cannot give them any advantage in the deployment of this
emerging technology. In fact, earlier this year, Senator Ossoff
and I convened a hearing on the Judiciary Subcommittee on Human
Rights and the Law, where our witnesses discussed the Chinese
Communist Party's interest in rolling out AI systems to enhance
the regime's ability to surveil their citizens, which brings us
here today.
This afternoon, we will explore ways to mitigate consumer
harm, promote trust, and transparency, and identify potential
risk stemming from AI technologies. And I am looking forward to
hearing from each of you on these.
But before we turn to our witnesses, I wanted to remind my
fellow lawmakers that amidst the hyper focus on AI, we must not
forget about other issues that are as critical to U.S.
technological advancement and global leadership.
First, for a decade I have worked to bring about a
comprehensive data privacy law. As you mentioned, that is
something that should be a first step. And I know Madam
Chairman is well aware, and joins me in wanting to see a
Federal standard. And it is vital that my colleagues keep in
mind the need for that Federal privacy standard, as we look at
AI. In our judiciary hearings that we have had on AI, everybody
mentioned the need to have that so people can protect their
data.
It is virtually impossible to talk about these new and
advanced systems without a real discussion about how online
consumers will be able to protect what I term, ``their virtual
you'', which is their presence online.
Second, as AI systems require more and more computing
power, the need for high performance in quantum computing will
become vital. This is why I have already introduced two
bipartisan bills on this topic, and I encourage this committee
to move on reauthorizing the National Quantum Initiative Act.
We need to do that this year.
So thank you to each of you, for the leadership you have on
the issue for being here as our witnesses, and thank you for
the hearing.
Senator Hickenlooper. Thank you, Senator Blackburn.
I am now going to turn it over to the Chair of the Commerce
Committee, who has a long, a long history, probably more direct
history with AI than any other senator. Senator Cantwell from
Washington.
STATEMENT OF HON. MARIA CANTWELL,
U.S. SENATOR FROM WASHINGTON
The Chair. Thank you, Mr. Chairman. And thank you to
yourself; and to Senator Blackburn at the Subcommittee level
for holding this important hearing. I think we are
demonstrating that just as AI needs to be open and transparent,
we are going to have an open and transparent process as we
consider legislation in this area.
And I want to thank Senator Blackburn for her comments
about privacy because I do think these things go hand in hand,
having good, strong privacy protections certainly prevents the
kind of abuse or misuse of information that could cause
substantial harm to individuals.
And I thank the witnesses for being here today to help us
in this discussion.
I recently was informed about a situation in my state that
I found quite alarming. A family in Pierce County, Washington,
received a phone call. A scammer used AI to spoof the voice of
their daughter, telling them that she had been in a car
accident and that a man was threatening to harm her if they
didn't wire $10,000. So I can't imagine what this deepfake
meant to that family or the concerns that they have.
And a recent deepfake image claimed a bombing occurred at
the Pentagon, and that fake image sparked a dip in the stock
market.
DARPA is leading the way on important developments to
approach detecting AI-generated media. And I plan to introduce
legislation in this area.
I think that AI, as was discussed by my two colleagues, has
amazing potential. I held an AI Summit in my state and saw some
of those amazing technologies already being pushed by the Allen
Brain Institute and some of their early technologies; certainly
helping in things like climate, and farming, and detecting
illegal activities, and helping us move forward in important
areas of research.
So we know that we have choices here. We know we want to
continue to empower consumers and make sure that we are
stopping the fraudsters. And we want to make sure that any
misuse of AI that we are stopping that, whatever we can do to
make sure that we are protecting American's privacies.
So I hope that today's hearing will give us some ideas
about how to drive innovation and maintain U.S. leadership in
this very important security-related technology, and the issues
of global competitiveness, that we talk and discover ideas,
about deepfakes and potential national security issues, the
framework for legislation, protect online privacy, and combat
discrimination.
I know that we need to grow education, in general, in our
workforce. And the information age has already put great
transformations in place. The jobs of tomorrow are here today,
but the skill levels for people to do them are not.
We know that we need to invest more from the CHIPS and
Science Act in skilling a workforce for tomorrow. That was
before AI. With AI, there is an accelerant on that. And that is
why I believe we need something as grand as the GI Bill was
after World War II in empowering Americans for new
opportunities in this area.
So I look forward to hearing the comments from our
witnesses.
And thank you again Mr. Chairman, for holding this very
important hearing about the potential and challenges facing us.
But clearly, we need an open and transparent system, just as we
did for the internet, so that innovation can flourish. Thank
you.
Senator Hickenlooper. Thank you, Madam Chair. I appreciate
your being here and helping create this hearing. Now we will go
and we will have the opening statements from each of the
witnesses.
Ms. Espinel.
STATEMENT OF VICTORIA ESPINEL, CHIEF EXECUTIVE OFFICER, BSA |
THE SOFTWARE ALLIANCE
Ms. Espinel. Good afternoon, Chair Hickenlooper, Ranking
Member Blackburn, Chair Cantwell, and Members of the
Subcommittee.
My name is Victoria Espinel, and I am the CEO of BSA | The
Software Alliance.
BSA is the advocate for the global enterprise software
industry. BSA members are at the forefront of developing
cutting-edge services including AI, and their products are used
by businesses across every sector of the economy. I commend the
Subcommittee for convening today's hearing, and I thank you for
the opportunity to testify.
There are two things that need to be done. Companies that
develop and use AI must act responsibly to identify and address
risks, and Congress needs to establish thoughtful, effective
rules that protect consumers, and promote responsible
innovation.
AI has real world benefits. Think about extreme weather
events, hurricanes, wildfires, tornadoes, that have affected
many states this year, as we know, there are families wondering
whether the eye of a hurricane will hit their hometown, and
whether they will be safe if they stay, or if they should pack
up and go. How will they know whether they should leave? And if
they do, which nearby destination is the safest to ride out the
storm?
AI is helping to provide these answers. With AI weather
forecasters are better able to predict extreme weather events,
helping people prepare before disaster strikes. And what
happens to those families who are in the storm's path? How do
they get food in the aftermath of a storm? How do rescue
workers know that they need help?
AI is helping relief workers anticipate where medical
equipment, food, water, and supplies are most needed in
response to natural disasters.
In the face of extreme danger, AI's predictions can save
lives. More needs to be done, however, so that we can see
greater benefits. And with thoughtful rules in place,
innovation in AI will continue to advance, and the responsible
use of artificial intelligence will serve society.
There has been a wave of attention on AI since ChatGPT
launched publicly nine months ago. But this committee began
studying the issue in a thoughtful manner years earlier. Nearly
six years ago, I testified here about the building blocks of
machine learning and artificial intelligence. Chair Cantwell
and Senator Young introduced one of the first AI bills in 2017.
We also appreciate the committee's recent work to establish the
National Artificial Intelligence Initiative, and your request
to BSA about how our member companies are using the NIST, AI
Risk Management Framework to responsibly develop and use AI.
The pace of AI development and use has increased
significantly since 2017. As with any new technology there are
legitimate concerns that need to be addressed, including the
risk of bias and discrimination.
This committee is well placed to move legislation that sets
rules around AI. The U.S. economy will benefit from responsible
and broad-based AI adoption. An important part of facilitating
that adoption is passing a strong national law. The countries
that best support responsible AI innovation will see the
greatest benefits of economic and job growth in the coming
years.
Moreover, other countries are moving quickly on regulations
that affect U.S. companies. The U.S. should be part of shaping
the global approach to responsible AI. The window for the U.S.
to lead those conversations, globally, is rapidly closing.
This is what we think legislation should do: It should
focus on high risk uses of AI, like those that decide whether a
person can get a job, a home, health care. It should require
companies to have risk management programs. It should require
companies to conduct impact assessments, and it should require
companies to publicly certify that they have met those
requirements.
This will include some concrete steps. I have set these out
in more detail in my written testimony, and I hope we have a
chance to discuss those.
It is important that legislation reflect different roles.
Some companies develop AI, some companies use AI, our companies
do both, and both roles have to be covered. Legislation should
set distinct obligations for developers and users because each
will know different things about the AI system in question, and
each will be able to take different actions to identify and
mitigate risks.
So my message to Congress is simple. Do not wait. AI
legislation can build on work by governmental organizations,
industry, and civil society. These steps provide a collective
basis for action. You can develop and pass AI legislation now
that creates meaningful rules to reduce risks and promote
innovation. We are ready to help you do so.
Thank you for the opportunity to testify. And I look
forward to your questions.
[The prepared statement of Ms. Espinel follows:]
Prepared Statement of Victoria Espinel, Chief Executive Officer,
BSA | The Software Alliance
Good afternoon Chairman Hickenlooper, Ranking Member Blackburn, and
members of the Subcommittee. My name is Victoria Espinel, and I am the
CEO of BSA | The Software Alliance.\1\
---------------------------------------------------------------------------
\1\ I am a member of the National AI Advisory Committee, but I am
testifying in my capacity as CEO of BSA.
---------------------------------------------------------------------------
BSA is the leading advocate for the global enterprise software
industry.\2\ Our members are at the forefront of developing cutting-
edge services--including AI--and their products are used by businesses
across every sector of the economy. I commend the Subcommittee for
convening today's hearing and thank you for the opportunity to testify.
I also appreciate this Committee's longstanding focus on AI, including
your efforts to establish the National Artificial Intelligence
Initiative and your outreach to BSA to learn about how our companies
are implementing the AI Risk Management Framework released earlier this
year by the National Institute of Standards and Technology.
---------------------------------------------------------------------------
\2\ BSA's members include: Adobe, Alteryx, Atlassian, Autodesk,
Bentley Systems, Box, Cisco, CNC/Mastercam, Databricks, DocuSign,
Dropbox, Elastic, Graphisoft, IBM, Informatica, Juniper Networks,
Kyndryl, MathWorks, Microsoft, Okta, Oracle, Palo Alto Networks,
Prokon, PTC, Rubrik, Salesforce, SAP, ServiceNow, Shopify Inc., Siemens
Industry Software Inc., Splunk, Trend Micro, Trimble Solutions
Corporation, TriNet, Twilio, Unity Technologies, Inc., Workday,
Zendesk, and Zoom Video Communications, Inc.
---------------------------------------------------------------------------
Nearly six years ago, I testified before this Committee at a
hearing focused on the building blocks of machine learning and
artificial intelligence.\3\ Chair Cantwell and Senator Young also
introduced one of the first AI bills that year. Since then, the
building blocks we discussed in 2017 have come together at a rapid
pace. Traditional benchmarks for measuring how AI performs tasks like
recognizing and classifying images or understanding text are becoming
obsolete, as researchers launch new methods to measure progress.\4\
---------------------------------------------------------------------------
\3\ Testimony of Victoria Espinel, Hearing on Digital Decision-
Making: The Building Blocks of Machine Learning and Artificial
Intelligence, Before the Senate Committee on Commerce, Science, &
Transportation Subcommittee on Communications, Technology, Innovation,
and the Internet (Dec. 12, 2017), available at https://www.bsa.org/
files/policy-filings/12122017BSAAI
Testimony.pdf.
\4\ AI Index Steering Committee at the Institute for Human-Centered
AI at Stanford University, The AI Index 2023 Annual Report (April
2023), page 114, available at https://aiindex
.stanford.edu/wp-content/uploads/2023/04/HAI_AI-Index-Report_2023.pdf
(AI Index 2023).
---------------------------------------------------------------------------
As I said then, AI is a foundational technology that drives
products and services that people use every day. It also raises
important policy issues, which are core to our work at BSA. We
undertook a year-long project to work with member companies to develop
the BSA Framework to Build Trust in AI,\5\ which was released in 2021
and is designed to help organizations mitigate the potential for
unintended bias in AI systems. Built on a vast body of research, the
BSA Framework sets out a lifecycle-based approach for performing impact
assessments to identify risks and highlights best practices for
mitigating those risks.
---------------------------------------------------------------------------
\5\ See Confronting Bias: BSA's Framework to Build Trust in AI
(June 2021), available at https://www.bsa.org/reports/confronting-bias-
bsas-framework-to-build-trust-in-ai. BSA has testified before the
United States Congress and the European Parliament on the Framework.
---------------------------------------------------------------------------
Best practice documents, like BSA's Framework, have moved the
policy debate forward, but I am here today to say that they are not
enough, and legislation is needed. Thoughtful AI legislation will
benefit the U.S. economy by creating new guardrails that build trust in
the use of AI technologies. It will protect consumers by ensuring AI
developers and deployers take required steps to mitigate risks; and it
will set the U.S. as a leader in the global debate about the right way
to regulate AI. Fortunately, Congress can leverage tools that already
exist to create legislation that requires companies to identify and
mitigate risks associated with high-risk AI systems.
I. Congress Should Act Now to Adopt Meaningful AI Legislation
Congress should not wait to enact legislation that creates new
obligations for companies that develop and use AI in high-risk ways.
Legislation will not only protect consumers from real risks of
harm, but will also create trust in AI technologies that will benefit
the economy broadly. Consumers and businesses already rely on a wide
range of AI-powered services, but they will only continue to adopt new
AI technologies if they trust that those products and services are
developed and deployed responsibly. Because companies of all sizes and
in all industry sectors can benefit from AI, thoughtful AI legislation
is important to promoting the United States economy. Countries that
support the broad adoption of AI will see the greatest growth in jobs
and economic prosperity.
To enact legislation, Congress should take advantage of the
considerable work that governmental organizations, civil society
advocates, and industry groups have put into identifying the risks of
using AI in different contexts and the concrete steps organizations can
take to mitigate those risks. Although these proposals have important
differences, they collectively form a basis for action. For example,
there are fundamental objectives on which everyone should agree: AI, in
any form, should not be used to commit illegal acts. It should not be
used to compromise privacy, facilitate cyberattacks, exacerbate
discrimination, or create physical harm. AI that is developed and
deployed responsibly, that improves our lives and makes us safer,
should flourish. But Congress should go farther, to create new rules
for companies that develop and deploy high-risk AI systems.
By enacting legislation, Congress will also ensure the United
States is not just a leader in developing AI technology but is a
leading voice in the global debate about regulating AI. The window to
lead conversations about AI regulation is rapidly closing, as other
governments are moving to shape the rules that will govern AI's future.
By the end of this year, the European Union is expected to finalize its
AI Act, which is on pace to be the most comprehensive AI law enacted.
Japan is leading a G7 effort to establish common standards for AI
governance. In November, the United Kingdom will host a global AI
summit, focused on the safe and responsible deployment of AI. These
governments are not alone. Australia, Brazil, Canada, China, Korea,
Singapore, and Thailand are among the long list of countries that are
examining the right policy approaches for addressing AI risks.\6\
---------------------------------------------------------------------------
\6\ See Australia: BSA Comments on Supporting Safe and Responsible
AI Innovation in Australia (July 26, 2023), available at https://
www.bsa.org/files/policy-filings/07262023safeai.pdf; Brazil: BSA
Recommendations to the Jurists Committee on AI Regulation in Brazil
(May 30, 2022), available at https://www.bsa.org/files/policy-
filingsen05302022airegbrazil.pdf; Korea: BSA Comments on Bill on
Fostering AI Industry and Securing Trust (Feb. 13, 2023), available at
https://www.bsa.org/files/policy-filings/en02132023kraitrust.pdf;
Singapore: BSA Comments on Public Consultation on Proposed Advisory
Guidelines on Use of Personal Data in AI Recommendation and Decision
Systems (Aug. 30, 2023), available at https://www.bsa.org/files/policy-
filings/08302023bsaaiguidelines.pdf; Thailand: BSA Comments on Draft
Bill on Promotion and Support of National AI Innovation (Aug. 18,
2023), available at https://www.bsa.org/files/policy-filings/
08182023bsanatlai.pdf.
---------------------------------------------------------------------------
The United States has recognized the enormous benefits of working
with other countries on AI governance issues, and is already
participating in a range of global efforts. We support the United
States' active involvement in these fora, including the US-EU Trade and
Technology Council (TTC), the Global Partnership on AI (GPAI), the
Organisation for Economic Co-Operation and Development (OECD), and the
G7 Hiroshima AI project. But the U.S. voice in those efforts will be
stronger if the United States adopts national AI legislation that
creates clear guardrails for how companies develop and deploy high-risk
AI systems.
My message to Congress is: Do not wait. You can adopt AI
legislation now that creates meaningful rules to reduce risks and
promote innovation. We urge you to do so.
II. The U.S. Economy Will Benefit from Legislation that Builds Trust in
AI
Businesses of all sizes across all industries are looking for ways
to adopt AI tools to grow. Indeed, consumers and businesses today
already rely on a wide range of services powered by enterprise AI--and
the AI our companies create is all around us. While these AI systems
may not get the most attention, they are increasingly integrated in the
economy and everyday life. For example:
Consumers. As consumers, AI-powered services remind us when
we forget to attach a document to an e-mail, or prompt us to
open a particular document so that we can pick up where we left
off. We expect to auto-complete forms, chat with virtual
assistants that are available even when customer service
representatives are not, and to use AI-powered apps to identify
and reach personal financial goals, or to predict how a sports
player will perform in an upcoming game.
Businesses. Businesses across every industry sector are
integrating AI into their products and services, including
retail stores that use AI to predict when demand for a product
will surge so that they can keep shelves stocked, warehouses
that rely on AI-powered logistics planning to minimize supply
chain shortages, and manufacturers that use AI to detect safety
concerns on factory floors. Across industry sectors, businesses
of all sizes also use AI to identify and manage common
documents, detect fraudulent transactions, and guard against
cybersecurity threats.\7\
---------------------------------------------------------------------------
\7\ For more examples of everyday uses of AI, see BSA, Everyday AI
for Consumers, available at https://www.bsa.org/files/policy-filings/
08012023aiconsumers.pdf, and BSA, Everyday AI for Businesses, available
at https://www.bsa.org/files/policy-filings/08012023aibusiness.pdf.
The economic benefits of AI are not limited to one industry sector
or one business model. Instead, the promise that AI may one day impact
every industry is quickly turning into a commercial reality and driving
digital transformation across sectors. Airlines now use AI systems to
more efficiently clean planes between flights; farmers use AI to
analyze large amounts of weather information to maximize their harvest;
manufacturers use AI to test new prototypes, and construction companies
build AI-generated ``digital twins'' of real-life cities to understand
the impacts of a proposed design.
Enterprise software companies are at the leading edge of this
transformation, creating the products and services relied on by other
companies.\8\ I have included an extensive list of examples from BSA's
members in an annex to this testimony, but want to highlight a handful
of ways that companies in all industries are using enterprise-powered
AI:
---------------------------------------------------------------------------
\8\ BSA's policy and educational resources on AI are available at
https://ai.bsa.org/. For additional information about AI's adoption
across industry sectors see BSA, Artificial Intelligence in Every
Sector, available at https://www.bsa.org/files/policy-filings/
06132022bsaaieverysector
.pdf.
In healthcare, a large pharmacy chain uses an AI-powered
platform to forecast demand and redistribute medications across
thousands of store locations and to deliver near real-time
insights and recommendations for pharmacists to provide more
personalized advice to patients. This helps managers understand
the supply chain, store labor and productivity, patient vaccine
---------------------------------------------------------------------------
scheduling, and prescription pickup processes.
In manufacturing, a car maker used AI-based generative
technology to redesign a seat bracket, which secures seat belt
fasteners to seats and seats to floors, that is 40 percent
lighter and 20 percent stronger than the previous iteration.
Changes like these can help reduce the amount of material
needed to build a car and make vehicles more fuel efficient.
In agriculture, the research division of an enterprise
software provider partnered with a climate risk company to
develop AI models capable of providing more accurate long-range
weather predictions. Traditional weather forecasting methods
can provide accurate predictions for a seven-day window. By
leveraging AI, the researchers are developing new forecasting
models to provide accurate predictions of weather trends two-to
six-weeks out from a given date. By providing reliable extended
forecasts, these tools will help water managers predict
snowpack and water availability for irrigation, hydropower, and
other critical agricultural and environmental uses.
The importance of AI to the U.S. economy is inescapable.
Innovations in electricity and personal computers created investment
booms of as much as 2 percent of U.S. GDP as those technologies were
adopted into the broader economy.\9\ Economic forecasts estimate that
AI could have an even bigger impact on GDP. By 2025, investment in AI
is expected to approach $100 billion in the United States and $200
billion globally.\10\ Generative AI alone could add up to $4.4 trillion
of value to the global economy every year.\11\
---------------------------------------------------------------------------
\9\ Goldman Sachs, AI Investment Forecast to Approach $200 Billion
Globally By 2025 (Aug. 1, 2023), available at https://
www.goldmansachs.com/intelligence/pages/ai-investment-forecast-to-
approach-200-billion-globally-by-2025.html.
\10\ Id.
\11\ McKinsey & Company, The Economic Potential of Generative AI
(June 2023), available at https://www.mckinsey.com/capabilities/
mckinsey-digital/our-insights/the-economic-potential-of
-generative-ai-the-next-productivity-
frontier?gclid=Cj0KCQjwl8anBhCFARIsAKbbpyTeLnz0c6i4X
2UTnmWdO1KGQnE1mUR8ErJSrM0eMnWDxgfaZukt_L0aAqP7EALw_wcB#/.
---------------------------------------------------------------------------
Why does this matter for policymakers?
To realize these economic benefits, consumers and businesses must
trust that AI is developed and deployed responsibly. While the adoption
of AI can unquestionably be a force for good, it can also create real
risks if not developed and deployed responsibly. Setting thoughtful
rules for AI is therefore central to the vitality of our economy.
Industries of all kinds and businesses of all sizes are looking for
ways to use AI to grow, but they will only adopt AI-powered products
and services that they trust. Countries that best facilitate
responsible and broad-based AI adoption will see the greatest economic
and job growth in the coming years.
III. Congress Should Enact Legislation That Builds Trust in AI
Now is the time for Congress to adopt thoughtful legislation that
addresses known risks of AI systems.
There will continue to be a range of significant and evolving
policy debates around AI. But there's no need to wait to pass
legislation that creates meaningful guardrails against the AI risks
that exist today. You can--and should--build on existing regulatory
efforts by setting rules across the economy to address concerns about
the potential for bias when AI systems are used in high-risk ways.
A. Legislation Should Create New Obligations for High-Risk Uses of AI
AI legislation can build on work that's already been done.
To anticipate and address potential harms, developers of AI used in
high-risk situations should understand the practical implications of
its use, including through the use of impact assessments. Those who
train AI systems for high-risk uses, or are deploying AI systems in
high-risk contexts, should also be required to understand how the tools
they create or deploy might result in unintended outcomes and act to
mitigate those risks.
We urge Congress to adopt legislation that creates meaningful
guardrails for high-risk uses of AI. That legislation should require
companies to:
(1) establish risk management programs to identify and mitigate
risks across AI systems;
(2) conduct annual impact assessments for high-risk uses of AI, and
(3) publicly certify that they have met these requirements.
Congress can build on tools that already exist today--and require
companies to adopt those tools to identify and mitigate risks
associated with high-risk AI systems. My testimony focuses on two of
those tools: risk management programs and impact assessments.
B. Risk Management Programs
The goal of risk management is establishing repeatable processes
for identifying and mitigating potential risks that can arise
throughout the lifecycle of an AI system. Risk management is
particularly important in contexts like AI, privacy, and cybersecurity,
where the combination of quickly evolving technologies and highly
dynamic threat landscapes can render traditional approaches
ineffective. Rather than evaluating a product or service against a
static set of requirements that can rapidly become outdated, risk
management programs integrate compliance responsibilities into the
development process to help identify and mitigate risks throughout a
product or service's lifecycle.
Risk management programs have two key elements: (1) a governance
framework to support the organizations' risk management functions, and
(2) a scalable process for performing impact assessments that identify
and mitigate risks of an AI system. The governance framework is
critical because it promotes collaboration between an organization's
development team and its compliance team at key points in the design,
development, and deployment of an AI system.
One way for companies to establish risk management programs is by
using the AI Risk Management Framework (AI RMF), which was released
earlier this year by the National Institute of Standards and Technology
(NIST).\12\ Creating the AI RMF was a significant achievement and
builds on NIST's work creating frameworks for managing cybersecurity
and privacy risks. For example, the NIST Cybersecurity Framework (CSF)
is widely used by private and public-sector organizations worldwide;
since 2017, it has been mandatory for Federal agencies to use the CSF
to improve their cybersecurity risk management programs.\13\ Like the
CSF, the AI RMF is a voluntary tool. It helps organizations incorporate
trustworthiness considerations into the design, development, use, and
evaluation of AI products.
---------------------------------------------------------------------------
\12\ NIST AI Risk Management Framework, available at https://
nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf.
\13\ See NIST, Cybersecurity Framework, Questions and Answers,
(discussing Federal agency use of the NIST CSF), available at https://
www.nist.gov/cyberframework/frequently-asked-
questions/framework-basics#agency.
---------------------------------------------------------------------------
The AI RMF can help companies establish risk management programs
that guard against a range of potential AI-related harms. It focuses on
corporate practices around four functions: govern, map, measure, and
manage. The AI RMF is designed to be usable by large and small
organizations alike and can be applied in varied circumstances, such as
for procurement purposes, for organizations with existing governance
programs, and for organizations just beginning to think about risk
management. It also identifies several indicia of trustworthy AI, which
include privacy, explainability, and fairness, and incorporates an
assessment of these characteristics as part of the measure function.
Importantly, the AI RMF acknowledges that tradeoffs among
trustworthiness characteristics may exist. For example, the use of
privacy-enhancing technologies may decrease accuracy, which could
affect decisions about fairness in certain domains, and there may also
be tradeoffs between interpretability and privacy.
The AI RMF encourages:
Consultation with diverse stakeholders;
Establishing processes to identify, assess, and mitigate
risks;
Defining individual roles and responsibilities to people
throughout an organization;
Identifying metrics for evaluation;
Evaluating fairness and bias;
Maintaining post-deployment feedback mechanisms; and
Establishing incident response plans.
Ultimately, effective AI risk management programs should be
underpinned by a governance framework that establishes the policies,
processes, and personnel that will be used to identify, mitigate, and
document risks throughout the system's lifecycle. The purpose of such a
governance framework is to promote understanding across organizational
units--including product development, compliance, marketing, sales, and
senior management--about each entity's role and responsibilities for
promoting effective risk management during the design, development, and
deployment of AI systems.
C. Impact Assessments
Impact assessments have three purposes: (1) identifying potential
risks that an AI system may pose, (2) quantifying the degree of
potential harms the system could generate, and (3) documenting steps
taken to mitigate those risks.\14\ As noted earlier, performing impact
assessments is a key part of creating a meaningful risk management
program.
---------------------------------------------------------------------------
\14\ See BSA, Impact Assessments: A Key Part of AI Accountability,
available at https://www.bsa.org/files/policy-filings/
08012023impactassess.pdf.
---------------------------------------------------------------------------
Impact assessments are already widely used in a range of other
fields, including data protection, as an accountability mechanism that
demonstrates a system has been designed in a manner that accounts for
the potential risks it may pose to the public. Because impact
assessments already exist today, they can be readily adapted to help
companies identify and mitigate AI-related risks. For example, three
state privacy laws already require companies to conduct impact
assessment for specific activities, including processing sensitive
personal data, engaging in targeted advertising, or selling personal
data; seven more state privacy laws will soon do so.\15\ Globally,
privacy and data protection laws worldwide use impact assessments as a
tool for improving accountability.\16\
---------------------------------------------------------------------------
\15\ Colorado, Connecticut, and Virginia already impose these
requirements. See Colorado Privacy Act, Colo. Rev. Stat. Tit. 6, Art.
1, Pt. 13 Sec. Sec. 6-1-1301-6-1-1313; Connecticut Data Privacy Act
Conn. Gen. Stat. Tit. 42, Ch. 743jj, Sec. 42-515-525; Virginia Consumer
Data Protection Act; Va. Code Tit. 59.1, Ch. 53, Sec. 59.1-575-585.
Recently passed state privacy laws in Florida, Indiana, Montana,
Oregon, Tennessee, and Texas will also require impact assessments for
certain activities.
\16\ Data protection impact assessments are an established
accountability tool under privacy laws worldwide. Under the European
Union's General Data Protection Regulation (GDPR), impact assessments
are required for activities likely to result in a ``high risk'' to the
rights and freedoms of individuals. Brazil's General Data Protection
Law (LGPD) also allows the national data protection agency to require a
company to prepare a data protection assessment, and in Singapore
organizations are required to conduct impact assessments if they
process data for certain purposes. There is also significant regulatory
guidance on how organizations can conduct impact assessments to
identify and mitigate privacy risks. See, e.g., Office of the
Australian Information Commissioner, Guide to Undertaking Privacy
Impact Assessments (Sept. 2, 2021), available at https://
www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-
government-agencies/privacy-impact-assessments/guide-to-undertaking-
privacy-impact-assessments; European Data Protection Board, Guidelines
on Data Protection Impact Assessment (Oct. 13, 2017), available at
https://ec.europa.eu/newsroom/article29/items/611236; Singapore
Personal Data Protection Commission, Guide to Data Protection Impact
Assessments, (Sept. 14, 2021) available at https://www.pdpc.gov.sg/-/
media/Files/PDPC/PDF-Files/Other-Guides/DPIA/
Guide-to-Data-Protection-Impact-Assessments-14-Sep-2021.pdf.
---------------------------------------------------------------------------
The use of impact assessments can be readily expanded to require
assessments for high-risk AI systems. Conducting impact assessments can
help companies identify and mitigate risks, including risks of
unintended bias. In our view, when AI is used in ways that could
adversely impact civil rights or access to important life
opportunities, the public should be assured that such systems have been
thoroughly vetted and will be continuously monitored to account for the
risks associated with unintended bias. Requiring impact assessments for
companies that develop and deploy these high-risk systems is an
important way to do that.
D. Legislation Should Focus on High-Risk Uses and Create Role-
Appropriate
Requirements.
Legislation that requires risk management programs and impact
assessment will create new safeguards for high-risk AI systems. Any
legislation incorporating these requirements should:
Focus on high-risk AI uses. Legislation should place
guardrails around high-risk uses of AI, namely AI systems that
determine an individual's eligibility for housing, employment,
credit, education, access to physical places of public
accommodation, healthcare, or insurance. These systems have the
potential to affect important life opportunities--and are a key
area for policymakers to address. In contrast, many everyday
uses of AI present few risks to individuals and create
significant benefits, like helping organize digital files,
auto-populate common forms for later human review, or predict
the fonts to use in a template document. Indeed, Americans have
reported being most excited about the potential for AI to
perform household chores and conduct repetitive workplace
tasks, while expressing significant concerns about the use of
AI to make important life decisions for people.\17\ Legislation
should address these high-risk uses.
---------------------------------------------------------------------------
\17\ AI Index 2023, page 329.
Recognize the different roles of companies that develop AI
systems and companies that deploy AI systems. Legislation
should recognize the different roles of companies that create
and use AI systems. Developers are the companies that design,
code, or produce an AI system, such as a software company that
develops an AI system for speech recognition. In contrast,
deployers are the companies that use an AI system, such as a
bank that uses an AI system to make loan determinations.
Legislation should recognize the different roles of developers
and deployers, because these two types of companies will have
access to different types of information and will be able to
---------------------------------------------------------------------------
take different actions to mitigate risks.
For example, the developer of an AI system is well positioned to
describe features of the data used to train that system, but it
generally will not have insight to how the system is used after
it is purchased by another company and deployed. A company that
deploys an AI system is well positioned to understand how the
system is actually being used, what type of human oversight is
in place, and whether there are complaints about how the system
works in practice. Legislation should recognize these different
roles, so that the appropriate company in the real-world AI
supply chain can identify and mitigate risks. For the same
reasons, this kind of distinction is considered best practice
in privacy and security legislation around the world.\18\
---------------------------------------------------------------------------
\18\ See BSA, AI Developers and Deployers: An Important
Distinction, available at https://www.bsa.org/files/policy-filings/
03162023aidevdep.pdf.
Legislation that leverages these existing tools can crate
meaningful safeguards for the development and deployment of AI in high-
risk uses.
E. Benefits of Legislation
If Congress passes a law that requires risk management programs and
impact assessments, it can require companies to identify and mitigate
across the lifecycle of an AI system. These requirements will have real
benefits to consumers, by making companies evaluate and address
potential risks that can arise across the lifecycle of an AI system.
These risks can arise in contexts including:
Problem Formulation. The initial step in building an AI
system is often referred to as ``problem formulation.'' It
involves the identification and specification of the
``problem'' the system is intended to address, an initial
mapping of how the model will achieve that objective, and the
identification of a ``target variable'' the system will be used
to predict. Because many AI systems are designed to make
predictions about attributes that are not directly measurable,
data scientists must often identify variables that can be used
as proxies for the quality or outcome it is intended to
predict.
While the use of proxy target variables can be entirely reasonable,
the assumptions underlying the choice of proxies must be
closely scrutinized to ensure that it does not introduce
unintended bias to the system. The risk that can arise in
problem formulation is exemplified by a study of a widely used
healthcare algorithm that hospitals rely on to identify
patients in need of urgent care. The research team concluded
that the algorithm was systematically assigning lower risk
scores to black patients compared to similarly sick white
counterparts because it relied on data about historical
healthcare costs as proxy for predicting a patient's future
healthcare needs. Unfortunately, because black patients have
historically had less access to healthcare, the reliance of
spending data painted an inaccurate picture and led to
dangerously biased outcomes.\19\
---------------------------------------------------------------------------
\19\ Heidi Ledford, Millions of Black People Affected by Racial
Bias in Health-Care Algorithms, Nature (Oct. 24, 2019), available at
https://www.nature.com/articles/d41586-019-03228-6.
Training Data. The data used to train an AI system is a
second major vector for bias. If the data used to train a
system is misrepresentative of the population in which it will
be used, there is a risk the system will perform less
effectively on communities that may be underrepresented in the
training data. Likewise, reliance on data that itself may be
the product of institutional or historical biases can entrench
---------------------------------------------------------------------------
those inequities in an AI model.
Labeling. The process of ``labelling'' training data can
also introduce bias. Many AI systems require training data to
be ``labeled'' so that the learning algorithm can identify
patterns and correlations that can be used to classify future
data inputs. Because the process of labeling the data can
involve subjective decisions, there is the potential for
introducing unintended bias into the training data.
Deployment. Even a system thoroughly vetted during
development can begin to exhibit bias after it is deployed. AI
systems are trained on data that represents a static moment in
time and that filters out ``noise'' that could undermine the
model's ability to make consistent and accurate predictions.
Upon deployment in the real world, AI systems inevitably
encounter conditions that differ from those in the development
and testing environment. Because the real world changes over
time, the snapshot in time that a model represents may
naturally become less accurate. If the input data for a
deployed AI system differs materially from its training data,
there is a risk that the system could ``drift'' and that the
performance of the model could be undermined. For instance, if
an AI system is designed (and tested) for use in a specific
country, the system may not perform well if it is deployed in a
country with radically different demographics. Bias can also
arise if an AI system is deployed into an environment that
differs significantly from the conditions for which it was
designed or for purposes that are inconsistent with its
intended use.
Congress should act now to require companies that develop or deploy
high-risk AI to adopt risk management programs and to conduct impact
assessments. These requirements will protect consumers and create
meaningful rules that reduce risks and promote innovation.
* * *
We appreciate Congress's leadership on the important policy issues
raised by AI. We are ready to help as you craft and pass legislation.
Thank you and I look forward to your questions.
______
Annex: AI In Every Sector
Improving Healthcare and Quality of Life
The rapid digitalization of health information has created
tremendous opportunities for AI to transform how clinicians care for
patients, how consumers manage their health, and how researchers
discover breakthroughs in the treatment and prevention of diseases.
Helping Pharmacies Redistribute Medication and Provide Personalized
Advice to
Patients
Walgreens uses the Databricks Lakehouse platform to run an
intelligent data platform incorporating AI to forecast demand and
redistribute medications across Walgreens' network of nearly 9,000
pharmacies, while delivering near real-time insights and
recommendations for pharmacists to help provide more personalized
advice to patients. This integrated AI-driven platform allows
Walgreens' different data teams to work better together, create smarter
algorithms and generate new types of reporting to help managers
understand the supply chain, store labor and productivity, patient
vaccine scheduling, and prescription pickup processes.
Advancing Accessibility
For people with visual impairments, AI is turning the visual world
into an audible experience. Microsoft's Seeing AI app helps people who
are blind or visually impaired recognize objects, people, and text via
a phone or tablet's camera and describes what it recognizes to the
user. With this new layer of information, users can navigate the world
more independently.
Strengthening Security
Although data security is core to the management of most
organizations, cyber threats continue to evolve at a breakneck pace. AI
helps organizations stay a step ahead of hackers by predicting
potential attacks, mitigating attacks in real-time, managing access to
resources, and encrypting sensitive data.
Enabling Fast Action Against Security Threats
Palo Alto Networks' AI-driven Security Operations Center automation
engine, XSIAM, is delivering never-before-seen cybersecurity outcomes.
The company's own instance of this tool ingests 36 billion events every
day from across all network layers and attack surfaces and triages just
8 of those for human analysis. This empowers their most precious
resources--people--to focus on the most sophisticated attacks that
uniquely require human analysis. Importantly, this AI-driven tool has
reduced overall Mean Time to Detection (MTTD) to 10 seconds and Mean
Time to Response (MTTR) to one minute for high priority incidents. This
more resilient and automated cyber future would not be possible without
AI.
Protecting Business Transactions
Splunk is helping financial institutions to leverage AI and data
analytics to strengthen their cybersecurity and their ability to serve
customers. For example, consumer report and risk scoring provider
TransUnion uses data analytics and machine learning capabilities
provided by Splunk to monitor customer traffic and transactions.
TransUnion monitors and manages customer traffic to its website and
detects when unusual activity takes place so it can alert customers
about security concerns and ensure seamless customer experiences.
Building 21st Century Infrastructure
Whether it's creating smarter and safer cities by integrating
sensors in bridges and highways to monitor their safety or increasing
efficiency by cutting travel time and fuel expenses, AI plays an
instrumental role in creating an infrastructure designed for the 21st
century.
Optimizing Manufacturing
Generative design tools can optimize the manufacturing process to
reduce waste and improve products. Autodesk teamed up with Michigan-
based foundry Aristo Cast to develop an ultralightweight aircraft seat
frame. The team used generative design, 3D printing, lattice
optimization, and investment casting to ultimately create a seat frame
that weighs 56 percent less than typical current models. For a 615-seat
Airbus A380 plane, that would mean saving $100,000 in fuel per year, as
well as more than 140,000 fewer tons of carbon in the atmosphere.
Streamlining Building Projects
Companies are using AI to streamline the building design and
construction processes. Bentley Systems has teamed with Hyundai
engineering on an AI system that automates design processes for steel
and concrete structures, reducing the time needed to create designs and
the cost of building a structure.
Monitoring Vehicle Fleets
Oracle's anomaly detection software uses AI to monitor the
operation of complex systems and detect potentially concerning
incidents. Transportation and logistics company SS Global LLC uses
Oracle's software to monitor their fleet of vehicles and get alerts
when there are early signs of potential safety issues. By detecting the
early onset of tire baldness and air leaks, the system helps SS Global
perform predictive maintenance that keeps their fleet safer and more
efficient.
Creating New Ways to Learn
AI applications are enabling personalized learning resources for
every stage of life, including adaptive learning programs, digital
tutoring, curriculum recommendations, and more. There are more digital
resources available to instructors and students than ever before, and
AI is affording them the ability to access relevant tools quickly and
easily.
Enriching Math Education
Educators are using IBM's Teacher Advisor With Watson AI to access
the math resources they need in seconds, including proven lesson plans,
activities, standards information, and teaching strategies for students
with varying degrees of preparation and ability. This can save valuable
time for teachers throughout the school year.
Tailoring Workplace Learning
Employers are using Workday Learning, an application that uses
machine learning to personalize workplace learning for individuals, to
recommend professional development content and courses based on
employee position, tenure at the company, interactions with the
content, and other factors. This helps companies adjust learning
strategies and programming to ensure employees learn new skills,
continue to grow in their roles, and prepare for what's ahead.
Enhancing the Customer Experience
For businesses with large customer bases that are processing a high
volume of purchases--such as banks, restaurant chains, and large
retailers--analyzing the massive amount of data collected every day is
impossible without the computing and predictive power of AI. By using
machine learning tools, businesses across a wide range of industries
can analyze customer preferences and their own business performance to
improve end-user experiences and increase efficiencies. Software also
helps businesses generate optimal product designs by using data to
produce and analyze far more iterations than humans alone could create.
Customizing Care Experiences
Powered by Salesforce AI technology, Eli Lilly has reimagined
patient care with its Patient Connect Platform app. The app helps
customers learn to use products, access information about their
medications, and record how well they are feeling. The desktop and
mobile apps also allow patients to consult with a healthcare
concierge--a specialist who provides one-on-one support to guide
patients toward beneficial health outcomes.
Improving Customer Service Experience
Zendesk is using AI to improve the customer service experience for
both customers and the agents that interact with them. Using Zendesk's
intelligent triage functionality, a company can automatically detect a
customer's intent (for example, whether a customer is making a return
or checking on shipping status), the language the customer is using,
and the customer's overall sentiment so that the inquiry can be quickly
routed to the best agent for the job. Several of Zendesk's business-to-
consumer customers are using this Zendesk AI feature to automatically
classify and route incoming tickets to the right agents at the right
time, which has resulted in higher customer satisfaction and more one-
touch tickets.
Scaling Community Impact
Twilio provides AI chatbot services to help businesses interact
with customers. The United Way Worldwide worked with Twilio to help
scale and route inbound calls and texts to more than 200 agencies
nationwide that use their 211 system to help people locate essential
needs like housing, financial assistance, food, childcare,
transportation, and more. Using the AI-assisted interactive voice
response menu built with Twilio Autopilot, the United Way and Twilio
built a system that enables a caller to access a single 1-800 number or
be transferred by their local 211 to access assistance. The result is a
centralized system that efficiently reduces the call volume nationwide
but increases the time staffers are able to devote to mission critical
calls.
Improving Business Operations
AI is helping to streamline business operations and increase
productivity.
Enhancing Business Functions
SAP provides chatbot solutions that are seamlessly integrated into
other business functions, giving customers, partners, and employees a
bird's-eye view of business operations. For example, SAP provides
software services to Hewlett Packard Enterprise Company, including an
AI-based chatbot system that can reference serial numbers, packing
slips, and shipment dates drawn from cloud services, thereby getting
the right information to the right people at the right time.
Improving Contract Analysis
DocuSign has been helping organizations use AI-based technologies
including natural language processing and rules-based logic to manage
and analyze agreements for several years now. Using AI-powered contract
analysis can increase productivity in the contract process by helping
to speed up contract reviews, increase contract visibility, and
identify opportunities and risks.
Empowering Creativity
AI and machine learning within Adobe's Creative Cloud tools help
artists, photographers, designers, and content creators around the
world handle the time-consuming aspects of their work that can easily
be automated, so they have more time to be creative. From removing
unwanted objects like mics and logos from videos in Adobe After
Effects, to colorizing black-and-white photos in just a few clicks in
Adobe Photoshop, to painting with digital brushes that look, feel, and
act like the real thing in Adobe Fresco, Adobe's AI and machine
learning features empower creators to focus their energy on what they
love--ideating, experimenting, and creating.
Helping in Times of Crisis
In times of humanitarian crises, fast response is essential.
Researchers are developing ways to use AI to help first responders in
the critical hours and days after a natural disaster, and to track
pathogens that could lead to outbreaks of disease and mitigate the
spread.
Navigating the COVID-19 Pandemic
Siemens' Dynamic VAV Optimization (DVO) is a software solution for
building management systems that uses machine learning and AI to
configure HVAC controls according to a building's priorities, whether
that's minimizing virus transmission or minimizing energy consumption.
In direct response to the challenges of the pandemic, DVO was launched
with a new operating Defense Mode in late 2020 to reduce the risk of
viral spread in indoor spaces. DVO adjusts ventilation, temperature,
and humidity conditions to minimize risk of viral spread indoors while
also maximizing energy efficiency.
Enriching Our Lives
Leveling Up Gaming and Entertainment
AI can be used to create sophisticated 3-D environments and train
autonomous characters in our favorite games and movies. Unity's AI
products are used to develop video games, animations, and other
detailed virtual environments. By training computer-based characters in
Unity's software, game designers can create more realistic environments
that capture a player's imagination and enhance the gaming experience.
Senator Hickenlooper. Thank you, Ms. Espinel.
Dr. Krishnan.
STATEMENT OF DR. RAMAYYA KRISHNAN, W. W. COOPER AND
RUTH F. COOPER PROFESSOR OF MANAGEMENT SCIENCE
AND INFORMATION SYSTEMS; DEAN, HEINZ COLLEGE OF
INFORMATION SYSTEMS AND PUBLIC POLICY; FOUNDING
FACULTY DIRECTOR, THE BLOCK CENTER FOR
TECHNOLOGY AND SOCIETY, CARNEGIE MELLON UNIVERSITY
Dr. Krishnan. Good afternoon. Chair Cantwell, Chairman
Hickenlooper, Ranking Member Blackburn, and Members of the
Committee, I am grateful for this opportunity to testify today.
My name is Ramayya Krishnan. I am the Dean of the Heinz College
of Information Systems and Public Policy. I have served as
President of INFORMS, a Global Operations Research and
Analytics Society, and my perspective is shaped by my own work
in this area, as well as my leadership of The Block Center for
Technology and Society, a university-wide initiative that
studies responsible use of AI and the future of work.
I am a member of the NAIAC, but I not here representing the
NAIAC.
You have already heard about the vast number of
applications and the potential for AI. I would like to
highlight that in addition to all of that, AI has the capacity
to enable breakthroughs in science and drug discovery that
unlock solutions that are currently intractable--currently
intractable problems in human health, and beyond.
These are among the many important socially and
economically beneficial uses of the technology. As AI
technologies are considered for use in high-stakes
applications, such as autonomous vehicles, health care,
recruiting, criminal justice, the unwillingness of leading
vendors to disclose the attributes and provenance of the data
that they have used to train and tune their models, and the
processes they have employed for model training and alignment
to minimize the risk of toxic or harmful responses, needs to be
urgently addressed.
This lack of transparency creates threats to privacy,
security, and the uncompensated use of intellectual property to
copyrighted content. In addition to the harms caused to
individuals and community due to bias and unreliable
performance, there is a need for greater accountability and
transparency.
In my testimony I would like to propose four decisive
recommendations. The first, is on promoting responsible AI,
Congress should require all federal agencies to use the NIST AI
RMF in design, development, procurement, use and management of
their AI use cases. The NIST AI RMF was developed with multiple
stakeholder inputs, and establishing it as a standard that will
have numerous benefits at home and abroad.
My next recommendation is a two-part recommendation. The
first relates to advancing data transparency in the AI
pipeline. The AI pipeline consists of training data, models,
and applications, and data transparency, Congress should
require standardized documentation, and like audited financial
statements, they should be verifiable by a trusted third party,
like an auditor, about the AI training data.
The metaphor to think about is to think of these as akin to
nutrition labels, so it is clear what went into producing the
model. The details about what these should contain, but at the
minimum it should document the sources and rights that the
model developers have, to be able to use the data that they
did, and the structural analysis that they have done to check
for biases, and the like. And perhaps even for adversarial
attacks against this data.
The second part of this recommendation is promoting model
validation and evaluation of the AI system. Congress should
direct NIST to develop standards for high-stakes domains, such
as health care, recruiting, criminal justice, and require model
validation report for AI systems deployed in high-stakes
applications.
The metaphor here is to think of this being akin to an
underwriter's lab-like report that will objectively assess the
risk and performance of an AI system in these high-stakes
applications.
The third recommendation is on content transparency, it is
about content labeling and detection. The main idea here is
that, as we have just heard, the generative AI has created--
increased the capacity to create multi-modal content, audio,
video, text, that can be--that is indistinguishable from human
created output, and currently there is no standardized way to
label the content as being AI generated, or human generated.
There are consortia like C2PA that are coming around with
standards, but we need standards here, and Congress should
require all AI models, open source, or otherwise, that produce
content to be able to label their content with watermarking
technology, and provide a tool to detect that label.
While the usual concern about labeling is with regard to
consumers, this is going to be equally important for model
developers to know if the data that they are using in their
models, is human-generated or AI-generated.
The last recommendation is about investing in a trust
infrastructure for AI, much like we did in the late-1980s, when
we stood up a trust infrastructure in response to cybersecurity
attacks, the Morris worm in 1988, and we set up the CERT, the
Computer Emergency Response Team. We need to do something
similar for AI.
Congress should stand up a capability which could be done
relatively quickly using the capacity of FFRDCs, and NIST in
the SEI MITRE, and other agencies, Federal Government agencies
that connect vendors, catalog incidents, record
vulnerabilities, and test and verify models, and disseminate
best practices.
This will go a long way towards improving our trust
capability, especially since the technology is moving so
quickly, that we will always have to meet this quick-response
capability.
Finally, in closing, the success of these recommendations
will in part rest on a comprehensive approach to enhance AI
skills across K through 12, and community colleges, as well as
policies and strategies, like wage insurance to address the
impact of AI.
Thank you for this opportunity to testify to the committee.
[The prepared statement of Dr. Krishnan follows:]
Prepared Statement of Dr. Ramayya Krishnan, W. W. Cooper and Ruth F.
Cooper Professor of Management Science and Information Systems; Dean,
Heinz College of Information Systems and Public Policy; Founding
Faculty Director, The Block Center for Technology and Society,
Carnegie Mellon University
Chair Cantwell, Ranking Member Cruz, Subcommittee Chair
Hickenlooper, Ranking Member Blackburn and Members of the Committee, I
am grateful for the opportunity to testify today. My name is Ramayya
Krishnan and I serve as Dean of the Heinz College of Information
Systems and Public Policy, a multidisciplinary academic unit that spans
information technology, data analytics and public policy at Carnegie
Mellon University. My perspective is shaped by my work over several
decades on the use of data analytics and advanced model-based
approaches to support decision making in consequential applications and
by my role as the faculty director of the Block Center for Technology
and Society, a university wide initiative that studies the responsible
use of AI and its consequences for the future of work. I am a member of
the National AI Advisory Committee (NAIAC). However, I am here in my
own capacity and not representing the NAIAC.
Artificial intelligence (AI) is a transformative technology. Its
application to create personalized tutors (See https://youtu.be/
yEgHrxvLsz0), create operational and clinical decision support tools in
health care (https://www.nature.com/articles/s41586-023-06160-y),
promote health literacy among citizens https://www.cmu.
edu/news/stories/archives/2023/august/revolutionizing-health-care-
harnessing-artificial-intelligence-for-better-patient-care), and enable
breakthroughs in science and drug discovery that will unlock solutions
to currently intractable problems in human health and beyond are among
the many economically and societally beneficial uses of the technology.
And likely many of us have used an AI chatbot like chatGPT, a
generative AI technology, and seen both its immense potential and its
failures.
As AI technologies are considered for use in high stakes
applications such as in health care, recruiting and criminal justice,
the unwillingness of the leading vendors to disclose the attributes and
provenance of the data they have used to train and tune their models,
and the processes they have employed for model training and
``alignment'' to minimize the risk of toxic or harmful responses needs
to be urgently addressed (https://arxiv.org/pdf/1901.10002.pdf). This
lack of transparency creates threats to privacy, security, and
uncompensated use of intellectual property and copyrighted content in
addition to harms caused to individuals and communities, due to biased
and unreliable performance. There is a need for greater accountability
and transparency in the development and deployment of AI to spur its
responsible adoption and use.
In my testimony, I propose four decisive recommendations for
Congress to consider to address these challenges. Investing in these
recommendations will provide near term impact on trusted adoption of AI
and, when combined with a focused research program, will ensure U.S.
leadership in responsible and trustworthy AI.
The recommendations include foundational actions to advance broad
based adoption of Responsible AI practices as well as measures to
accelerate the adoption of practices and technologies to mitigate the
threat of deep fakes and protect the rights of creators. My final
recommendation is to create an infrastructure for AI trust by
establishing a capability to monitor and respond to AI vulnerabilities
and failures and support the development and dissemination of solutions
and best practices. These measures will need to be closely aligned with
a focused research program that advances the development of tools for
watermarking and labeling as well as research into measurement, metrics
and evaluation of reliability and quality of the AI supply chain. A
detailed policy memo co-authored by my colleagues and myself from
Carnegie Mellon on accountable AI is available at https://www.cmu.edu/
block-center/responsible-ai/cmu_blockcenter_rai-memo_final.pdf.
1. Promoting Responsible AI
Congress should require all Federal agencies to use the NIST
(National Institute of Standards and Technology) AI Risk Management
Framework during the design, development, procurement, use, and
management of their AI use cases. This will promote responsible
adoption and deployment of AI in government and more broadly in
society. Investing in workshops such as the NIST-Carnegie Mellon AI RMF
workshop which convened academics and industry representatives from
sectors such as banking, health care and consulting to discuss the gaps
that need to be addressed to operationalize responsible AI in the
economy will be particularly valuable. The NIST AI RMF was developed
with multiple stakeholder inputs and establishing it as a standard will
have numerous benefits at home and abroad.
2. Promoting Greater AI Transparency
a. Content transparency: Content Labeling and Detection:
A significant concern with the advent of generative AI is the ease
with which multi-modal content (audio, video text) can be created that
is indistinguishable from human created content. Multiple examples
document why this is a problem. Students submitting AI-produced content
in lieu of their own work is an academic integrity issue and hurts
learning outcomes. Audio and video deep fakes raise concerns from
multiple standpoints--from affecting the economic outcomes and
reputations of well-known artists to concerns for human rights and
democracy.
Currently, there is no standardized way to label content as AI
generated and no standardized tool that can use such labels to help
consumers recognize AI generated content. Recent commitments by major
vendors to develop watermarks for AI generated content and the
emergence of content provenance standards (e.g., C2PA) is a step in the
right direction. While proposals exist for audio and visual content,
watermarking and provenance for AI generated text remains a challenge.
As with all security technologies, watermarking will need to stay one
step ahead of attempts to defeat it (https://
legaljournal.princeton.edu/the-high-stakes-of-deepfakes-the-grow
ing-necessity-of-federal-legislation-to-regulate-this-rapidly-evolving-
technology/ and https://gjia.georgetown.edu/2023/05/24/should-the-
united-states-or-the-european-union-follow-chinas-lead-and-require-
watermarks-for-generative-ai/).
While the usual concern about this issue has been from the point of
view of the consumer, this inability to distinguish AI generated
content from human produced content and knowledge of its provenance is
relevant to model developers as well. Since Internet content is used at
scale to train models and as AI produced content proliferates on the
internet, model developers will need the capacity to differentiate AI
produced content from human produced content since this has
implications for model performance.
Congress should require all AI models (open source and closed
source models) that produce content to label their content with
watermarking and provenance technology and provide a tool to detect the
label.
b. Advancing Transparency in the AI pipeline for high stakes
applications
The AI pipeline or value chain (Hosanagar and Krishnan, 2023)
consists of training data, models and applications. It is this pipeline
that is used to create AI systems in high stakes applications such as
in autonomous vehicles, health care, recruiting, and criminal justice.
The leading vendors of closed source models do not disclose the
attributes and provenance of the data they have used to train and tune
their models, and the processes they have employed for model training
and ``alignment'' to minimize the risk of toxic or harmful responses.
When these AI pipelines are used in high stakes applications, greater
transparency around how AI is integrated into the broader system is
needed. We can learn from prior work in accountable AI as well as in
modeling reliability of societal and engineered systems to address
these transparency questions. In the following, I will highlight key
recommendations that pertain to data transparency and AI model
validation and evaluation. These measures are vital to address critical
challenges created by Large Language Models. It will also be vital in
enabling U.S. companies to remain globally competitive as international
standards are developed.
Advancing Data Transparency
Model developers need to document the rights they have to
work with the data they are using to train the model. This
documentation should also provide information about the source
of the data, whether it was public or private, etc.
Model developers should respect the right of data owners to
opt out of data crawling (robots.txt file) and also provide
data owners the opportunity to opt out of the use of their
already collected data in model training or tuning.
Model developers need to document the standards that were
used in bias assessment and demonstrate the analysis that was
conducted to assess structural bias in the data.
Congress should require standardized documentation and, like
audited financial statements, they should be verifiable by a trusted
third party (e.g., an auditor). The metaphor is to think of these as
akin to ``nutrition labels.'' so it is clear what went into producing
the model.
Promoting Model Validation and Evaluation of the AI system
Develop clear standards for articulating intended use cases
and metrics for reliability and utility so that users can have
clear expectations of performance under well-defined
conditions. Congress should direct NIST to develop standards
for these high stakes applications such as health care,
recruiting, and criminal justice.
Define AI sandboxes and test data sets, evaluation
frameworks, measurement and metrics, and continuous monitoring
standards based on the assessed risk of the application space
or use case.
Require the auditor to use these standards and validation
infrastructure to evaluate the AI system and provide the
required assurance prior to deployment.
Congress should require a model validation report for AI systems
deployed in high stakes applications. The metaphor is to think of this
as akin to an ``underwriters lab'' that objectively assesses the risk
and performance of an AI system.
c. Investing in a trust infrastructure for AI (AI CERT), or ALRT (AI
lead response team)
AI is developing rapidly. Even with the proposed transparency
measures, there will be a need to respond rapidly to newly discovered
AI vulnerabilities, exploits and failures. The safety and reliability
of the AI ecosystem is a necessary condition to engender trust and spur
widespread adoption and deployment of AI. While AI Incident databases
(https://incidentdatabase.ai/) from the Responsible AI Collaborative,
Project Atlas from MITRE (see https://atlas.mitre.org/) and the
recently organized DEFCON red teaming event and voluntary commitments
are important steps forward, an institutional solution is required.
The proposed solution, an ALRT, would connect vendors, AI system
deployers and users. It would catalog incidents, record
vulnerabilities, test and verify models, and recommend solutions and
share best practices to minimize systemic risks (https://
www.forbes.com/sites/rosecelestin/2023/08/23/the-ai-financial-crisis-
theory-demystified-how-to-create-resilient-global-ecosystems/
?sh=27282b4d51ce) as well as harm stemming from vulnerability exploits.
This is modeled after the computer emergency response team (CERT) that
the U.S. Government stood up in response to cyber security
vulnerabilities and threats in late 1980s. The following capabilities
are required to serve the needs of CERT for AI)
Catalog incidents, record vulnerabilities, recommend
solutions and, share best practices to minimize risks
Coordinate commercial and public government focus with the
need to rapidly respond to national security challenges (e.g.,
chemical/bio weapons https://arxiv.org/ftp/arxiv/papers/2306/
2306.03809.pdf). Be able to respond to threats that affect
.com, .gov and .mil domains. In effect, combine open,
restricted and classified work
Possess deep technical capabilities spanning core AI and
computing and an understanding of how the core AI is
operationalized to meet application needs
Maintain domain knowledge connected to applications
Convene industry, government and academic partners around
core tech AI technology as well as operationalization of AI
Congress should stand up these capabilities quickly via existing
FFRDCs and harness the strengths at NIST and other Federal agencies.
implementation of these recommendations will have an immediate impact
on trusted adoption of AI (e.g., standing up ALRT). Combining
implementation with investments in a focused program of research on
Responsible AI and AI transparency will ensure U.S. leadership in
trustworthy AI.
Finally, in closing, the success of these recommendations will in
part rest on comprehensive strategies that enhance AI skills across the
continuum from K-12 (https://www.tutors.plus/) and community college
(https://sailplatform.org/) education to new tools, strategies and
policies (https://tinyurl.com/2528ke6z) to support workers in virtually
all industries to adapt to the impact of AI. Thank you for this
opportunity to testify to your committee.
Senator Hickenlooper. Thank you, Dr. Krishnan.
Mr. Gregory.
STATEMENT OF SAM GREGORY, EXECUTIVE DIRECTOR, WITNESS
Mr. Gregory. Chairman Hickenlooper, Ranking Member
Blackburn, and Members of the Subcommittee; I am Sam Gregory,
Executive Director of WITNESS, a human rights organization.
Since 2018, WITNESS has led a global effort, Prepare, Don't
Panic, to inclusively prepare society for deepfakes in
synthetic media technologies, and more recently generative AI.
Now, the moment to act has come.
Increased volumes of easily made realistic synthetic
photos, video, and audio, more targeted and personalized is a
paradigm shift. Alongside creative and commercial benefits from
generative AI, there are already harms in the U.S. and
globally, with disproportionate impacts on at-risk groups.
Women are targeted with nonconsensual sexual images;
simulated audio scams are proliferating, as are AI generated
child sexual abuse images. Claims of AI generation are used to
dismiss critical human rights and journalistic content that is
real. Text-to-image tools perpetuate discriminatory patterns,
and creatives see their work used for training AI models
without consent.
As you develop legislation, first consult broadly with the
communities most impacted by AI in the U.S. and globally. Given
differential impacts, a resilient approach grounded in civil
and human rights, will best future proof legislative responses
to AI.
Elections are poised to be deeply influenced by generative
AI, and recent polls find the American public is fearful of its
impact. Underresourced newsrooms and community leaders are
under pressure, and do not have access to reliable tools that
can detect deepfakes and AI-manipulated content.
It is also unreasonable to expect consumers to spot
deceptive yet realistic imagery and voices. Guidance to look
for the six-fingered hand, or spot virtual errors in the Pope
in the puffer jacket, do not help in the long run.
I believe that Congress should think about AI governance as
a pipeline of responsibility that is distributed across
technology actors from the foundation models to those designing
and deploying software and apps, and to platforms that
disseminate content. This should be supported by testing,
auditing, transparency, and pre-post-impact assessment.
Within that, solutions that help show the provenance of AI,
and if desired, human-generated content can use this pipeline
responsibility to bring transparency to consumers. This may
take the shape of direct visible, or indirect machine readable
disclosure around how media is made, created, and/or
distributed. Approaches like this were included in the White
House Voluntary Commitments, and in the EU's AI Act.
As I note in my written testimony, visible watermarks and
direct disclosure have use cases. For example, potentially, in
marking election-related materials, but they are easily
removed, and not nuanced for the complex production era that we
are about to enter.
Invisible watermarks at the pixel and dataset level are
another option.
Cryptographically signed metadata like the C2PA standard
show the creation, production, and distribution process over
time, which is important--is important as we increasingly
intermingle AI and human generated.
Approaches like this also allow creators to indicate how
and if their content may be used for training AI models. These
disclosure approaches are not a punitive measure to single out
AI content, nor indicate deception. This provenance data only
provides additional information signals, but does not provide
any truth to the content.
To safeguard constitutional and human rights approaches to
provenance could meet some core criteria. They should first
protect privacy. Personally identifiable information should not
be required. Knowing the how of AI-based production is key, but
does not require correlating the identify of who made the
content, or instructed the tool. Allowing redaction is also
key, particularly when combining AI-based media with other
human-generated media.
Secondly, opt-in. While disclose indicating content was AI-
generated could be a legal requirement in certain cases, it
shouldn't be a requirement for all provenance tools, especially
those for non-AI content, which should always be opt-in.
And thirdly, standards should be developed attentive to
potential authoritarian misuse in a global context.
Another response complementary to content provenance, is
after-the-fact detection for content believed to be AI-
generated. From witnesses' experience, the skills and tools to
detect AI-generated media remain unavailable to the people who
need them most. Journalists, rights defenders, and election
officials domestically and globally; it remains critical to
support Federal research and investment in this area to close
this detection access and equity gap.
To conclude, I encourage you to incorporate broad
consultation with groups disproportionately experiencing
existing harms and AI impacts into upcoming legislative
approaches, to go beyond risk-based and voluntary approaches,
and support a rights-based framework for action.
And finally, to support research and legislation on
standardized watermarking and provenance that takes into
account global implications, and centers privacy and
accessibility.
Thank you for the opportunity to testify today.
[The prepared statement of Mr. Gregory follows:]
Testimony of Sam Gregory, Executive Director, WITNESS
Chairman Hickenlooper, Ranking Member Blackburn and members of the
Senate Commerce Subcommittee on Consumer Protection, Product Safety and
Data Security, thank you for the opportunity to testify today about
transparency in AI.
I am Sam Gregory, Executive Director of WITNESS, a human rights
organization.\1\ Since 2018, WITNESS has led a global effort, Prepare,
Don't Panic, to understand how deepfake and synthetic media
technologies, and more recently large language models (LLMs) and
generative AI, are impacting consumers, citizens and communities in the
U.S. and globally, and to prepare accordingly.\2\ These efforts have
included contribution to technical standards development,\3\ engagement
on detection and authenticity approaches that can support consumer
literacy,\4\ analysis and real-time response to contemporary usages,\5\
research,\6\ and consultative work with rights defenders, journalists,
content creators, technologists and other members of civil society.\7\
Our experience is further informed by three decades of experience
helping communities, citizens, journalists and human rights defenders
create trustworthy photos and videos related to critical societal
issues and protect themselves against the misuse of their content.
---------------------------------------------------------------------------
\1\ WITNESS https://www.witness.org/
\2\ For our work on generative AI and deepfakes see: https://
www.gen-ai.witness.org/
\3\ Jacobo Castellanos, WITNESS and the C2PA Harms and Misuse
Assessment Process, WITNESS, December 2021, https://blog.witness.org/
2021/12/witness-and-the-c2pa-harms-and-mis
use-assessment-process/
\4\ WITNESS Media Lab, How do we work together to detect AI-
generated media? https://lab.witness.org/projects/osint-digital-
forensics/
\5\ Nilesh Christopher, An Indian politician says scandalous audio
clips are AI deepfakes: We had them tested, Rest of World, July 2023,
https://restofworld.org/2023/indian-politician
-leaked-audio-ai-deepfake/
\6\ Gabriela Ivens and Sam Gregory, Ticks or It Didn't Happen:
Confronting Key Dilemmas in Authenticity Infrastructure for Multimedia,
WITNESS, December 2019, https://lab.witness.org/ticks-or-it-didnt-
happen/
\7\ Raquel Vazquez Llorente, Jacobo Castellanos and Nkem Agunwa,
Fortifying the Truth in the Age of Synthetic Media and Generative AI.
WITNESS, June 2023, https://blog.witness.org/2023/05/generative-ai-
africa/; Sam Gregory, Deepfakes, misinformation and disinformation and
authenticity infrastructure responses: Impacts on frontline witnessing,
distant witnessing, and civic journalism. Journalism, December 2021,
https://www.researchgate.net/publication/356976
532_Deepfakes_misinformation_and_disinformation_and_authenticity_infrast
ructure_responses_
Impacts_on_frontline_witnessing_distant_witnessing_and_civic_journalism.
Also, see: Deepfakes: Prepare Now (Perspectives from Brazil), WITNESS,
2019, https://lab.witness.org/brazil-deepfakes-prepare-now/; Deepfakes:
Prepare Now (Perspectives from South and Southeast Asia), WITNESS, 2020
https://lab.witness.org/asia-deepfakes-prepare-now/; Corin Faife, What
We Learned from the Pretoria Deepfakes Workshop, WITNESS, 2020, https:/
/blog.witness.org/2020/02/report-pretoria-deepfakes-workshop/; Corin
Faife, How Can U.S. Activists Confront Deepfakes and Disinformation?
WITNESS, 2020, https://blog.witness.org/2020/12/usa-activists-disinfor
mation-deepfakes/
---------------------------------------------------------------------------
Today, I will have a particular focus on how to optimize the
benefits, and minimize the harms and risks from multimodal audiovisual
generative AI. These tools, with their potential to create realistic
image, audio and video simulations at scale, as well as personalized
content, will have far-reaching implications for consumers, creative
production and generally, our trust in the information we see and hear.
Executive Summary
My organization, WITNESS, has for a number of years promoted a
perspective of Prepare, Don't Panic in relation to deepfakes and
generative AI. But the moment to act and prepare society for
audiovisual generative AI and its impacts has come.
Transparency around AI's role in the production of information that
consumers and citizens engage with is a critical area. In this
testimony, I will focus on questions of disclosure and provenance in
audiovisual content, and how these relate to the responsibility of
actors in the AI pipeline. WITNESS, in consultation with global experts
and communities affected by technology development, focuses on three
overarching principles to guide the assessment of the opportunities and
risks that generative AI brings to society.
1. Place firm responsibility on stakeholders across the AI,
technology and information pipeline.
2. Center those who are protecting human rights and democracy, and
communities most impacted by AI, domestically and globally, in
the development of solutions.
3. Embed human rights standards and a rights-based approach in the
response to AI.
With these principles in mind, U.S. policy makers and legislators
have a range of options to promote transparency in AI and protect
consumers and their data:
1. Ensure broad consultations with communities impacted by AI when
developing solutions to watermarking, provenance and
disclosure, and in broader processes of transparency common to
all AI systems-including documentation, third-party auditing,
pre-release testing, evaluation, and human rights impact
monitoring.
2. Push for a rights-based approach to transparency in AI, that
promotes standardized systems for disclosing when content has
been made with AI, while supporting opt-in solutions that track
the provenance of non-synthetic media.
3. Prohibit personal data from being included by default in any
approaches to provenance, disclosure and watermarking, for all
types of content.
4. Enact comprehensive data privacy legislation, as well as
integrate data privacy rights into broader AI legislation that
also includes solutions for opting out of models' datasets.
5. Support research and investment in technologies that can detect
AI manipulation and generation and are accessible domestically
and globally, as well as consumer-facing tools that provide
information on content provenance.
The domestic and global context of audiovisual generative AI and
deepfakes
While there are creative and commercial benefits to generative AI
and synthetic media, these tools are also already connected to a range
of harms to U.S. consumers and global users. Chatbots provide
incorrect, factual-appearing information. Audio scams using simulated
audio are proliferating. Non-consensual sexual images are used to
target private citizens and public figures, particularly women, and AI-
generated child sexual abuse images are increasing. Claims of AI-
generation are used to dismiss verifiable content. Text-to-image tools
perpetuate existing patterns of bias or discriminatory representation
present in their training data. Creatives and artists have had their
production incorporated into training for AI models without consent,
and no-one has access to reliable ways to opt their images out of these
training data sets.\8\
---------------------------------------------------------------------------
\8\ Rhiannan Williams, Melissa Heikkila, You need to talk to your
kid about AI. Here are 6 things you should say, MIT Tech Review,
September 2023, https://www.technologyreview.com/2023/09/05/1079009/
you-need-to-talk-to-your-kid-about-ai-here-are-6-things-you-should-say/
; Matt O'Brien, Chatbots sometimes make things up. Is AI's
hallucination problem fixable?, AP, August 2023, https://apnews.com/
article/artificial-intelligence-hallucination-chatbots-chatgpt-
falsehoods-ac4672c5b06e6f91050aa46ee731bcf4; FTC Consumer Alert,
Scammers use AI to enhance their family emergency schemes, March 2023,
https://consumer.ftc.gov/consumer-alerts/2023/03/scammers-use-ai-
enhance-their-family-emergency-schemes; Benj Edwards, AI-generated
child sex imagery has every U.S. attorney general calling for action,
Ars Technica, September 2023, https://arstechnica.com/information-
technology/2023/09/ai-generated-child-sex-imagery-has-every-us-
attorney-general-calling-for-action/; Nilesh Christopher, ibid; Rida
Qadri, Renee Shelby, Cynthia L. Bennett and Emily Denoton, AI's Regimes
of Representation: A Community-Centered Study of Text-to-Image Models
in South Asia, 2023, https://arxiv.org/abs/2305.11844 ; Harry H. Jiang,
Lauren Brown, Jessica Cheng, Mehtab Khan, Abhishek Gupta, Deja Workman,
Alex Hanna, Johnathan Flowers, and Timnit Gebru, AI Art and its Impact
on Artists, August 2023, https://dl.acm.org/doi/fullHtml/10.1145/
3600211.3604681;
---------------------------------------------------------------------------
In the area that I focus on, audiovisual generative AI and
deepfakes, research indicates that humans do trust the realism cues of
audio and video,\9\ cannot identify machine-generated speech cloning
accurately,\10\ do not recognize simulated human faces,\11\ do not fare
well spotting face-swapped faces,\12\ and retain false memories of
deepfakes.\13\ In the direct experience of my own organization in
analyzing high-profile suspected deepfakes encountered globally, it is
challenging to support rapid, high-quality media forensics analysis;
detection resources are not widely available to the media or the
public; and the gap between analysis and timely public understanding is
wide and can be easily exploited by malicious actors.\14\
---------------------------------------------------------------------------
\9\ Steven J, Frenda, Eric D. Knowles, William Saletan, Elizabeth F
Loftus, False memories of fabricated political events, Journal of
Experimental Social Psychology, 2013,https://papers.ssrn.com/sol3/
papers.cfm?abstract_id=2201941
\10\ Hibaq Farah, Humans can detect deepfake speech only 73 percent
of the time, study finds, The Guardian, August 2023, https://
www.theguardian.com/technology/2023/aug/02/humans-can-detect-deepfake-
speech-only-73-of-the-time-study-finds
\11\ Sophie J. Nightingale and Hany Farid, AI-synthesized faces are
indistinguishable from real faces and more trustworthy, PNAS, February
2022, https://www.pnas.org/doi/10.1073/pnas
.2120481119
\12\ Nils C. Kobis, Barbora Dolezalova and Ivan Soraperra, Fooled
twice: People cannot detect deepfakes but think they can, IScience,
November 2021, https://www.sciencedirect.com/science/article/pii/
S2589004221013353
\13\ Nadine Liv, Dov Greenbaum, Deep Fakes and Memory Malleability:
False Memories in the Service of Fake News, AJOB Neuroscience, March
2020, https://www.tandfonline.com/doi/abs/10.1080/
21507740.2020.1740351?journalCode=uabn20
\14\ Sam Gregory, Pre-Empting a Crisis: Deepfake Detection Skills +
Global Access to Media Forensics Tools, WITNESS, https://
blog.witness.org/2021/07/deepfake-detection-skills-tools-access/ ;
Nilesh Christopher, ibid.
---------------------------------------------------------------------------
These AI tools create accidental harms when they don't work as
promised or anticipated including when they `hallucinate' information.
Deceptive information is a feature, not a bug of systems. Consumers
also face misuse harms when generative AI tools work as intended to,
but are exploited deliberately for criminal or deceptive purposes, such
as cloning voices for scams.
There are supply chain harms derived from representational biases
that are a reflection of both developers' choices and prejudices
embedded in the training data, as well as harms that come from the
inappropriate incorporation of personal data, creative production or
intellectual property into the development processes of AI.
Finally, given the lack of public understanding of AI, the rapidly
increasing verisimilitude of outputs, and the absence of robust
transparency and accountability, the combination of poorly functioning
and misused technology brings structural harms--in this case
undermining broader trust in information and public processes.\15\
---------------------------------------------------------------------------
\15\ Matt Davies and Michael Birtwistle, Seizing the `AI moment':
making a success of the AI Safety Summit, Ada Lovelace Institute,
September 2023, https://www.adalovelaceinstitute.org/blog/ai-safety-
summit/; for additional discussion of evaluating impacts, Irene
Solaiman et al., Evaluating the Social Impact of Generative AI Systems
in Systems and Society, June 2023, https://arxiv.org/abs/2306.05949
---------------------------------------------------------------------------
High-risk usages of synthetic media and generative AI may not be
easily defined and will depend on context, and the potential for
differential impact. Risk-based approaches to regulation transfer a lot
of responsibility to the private sector and may result in regulatory
gaps impacting those consumers already most at risk. A rights-based
approach--not only a risk-based approach--that is grounded in U.S.
Constitutional values is key to protecting the interests of consumers
and citizens in the US. Existing legal frameworks on data protection
and sector specific regulations for health or financial markets, for
instance, provide a basis for action, as do the protections in the
White House AI Bill of Rights in relation to broader AI issues,
Executive Order 13960 on Promoting the Use of Trustworthy Artificial
Intelligence in the Federal Government, and the Organization for
Economic Co-operation and Development's (OECD's) 2019 Recommendation on
Artificial Intelligence that the U.S. adopted.
As I highlight later, comprehensive data privacy legislation and if
needed AI-specific regulation that incorporates strong privacy
protections would provide a core bedrock for addressing both strong
implementation of safeguards for consumers such as transparency, as
well as supporting future developments in AI technologies.
Synthetic media tools are now able to produce images of real-life
events and realistic audio of individuals with limited input data, and
at scale. Increased volume of easily made, realistic synthetic photos,
audio and eventually audio, of specific real individuals and contexts
is a paradigm shift. In the future, this will include more accurate,
targeted and interactive personalization for a given context,
individual consumer, specific user or audience in existing social media
contexts, as well as emerging formats for communications such as VR and
AR.\16\ Generative AI tools are increasingly multimodal, with text,
image, video, audio and code functioning interchangeably as input or
output.
---------------------------------------------------------------------------
\16\ Eric Horvitz, On the Horizon: Interactive and Compositional
Deepfakes, 2022, https://arxiv.org/abs/2209.01714; Thor Benson, This
Disinformation is Just for You, WIRED, August 2023, https://
www.wired.com/story/generative-ai-custom-disinformation/
---------------------------------------------------------------------------
It is unreasonable to expect consumers and citizens to be able to
`spot' deceptive and realistic imagery and voices. As the Federal Trade
Commission (FTC) has already noted,\17\ most of the challenges and
risks with generative AI cannot be addressed by the consumer acting
alone.
---------------------------------------------------------------------------
\17\ FTC Business Blog, Chatbots, deepfakes, and voice clones: AI
deception for sale, March 2023, https://www.ftc.gov/business-guidance/
blog/2023/03/chatbots-deepfakes-voice-clones-ai-deception-sale
---------------------------------------------------------------------------
Similarly, responses to the risks of these tools cannot be
adequately addressed by regulatory agencies or laws without a pipeline
of responsibility across foundation models, developers and deployers of
AI models.
Single technical solutions will not be sufficient either. In the
case of audiovisual generative AI, deepfakes and synthetic media,
technical approaches to detection will need to be combined with
privacy-protecting, accessible watermarking and opt-in provenance
approaches, and with mandatory processes of documentation and
transparency for foundation models, pre-release testing, third-party
auditing, and pre/post-release human rights impact assessments.
Harms and risks of deepfakes, generative AI and synthetic media
identified by WITNESS
Through the past five years of WITNESS consultations, civil society
leaders have consistently identified a set of existing harms and
potential threats from synthetic media and deepfakes. As tools have
become more accessible and personalizable, and easier to use, a higher
number of people have had the ability to engage with them. They have
been able to imagine--or experience--how these technologies could
impact their lives.
The main overarching concern echoes across countries: threats from
synthetic media will disproportionately impact those who are already at
risk, because of their ethnicity, gender, sexual orientation,
profession, or belonging to a social group. Women particularly already
face widespread threats from non-consensual sexual images that do not
require high-quality or complex production to be harmful. Many
marginalized and vulnerable populations have already been affected by
the existing AI-driven dynamics of the information ecosystem. They have
experienced AI and other forms of technology that have brought
differential and/or disparate impact to them. This reflects both the
biases in these tools (e.g., representational bias), as well as their
use and misuse to disproportionately target these populations.
Elections in the coming year are poised to be deeply influenced by
the malicious or deceptive use of generative AI. We hear how the fear
of synthetic media, combined with the confusion about its capabilities
and the lack of knowledge to detect AI-manipulation, are misused to
dismiss authentic information with claims it is falsified. This is so-
called plausible deniability or the ``liar's dividend''.\18\ In our
work analyzing claims of deepfakes, incidents of the liar's dividend
are highly prevalent.
---------------------------------------------------------------------------
\18\ Robert Chesney and Danielle Keats Citron, Deep Fakes: A
Looming Challenge for Privacy, Democracy, and National Security, 107
California Law Review 1753, July 2018, https://papers.ssrn.com/sol3/
papers.cfm?abstract_id=3213954
---------------------------------------------------------------------------
Similarly, these tools could be used by foreign governments to
close civil society space by, for instance, incorporating them into
patterns of criminalization and harassment of journalists and human
rights defenders, and disinformation targeting their activities and
those of political opponents at home and abroad. The potential threats
brought by synthetic media and generative AI have motivated governments
to enact laws suppressing free expression and dissent, posing a threat
to the principles of free expression, civic debate and information
sharing. Proposed rule-making and legislation on generative AI and
deepfakes in China is indicative of this trend.\19\
---------------------------------------------------------------------------
\19\ Karen Hao, China, a Pioneer in Regulating Algorithms, Turns
Its Focus to Deepfakes, Wall Street Journal, January 2023 https://
www.wsj.com/articles/china-a-pioneer-in-regulating-algorithms-turns-
its-focus-to-deepfakes-11673149283
---------------------------------------------------------------------------
Lastly, pressures to understand complex synthetic content, and
claims that content is synthesized, place additional strain on already
under-resourced local and national newsrooms and community leaders
responsible for verifying digital content. With hyperbolic rhetoric as
well as the realities of advances in generative AI undermining trust in
content we encounter, human rights defenders, journalists and civil
society actors will be among the most impacted by generative AI.
These technologies need to be developed, deployed, or regulated
with an in-depth understanding of a range of other local and national
contexts. The voices of those impacted by AI, need to be central to the
discussion and prioritization of solutions.\20\ Yet, emerging
technologies are designed and deployed without the input of those most
impacted, ignoring the threats and risks these technologies bring to
communities already at a disadvantage.
---------------------------------------------------------------------------
\20\ Sam Gregory, Journalism, December 2021, ibid.
---------------------------------------------------------------------------
Why consumer-facing transparency matters in an information environment
with more complex creative and communicative production
Media, communication and content production are increasingly
complex. Increased access to tools for creative generation and
knowledge production will bring benefits to society. However, to
realize this, one key component is transparency across the pipelines of
AI design, content production and information distribution.\21\
Transparency approaches can also support better control for individuals
and others on how their data is used in AI models.
---------------------------------------------------------------------------
\21\ Sam Gregory, Synthetic media forces us to understand how media
gets made, Nieman Lab, December 2022, https://www.niemanlab.org/2022/
12/synthetic-media-forces-us-to-understand-how-media-gets-made/
---------------------------------------------------------------------------
Frequently in my work, I am asked to provide advice to consumers on
how to spot an AI-generated image--for example, to look for `the
distorted hands', or in the case of a deepfake, to see if it does not
blink. I discourage this as these heuristics are the current Achilles
heel or temporary failings of a process, not long-term durable or
scalable guidance. Most audiovisual content we create and consume
involves AI. In a world with wider access to tools that simplify the
generation or edition of photos, videos, and audio, including photo and
audio-realistic content, it is important for the public to be able to
understand if and how a piece of media was created or altered using AI.
We refer to this as `content provenance'. Such labeling, watermarking
or indications of provenance are not a punitive measure to single out
AI content or content infused with AI, and should not be understood as
a synonym of deception, misinformation or falsehood. The vast majority
of synthetic media is used for personal productivity, creativity or
communication without malice. Satirical media made using AI is also a
critical and protected form of free speech.\22\
---------------------------------------------------------------------------
\22\ Henry Ajder and Joshua Glick, Just Joking! Deepfakes, satire,
and the politics of synthetic media, WITNESS and MIT, December 2012,
https://cocreationstudio.mit.edu/just-joking/
---------------------------------------------------------------------------
We have heard repeatedly from information consumers around the
world that responsibility should not be placed primarily on end-users
to determine if the content they are consuming is AI-generated, created
by users with another digital technology or, as in most content, a mix
of both.\23\ To ensure disclosure--and more broadly, to promote
transparency and accountability--all actors across the AI and media
distribution pipeline need to be engaged. These include:
---------------------------------------------------------------------------
\23\ WITNESS, Synthetic Media, Generative AI And Deepfakes Witness'
Recommendations For Action, 2023, https://www.gen-ai.witness.org/wp-
content/uploads/2023/06/Guiding-Principles-and-Recs-WITNESS.pdf
Those researching and building foundation or frontier
---------------------------------------------------------------------------
models;
Those commercializing generative AI tools;
Those creating synthetic media;
Those publishing, disseminating or distributing synthetic
media (such as media outlets and platforms); and
Those consuming or using synthetic media in a personal
capacity
There is now a significant trend in AI governance towards a
pipeline approach and a focus on labeling and disclosure. In July 2023,
seven leading AI companies agreed with the White House to a number of
voluntary commitments to help move toward safe, secure and transparent
development of AI technology, including committing to earning people's
trust by disclosing when content is AI-generated.\24\ In the European
Union, companies who have signed on to the voluntary EU Code of
Practice on Disinformation have agreed to a similar commitment, with
the EU's Commissioner Vera Jourova calling on these companies to label
AI-generated content.\25\ The EU AI Act includes significant
requirements for disclosing deepfakes and machine-generated content
from foundation models.
---------------------------------------------------------------------------
\24\ The White House, FACT SHEET: Biden-Harris Administration
Secures Voluntary Commitments from Leading Artificial Intelligence
Companies to Manage the Risks Posed by AI, July 2023 https://
www.whitehouse.gov/briefing-room/statements-releases/2023/07/21/fact-
sheet-biden-harris-administration-secures-voluntary-commitments-from-
leading-artificial-intelligence-companies-to-manage-the-risks-posed-by-
ai/ and The White House, Ensuring Safe, Secure, and Trustworthy AI,
https://www.whitehouse.gov/wp-content/uploads/2023/07/Ensuring-Safe-
Secure-and-Trustworthy-AI.pdf
\25\ Foo Yun Chee, AI-generated content should be labelled, EU
Commissioner Jourova says, Reuters, June 2023 https://www.reuters.com/
technology/ai-generated-content-should-be-labelled
-eu-commissioner-jourova-says-2023-06-05/
---------------------------------------------------------------------------
Most provenance systems will require methods that explain both AI-
based origins or production processes, but also document non-synthetic
audio or visual content generated by users or other digital processes-
like footage captured from `old fashioned' mobile devices.\26\ As the
White House notes in its statement on the voluntary commitments,
``companies making this commitment pledge to work with industry peers
and standards-setting bodies as appropriate towards developing a
technical framework to help users distinguish audio or visual content
generated by users from audio or visual content generated by AI.'' It
will be hard to address AI content in isolation from this broader
question of media provenance. Implementing this approach to
transparency will require standards that focus on how to design
durable, machine-readable shared standards that provide useful signals
to consumers, as well as other actors in the information pipeline
(e.g., content distributors and platforms).
---------------------------------------------------------------------------
\26\ Sam Gregory, To battle deepfakes, our technologies must track
their transformations, The Hill, June 2022, https://thehill.com/
opinion/technology/3513054-to-battle-deepfakes-our-technologies-must-
lead-us-to-the-truth/
---------------------------------------------------------------------------
The opportunity in transparency and disclosure
It is crucial for democracy that people are able to believe what
they see and hear when it comes to critical government, business and
personal communications, as well as documentation of events on the
ground. It is also critical for realizing the creativity and innovation
potential that generative AI holds, that consumers are informed about
what they see and hear.
WITNESS has actively participated in the Partnership on AI's
Responsible Practices for Synthetic Media Framework.\27\ This Framework
describes direct forms of disclosure as those methods that are `visible
to the eye', such as labels marking the content, or adding context
disclaimers. Indirect forms of disclosure are not perceptible to the
human eye and include embedded metadata or other information that is
machine readable or presentable, such as cryptographic provenance or
embedding durable elements into either or both the training data and
the content captured or generated.\28\ Importantly, the Framework also
offers a useful breakdown of how responsibility for supporting this
disclosure should be considered at different stages across the AI
pipeline.
---------------------------------------------------------------------------
\27\ Partnership on AI, Responsible Practices for Synthetic Media
Framework, https://syntheticmedia.partnershiponai.org/ See also, Jacobo
Castellanos, Building Human Rights Oriented Guidelines for Synthetic
Media, WITNESS, February 2023, https://blog.witness.org/2023/02/
building-human-rights-oriented-guidelines-for-synthetic-media/
\28\ For synthetic content, the most recent example is SynthID,
released by Google on August 29, 2023. https://www.deepmind.com/blog/
identifying-ai-generated-images-with-synthid
---------------------------------------------------------------------------
There is significant and unhelpful confusion around terms used to
show use of AI in content.\29\ `Watermarking' is used as a catch-all
term that includes:
---------------------------------------------------------------------------
\29\ Claire Leibowicz, Why watermarking AI-generated content won't
guarantee trust online, MIT Tech Review, August 2023, https://
www.technologyreview.com/2023/08/09/1077516/water
marking-ai-trust-online/
a. Visible watermarks, signals or labels (e.g., a `Made with AI'
---------------------------------------------------------------------------
description on an image);
b. Invisible watermarks and technical signals that are imperceptible
to the human eye and can be embedded at pixel level, or as
early as the training stage of AI processes; and
c. Cryptographically-signed metadata that shows the production
process of content over time, like the C2PA standards.
Visible signals or labels can be useful in specific scenarios such
as AI-based imagery or production within election advertising, as
proposed in the REAL Political Ads Act. However, visible watermarks are
often easily cropped, scaled out, masked or removed, and specialized
tools can remove them without leaving a trace. Visible watermarks are
hence inadequate for reflecting the `recipe' for the use of AI in an
image or video, and in a more complex media environment fail to reflect
how AI is used in a meaningful way for consumers.
Technical interventions and signals at the dataset level can
indicate provenance as well as embed `Do Not Train' restrictions that
could give consumers more say in who is allowed to build AI models
using people's data and content. However, many datasets are already in
use and do not include these marks. Additionally, small companies,
independent developers, and open-source libraries and tools may not
have the capacity and ability to develop reliable and sustainable
invisible watermarks. Without accessible and standardized standards,
there is a risk of excluding a significant part of the AI innovation
ecosystem from the conversation. This, in turn, could lead to a handful
of AI companies' dominance becoming further entrenched. Dataset-level
watermarks also require their application across broad datasets, which
brings questions around ownership and responsibility regarding the
content and the repurposing of that content for training purposes.
Since, in most cases, the original content creators are not involved in
the decision to add their content to a training dataset, they are
unlikely to be involved in the decision to watermark their content as
well.
Cryptographic signature and provenance-based standards such as the
C2PA are built to make it very hard to tamper with the cryptographic
signature without leaving evidence of the attempt, and to enable the
reconnection of a piece of content to a set of metadata if that is
removed. These methods can allow people to understand the lifecycle of
a piece of content, from its creation or capture to its production and
distribution. In some cases, they are integrated with capture devices
such as cameras, utilizing a process known as `authenticated capture'.
Microsoft has been working on implementing provenance data on AI
content using C2PA specs,\30\ and Adobe has started to provide it via
its Content Credentials approach.\31\ While I do not speak for the
C2PA, WITNESS is a member of the C2PA, has participated in the
Technical Working Group and acted as a co-chair of the C2PA Technical
Working Group Threats and Harms Taskforce. In this context WITNESS has
advocated for globally-driven human rights perspectives and practical
experiences to be reflected in the technical standard.\32\
---------------------------------------------------------------------------
\30\ Kyle Wiggers, Microsoft pledges to watermark AI-generated
images and videos, Techcrunch, May 2023 https://techcrunch.com/2023/05/
23/microsoft-pledges-to-watermark-ai-generated-images-and-videos
\31\ Adobe Content Credentials, https://helpx.adobe.com/creative-
cloud/help/content-credentials.html
\32\ The Coalition for Content Provenance and Authenticity, C2PA
Harms Modelling, https://c2pa.org/specifications/specifications/1.0/
security/Harms_Modelling.html; Jacobo Castellanos, WITNESS, 2021, Ibid.
---------------------------------------------------------------------------
An approach like the C2PA can also allow creators to choose whether
their content may be used for training AI models or other data
purposes. Agency over their data is a critically needed response to
concerns from creators and others about the incorporation of their
content, personal images and other data into AI models without their
consent.
In reality, any effective shared standard, regulation, or
technological solution to provenance, disclosure and transparency is
likely to require a combination of cryptographically-signed provenance
metadata that reflects how both AI, non-AI and mixed media are created
and edited over time, as well as visible watermarking and/or technical
signals for synthetic content that confirm the use of AI specifically.
How to provide rights-respecting disclosure
To safeguard Constitutional and human rights, approaches to
provenance and disclosure need to meet at least three core criteria.
They need to:
Protect privacy:
Be accessible with modular opt-in or out depending on the
type of media and metadata, and;
Avoid configurations that can be easily weaponized by
authoritarian governments.
People using generative AI tools to create audiovisual content
should not be required to forfeit their right to privacy to adopt these
emerging technologies. Personally-identifiable information should not
be a prerequisite for identifying either AI-synthesized content or
content created using other digital processes. The `how' of AI-based
production elements is key to public understanding; this should not
require a correlation to the identity of `who' made the content or
instructed the tool.
Since 2019, WITNESS has been raising concerns about the potential
harms that could arise from the inclusion of personal data in solutions
that track the provenance of media.\33\ The U.S. government has the
opportunity to ensure that provenance requirements and standards are
developed in-line with global human rights standards, protect civil
rights and First Amendment rights, and do not include the automated
collection of personal data. While a requirement to include disclosure
indicating content was AI-generated could be a legal requirement in
certain cases, this obligation should not extend to using tools for
provenance on content created outside of AI-based tools, which should
always be opt-in.
---------------------------------------------------------------------------
\33\ Gabriela Ivens and Sam Gregory, Ibid; Sam Gregory, Tracing
trust: Why we must build authenticity infrastructure that works for
all, WITNESS, 2020, https://blog.witness.org/2020/05/authenticity-
infrastructure/
---------------------------------------------------------------------------
Building trust in content must allow for anonymity and redaction.
Immutability and inability to edit do not reflect the realities of
people, or how and why media is made-nor that certain redaction may be
needed in sensitive content.\34\ Flexibility to show how media
evolves--and to conduct redaction--is a functional requirement for
disclosure particularly as it relates to edited and produced content.
Lessons from platform policies around `real names' tell us that many
people--for example, survivors of domestic violence--have anonymity and
redaction needs that we should learn from.\35\ While specifications
like the C2PA focus on protecting privacy and don't mandate identity
disclosures, this privacy requirement needs to be protected during
widespread adoption. We should be wary of how these authenticity
infrastructures could be used by governments to capture personally
identifiable information to augment surveillance and stifle freedom of
expression, or facilitate abuse and misuse by other individuals.
---------------------------------------------------------------------------
\34\ Raquel Vazquez Llorente, Trusting Video in the Age of
Generative AI, Commonplace, June 2023, https://
commonplace.knowledgefutures.org/pub/9q6dd6lg/release/2
\35\ Jillian York and Dia Kayyali, Facebook's 'Real Name' Policy
Can Cause Real-World Harm for the LGBTQ Community, EFF, 2014, https://
www.eff.org/deeplinks/2014/09/facebooks-real-name-policy-can-cause-
real-world-harm-lgbtq-community
---------------------------------------------------------------------------
We must always view these credentials through the lens of who has
access and can choose to use them in diverse global and security
contexts, and ensure they are accessible and intelligible across a
range of technical expertise.\36\ Provenance data for both AI and user-
generated content provides signals--i.e. additional information about a
piece of content--but does not prove truth. An `implied truth' effect
simply derived from the use of a particular technology is not helpful,
nor is an `implied falsehood' effect from the choice or inability to
use them.\37\ Otherwise we risk discrediting a citizen journalist for
not using tools like these to assert the authenticity of their real-
life media because of security or access concerns, while we buttress
the content of a foreign state-sponsored television channel that does
use it. Their journalism can be foundationally unreliable even if their
media is well-documented from a provenance point of view.
---------------------------------------------------------------------------
\36\ Sam Gregory, Ticks Or It Didn't Happen, WITNESS, December
2019, https://lab.wit
ness.org/ticks-or-it-didnt-happen/
\37\ Sam Gregory, Journalism, December 2021, ibid.
---------------------------------------------------------------------------
Any credential on content must be an aid to help make informed
decisions, not a simplistic truth signal. They work best as a signal in
complement to other processes of digital and media literacy that
consumers choose to use, to help them triage questions they may have,
and that are available to other parties engaging with the content,
including potentially platforms.
The role of detection alongside provenance
`Seeing' both invisible watermarks and provenance metadata that are
imperceptible to the eye will require consumer-facing tools. However,
the average citizen shouldn't be required to keep up with watermarking
advances and detection tools, and cannot be expected to deploy multiple
tools to ascertain if a particular commercial brand, watermarking
approach, or mode of synthesis has been used.
Detection tools are also necessary for content believed to be AI-
generated that does not have provenance information or that has been
manipulated with counter-forensics approaches. There is justifiable
skepticism about whether after-the-fact detection tools are useful for
consumer transparency and consumer usage to identify generative AI and
deepfake outputs.\38\ Detection of audiovisual generative AI and
deepfakes outputs is flawed. Existing detection models frequently
require expert input to assess the results and often they are not
generalisable across multiple synthesis technologies and techniques or
require personalization to a particular person to be protected from
fraudulent voices or imagery. As such, detection tools can lead to
unintentional confusion and exclusion. We have seen how use by the
general public of detection tools has contributed to increased doubt
around real footage and enabled the use of the liar's dividend and
plausible deniability around real content, rather than contributing to
clarity.\39\
---------------------------------------------------------------------------
\38\ Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni
Poggi, Koki Nagano, Luisa Verdoliva, 2023 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), On The Detection
of Synthetic Images Generated by Diffusion Models, https://arxiv.org/
abs/2211.00680; Luisa Verdoliva, Media Forensics and Deepfakes: An
Overview, 2020, https://arxiv.org/abs/2001.06564
\39\ Sam Gregory, Pre-Empting a Crisis: Deepfake Detection Skills +
Global Access to Media Forensics Tools, WITNESS, https://
blog.witness.org/2021/07/deepfake-detection-skills-tools-access/;
Nilesh Christopher, ibid; Sam Gregory, The World Needs Deepfake Experts
to Stem This Chaos, WIRED, June 2021, https://www.wired.com/story/
opinion-the-world-needs-deepfake
-experts-to-stem-this-chaos/
---------------------------------------------------------------------------
However, from WITNESS's experience they are a critical element--
alongside the incorporation of provenance data and media literacy--when
it comes to real-world scenarios where journalists, civil society and
governments are attempting to discern how content has been created and
manipulated. As we have seen in our work supporting forensic analysis
of high profile global cases, there is a gap between on one side the
needs of journalists, civil society leaders and election officials, and
on the other side the availability of detection skills, resources and
tools that are timely, effective and grounded in local contexts. These
issues highlight the `detection equity' gap that exists--the tools to
detect AI-generated media are not available to the people who need them
the most. Further research into improving detection capabilities
remains critical as well as ensuring those who access tools also have
the knowledge and skills to use them.
Conclusion
Significant evolutions in volume, ease of access, personalization
and malicious usage of generative AI reflect both the potential for
creativity but also the heightened harms from audiovisual generative AI
and deepfakes-including the plausible deniability that these tools
enable, undermining consumers' trust in the information ecosystem. I
have highlighted in this statement the need to focus on existing harms
as identified by those on the frontlines of deepfakes and synthetic
media, and to center the voices of those affected by an AI-powered
information landscape. I encourage this Subcommittee and legislators to
go beyond a risks-based approach, and push for a rights-based framework
in order to prevent and mitigate accidental harms, misuse harms, supply
chain harms and structural harms. In this regard, approaches to
transparency in audiovisual content production that incorporate strong
privacy measures can protect personal information, safeguard democracy
around the world, and promote creative production.
Senator Hickenlooper. Thank you, Mr. Gregory.
Mr. Stair--Strayer, I am sorry.
STATEMENT OF ROB STRAYER,
EXECUTIVE VICE PRESIDENT OF POLICY,
INFORMATION TECHNOLOGY INDUSTRY COUNCIL (ITI)
Mr. Strayer. Thank you. Chair Hickenlooper, Ranking Member
Blackburn, and Members of the Committee; thank you for the
opportunity to testify today.
My name is Rob Strayer, and I lead the Global Policy Team,
at the Information Technology Industry Council, or ITI. ITI
represents companies from all corners of the technology sector,
and from across the AI ecosystem, including those involved in
both developing AI models, and deploying cutting-edge AI
applications.
ITI was pleased to provide a very detailed response to the
Subcommittee's inquiry earlier this year, about how our
companies are operationalizing the NIST AI Risk Management
Framework, as a means of building public trust.
We are encouraged by the bipartisan efforts in Congress to
address the challenges and opportunities from AI. In my
remaining time, I will address the competitive, global context
for AI, and then turn to transparency and accountability.
The development and adoption of AI technology will be
transformational across all sectors of the economy, estimates
the total global economic benefits of AI in the years ahead
range from $14 trillion to $25 trillion. That is absolutely
massive.
And just one example of the transformational nature of AI,
the cybersecurity industry is able to deploy machine learning
to detect and stop the most sophisticated attacks, using zero-
day exploits. AI can defeat these pernicious attacks using
insights about activity rather than having to rely only on
known malware signatures.
Adversaries will certainly be using AI to improve their
attacks. And we need to be able to leverage AI to protect our
critical infrastructure and IT systems; AI also will play an
essential role in future national security applications for the
military and for the intelligence community.
The good news is that today, the United States is leading
AI development, deployment, and innovation globally.
Nonetheless, foreign competitors are working hard on AI
breakthroughs and to deploy AI and new use cases in their
markets. And with access to open source models, and decreasing
model training compute costs, AI developers and deployers will
be able to move anywhere in the world with interconnections to
avoid stifling regulations.
Therefore, U.S. policymaking involving AI needs to be
understood in a global context to consider how new policies
affecting AI will help the United States maintain its
technological leadership rather than cede it to competitors,
including authoritarian states.
So how does the United States create a pro-innovation AI
policy framework that manages risk? U.S. AI policies should
have two primary components; one, promoting innovation and
investment, and two, building public trust and accountability.
My written testimony covers the importance of investment, and
so I will focus on public trust and accountability here.
Transparency is a key means by which to establish public
trust. Consumer trust will increase adoption of AI and expand
the AI ecosystem in the United States. ITI companies are
working to ensure that users understand when they are
interacting with an AI system, and generally how the system
works. ITI companies also are producing AI model cards, so that
consumers have access to the information about features and
limitations of AI models in clear, plain language.
So what is the government's role? To avoid regulations
being overly broad, risk should be identified in the context of
a specific AI use case. With risk identified it is then
imperative that the government review the existing regulatory
landscape. Legal regimes such fraud, criminal law, as well as
statutes like the Civil Rights Act, can address AI-related
risk.
It is critical to understand how these legal regimes
function, and where they may not be fit for purpose to address
AI risk before creating new legislation or regulatory
frameworks.
Finally, before adopting new AI regulatory requirements,
policymakers should understand the status of international
consensus-based standards, and the ability of those standards
to meet regulatory requirements. Without specific standards for
risk management processes, such as the measurement and
evaluation of the risk in models, it will not be possible to
implement regulations effectively, or harmonize rules globally.
To wrap up, Congress and private sector stakeholders can
work together to ensure the United States builds on its
competitive lead in AI, as AI transforms all sectors of the
economy, and generates trillions of dollars in economic growth.
This will benefit U.S. companies and citizens for decades into
the future.
Thank you. And I look forward to your questions.
[The prepared statement of Mr. Strayer follows:]
Prepared Statement of Rob Strayer, Executive Vice President of Policy,
Information Technology Industry Council (ITI)
Chairman Hickenlooper, Ranking Member Blackburn, and Distinguished
Members of the Committee and Subcommittee, thank you for the
opportunity to testify today.
My name is Rob Strayer, and I'm the Executive Vice President of
Policy at the Information Technology Industry Council (ITI). I lead
ITI's global policy team, driving ITI's strategy and advocacy efforts
to shape technology policy around the globe and enable secure
innovation, competition, and economic growth, while supporting
governments' efforts to achieve their public policy objectives. ITI is
the premier advocate and thought leader in the United States and around
the world for the information and communications technology (ICT)
industry. We represent leading companies from across the ICT sector,
including hardware, software, digital services, semiconductor, network
equipment, cybersecurity, Internet companies, and other organizations
using data and technology to evolve their businesses. Our members stand
at the forefront in developing and deploying consumer-facing, business-
to-business, and government-focused AI solutions.
We are encouraged by the bipartisan efforts in Congress to address
the challenges and opportunities from AI. This subcommittee's
jurisdiction over issues ranging from data privacy and consumer
protection to standards gives you an important role to play in AI
policy discussions. To that end, ITI was pleased to provide information
to Chairman Hickenlooper and Ranking Member Blackburn's inquiry earlier
this year about how ITI members are operationalizing NIST's AI Risk
Management Framework (AI RMF) to build and foster public trust.\1\
Congress and the Administration should work together to ensure any
legislation or regulatory proposals encourage future innovation and
investment in the United States, protect consumers and businesses,
mitigate foreseeable risks, and do not complicate or duplicate existing
standards, laws, and sector-specific regulatory regimes. ITI looks
forward to being a partner in those efforts.
---------------------------------------------------------------------------
\1\ ITI's June 2023 response to Chairman Hickenlooper and Ranking
Member Blackburn's April 2023 letter (June 1, 2023) available at:
https://www.itic.org/documents/artificial-intelligence
/ITIJune2023ResponsetoSens.HickenlooperandBlackburnAIRMFLetter.pdf
---------------------------------------------------------------------------
I. Transformational Impact of AI
The development and adoption of AI technologies will be
transformational across a variety of critical sectors, including health
care, telecommunications, aerospace, manufacturing, transportation, and
other sectors under the Committee's jurisdiction. It will help
companies be more effective and efficient, particularly at addressing
business operations challenges, research and development, and software
engineering. In fact, an Accenture survey of 1,500 executives across
all sectors found that 84 percent believed AI is critical to meeting
their growth objectives and 73 percent said they risk going out of
business if they cannot scale AI.\2\
---------------------------------------------------------------------------
\2\ See Accenture AI investment study (November 14, 2019),
available at https://www
.accenture.com/us-en/insights/artificial-intelligence/ai-investments
---------------------------------------------------------------------------
As a testament to AI's revolutionary impact, credible estimates of
the total global economic benefits of AI in the years ahead, which now
includes the impact of generative AI, range from $14 trillion to $25
trillion.\3\
---------------------------------------------------------------------------
\3\ See McKinsey and Company The Economic Potential of Generative
AI: The Next Productivity Frontier (June 2023), available at https://
www.mckinsey.com/ /media/mckinsey/business%20
functions/mckinsey%20digital/our%20insights/
the%20economic%20potential%20of%20generative
%20ai%20the%20next%20productivity%20frontier/the-economic-potential-of-
generative-ai-the-next
-productivity-frontier-vf.pdf?shouldIndex=false
---------------------------------------------------------------------------
Today, the United States is leading AI development, deployment, and
innovation. The United States employs the best and the brightest AI
researchers and experts working to advance American leadership in AI
innovation. Other nations have recognized the United States as the
center for AI excellence and are working harder than ever to develop
the next major technological developments in AI and to deploy AI in new
use cases in their countries.
Policymaking and regulation involving AI needs to be understood in
the global context of technology competition. The United States has the
potential to build on its lead as AI transforms all sectors of the
economy, generates trillions of dollars in economic growth, and
benefits U.S. companies and citizens for decades into the future.
Overly broad and prescriptive regulation, however, could undermine that
leadership position and cede it to U.S. competitors, including
authoritarian nations.
AI will play an essential role in future national security
applications for the military and intelligence communities and in the
cybersecurity defense of critical infrastructure. It is not an
exaggeration to say that U.S. national security depends on continued
U.S. technological leadership in AI. It is more important than ever
that the United States considers how any new policy affecting AI will
help it maintain its technological leadership in AI.
Below are some of the use cases that AI will empower:
Cybersecurity
Threat Mitigation: AI and machine learning can be
leveraged to improve cybersecurity. Indeed, defensive
cybersecurity technology must embrace machine learning and
AI as part of the ongoing battle between attackers and
defenders. The threat landscape constantly evolves, with
cyberattacks that are complex, automated and constantly
changing. Attackers continually improve their sophisticated
and highly automated methods, moving throughout networks to
evade detection. The cybersecurity industry is innovating
in response: making breakthroughs in machine learning and
AI to detect and block the most sophisticated malware,
network intrusions, phishing attempts, and many more
threats.\4\ AI is the best tool defenders have to identify
and prevent zero-day attacks and malware-free attacks
because AI can defeat novel threats based on behavior cues
rather than known signatures. Leveraging these technologies
is essential to meeting constantly evolving threats.
---------------------------------------------------------------------------
\4\ Testimony of Rob Strayer, Hearing on Securing the Future:
Harnessing the Potential of Emerging Technologies while Mitigating
Security Risks, Before the U.S. House Homeland Security Committee (June
22, 2022) available at https://www.itic.org/documents/cybersecurity/
20220622ITIHouseHomelandCmteTestimonyonEmergingTechandCyber.pdf
---------------------------------------------------------------------------
Manufacturing
Predictive Maintenance: AI can analyze real-time
sensor data from manufacturing equipment to predict
maintenance needs accurately. By identifying potential
equipment failures in advance, manufacturers can schedule
maintenance proactively, minimizing unplanned downtime and
reducing costs.
Supply Chain Management: AI can optimize supply chain
operations by analyzing data from multiple sources,
including demand forecasts, inventory levels, and
logistical constraints. AI algorithms can optimize
inventory management, improve demand forecasting accuracy,
and enable efficient routing and scheduling of shipments.
Health Care
Medical Imaging Analysis: AI can analyze medical
images such as X-rays, CT scans, and MRIs, helping doctors
detect and diagnose diseases more accurately and
efficiently. AI can assist in identifying anomalies,
tumors, or other abnormalities, leading to earlier
detection and treatment.
Drug Discovery and Development: AI accelerates the
drug discovery process by analyzing massive datasets and
identifying potential drug candidates. AI algorithms can
predict the efficacy of drugs, design molecules, and
optimize clinical trials, reducing the time and cost of
bringing new drugs to market.
Telecommunications
Network Planning and Deployment: AI can help analyze
data to assist in planning the deployment of
telecommunication networks. AI can help determine optimal
tower locations, estimate coverage areas, and predict
network capacity requirements, enabling providers to make
informed decisions during network expansion.
Network Security: AI can monitor network traffic,
detect anomalies, and identify potential cybersecurity
threats. AI algorithms can analyze patterns, identify
malicious activities, and take immediate action to protect
the network and customer data from cyberattacks.
The United States needs to develop a pro-innovation policy framework
that appropriately manages risk while maintaining U.S.
technological leadership.
ITI's AI policy framework has four key pillars: (1) fostering
innovation and investment, (2) facilitating public trust in and
understanding of the technology, (3) ensuring security and privacy, and
4) maintaining global engagement.\5\ My testimony today primarily
focuses on the first two, although a comprehensive framework should
seek to address all the above policy pillars.
---------------------------------------------------------------------------
\5\ See ITI Global AI Policy Principles, available at https://
www.itic.org/documents/artificial-intelligence/
ITI_GlobalAIPrinciples_032321_v3.pdf
---------------------------------------------------------------------------
A. A pro-innovation policy framework should support innovation and
investment.
While much of the conversation of late has focused on ways in which
to foster accountability, there needs to be at least equal attention
given to fostering innovation and investment. Continued investment in
AI research and development, by both the government and private sector,
is essential for the United States to maintain its leadership position.
ITI applauds the funding and authorizations in the CHIPS and Science
Act for Federal efforts to enhance U.S. technological leadership in AI
and other emerging technologies.
Regulatory policies that encumber the ability of researchers and
developers in the United States will drive investments and research
activities into other countries. Through open-source models and
platforms, access to AI capabilities will be placed increasingly in the
hands of a growing number of innovators of all sizes. This combined
with decreasing costs for AI compute training resources, which are
estimated to decrease 70 percent annually, will allow innovators to
migrate away from jurisdictions with stifling regulations.
Most funding for AI research and development will come from the
private sector. Smart technology investment-related tax policies and
market incentives can encourage greater investments by the private
sector that will produce AI innovations. ITI has detailed its views on
these tech policies elsewhere.\6\
---------------------------------------------------------------------------
\6\ See, e.g., https://www.itic.org/news-events/techwonk-blog/
congress-must-act-to-support-us-research-and-development
---------------------------------------------------------------------------
Government sponsored research in AI also has a role. Government
investments in foundational science and AI-specific program research
are important to fill gaps. Research by academia and the private sector
into privacy enhancing technologies (PETs) and in measurement science
to test, evaluate, validate, and verify (TEVV) model performance are
critical to effectively implementing a risk management approach.
Innovations in measurement tools for AI will make risk management more
concrete and objective and improve accountability and transparency.
The government also has a role in incentivizing professional and
technical apprenticeships, education and training programs in STEM
fields, and promoting access to external and online reskilling
programs. AI is not just a function of STEM or advanced technical
training; the best way to ensure access to an AI workforce is to invest
broadly across all relevant disciplines and teach flexible skills and
problem solving from early childhood education.
B. A pro-innovation policy framework should be risk-based, evaluate the
existing regulatory landscape, and clearly delineate risk areas
that are not adequately addressed.
In seeking to support innovation, it is important that we
understand the risks that we are seeking to address with a regulatory
framework. AI will continue to evolve, and we need to address risks as
they develop, while not suppressing the advancement of AI.
Risks need to be identified and mitigated in the context of the
specific AI use. This will help policymakers determine use cases or
applications that are of particular concern, avoiding overly
prescriptive approaches that may serve to stifle innovation. Beyond
that, context is key. Not all AI applications negatively impact humans,
and thus, they cannot inflict harm that would warrant regulation.
With those risks identified, the next step is to consider the role
for existing statutory and regulatory authorities to address discrete
risk. We don't want layers of regulation that conflict with one
another, create undue burdens on innovators, and slow advancement.
Therefore, it is imperative that the government review the existing
regulatory landscape to assess where there might be gaps. There are
existing laws and regulatory frameworks that can address AI-related
risks, so it is critical to understand how those laws apply, and where
they may not be fit-for-purpose, prior to creating new legislation or
regulatory frameworks pertaining to AI.
As an initial step, policymakers should evaluate how NIST's AI RMF
is being adopted and how it can be used to manage risk. The AI RMF
provides companies with a comprehensive way to think about risk
management practices, which is fundamental to fostering long-term
public trust. It captures many of the outcomes and best practices that
companies are already undertaking, such as framing and prioritizing
risks and addressing AI trustworthiness characteristics (e.g.,
reliability, safety, explainability, privacy, fairness, accountability,
and transparency). ITI and its member companies were active in the
development of this Framework and are actively adopting it. We
appreciate that NIST has also launched the AI RMF Playbook as a
complement to the AI RMF. Indeed, this tool is instrumental to ensuring
that the Framework is actionable and implementable, particularly for
organizations that may be less familiar with the scope of guidelines
and best practices that are available to them. In recent comments to
the Office of Science and Technology Policy, we encouraged the
Administration to explore how the AI RMF might be integrated into
Federal contracts and encouraged the government to leverage the AI RMF
in crafting forthcoming guidance. \7\
---------------------------------------------------------------------------
\7\ See ITI's July 2023 response to OSTP RFI, re: ITI Response to
Office of Science and Technology Policy Request for Information on
National Priorities for Artificial Intelligence, available at https://
www.itic.org/documents/artificial-intelligence/
ITIResponsetoOSTPRFIonNation
alAIPrioritiesFINAL%5B25%5D.pdf
---------------------------------------------------------------------------
Conducting a robust gap analysis of existing legal authorities
relevant to AI's potential harms is critical because there are many
laws and regulations that can address the diversity of impacts
implicated by the technology. Some of these relevant bodies of law and
regulation, coupled with relevant potential AI-related harms, include:
intellectual property law, especially the Copyright Act of 1976, to
address issues related to the use of copyrighted material in training
data and questions regarding the IP rights in AI generated content; the
Federal Trade Commission Act to address unfair, deceptive or abusive
practices related to AI-enabled misrepresentations or harmful content;
product liability common law to address potential safety issues related
to products containing AI technology that may cause physical injury;
First Amendment jurisprudence and Section 230 of the Communications
Decency Act to address issues related to AI-generated content and
freedom of expression interests; Title VII of the Civil Rights Act of
1964 and related laws to address issues related to bias,
discrimination, or other civil rights harms; and relevant Federal
sector-specific privacy provisions, such as in the Health Insurance
Portability and Accountability Act, to address potential privacy harms
related to AI that include the accuracy of data.
In our view, it makes sense to proceed with creating new
legislation only if there is a specific harm or risk where existing
legal frameworks are either determined to be insufficient or do not
exist.
Regarding privacy protections, as noted above, privacy laws do
exist that can address some AI-related privacy harms. Yet, there is
also an undeniable regulatory gap given the absence of Federal privacy
legislation. ITI testified before the House Energy and Commerce
Committee in 2022 in favor of preemptive Federal comprehensive privacy
legislation, which we consider critical to protecting consumers from
data related harms and a necessary complement to any potential AI
legislation or regulation.\8\ However, ITI urges the Committee not to
conflate potential AI legislative provisions with comprehensive privacy
legislation. For example, in ITI's testimony on the American Data
Privacy and Protection Act, we expressed concerns that the bill
conflated the two issues by prematurely including prescriptive
requirements to conduct algorithmic design evaluations and impact
assessments, and that the scope of those requirements, which would have
potentially covered all algorithms, were overbroad and would have swept
in a vast array of technologies well beyond AI.
---------------------------------------------------------------------------
\8\ Testimony of John Miller, Hearing on Protecting America's
Consumers: Bipartisan Legislation to Strengthen Data Privacy and
Security, Before the U.S. Energy and Commerce Committee (June 14, 2022)
available at https://docs.house.gov/meetings/IF/IF17/20220614/114880HH
RG-117-IF17-Wstate-MillerJ-20220614.pdf
---------------------------------------------------------------------------
C. A pro-innovation policy framework should aim to foster public trust
in the technology.
The guiding goal of an AI policy or regulatory framework should be
fostering public trust in AI technology. Fostering trust in AI systems
\9\ requires AI model developers, deployers, and policymakers to work
together. If we are successful in achieving that trust, adoption of AI
by consumers and businesses will increase. AI adoption will benefit the
users of new services, and it will encourage further development and
experimentation in AI and other emerging technology fields, such as
quantum and high-performance computing. Commercial successes will
provide resources to companies that they can invest in AI and other
innovations. In that way, trust is an essential element of a beneficial
research and development cycle for technology and the expansion of the
AI ecosystem.
---------------------------------------------------------------------------
\9\ We define an AI system as a machine-based system that can, for
a given set of human-defined objectives, make predictions,
recommendations, or decisions influencing real or virtual environments.
AI systems are designed to operate with varying levels of autonomy.
This is based on the OECD definition of AI.
---------------------------------------------------------------------------
Transparency is a key means by which to achieve that trust. To
support those efforts, ITI developed AI Transparency Policy
Principles.\10\ Indeed, ITI members are actively taking steps to build
and deploy safe and transparent AI technologies for products and
systems. Transparency is paramount for our member companies,
particularly when it comes to fostering trust in AI technology. They
have placed a premium on these activities. While transparency can take
different forms, our companies are working to ensure that users
understand when they are interacting with an AI system and broadly how
that system works.
---------------------------------------------------------------------------
\10\ https://www.itic.org/documents/artificial-intelligence/
ITIsPolicyPrinciplesforEnabling
TransparencyofAISystems2022.pdf. Transparency is not the only means by
which to foster public trust, and so any approach should consider the
role that transparency can play, as well as other tools.
---------------------------------------------------------------------------
In general, transparency can be understood as being clear about how
an AI system is built, operates, and functions. When appropriately
configured, transparency mechanisms can help to comprehend outputs of
an AI system and foster accountability. Transparency is an overarching
concept, with both explainability and disclosure falling under this
umbrella. In contemplating policy approaches to transparency, we
highlight several key considerations that legislators should consider.
First, like with policy approaches to AI generally, transparency
requirements should be risk-based. It is important to consider the
diversity of possible AI use cases and applications, given that the
demand for transparency requirements from various users may vary
significantly based on the AI application or intended use. Many use
cases present little to no risk to the user, and so imposing
transparency requirements in such situations will likely add little
value to the user and hinder innovation by adding onerous,
disproportionate requirements.
Second, in thinking about transparency, it is important to consider
the objective and intended audience. The target audience at which
transparency requirements are directed, including their level of
expertise, plays a key role. For example, transparency could be useful
to several different audiences (e.g., regulators, consumers,
developers, etc.), which will in turn influence the development of
potential provisions. Understanding the intended audience will also
inform the type of information presented, the way it is presented, and
the amount of information presented.
Third, disclosure can play an important role in facilitating
transparency. AI systems should be disclosed when they are playing a
significant role in decision-making or interacting directly with users.
In the context of disclosure, language should be plain and clear so
that it is understandable to a wide audience. Disclosures should
generally include high-level information, such as a topline explanation
of how an AI system works, capabilities, and known limitations on
performance. Additionally, disclosure should be the responsibility of
the deployer to ensure that disclosure and other consumer-facing
obligations are met. That said, the developer of the AI system should
ensure that terms of sale do not prohibit disclosure. Relatedly, we are
supportive of information-sharing in the value chain to facilitate
cooperation between developers and deployers.
Finally, Congress should avoid an overly prescriptive approach to
transparency and maintain appropriate IP protections. It is important
that transparency requirements allow for flexibility because it may not
be appropriate or useful to provide the same type of details in every
context or for every target audience.\11\ Organizations should have the
ability to tailor such information, depending on context, use of, and
level or risk associated with the system. Also important is that
transparency requirements do not require the disclosure of source code
and sensitive IP, or otherwise reveal sensitive individual data. Any
requirements around transparency should avoid requiring companies to
divulge sensitive IP or source code. Disclosure of source code could
seriously put at risk trade secrets, undermine IP rights, and
contravene widely accepted best practices for digital trade. It could
also pose risks to safety and security and allow malicious actors to
manipulate an AI system.
---------------------------------------------------------------------------
\11\ Using AI to limit fraud, spam, illegal, or malicious
information are some examples of where including technical details or
too prescriptive of a disclosure may be inappropriate.
---------------------------------------------------------------------------
D. A pro-innovation policy framework should consider the views and
input of all stakeholders.
Both developers and deployers of AI systems should be consulted in
the development of policy frameworks, as well as civil society,
academia, and companies operating across different sectors. Small,
medium, and large companies in all sectors will be using AI to be more
efficient and offer better quality services and products, so it is
imperative to cast a wide net to obtain a diverse set of perspectives.
Additionally, a pro-innovation policy framework should seek to
appropriately delineate the roles and responsibilities of different
stakeholders in the AI value chain. Stakeholders throughout the value
chain, including small businesses across a variety of sectors, play a
role in the development and deployment of AI in a responsible manner.
As such, responsibilities should reflect the important distinction
between developers and deployers and be allocated among actors based on
their role and function in the AI value chain.
E. A pro-innovation policy framework should rely upon globally-
recognized international standards.
Before adopting new AI regulatory requirements, policymakers should
understand the status of international consensus-based standards and
the ability of those standards to meet regulatory requirements. AI
standards are essential to increase interoperability, harmonization,
and trust in Al systems. They can inform AI regulation, practical
implementation, governance, and technical requirements. Without
specific standards for risk management processes, such as the
measurement and evaluation of models, it will not be possible to
implement regulations effectively. Moreover, regulations that are not
aligned with international standards will undermine the leadership of
companies doing business in the United States that seek to scale into
other jurisdictions that adopt international standards.
The International Standards Organization and International
Electrotechnical Commission have formed a Joint Technical Committee for
IT standards, and it has established a subcommittee on Artificial
Intelligence (SC 42). That subcommittee has published several AI
standards and is working on an Artificial Intelligence Management
System (AIMS) standard that will cover processes for the development or
use of Al, such as bias, fairness, inclusiveness, safety, security,
privacy, accountability, explicability, and transparency.\12\ This
management system standard will help advance innovation and technology
development through structured governance and appropriate risk
management. SC 42 currently also has other standards under development,
focused on terminology, reference architecture, governance of AI, and
trustworthiness. We encourage Congress to consider how standards can
address risk management requirements and ensure international
harmonization.
---------------------------------------------------------------------------
\12\ See ISO/IEC 42001 Information technology--Artificial
intelligence--Management system.
---------------------------------------------------------------------------
III. Conclusion
To lead in AI, the United States needs a pro-innovation policy
framework that prioritizes innovation and investment as well as
building public trust in AI. Public trust will increase AI adoption by
businesses and consumers, and accelerate the flourishing of the AI
ecosystem of developers and deployers.
Building public trust through risk management will be critical, and
the private sector is already leading in this area, such as through
adoption of the NIST AI RMF. The government's role should be limited to
addressing critical risks in specific use cases. Where those risks are
identified, Congress should evaluate the existing legal and regulatory
landscape, and clearly delineate risk areas that are not adequately
addressed, before enacting new legislation to address any gaps. Future
requirements should be aligned with international consensus standards
wherever possible to ensure that risk management is effective and to
harmonize the global marketplace for technology.
Thank you, and I look forward to your questions.
Senator Hickenlooper. Great. Thank each of you for being
here, for your opening statements, for all the work you have
done on these issues already.
This first question I have might be considered obvious. AI
systems learn about the world by processing written, visual,
and audio files, works created by people, by humans.
Dr. Krishnan, just to kind of lay out, because a lot of
people don't--or come into this issue fairly freshly; what
rights already exist for consumers to decide how AI systems
access their creations, and what additional rights do they
have?
Dr. Krishnan. Can you hear me now? Okay. Thank you. Thank
you for the question. So perhaps a place to start is to think
about the AI pipeline first, and then I will focus in
particular on the creators of content, and the need to
appropriately balance what a typical creator's interests are,
and how that might be protected.
So we think about the AI pipeline, it includes training
data, models, and applications. When you have data that models
use, that involves creative artifacts, be it music, images,
video, that is copyrighted, a group of creators may actually
want to seek advertising off this content that they have
created, and therefore may pull such information on the web,
with the full expectation that consumers interested in this
content may sample it, and then they may be able to earn income
from it.
The concern here is that if this content is then scraped
and used in AI models, which then produce output in the style
of the same kind of content that the creators have created,
that could potentially take away revenue from the individual
creators who have created this content. So this issue on the
one hand of protecting creators, who would like to have income
be generated from the creative acts that they are engaged in,
on the one hand.
The other is the capacity of these models to use these
types of data for the purposes of creating the capabilities
that we have witnessed. So one potential path forward, on the
one hand, you want to have to copyright, and the benefits that
accrue from licensing that come from that, is perhaps to use
technology, as work from the University of Chicago, that allows
for individuals to upload their creative content, and the
technology makes modifications for that content which is not
visible to the human eye.
So the human sees it as an image, just like the artist
intended it to, but it is not trainable by an AI model, so that
the AI model can't produce it in the likeness of the artist.
And if the model developer wants to obtain access to that
content they can licenses it, and that might be, potentially, a
way of, on the one hand, providing notice and disclosure, which
currently doesn't exist through those people who have created
this content, whose content got scraped, while at the same
time, meeting the needs both of the model developer, and of the
artist.
Senator Hickenlooper. Right. Got it.
Mr. Gregory, what would you add to that, as the next step?
That is a broad brush, to start with.
Mr. Gregory. I think there are also ways in which content
developers creating content--excuse me--thank you--people
creating content can also choose to add information to that
data, so there are ways we can do this at the level of very
minor modifications. There are also ways in which you could be
tracking those desired usages, for example, using the C2PA
standard.
So I think the more options we can give people that are
standardized for understanding the choices they make about the
information they consume, but also the information they place
online, would be appropriate.
Senator Hickenlooper. Great. Thank you.
Ms. Espinel, what are some evidence-based steps that your
member companies have been using to develop AI with the safety
of consumers in mind, but in that notion of the AI Risk
Management Framework?
Ms. Espinel. Thank you. So BSA members have--many of them
have implemented very extensive risk management programs that
include, as part of that, impact assessments. And I will talk a
little bit about what they do, and how they can use evidence to
make the kinds of determinations that you are talking about, to
both increase transparency, but also reduce the risk of bias
and discrimination.
So as an example, if a BSA member is acting as a developer
of AI, they can assess the training data that they are using to
ensure that it is representative of the community. And they can
use the evidence that is gathered from that assessment to
ensure that the risk of bias and discrimination is as low as
possible.
That is certainly in line with the NIST AI--the risk
management framework that was developed by NIST. But I would
say, as important as the NIST Risk Management Framework is, and
as much commendation I give to the Department of Commerce for
coming up with it, we don't think it is sufficient to--we think
it would be best if legislation required companies in high risk
situations to be doing impact assessments and have internal
risk management programs.
So yes, there is much that our companies have been doing. I
think there is probably much that many companies have been
doing, but we think in order to bring clarity and
predictability to the system, and to ensure that use of
artificial intelligence is as responsible as possible for
Congress to require impact assessment, and require risk
management programs, is essential.
Senator Hickenlooper. All right. Thank you very much. I am
going to come back for more questions later.
Ranking Member Blackburn.
Senator Blackburn. Thank you so much. Ms. Espinel, I want
to come to you first. Because I know that the EU is looking at
implementing their AI Act later this year. I was recently over
there, and working on the privacy issue, and holding some
meetings, and I think they are a little bit frustrated that we
haven't moved forward on some of these guidelines and
governance that would help our innovators here in the country.
So talk for just a minute about how your members would
navigate a patchwork system around the globe when it comes to
AI and the guardrails that are going to be put around it?
Ms. Espinel. Thank you very much. So let me start off my
thanking you. When you were in Brussels, you had an opportunity
to visit us at BSA.
Senator Blackburn. Right.
Ms. Espinel. And visit many of our member companies. You
led a fantastic roundtable there. So thank you very much for
that. But also thank you for that question, because I think
that is a very important issue.
So as you said the EU is moving forward with legislation.
They are not the only country that is moving forward with
legislation; governments around the world are moving forward
with legislation as well. And I think one of the challenges,
but an avoidable challenge, is if we end up in a situation
where there is, as you said, an inconsistent patchwork of
regulations around the world.
I think because there is so much policymaker focus around
the world on artificial intelligence, as you said, in part,
because of the launch of ChatGPT, there is a real opportunity
right now to move towards a harmonized approach globally. And
that may not include every country in the world, but I think it
could include a lot.
And I think if the United States--I think the United States
has a very important role to play there, in terms of moving a
large number of countries to a harmonized approach.
Senator Blackburn. Should we be moving to set standards and
lead this?
Ms. Espinel. In my opinion, yes.
Senator Blackburn. Okay.
Ms. Espinel. The United States is leader in innovation; we
should be a leader here as well.
Senator Blackburn. Absolutely. Mr. Strayer, when you are
working with your members, the National Data Privacy Law, how
did they put importance on that before we move forward with
some of these other components dealing with AI?
Mr. Strayer. We believe a comprehensive national privacy
law is absolutely critical, that ensures a lot of the issues
that will come about, about data training sets, and other data
that emerges from AI systems that are being used by businesses
every day. So we pretty much support that--you know, acting
quickly, we don't think those need to be done--but needs to be
done first before moving on AI regulation, but we think both
have to be done.
The thing I would say about standards is, U.S.-based
companies and western companies, generally, are leading in
developing standards through the International Standards
Organization. They are working now on an AI management system
standard. Those will be, hopefully, the bedrock for when EU
finishes their legislation, the standards that should apply
globally, but that has not been fully resolved by the European
Union. But those standards should be to harmonize global
standards for the future.
Senator Blackburn. Yeah. Mr. Gregory, do you have anything
that you want to add on that?
Mr. Gregory. I would note that one of the areas that the EU
is focused on is labeling and disclosure of AI-based content. I
think there is a real opportunity for the U.S. to lead on this,
to come up with a standardized way that respect privacy, that
presents information that is useful to consumers, and to set a
standard there that is applicable as well.
Senator Blackburn. Yeah. Dr. Krishnan?
Dr. Krishnan. I think the missed AI RMF offers an
opportunity here through what is called the NIST AI RMF
profiles, and through the use of these profiles I think we
could, with the appropriate standard-setting for high-risk
applications, both on the data--input side, the data
transparency side, as well as with model validation. We can
actually come up with standards that then get adopted, because
it is considerable support for AI RMF, both here at home and
abroad.
Senator Blackburn. Thank you. Ms. Espinel, let me come back
to you. I have got just a little bit of time left. I would like
for you to just briefly talk about how your companies have
worked to have policies that are transparent, that are
interpretable, that are explainable, when it is--when you are
dealing with AI?
And Mr. Strayer, I will come to you for the follow-on on
that.
Ms. Espinel. So it is very important that we have a system
that builds trust in AI, and transparency is clearly part of
what is important in order to build trust. Let me give you just
a few examples of ways that there could be transparency. One is
to let consumers know if they are interacting with an AI
service, such as the chatbot.
Another example would be, let consumers know if they--if
the image that they are seeing has been generated by artificial
intelligence. There is important work that is being done by
other witnesses at this table, in terms of content
authenticity, and letting consumers know if images have been
altered or manipulated in some ways.
Those are all--those are just three examples. But I want to
end by saying, in addition to transparency practices I do think
it is very important that we have regulation, and that we have
regulation that requires high-risk uses for companies that are
developing or using AI to be doing impact assessments, to try
to mitigate those risks.
Senator Blackburn. Mr. Strayer.
Mr. Strayer. I would just add that companies are also
setting up fact sheets for what they call model cards that
explain the features and limitations, to talk about where the
datasets came from, intended uses. So these are, you know,
pretty fulsome explanations. It is important in the area of
transparency to think about who the intended audience is, and
for what purpose.
So is it a consumer, is it business as well on the chain?
Is it for the deployer? So one should think about all those
when they think about what requirements should be set in this
area.
Senator Blackburn. Thank you.
Senator Hickenlooper. All right. Thank you.
Senator Moran.
STATEMENT OF HON. JERRY MORAN,
U.S. SENATOR FROM KANSAS
Senator Moran. Chairman Hickenlooper, thank you. Thank you
all for your presence and testimony today. It is annoying that
we are here now on AI when we have been unsuccessful in
reaching conclusions on data privacy legislation. It just seems
like one issue piles up after another, both of huge
significance.
Ms. Espinel, let me start with you. I want to talk about
the NIST AI Risk Management Framework, launched after NDAA
authorization in 2023. I am working on legislation, in fact,
authored an amendment to NDAA this year, that would require
federal agencies to apply the AI framework when using AI
systems. An attempt to ensure that government acts responsibly
in implementing AI systems, and in a manner that limits
potential risk, not only to Americans and their data, but to
governmental agencies and their missions.
Can you talk about, Ms. Espinel, about the implementation
of policies based on NIST AI Risk Management Framework that can
establish a baseline of good behavior when implementing
artificial intelligence systems, which can actually unleash
beneficial AI technologies instead of just hindering the
development of AI?
Ms. Espinel. I would be delighted to. It is very important
the NIST AI framework is flexible, it provides a roadmap for
companies in terms of how they can put practices and policies
in place, to responsibly develop and use AI. We support it
being used by the U.S. Government, we support it being used in
the context of procurement.
And I will close by saying, I think it is the place, that
you kind of alluded to at the end, and as Ranking Member
Blackburn raised, where the U.S. can show leadership. I think a
model or an approach similar to the NIST AI Risk Framework, is
one that could be usefully adopted by other countries as well;
and so very supportive of that. Thank you.
Senator Moran. Thank you. Mr. Strayer, I am going to go to
you based upon your past history at--as Deputy Assistant
Secretary for Cyber and International Communications and
Information Policy at the State Department. That is a long
title.
Mr. Strayer. They give you long titles at the State
Department.
Senator Moran. Yes, sir. I hope the pay is commiserate.
[Laughter.]
Senator Moran. Let me suggest the story to you of Huawei
launching a phone last week, containing a suspected homegrown
semiconductor that represents a leap forward in their ability
to produce advanced microprocessors. Despite the efforts by the
U.S. Department of Commerce to deprive that company of U.S. and
partner technology to develop and manufacture advanced
semiconductors, a lot of details yet to be known about that.
As part of that related effort to deny China the ability to
develop these technologies, in August, President Biden issued
an Executive Order limiting outbound investment of Chinese
companies that develop advanced technologies, including AI.
What are the national security implications for the U.S. if
adversarial nations take a leading position in the advancement
of artificial intelligence? Do you believe the U.S. can
appropriately mitigate this risk through the current strategy
of denying access to key technologies and investments? And what
can we learn from our past and present efforts at restriction
in the development of semiconductors?
Mr. Strayer. Thanks Senator. That is quite a compound
question I will try to do my best job to----
Senator Moran. It goes with your title.
Mr. Strayer. Touche. So first is, to maintain our
leadership, we really also--we need to focus on running faster,
that is how we innovate and make it to the next cycle of R&D in
a position that puts us ahead of our adversaries, the United
States, that is. So we need to continue to develop the best
technology for all purposes, which will obviously benefit our
military and national security uses for those.
On the defensive side, we have seen the October 7th, of
last year, export control executive order regulation, and now
this most recent outbound investment restriction.
We are still seeing those play out, there is open comment
periods on these, they are trying to tighten them up. So we
will see all those play out. There is a very important issue,
though, with these. And that is, we really need to have a
strong discussion with the private sector about how you can do
the minimum amount of damage to future R&D and revenues for
U.S.-based companies, while ensuring that none of these tools
end up in the hands of adversaries, where we are going to use
those for military purposes.
So really sharpen your focus on where it might be used for
dual use or military purposes, while benefiting U.S. and
Western companies to our asymmetric advantage, because they
need to keep maintaining those revenues over the long term. So
I think we need a stronger focus on working with the private
sector to get that balance, right, that is stopping the
military uses of other countries, and enhancing our own use and
market competitive development of the technology.
Senator Moran. Thank you. I have a long list of questions
as well. But my time, at least at this moment, has expired.
Senator Hickenlooper. We will have a second round. Did you
feel--did you get an answer on the restriction--is the
restriction effective? Because I am not quite sure we all--you
know, I would want to hear a little more about that?
Senator Moran. As long as that is on your time, it is a
great question, Chair.
[Laughter.]
Senator Hickenlooper. My time is your time.
Senator Moran. Good.
Mr. Strayer. So these restrictions are quite robust, in the
case of export controls from the October 7th regulation that
the Commerce Department issued on more advanced semiconductors.
With regard to the outbound investment restrictions, they are
starting more narrow, and they are doing rulemaking through the
Treasury Department on these. I think that is a smart way to
start, and then start to expand if they need to beyond that.
The really key issue with these restrictions is that we
don't want to force all technology--or key innovative
technology to move outside the United States, so we need to
bring allies and partners along with the United States, in the
case of export--and the export controls for semiconductors,
Japan, and others have come along, Taiwan, come along with the
United States, the Netherlands, on the equipment as well.
That has not yet occurred on the outbound investment
restriction, so I think one needs to think about how that is
done in a multilateral way, so that we are not making United
States a place where you don't do investments in these areas,
and we are ceding that leadership to other--even Western
countries.
Senator Hickenlooper. Isolating ourselves. Got it. Thank
you.
Senator Klobuchar, we have remotely, for a few questions. I
can't hear you.
STATEMENT OF HON. AMY KLOBUCHAR,
U.S. SENATOR FROM MINNESOTA
Senator Klobuchar. -- Judiciary, so I truly appreciate you
taking me remotely, and with many of the same issues. We are
focused on many things with AI, and we should be from the
security risk to the effects on innovation to, of course, the
potential that we have to use the technology for good. And one
of the things that we are really concerned about, which is off
this topic a little, but it is just, to always keep in mind, as
the experts that you are, is just the effect this is going to
have on our democracy and our elections.
And so two things, one is that Senator Hawley, and Senator
Collins, and Senator Coons, and I just introduced a bill in the
last hour about deceptive AI-generated content. And that is the
most extreme, right. That is the stuff where you have got
people acting like they are the candidate when they are not,
which is going to AI generated images acting like they are a
candidate when they are not, which is going to create complete
confusion, regardless of whether it is labeled or not.
And I thought it was really important to do this on a
bipartisan basis, so I hope my colleagues will, and you can't
get much more bipartisan than Hawley, and Collins, and
Klobuchar, and Coons. And so I hope that my colleagues will
look at this. It is about fraudulent AI-generated content in
political ads.
The second thing that we are looking at, and for another
class of AI-generated political ads, would be disclaimers, and
watermarks, and the like, for the ones that don't meet the
standard for the most extreme, deceptive, with of course,
exception for satire, because we all like a lot of satire
around here. Okay. So I am just going to just ask you about the
watermark piece, because we just introduced the other bill, the
disclosure piece.
Mr. Gregory, and Dr. Krishnan, do you agree that without
giving people information to determine whether an image or a
video is created via AI, the generative AI poses a risk to our
ability to hold free and fair elections?
Mr. Gregory. The evidence already suggests that this is a
problem both in the U.S. and globally given the capacities of
these tools, so yes. I also believe that election content is a
first place where it is possible to start with both visible
disclosure and particularly indirect disclosure by labeling a
piece of content, and also providing metadata that could
explain it. The key part would be to protect the capacity for
satire, as you note, that is essential to protect, yes.
Senator Klobuchar. Okay. Well, I really do appreciate that
answer, and also the timeliness of it, given that we are in
election season. We have already seen the use of it against
some of my colleagues on both sides of the aisle. And so people
are very aware. I think it also, by the way, extends, and I
will get your answer, Dr. Krishnan, in writing, if that is
okay, because I want to move on to another question.
I wanted to move into the AI risk assessment and mitigation
issue. we know that these systems have the potential to impact
individuals in many key areas, especially if it evaluates
rental insurance applications. I am working with Senator Thune
on a bill to require companies that develop and use AI to
identify risk and implement procedures to mitigate risk. This
involves Department of Commerce Oversight.
Ms. Espinel, do you agree that both developers and
deployers of the AI systems bear responsibility to mitigate
risk before they release the AI on the market?
Ms. Espinel. I believe that both developers and deployers,
users of AI, should have obligations. I believe they both have
responsibility. I would emphasize that I think the obligations
that are put on them should differ, depending--so that it
reflects what they do. So a developer of AI is going to have
information about how the data--the data that was used and how
the AI was trained.
To use an example of a deployer, a bank is going to have
information about how loans were actually made, and whether or
not loans were made in a way that was disproportionately
negatively impacted a part of the community.
So 100 percent, I agree with you. And thank you for
thinking about the distinction between developers and
deployers, the fact that both should have obligations, and that
those obligations and requirements should reflect what they do;
the information that they are going to have about the AI system
in question, and the different steps that each can and should
take to identify and mitigate risk.
Senator Klobuchar. Thank you. We have also seen
intellectual property issues with AI, songs, and the like,
copyright, only around half the states have laws, that give
individuals control over these, their name, image, and voice.
Do you agree, Ms. Espinel, that we need stronger
protections for the image, likeness, and voices of creators?
Ms. Espinel. I know there have been instances where AI--
there has been AI-generated information or content that has
pretended to be someone that it wasn't. That is clearly wrong,
and should be stopped. I am thinking about writer publicity, as
you point out, that is not something that exist consistently
throughout the United States. And so I think thinking about
solutions to that problem, which is clearly wrong and needs to
be addressed, is very important.
Senator Klobuchar. And just one quick other question of
you, Mr. Gregory. In addition to some of the copyright issues
we are talking about, we also have journalism issues, and we
have--Senator Kennedy and I have a bill that allows the
companies to--the companies to have to negotiate with news
organizations on the issue of their content. And in addition to
newspapers, AI systems are trained on other content like
lifestyle magazine, most of which were not compensated for that
content.
Do you agree, Mr. Gregory, that there is more we need to do
to ensure that content creators are fairly compensated for
their contributions to AI models?
Mr. Gregory. Yes. I think there needs to be stronger ways
to understand which content is being ingested into AI models,
and the decisions that are made around that. And I would
particularly highlight that journalists already face
significant pressures, including of course, that they are
unable to detect AI generated medias; they face pressures both
if their content is ingested, and also that they are on
frontlines of defending the truth in the context we face now.
Senator Klobuchar. Thank you very much. Appreciate it, all
of you. And thank you to the Chairman and Ranking Member for
having this hearing.
Senator Hickenlooper. Thank you, Senator. Appreciate that.
We don't have the next video--senator to question, so I will
weigh in--oops--there we have Senator Young, just in the nick
of time.
STATEMENT OF HON. TODD YOUNG,
U.S. SENATOR FROM INDIANA
Senator Young. Thank you, Chairman, for acknowledging my
presence in this room, and for chairing this hearing. I thank
our witnesses for being here. I will just dive in. I know you
have been responding to a number of inquiries.
I will begin with the observation that artificial
intelligence wasn't something that we invented yesterday; it
has been around for decades now. In fact, for years we have
seen AI technology in products and services across nearly every
sector of our economy, and in a wide variety of use cases.
Analysis of each of these use cases, and concerns of an AI-
enabled society should, in my view, start with the same litmus
test. Does existing law address whatever potential
vulnerability we are concerned about?
I have found through my interactions with a number of
experts in this area, that existing law would address the vast
majority of concerns that we have, not every one though.
We have to closely evaluate and carefully target areas
where existing law doesn't address these vulnerabilities, that
is why, of course, we are here today, to identify high risk use
cases of AI, and discuss potential guardrails to minimize those
risks. Recent advancements, in generative AI platforms, like
ChatGPT, have raised concerns among my constituents, and many
others, about a dystopian future. Straight out of science
fiction, that could be right around the corner.
The truth is, nobody knows what future this incredible
technology will usher in, and it is human nature to fear
uncertainty. History shows innovations that democratize access
to information, into media, the printing press, recorded sound,
film, have been met with similar concerns, usually exaggerated
concerns. But history also shows these innovations have brought
great benefits to society, to national security, and the
economy.
As we evaluate the need for any level of AI regulation it
is important we don't lose sight of the many potential benefits
that harnessing the power of AI presents. These benefits
include: self-driving cars, medical advances, immersive
technology, educational opportunities, and more. So I sort of
want to get on record, a high-level perspective.
With that said, you are here today, and let us not focus on
the unknowns but rather the knowns, the here and now risks. As
we think about trust, transparency, and exploitability within
AI, the goal is not to stifle growth, but rather to increase
adoption, and innovation in the long term.
Ms. Espinel, can you briefly discuss two things: first, the
important distinction between a developer and a deployer? And
then second, how should Congress think about the role of
transparency between businesses and consumers as opposed to
transparency between businesses and government? And I ask that
you answer these pretty tightly, if you could, so I can ask
some follow-up questions. Thanks.
Ms. Espinel. Thank you. So developers and deployers,
developers and users, developers of AI systems are the ones
that are producing the AI system, they are creating the AI
systems, and the deployers are the users. To give an example, a
developer is a software company that is developing a speech
recognition system, and a bank is using, a deployer, using an
AI system to help make determinations about who should get
loans.
Those are very distinct roles, and the developer and the
deployer will know very different things about the AI system,
both how it is being developed, and how it is being used. And
because they know different things, they will be able to do
very different things in terms of addressing and mitigating
that risk. So as you were thinking about legislation, clearly
distinguishing between developers and deployers is very--is
critical, in order for the legislation to be effective and
workable.
In terms of transparency, you also alluded to the fact
that--or mentioned the fact that AI has been used for a long
time, right. It has. It is also used in many different types of
circumstances, and some of those are high risk, and some of
them are lower risk. And it is our believe that in high risk
situations, so for example, making a--a government making a
determination about access to public benefits as an example; if
there are consequential decisions that impact a person's
rights, we believe there should be legislation requiring that
an impact assessment be done, and that those risks be
mitigated.
But there are also uses, as you have said, that have been
around for a long time, that are relatively low risk. So
reminding me when I send an email that I may have left off an
attachment. Or one that has been quite popular lately,
adjusting, you know, the background, if you were on a video
conferencing call. Those are relatively low risks, and having
impact assessment required in those cases, we believe would be
overly burdensome, and not add a lot of value.
But where there are other consequential decisions, whether
by--whether companies to consumers, or government to its
citizens, we believe impact assessment should be required.
Senator Young. Well, thank you. Does anyone else want to
chime in on any of those points? Otherwise, I will turn to Dr.
Krishnan. Okay.
Doctor, what are the clear high risk use cases, to your
mind, for AI that Members of Congress should be thinking about
right now?
Dr. Krishnan. The ones that come to mind immediately are
autonomous vehicles, health care, hiring, recruiting, housing,
you know, there are--important scarce resources are being
allocated via AI, those would all represent where there is harm
either to the individual, or to society if things didn't go
well.
Senator Young. Right, right. And so your bringing up the
use case, I am not surprised by. And then you would probably
acknowledge there some in the national security realm----
Dr. Krishnan. Without a doubt, yes.
Senator Young. Okay. Okay. I guess the last thing I would
like to ask is, stepping back, AI has of course garnered all
sorts of attention over the last number of months. Is AI all
that different, and I will ask Dr. Krishnan, from other major
technological advances? Or is it just the latest shiny object
that we are attentive to? Why is this so important, or perhaps
you would just say, no, this is like every other technology?
Dr. Krishnan. At one level, it is like other technologies
we have dealt with in terms of having this impact on the way
work is done, tasks that are impacted. And others are special
characteristics of the technology in terms of working with the
kinds of modality, audio, video, text, things that we haven't
typically seen has not been part of a technological capability.
Like you mentioned ChatGPT in your opening remarks, that
kind of capability was not something that--at least the typical
citizen was thinking that this was something that a computer
could do. And so the difference I guess, is also in terms of
how the technology is able to learn over time, with data. So
there are some differences that are technical differences with
regard to this technology, and then there are differences with
regard to how to govern the use of this technology.
And that is why in my testimony I talk about data
transparency, and model transparency, and having standards for
high-risk applications, then also having this trust
infrastructure, because you can't predict exactly how this
technology is going to evolve, to ensure that we are able to
capture vulnerabilities, deal with failures, and come up with
solutions that can be disseminated.
Senator Young. On the fly, right; the trust infrastructure?
Dr. Krishnan. Yes.
Senator Young. I guess the----
Dr. Krishnan. It is like the search for AI is how I think
about it. Like what we have done for cybersecurity, sir.
Senator Young. Maybe I can pull on that thread, just ever
briefly, and end--you can respond to what for me has been an
observation. I will leave it to you to determine whether or not
it has been an insight. But my conclusion is, we are going to
need a lot more expertise on this technology, a lot more
sophistication within government, in individual agencies,
perhaps at the White House, so that on an ongoing basis, we can
figure out how to apply existing statutes to emerging threats,
or concerns, or challenges, or opportunities.
Dr. Krishnan. Mm-hmm.
Senator Young. Or to flag when new legislative authorities
may be needed. Is that your estimation? The human resources
challenge within government?
Dr. Krishnan. Yes. And in industry as well, so I think a
scholarship for a service-type program, for AI, would be very,
very valuable.
Senator Young. Thank you.
Senator Hickenlooper. That last point was worth the price
of admission. Thank you, Senator.
And I couldn't agree more. I think that if you look at
the--try to estimate the cost of government keeping up with the
rate of change and the innovation that is going to be required,
it is a staggering thought.
And I have a son in college, and all the kids that are in
STEM are looking at high-paying jobs right out of school just
to start without the experience to be able to help government
keep pace, and that it is that the competition is going to just
fuel greater intensity and greater inflation among the wages,
which again, is a good thing for the kids, but hard for the
success of government relations within the industry. Thank you.
I am going to go into the second round. I could do a third
and a fourth round. I will probably try and let you out of here
by 4:00, a little bit after.
Mr. Gregory, you recently co-chaired the Threats and Harms
Task Force within the Coalition for Content Provenance and
Authenticity, my staff tells me it is referred to as C2PA. C2PA
refers to the provenance as the basic trustworthy facts about
the origins of a piece of digital content. This could help
users distinguish between human- and AI-generated content. It
could reveal personal information about the creator of that
content as well.
So the question I have is, how do you--how do we protect,
how do you protect, how do we protect the privacy of content
creators while being transparent with consumers about the
content they see when they are online?
Mr. Gregory. As the Senator notes, I co-chaired the Threats
and Harms Task Force. I don't speak for the C2PA here, but I
will make some observations about how we protect privacy in
this context.
I think this is absolutely critical. A starting point is to
recognize we are moving into a much more complex media
ecosystem. So the idea of understanding how media is made, how
it has evolved, right where the AI has come in, where the human
element has come in, I think is going to become increasingly
important.
When we think about that future, though, we need to make
sure that these systems are not either accidentally or
deliberately, perhaps by authoritarian governments who might
adopt them, systems, which they can track people's activity.
That is why when we start looking at these types of approaches,
we have to start from the principle that they do not oblige
personal information or identity to be part of it.
That is particularly easy with AI-generated content because
really what matters with AI generation is ``how'', not ``who'',
right, the AI was used. The ``who'' could be helpful, but it is
not necessary, and it could be helpful to the wrong people. So
when we start from that premise, I think that is very important
as Congress looks at how to standardize this, and how to
approach this, that they understand how we are going to reflect
the evolving nature of media production in a way that protects
privacy, doesn't oblige personally identifiable information,
and will be usable worldwide.
Senator Blackburn's question earlier about how the U.S. can
compete. We should be developing standards that can be
applicable globally. That means they need to be accessible,
privacy-protecting, and usable in a variety of contexts
globally.
Senator Hickenlooper. Absolutely. Ms. Espinel, increasing
the transparency of AI systems is one of the key vehicles by
which we gain confidence and trust among the users. Without
appropriate guardrails around the risks from AI, I think
developers will struggle to compete in the U.S., and certainly
internationally as well. So it is in our best interest to
demonstrate leadership, and safe and responsible deployment.
What transparency steps do you prioritize; do you think are
most crucial in terms of gaining the trust of consumers?
Ms. Espinel. So what if I start off by saying that I think
building trust, as you say, is important for our ability to
compete. I think it is important for the U.S. economy, and it
is obviously important in terms of protecting consumers. So I
think that is an absolutely critical step.
Impact assessments are a tool that we think that
organizations should be using whether--again, whether they are
creating the AI or they are using the AI. If they are in a
high-risk situation, if the AI is being used to make a
consequential decision, then impact assessments are an
accountability tool that should be required. And by requiring
impact assessments, you will increase transparency.
Consumers need to have confidence that if AI is being used
in a way that could have an impact on their rights, or have,
you know, significant consequence for their life, that that AI
is being vetted, and that it is being continuously monitored to
be as safe, as secure, as nondiscriminatory as possible.
And so I would go back to saying having a requirement for
impact assessments by developers or deployers in high-risk
situations, having a strong national law from the United
States, I think is very important in terms of protecting
consumers and our economy.
And then going to your last point, I think it is also very
important for the United States to have an affirmative model of
effective legislation when we--when other countries are moving
quickly to regulate. And I think having the United States be a
leader in shaping that conversation, and the global approach to
responsible AI is critically important.
Senator Hickenlooper. And an affirmative model; what a
concept.
Ms. Espinel. Mm-hmm.
Senator Hickenlooper. Dr. Krishnan, you have done a lot of
research with consumers and social behavior within digital
environments, so on that same subject, what information should
be disclosed to consumers to establish trust in online services
around AI?
Dr. Krishnan. Well, first and foremost, I think when you
are interacting with an AI system I think you need to know that
you are interacting with an AI system.
Senator Hickenlooper. Okay.
Dr. Krishnan. So disclosure; that is the first step. The
second is if data is going to be collected by the -- by, let us
say, a ChatGPT or Bard during your interaction, you should be
explicitly given the option of opting in for the purposes of
saying, ``Is my data then going to be used by the AI to
further--for training purposes?'' That, I think, we can learn
from much of our prior work in privacy to apply it in this kind
of context. So the opt-in.
And then the third, I think, is with regard to the trust
interaction that individuals have. You know, to a large extent,
individuals build trust based on their own experience, much as
we talk about data transparency, model transparency. My
interaction with this tool: Does it actually behave the way I
expect it to? That actually builds trust over time. And I think
it is a combination of these that will result in the intended--
you know, what we would like to see as an outcome.
One quick additional point I want to make is, while we have
been talking about NIST, RMF, and the like, I think it would be
great to have demonstration projects for citizens to recognize
the value that AI can actually bring to the table. ChatGPT was
perhaps the first time that they got to interact with AI on the
kind of scale that we thought they were. It would be great to
see something like the Khan Academy's education products,
things of that nature. It gives them a clear indication of the
value that this brings. I think that would be very good too.
Senator Hickenlooper. Couldn't agree more. I am going to
sneak in one last question before the Chair comes back. I don't
know if you guys play bridge, but the Chair trumps every suit.
To be clear.
Mr. Gregory, let me go to my--switch to my question for Mr.
Strayer. The AI ecosystem can be generally viewed as those that
develop AI and those that use AI. As we have heard data is
nuanced--there is a lot of nuance around that; risk management
principles should be tailored both to developers and users,
employers, and certainly there is not going to be any one-size-
fits-all; there is no silver bullet. How can we create an
oversight and enforcement framework to make sure that we can
hold bad actors accountable? You know, people that use the AI
systems maliciously?
Mr. Strayer. Well, on the true malicious actors out there,
there is going to need to be, you know, law enforcement
cooperation and also enforcement of some of our existing laws.
When it comes to standard risk management, a number of the
appropriate risk management tools are going to make the model
more resilient, more robust, less susceptible to compromise and
manipulation. So those are important steps.
The other thing just to keep in mind is, at these risk
management steps, you know, there should be higher risk
management for the highest risk use cases, and lesser
requirements on something that is just doing predictive text in
an email.
And then finally, also to think a little bit about how
small businesses and those that might be just doing
experimentation that aren't planning for commercial deployment
might be required at a lower standard than those that are going
to make massive commercial deployments of things.
Senator Hickenlooper. That is such a good point; the small
business aspect gets overlooked so often.
I am going to have to go vote, but I am going to leave you
in the very capable hands of Senator Cantwell, who, as I said
earlier, really knows probably more about this issue than
anybody else in the Senate. So I am not worried that you are
going to get lost in the forest.
The Chair. Thank you, Chair Hickenlooper. And again, thank
you to you and Senator Blackburn for holding this important
hearing; and for all our witnesses participating in this; I am
sure it has been a robust discussion on many fronts.
I wanted to go back to the particulars of what you all
think we should do on the deepfake side, as we see technology
being developed, and DARPA playing a pretty key role as it is
today, in looking at deepfakes and deepfake information. What
is it you think is the landscape of a Federal role in
identifying? Some have described a system of a watermark, some
have described a--you know, immediate information system
similar to what the Amber Alerts are, or something of that
nature.
What do you all see as the key tools for effectiveness in
developing a system to respond to deepfakes? And we will just
go right down the----
Ms. Espinel. So it is a very important issue. I think there
is a lot of great work that is being done, some of it
spearheaded by a BSA member company named Adobe that has been
working on the Content Authenticity Initiative. And I think in
terms of giving--I know a lot of that is focused on making sure
that consumers have more accurate information that is truly
easily accessible, that they could access, and use, and take
into account about the generation of AI content, and about
whether or not that content has been manipulated or altered in
other ways.
But I also know that there are witnesses at this table that
are devoting a great deal of their life and energy to that
thought. So I am going to cede the mic to them.
Dr. Krishnan. Senator, first a broad comment about trust. I
think trust is a system-level construct, so when you think
about humans interacting with machines, machines interacting
with machines, one needs to think about what are the ways in
which we can enable trusted interactions, trusted transactions
to engage--to happen between them.
Deepfakes, as an example, I think content labeling and
detection tools to go along with content labeling is absolutely
essential to allow for individuals, so that when I am
interacting with a piece of content, for me to know whether it
was actually AI-produced, whether it is a deepfake, so to have
that information.
Equally well, beyond the technology piece, you need
education for individuals to know how to actually process this
information so that they can arrive at the right outcome with
regard to this interaction between human and machine.
Similarly, you could also have machine-to-machine exchanges of
data where you could have, you know--I produce a piece of video
content, and I pass it on to another machine, that has to be
a--this is where standards are important. This is where C2PA,
the standard you heard about, combined with watermarking, could
actually provide the trust infrastructure to address this
deepfake problem.
The Chair. Okay. Mr. Gregory.
Mr. Gregory. I believe there are a number of steps the
Federal Government can take. The first is to have a strong
understanding of the existing harms and impacts, and really be
able to understand where to prioritize with groups who are
impacted. That includes harm as we know already like
nonconsensual sexual images, but also the growing number of
scams.
The second area would be to focus on provenance and to come
up with a standardized way for people to understand both AI
provenance, and opt-in. Human-generated provenance.
The third would be to focus on detection. Detection is not
a silver bullet; it is flawed, but its availability is still
limited to the people who need it most on the frontlines of
journalism, human rights, and democracy, so continued
investment from DARPA, and others, to really resource and
support in diverse circumstances.
I believe there is a space for legislation around some
specific areas, such as non-consensual sexual images, AI-
generated CSAM, and potentially political ads that could be
taken. And I believe it is the role also to look ahead and
understand that this continuing ease of generation of synthetic
media means that we will get more and more personalized, and
this will have an impact in spaces like social media and
platforms. So we should look ahead to those dimensions and be
ready to consider those.
The Chair. Okay. Mr. Strayer.
Mr. Strayer. I won't repeat what has already been said, but
two things on the technical side very much to emphasize the
importance of having an open standard for provenance. And
secondly on the social dimension, you know, digital literacy is
going to be really important for these things to be
implemented. So bringing it to other stakeholders that include
the media platforms, consumers, on the digital literacy side,
for how these tools will be implemented effectively.
The Chair. So who do you think should be in charge of this;
anybody? Mr. Gregory, you look like you are going to volunteer
here.
Mr. Gregory. I am going to volunteer, but I am probably not
the best placed. So I will note that I see good leadership from
agencies like the FTC that have been doing good work to support
consumers to date; so supporting existing agencies that are
doing good work with the resourcing and the support. In terms
of the legislative gaps, I am not well placed to observe where
those should come from. In terms of the R&D, I think that is
broad support that ideally also goes outside of DARPA to other
research facilities; and facilities more broadly in the U.S.
Dr. Krishnan. In my testimony, I think with regard to the
content being produced, I think Congress should require close-
source and open-source models to actually create this
watermarking label and a detection tool to go with this label.
This is for images and video. Text is a huge issue as to what
it is--because you could have deepfakes with regard to text as
well. And I think research is needed there. So I think it is a
combination of things, but I think Congress should take a
leadership role.
Mr. Strayer. I will just add. Congress obviously has a very
important role to play. I also think that NIST is a place
where, over time, we have seen them deal with some very
difficult problems, come up with new profiles for addressing
very specific challenges and developing standards that are
universally accepted, through in this process. And so I think
NIST has a key role to play here too.
The Chair. Well that is why in the original legislation
that we did with the NAIAC was to establish, you know, getting
everybody together and figure out what we think the U.S.
Government's role and responsibility should be. And while they
haven't finished, you know, all of their findings, they have
certainly made a list of the directions and recommendations.
And so I think they are a good place to look for on this issue
as well, at least from a discussion perspective.
But today's hearing was about stimulating some input about
the issues around that. And what you basically are saying is
there is no fail-safe way to do this. It is going to need
constant participation, both on the side of making sure there
is not mistakes; it is one of the reasons why I support getting
a big--privacy bill that establishes a hard line against
discriminatory action because then you could always take that
action, again when somebody has had substantial harm given by a
direction.
I think the privacy framework we have already laid out to
basically stop that kind of activity and protect people. We
have heard a lot from the civil liberties community about this,
about what you might see as online redlining, and you worry
about something in the machine learning environment just
putting that into a system, and then it being there for years
and years without anybody even understanding that there was a
discriminatory tactic against somebody.
And all of a sudden, you know, all of these people don't
have the same kind of thing, alone, that they wanted. And so
this is something we definitely want to have a forceful bright
line, in my opinion, against, and say that if these kinds of
activities do exist, that we will stop them, and that we have a
strong law on the books to prevent them from happening.
What do you think on the collaboration level, from an
international basis, as it relates to deepfakes and
communication? Anybody given that thought about how that
framework should operate?
Mr. Strayer. Just to point out, one analogy of the past was
there was a lot of challenge with extremist -- violent
extremist content online in the--roughly in the mid-2000s post-
9/11, there was something formed called the Global Internet
Forum to Counter Terrorism, and that was really the major
platforms. But then many other players came together to form
practices and procedures for getting this extremist content off
the internet.
And so some kind of multi-stakeholder group coming together
to do this is probably one of the best ways we can see this
addressed expeditiously, as the problem will grow very quickly
as well.
The Chair. Didn't Interpol play a big role in the early
days of the internet in trying to do a similar thing? Trying to
police against pornography online, and catching, you know, bad
actors who were perpetrating content?
Mr. Strayer. Absolutely, yeah.
The Chair. And so that was a--where an international
organization was working, and organizations working with them
to try to police, I guess, or create standards or information
for people to stop those activities?
Mr. Strayer. Yeah. Sort of a Clearinghouse model, I think
that is how they pursued.
The Chair. And do you think that was successful?
Mr. Strayer. They were, I think, a big component of it. I
think the United States shouldn't shy away from taking credit
for a lot of work that it did, bilaterally, through the
Department of Justice to educate foreign partners about the
ways that they can address things like pornography that rise to
that level that is criminal. So I think the United States has
been a real leader in ensuring security and safety on the
internet.
The Chair. Thank you. Mr. Gregory.
Mr. Gregory. To add that one of the gaps that we see
frequently, and we support local journalists who are trying to
identify deepfakes as well as local civil society, as they
don't have access to skills and resources. So looking at
mechanisms to share skills, capacity, fellowship that would
bring that expertise closer to the people who need it. The
circumstance we see very frequently right now is people
claiming that real content is AI-generated, and people being
unable to prove it is real.
And that is corrosive in many contexts around the world,
and a lot of that has to do with the lack of access to skills
and resources; so thinking about opportunities for the U.S.
Government to support that.
The Chair: So what would that be? Because now you are
talking about a subject very near and dear to my heart; and
that is the erosion of local journalism by the commoditization
of advertising. And I would say the non-fair use of big
companies not giving media their fair value for content. You
are not really--you know, it is not your content to keep the
advertising revenue when it is within your browser instead of
going to The Seattle Times or some other website. So this is a
problem. And we have to fix that as well.
But you are saying their job is, you know, truth, justice,
and the American way. And how can they detect that if they
can't do the kind of investigations? Is that your point?
Mr. Gregory. Yes. That they don't have access to the tools
that they need, and so as DARPA and others build tools, making
sure they are accessible and relevant to journalism and others,
it is skills so that those are available, and that could be
facilitated through existing programs that provide skill
sharing.
I agree with you. There is a larger context where this is
but a small symptom of a broader challenge to journalism, where
AI increases those challenges, as well as provides
opportunities for journalists to use it.
The Chair. Well, we definitely heard that in Seattle at our
Summit, that we already have a problem as it relates to keeping
and saving local journalism, and I am very concerned about it,
because we have existed as a country for hundreds of years with
this kind of oversight to make sure that the process that we
all participate in works and functions, and the issues are
brought up.
And clearly, we are seeing places in the United States
where the journalism has--you know, ceased to have a credible
model, that is a financial model, and thus we have seen the
rise of a lot of very unfortunate issues, including corruption,
because there is no one there to cover and watch the day-to-
day.
So it is a very interesting question you are posing beyond
what we do as a government in detecting deepfakes. How do you
bring the oversight to those whose job it is to do oversight?
Mr. Gregory. And whose job will get even more complicated
in the coming years with the growth of AI-generated content?
The Chair. Yeah. And so do you think that is about
misinformation, or do you think it is bigger than just
misinformation?
Mr. Gregory. I believe it is a question of misinformation
to some extent. It is a question of the easy capacity to create
a volume of information that journalists have to triage and
interpret. It is a question of that against a backdrop of lack
of resources.
The Chair. Okay, and so what would you do about that?
Mr. Gregory. In the U.S. context it is very hard to work
out how to direct further resources towards local journalism.
One option would be to consider, as we look at the way in which
content is being ingested into AI models, is there any
financial support to journalistic entities as they do that;
this is obviously something that is being considered in the
social media context in other countries. I don't know whether
that would be a viable option to address local journalism's
needs.
The Chair. So how exactly would it work?
Mr. Gregory. I don't know the model that would work in our
context. We have certainly seen other contexts, globally, where
governments have looked for ways to finance journalism from
social media.
The Chair. Yes.
Mr. Gregory. But it is not a viable option here in the U.S.
The Chair. Well, okay, I like that--the phraseology should
be, local journalism is financing their--these websites and
their models. That is what is happening here.
Mr. Gregory. Yes.
The Chair. And we just haven't been able to find the tools
to claw that back. But if we have to go and look at this fair
use issue, we will go back and look at it, because we are not
going to keep going this direction. And AI is an accelerant. It
is an accelerant on everything. The information age is putting
challenges, and AI will accelerate that. But we have got to
harness the things that we care about and make sure that we get
them right, because we want the innovation, but we also want
these particular issues to be resolved. So we certainly,
certainly in Seattle, have had that discussion.
Dr. Krishnan. Can I briefly comment on this?
The Chair. Yes. Yeah, go ahead.
Dr. Krishnan. So on the first part, with regard to the
tools, I do think that the kind of infrastructure for trust
that we have built up with information security with the CERT,
with SISA for instance, that that kind of capability, if you
built it for AI as well, which could be fairly quickly stood up
with FFRDCs, that gives us the capacity even across countries
to track deepfakes, even if they don't necessarily adhere to a
content standard like C2PA, because I don't think any
individual organization has that capacity.
But something like the CERT could have that capacity
because it will span dot-mil, dot-com, dot-gov concerns, and
this capability and expertise will reside in something like
that. That is with regard to your first question, with regard
to how do we manage and harmonize standards across countries.
With regard to the second point, I think it is spot on with
regard to values on the one hand, the capacity to licenses
copyrighted content, and then how do you actually assess--that
is on the input side, so if you think of the AI models as
taking input data from, say, The Seattle Times, or things of
that nature, how they--how do they declare first that they are
using this data, and then compensating The Seattle Times fairly
for that use of that?
On the output side, the interesting question is, is it the
case that The Seattle Times is getting more traffic from the
ChatGPTs and the Googles of the world? Or is it the case that
the revenue that should have come to The Seattle Times is
really going to ChatGPT or Bard. I mean, the argument has been
that because they provide that entry point into a content that
they are actually directing traffic that otherwise would not
have found you.
So I think that requires, I think, analysis and research of
the traffic with regard to who is going where and who is
directing what to these sites, because I think that gets at
this revenue point that----
The Chair. Well, I am pretty sure about 25 percent of the
traffic that is generated online, that big sites are getting
from news organizations, are really revenue that belongs to
news organizations.
Dr. Krishnan. Right.
The Chair. Regardless of the commoditization of
advertising. It is still revenue that belongs to the
newspapers. And so to my point about this, is that our report,
that this committee, at least when we were the authors of a
report, we found that local journalism was the trusted news
source.
Dr. Krishnan. Mm-hmm.
The Chair. This is the point. And that you have many
voices, that that is the ecosystem that keeps the trust. I mean
somebody could go awry. But guess what? The rest of the
ecosystem keeps that trust. So you know, I think The Seattle
Times would say it is a very viable, identifiable source of
trust. If you were creating information off of their historical
database of all Seattle Times ever published stories, which is
a very long time, that is probably some of the most trusted
journalistic information you could ever get, because they had
to be in that business, right? But anybody who would then take
that content and then who knows do what with it, you know,
obviously is a very, very different equation.
So look, I think--I want to go back to the international
point for a second because I do think you mentioned a lot of
organizations. I am not sure everybody grasped, or maybe I
didn't grasp everything you were saying about that. What do you
think we should be--do you think the NAIAC should be working in
coordination right now with international organizations to
discuss what a framework looks like? Or are you thinking this
is more siloed within organizations like national security
issues versus consumer issues, versus other things?
Dr. Krishnan. So then NAIAC does have a group that Ms.
Espinel leads, as a working group. The AI Futures Working Group
that I lead with regard to this trust infrastructure point that
I was making, we have been focused on that, but it does have
international implications. But perhaps Ms. Espinel can speak
more to it.
Ms. Espinel. So I do--I have the honor of chairing the
International Working Group for the NAIAC Advisory Committee. I
would be--there are conversations that we are having internally
about ways that NAIAC, as a committee, could be helpful. And
either in terms of making recommendations to the
administration, which is our mandate, or perhaps NAIAC as a
committee, and I would be--some of them I can't talk about
publicly here, although I would be--I would be happy to have
follow-up conversations.
I can tell you about one though that I think goes to what
you are talking about, which is, I think we believe that it is
very important as governments are thinking about what the right
approach is to regulating AI, or trying to address some of the
concerns that have been raised by artificial intelligence, to
make sure that those conversations are happening not just with
the United States, not just with the United States, and the EU,
not just inside the G7, the OECD, but to try to have that be a
broad-based conversation, including bringing in emerging
economies that are--have not typically been as much a part of
some of these discussions as I think should be the case.
And so I think if we are going to end up with solutions
that are really effective, for example, in deepfakes, that is
going to have to be a global initiative, and I think it will be
stronger and more effective if it is--if those discussions are
happening with a group of companies--with a group of countries
that represent different perspectives.
So emerging economies are going to have slightly different
benefits and challenges, they need to be part of that
discussion and that is----
The Chair. Well, besides----
Ms. Espinel. Sorry. I am kind of probably overly passionate
about it. So I feel like I have gone on a bit too long.
The Chair. No, no. The question I was trying to get at as
people--listen, this committee passed this legislation, we
created the NAIAC, we said, here is your responsibilities, we
hope you have been thinking about this because we have given
you a few years to do so. And so I was wondering if the current
thinking was a divide over the complexity of dealing with
national security kind of deepfakes in, you know, commercial,
and citizen issues on deepfake.
And whether they--you would reach some conclusion on the
international side of, there is a lot to this, and a lot to
communicate and coordinate, because obviously the World Wide
Web is a big, you know, open system. So you could say the
United States is doing this, but you need others to
participate. But consumer issue is very different from how we
deal with national security issues. And so did the--has the
organization come to any conclusion on that?
Ms. Espinel. I think the short answer is no. Just not to be
overly legalistic, but there are significant restrictions on
what I am allowed to say in a public forum, and I don't want
to--I want to be very careful not to cross any lines.
So I can tell you that I think there are conversations
happening about national security and consumers, on the point--
I feel like it is fine for me to say, on the point that you
were talking about, I don't see there being a real challenge. I
don't see there being a lack of consensus on national security
versus consumer issues, and be able to engage internationally
on that.
The Chair. Well, they are just different organizations
within our government and I am pretty sure they are
internationally. So it just makes it challenging.
Ms. Espinel. It makes it challenging. And so I think one of
the--so this I could say, I will just say, in my capacity as
BSA, one of the--you have, for example, the U.K. government is
hosting a global summit in the beginning of November, and I
think one of the challenges they face is, who--if you are going
to have a global summit that is intended to address the safety
of artificial intelligence, which is what the UK has announced,
who are you going to have--who is going to be part of that
summit, and how many issues can they address?
Because there are a myriad of challenges, and as you say,
they are often addressed by different parts of government.
Speaking just for the United--speaking just in the context of
the United States, I think having effective coordination across
the Federal Government, I think there is more that could be
done there. And I think that would be very, very helpful
because you don't want these issues to get siloed, you don't
want agencies to be doing things that are duplicative or in
conflict.
The Chair. Okay.
Dr. Krishnan. And I will reach out to your office, Senator,
about the trust infrastructure point that I made. I am happy to
provide additional information.
The Chair. Well, we all know that we have lots of
international organizations that are working on coordination on
lots of internet issues, as it is today. I think the question
is, you know, what--has anybody with the NAIAC, come up with a
framework; before we start having these kinds of big
discussions? So anyway, we will get more information.
I want to turn it back over to Chair Hickenlooper. Thank
you so much for, again, holding this very important hearing.
I see that our colleague from New Mexico is here. And so
sorry to take up so much time; I thought I had a free
opportunity while you were off voting. Thank you all to the
panel too.
Ms. Espinel. Thank you. If I get to say briefly, I am
excited to be here to testify in my capacity as CEO of BSA, but
I would also be happy to follow up with you in my capacity as a
member of the NAIAC Committee.
The Chair. Thank you.
Senator Hickenlooper. And didn't I tell you I was leaving
you in good hands? I love coming in on the end of a discussion,
saying, God, how did I miss that? And certainly the traditional
news, the trust of traditional news, and how to make sure that
they are fairly compensated for the costs, but I don't think
any of us know any traditional news organization that is worth
a fraction of what it was worth 15 years ago. Just the nature
of what has happened.
I turn it over to the good Senator from New Mexico.
STATEMENT OF HON. BEN RAY LUJAN,
U.S. SENATOR FROM NEW MEXICO
Senator Lujan. Thank you, Mr. Chairman. Thank you to you,
and to Ms. Blackburn for this important hearing. And it is
great to be able to listen to Chair Cantwell as well.
To all the members of the panel, thank you so much for
being here today and offering your expertise.
One area I have spent a considerable amount of time during
my time in the Senate on this issue surrounding broadband, I
will say, as opposed to AI, is in the area of transparency and
making things easier to understand.
And what I mean by that is, Senator Cassidy and I, we
introduced something called the TL;DR, which was, Too Long;
Didn't Read Act. And we all know what those agreements look
like. Pages and pages of documents and, you know, before you
can download or use something, you know, you go down a little
box and it says, ``Accept terms of agreement'', and people
click that and they move on and they don't realize what is
there.
I was also proud to see and advocate during the Bipartisan
Infrastructure Bill, something called Nutrition Labeling for
Broadband, to make it easier for consumers to compare services
across the way.
Now, the same type of transparency and easy-to-understand
documentation is also needed so that consumers know where they
are interacting with AI, know how their information is going to
be used, and know when content is generated by AI. And I am
working on legislation to require this type of transparency and
disclosures and responsibilities.
Now, Mr. Gregory, you mentioned that threats from synthetic
media and disinformation most impact those already at risk,
like minority groups that face other forms of threats and
discrimination. Mr. Gregory, yes or no, are deepfakes and AI-
generated disinformation already being created in Spanish and
other non-English languages?
Mr. Gregory. Yes.
Senator Lujan. And, Mr. Gregory, yes or no, do you believe
technology companies invest enough resources into making sure
AI systems work equally well in non-English languages?
Mr. Gregory. Systems are not as adequate in non-English
languages, and there are less resources invested in making them
applicable in non-English languages; and when it comes to
deepfakes, many tools work less effectively for detection in
less dominant languages outside of English.
Senator Lujan. And with that being said, Mr. Gregory, would
the current system make more communities even more vulnerable
in the U.S. and globally through the risk of synthetic media?
Mr. Gregory. This is exactly the concern that my
organization has spent the last five years working on, is the
vulnerability of communities in the U.S. and globally to
synthetic media, because of their exclusion from access to
tools and support around it.
Senator Lujan. Well, thank you for your work in that space
and bringing attention to it. I introduced a piece of
legislation, called the LISTOS Act, this year to address the
lack of non-English investment in multilingual, large language
models. Any action Congress takes on AI transparency must
protect our most marginalized communities, including non-
English speakers. And for those that don't know what LISTOS
means, it means ``ready,'' but it is an acronym, so I credit
the staff for coming up with an acronym that calls to action to
ensure that we would all get ready.
Now, AI systems are useful in a huge variety of scenarios,
from making financial decisions, to medical diagnosis, and
content modernization, and I believe government should also
utilize AI for improving constituents' access to government
services. This just goes to show the broad application of what
AI can do.
Now, Ms. Espinel, AI systems are used in different sectors
and for different purposes, and these various systems can have
different kinds of outputs. For example, we have predictive AI
making best guesses, and generative AI creating content. I have
a hard time keeping up with a lot of this stuff; but maybe the
Insight Forum tomorrow will help a little for folks like
myself, but understandably. Consumers and other users of these
systems will have very different experiences interacting in
these systems.
Now, my question to you stems from, if people even know
they are interacting with an AI system, which is not a given,
so under that premise, isn't it then necessary that any
transparency and oversight requirements be specific to the
sector or use case of the AI system?
Ms. Espinel. So you are absolutely correct that AI is used
in many different ways, and has many different outputs, and has
many different consequences. So I would say a couple of things.
One is, I think, in terms of having transparency so that, for
example, a consumer knows if they are interacting with an AI
service like a chatbot, I think that is important.
But I would also say, to build on that, is that looking at
the specific uses, perhaps, as opposed to looking at a sector,
but looking to see whether or not that use is going to have a
consequential decision on someone's life. Is it going to impact
their ability to access government services, to access public
benefits, to get a loan, to get education, to get a job?
And if it is, if it is going to have a consequential
decision in someone's life, then I believe, we believe at BSA,
that companies should be required, by legislation, to do impact
assessments to identify those risks, and then take steps to
reduce those risks.
Senator Lujan. I appreciate that very much.
Now, Professor, under the thought of the explosion of
generative AI, which we know has serious implications for the
integrity of our information ecosystems; do you agree that
Congress should require tech companies to label any AI-
generated content with disclosures and metadata that includes
information on how the content was created?
Dr. Krishnan. Thank you, Senator. Yes, they should. And I
want to also thank you for your first remark about TL;DR and
the Privacy Nutrition labels. This is something that is
absolutely essential in this space. It creates something that
is easy for people to understand. And as I understand it, Apple
had introduced, two years ago, these labels that are privacy
schematics that are easy to understand for people, and those
need to be adapted for the purposes of what it is you were
introducing in your opening remarks.
Senator Lujan. I appreciate that very much, Professor.
Now, Mr. Chairman, I have several other questions, but I
see my time has expired. I have already exceeded it. I will
submit them into the record. But I just want to thank you all
for what you have been doing, what you are going to continue to
do, those that are participating in one form or another with
the forum tomorrow, and especially for the follow-up coming out
of this as well.
And so, for those of you I did not speak to very much, or
at all, I have some questions for you as well, and I will get
them to you. So, I look forward to hearing back from you, and
your expertise on the subject.
Thank you, Mr. Chairman.
Senator Hickenlooper. Thank you, Senator Lujan.
And once again, I thank all of you. I am going to wrap it
up, and unless we missed something urgent, speak now or forever
hold your peace; although that is from a different ceremony.
Today's discussion is important to help set the next steps for
working with consumers to understand the benefits, and
ultimately to trust AI.
Senators are going to be able to submit questions, not to
Senator Lujan, but all the senators, additional questions for
the record until the hearing record closes on September 26.
We ask witnesses to provide responses to the Committee by
October 10. And again, thank you very much for your time.
We are adjourned.
[Whereupon, at 4:30 p.m., the hearing was adjourned.]
A P P E N D I X
Prepared Statement of Hon. Ted Cruz, U.S. Senator from Texas
Thank you Chairman Hickenlooper and Ranking Member Blackburn. And
thank you Chairwoman Cantwell for calling this hearing. This Committee
should be at the forefront of discussions about artificial
intelligence, or AI.
It is important to keep in mind that while AI is becoming more
advanced, it is not new. It's been around for decades. You use AI every
day without realizing it--for example, when you receive online shopping
recommendations or have your text messages autocorrected. Beyond
improving mundane tasks, AI is already transforming our world for the
better. It's detecting cybersecurity threats to critical
infrastructure; improving agricultural yield; and with advancements,
potentially enhancing cancer detection and treatment. In these ways and
more, AI has already vastly improved Americans' quality of life.
Congress would do well to learn from the past. This is far from the
first time we've debated whether and how to regulate innovation. Take
the Internet as an example. In the 1990s, the United States made a
conscious, bipartisan decision to avoid heavy government intervention
that might stunt the internet's growth, including bureaucratic
organization under one agency head and influence over speech and
content issues. The results speak for themselves: The U.S. now has the
most successful Internet companies in the entire world--it isn't even a
close contest.
With AI, I'm afraid that we are allowing history to repeat itself--
only this time, we are following our European counterparts, who made a
mistake with their early regulation of the internet. You can't read the
news today without encountering Terminator-style fearmongering about AI
building a weapon of mass destruction or developing sentience that will
destroy humans.
Let's be clear: AI is computer code developed by humans. It is not
a murderous robot. Humans are responsible for where and how AI is used.
Unfortunately, the Biden Administration and some of my colleagues
in Congress have embraced doomsday AI scenarios as justification for
expanded Federal regulation. Some of these proposals are extremely
onerous: Licensing regimes, creating a new regulatory agency to police
computer code, and mandatory, NEPA-style impact assessments before AI
can be used. The fearmongering around AI has caused us to let our guard
down to accept so-called ``guardrails''--pseudo-regulations disguised
as safety measures. These are often cumbersome and infeasible,
especially for the startups so common in the tech sector.
I don't discount that there are some risks associated with the
rapid development and deployment of AI. But we must be precise about
what these risks are. I've heard concerns from my colleagues about
misinformation, discrimination, and security. These certainly present
challenges, but I have a hard time viewing them as existential risks,
or even worthy of new regulation.
To me, the biggest existential risk we face is ourselves. At this
point, Congress understands so little about AI that it will do more
harm than good.
It is critical that the United States continue to lead in AI
development--especially when allies such as the European Union are
charging toward heavy-handed regulation.
Let me propose an alternative. Instead of riling fears and pausing
AI development, let's pause before we regulate.
We can start by fully assessing the existing regulatory landscape
before burdening job-creating businesses--especially startups--with new
legal obligations. Many of our existing laws already apply to how AI
systems are used. For example, the Fair Credit Reporting Act protects
consumer information with reporting agencies and the Civil Rights Act
of 1964 prohibits discrimination. We should faithfully enforce these
laws, as written, without overstepping.
The FTC's investigation of OpenAI is a clear abuse of authority. As
I wrote to Chairwoman Khan this week, fearmongering and fanciful
speculation do not justify enforcement action against Americans
creating new AI tools. The FTC's unprecedented and aggressive policing
of AI would undoubtedly require statutory authority from Congress.
Leading the AI race is also important for national security. If we
stifle innovation, we may enable adversaries like China to out-innovate
us. I've been cautioning against ceding leadership on AI development to
China since 2016, when I held Congress's first-ever hearing on AI. We
cannot enact a regulatory regime that slows down innovation and lets
China get ahead of us.
Think about if we had let fear get the best of us with other
technological developments. Panics about new technology have occurred
throughout history--and the panics have generally not borne out. There
was widespread fear and calls to regulate around the advent of
innovations such as automobiles, recorded sound, typewriters, and
weaving machines. Every time, after the hysteria died down, we adapted
and allowed technology to improve our lives, spur economic growth, and
create new jobs.
The same opportunity exists today with AI. Let's not squander it.
______
R Street Institute
Washington, DC, September 12, 2023
Hon. John Hickenlooper,
Chair,
Subcommittee on Consumer Protection, Product Safety, and Data Security,
Commerce, Science, & Transportation Committee,
United States Senate,
Washington, DC.
Hon. Marsha Blackburn,
Ranking Member,
Subcommittee on Consumer Protection, Product Safety, and Data Security,
Commerce, Science, & Transportation Committee,
United States Senate,
Washington, DC.
Dear Chairman Hickenlooper, Ranking Member Blackburn and members of the
Subcommittee:
Thank you for your decision to hold a hearing on September 12, 2023
titled, ``The Need for Transparency in Artificial Intelligence.'' My
name is Adam Thierer and I am a senior fellow at the R Street
Institute. I also recently served as a commissioner on the U.S. Chamber
of Commerce's Commission on Artificial Intelligence Competitiveness,
Inclusion, and Innovation.\1\
---------------------------------------------------------------------------
\1\ U.S. Chamber of Commerce, Commission on Artificial Intelligence
Competitiveness, Inclusion, and Innovation: Report and Recommendations
(March 2023). https://www.uschamber.com/technology/artificial-
intelligence-commission-report.
---------------------------------------------------------------------------
Artificial Intelligence (AI) technologies are already all around us
and they are helping make our lives better in many ways. But the
potential for algorithmic systems is even greater and these
technologies also have important ramifications for our country's global
competitive standing and geopolitical security.
The United States must reject the regulatory approaches being
advanced by China, Europe and other nations, which are mostly rooted in
a top-down, command-and-control approach to AI systems.
Instead, America's approach to technological governance must
continue to be agile and adaptive because there is no one-size-fits-all
approach to AI that can preemptively plan for the challenges that we
will face even a short time from now.\2\
---------------------------------------------------------------------------
\2\ Adam Thierer, ``Getting AI Innovation Culture Right,'' R Street
Institute Policy Study No. 281 (March 2023). https://www.rstreet.org/
research/getting-ai-innovation-culture-right.
---------------------------------------------------------------------------
At this early stage of AI's development, government's role should
be focused on helping developers work toward consensus best practices
on an ongoing basis.\3\ In this regard, the National Institute of
Standards and Technology (NIST) has taken crucial steps with its AI
Risk Management Framework, which is meant to help AI developers better
understand how to identify and address various types of
---------------------------------------------------------------------------
\3\ Lawrence E. Strickling and Jonah Force Hill, ``Multi-
stakeholder Internet governance: successes and opportunities,'' Journal
of Cyber Policy 2:3 (2017), pp. 298-99. https://www.tandf
online.com/doi/abs/10.1080/23738871.2017.1404619.
---------------------------------------------------------------------------
potential algorithmic risk.\4\ NIST notes it ``is designed to
address new risks as they emerge'' instead of attempting to itemize
them all in advance.\5\ ``This flexibility is particularly important
where impacts are not easily foreseeable and applications are
evolving,'' the agency explains.\6\
---------------------------------------------------------------------------
\4\ National Institute of Standards and Technology, Artificial
Intelligence Risk Management Framework (AI RMF 1.0), U.S. Department of
Commerce, January 2023. https://nvlpubs
.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf.
\5\ Ibid., p.4
\6\ Ibid.
---------------------------------------------------------------------------
While it is always important to consider the dangers that new
technologies could pose, extreme regulatory solutions are not
warranted. Safety considerations are vital, but there is an equally
compelling public interest in ensuring that algorithmic innovations are
developed and made widely available to society.
Toward that end, AI governance should be risk-based and focus on
system outcomes, instead of being preoccupied with system inputs or
design.\7\ In other words, policy should concern itself more with
actual algorithmic performance, not the underlying processes.
Transparency and explainability are important values that government
can encourage, but these concepts must not be mandated in a rigid,
overly prescriptive fashion.\8\
---------------------------------------------------------------------------
\7\ Adam Thierer, ``The Most Important Principle for AI
Regulation,'' R Street Institute Real Solutions, June 21, 2023. https:/
/www.rstreet.org/commentary/the-most-important-principle-for-ai-
regulation.
\8\ Comments of Adam Thierer, R Street Institute to the National
Telecommunications and Information Administration (NTIA) on ``AI
Accountability Policy,'' June 12, 2023. https://www.rstreet.org/
outreach/comments-of-the-r-street-institute-to-the-national-
telecommunications-and-information-administration-ntia-on-ai-
accountability-policy.
---------------------------------------------------------------------------
Algorithmic systems evolve at a very rapid pace and undergo
constant iteration, with some systems being updated on a weekly or even
daily basis. If policy is based on making AI perfectly transparent or
explainable before anything launches, then innovation will suffer
because of endless bureaucratic delays and paperwork compliance
burdens. Society cannot wait years or even months for regulators to
eventually get around to formally signing off on mandated algorithmic
audits or impact assessments, many of which would be obsolete before
they were completed.
Converting audits into a formal regulatory process would also
create several veto points that opponents of AI advancement could use
to slow progress in the field. AI innovation would likely grind to a
halt in the face of lengthy delays, paperwork burdens and significant
compliance costs. Algorithmic auditing will always be an inexact
science because of the inherent subjectivity of the values being
considered.
Auditing algorithms is not like auditing an accounting ledger,
where the numbers either add up or don't. When evaluating algorithms,
there are no clear metrics that can quantify the scientifically correct
amount of privacy, safety or security in a given system.
This means that legislatively mandated algorithmic auditing or
explainability requirements could also give rise to the problem of
significant political meddling in speech platforms powered by
algorithms, which would raise free speech concerns. Mandated AI
transparency or explainability could also create some intellectual
property problems if trade secrets were revealed in the process.
This is why it is essential that America's AI governance regime be
more flexible, bottom-up, and driven by best practices and standards
that evolve over time.\9\ Beyond encouraging the private sector to
continuously refine best practices and ethical guidelines for
algorithmic technologies, government can utilize the vast array of laws
and regulations that already exist before adding new regulatory
mandates. The courts and our common law system stand ready to address
novel risks that are unforeseeable in advance. Many agencies are also
moving aggressively to consider how they might regulate AI systems that
touch their fields. Using various existing regulatory tools and powers
like product recall authority and unfair and deceptive practices law,
agencies can already address algorithmic harms that are proven. We
should not be adding another huge Federal bureaucracy or burdensome
licensing mandates to the mix until we have exhausted these other
existing solutions.\10\
---------------------------------------------------------------------------
\9\ Adam Thierer, ``Flexible, Pro-Innovation Governance Strategies
for Artificial Intelligence,'' R Street Institute Policy Study No. 283
(April 2023). https://www.rstreet.org/research/flexible-pro-innovation-
governance-strategies-for-artificial-intelligence.
\10\ Neil Chilson and Adam Thierer, ``The Problem with AI Licensing
& an `FDA for Algorithms,' '' Federalist Society Blog, June 5, 2023.
https://fedsoc.org/commentary/fedsoc-blog/the-problem-with-ai-
licensing-an-fda-for-algorithms.
---------------------------------------------------------------------------
The United States must create a positive innovation culture if it
hopes to prosper economically and ensure a safer, more secure
technological base. Policymakers must not try to micro-manage the
future or pre-determine market outcomes. It is essential that we strike
the right policy balance as our Nation faces serious competition from
China and other nations who are looking to counter America's early lead
in computational systems and data-driven digital technologies.
Sincerely,
/s/ Adam Thierer,
Senior Fellow,
R Street Institute.
______
Prepared Statement of Hodan Omaar, Senior Policy Analyst, Information
Technology and Innovation Foundation (ITIF)
Chairman Hickenlooper, ranking Member Blackburn, and members of the
subcommittee, we appreciate the opportunity to share with you our
thoughts on crafting policies to increase transparency in artificial
intelligence (AI) technologies for consumers. ITIF is a nonpartisan
think tank whose mission is to formulate and promote public policies to
advance technological innovation and productivity. In this statement,
we offer three considerations policymakers should keep in mind to
ensure consumers are protected from harm:
1. While policymakers should encourage companies to adopt the NIST
risk management framework, they should recognize that it is not
a silver bullet for trustworthy AI. There are a variety of
technical and procedural controls companies can employ to
mitigate harm and policymakers should encourage companies to
explore the full gamut of mechanisms to find those most
contextually relevant.
2. Because increasing AI transparency can make some systems less
accurate and effective, policymakers should fund research to
better understand this tradeoff and evaluate policies for
transparency against the impact on system accuracy.
3. Policymakers should hold AI systems to the same standard as human
decisions, which are not always transparent.
4. Policymakers should direct NIST to support work on content
provenance mechanisms, which are techniques that help users
establish the origin and source of content (both AI-generated
and human-generated), rather than create policies that simply
require systems to disclose when output is AI-generated.
AI offers significant societal and economic benefits in a wide
variety of sectors. The biggest risk to consumers is that the myriad
opportunities AI offers will not be translated into all the areas where
they can make a positive difference in people's lives.
However, there are several other areas of risk to consumers from
businesses using AI. One is the creation of unsafe AI products and
services, such as a company putting an AI chatbot that advises users to
do dangerous things on the market. Another is the use of AI to deceive
or manipulate unsuspecting consumers, such as a company using AI to
create and spread fake reviews about their goods or services, which
ITIF's Center for Data Innovation explores in its 2022 report ``How
Policymakers Can Thwart the Rise of Fake Reviews.'' \1\ A third is the
use of AI to commit crimes that harm consumers, such as using AI to
support cyberattacks that steal their sensitive information. While
there are other applications of AI that interact with consumers, such
as the use of AI to make lending or credit decisions or AI used in
employment decisions, we note that these are not in the scope of the
subcommittee and therefore keep our comments focused on those that are.
---------------------------------------------------------------------------
\1\ Morgan Stevens and Daniel Castro, ``How Policymakers Can Thwart
the Rise of Fake Reviews,'' (Center for Data Innovation, September
2022), https://datainnovation.org/2022/09/how-policymakers-can-thwart-
the-rise-of-fake-reviews/.
---------------------------------------------------------------------------
1. While policymakers should encourage companies to adopt the NIST risk
management framework, they should recognize that it is not a
silver bullet for trustworthy AI. There are a variety of
technical and procedural controls companies can employ to
mitigate harm and policymakers should encourage companies to
explore the full gamut of mechanisms to find those most
contextually relevant.
Chairman Hickenlooper and Ranking Member Blackburn are right to
state in their recent letter to technology companies that the National
Institute of Standards and Technology AI Risk Management Framework
(NIST AI RMF)--a framework that helps companies identify and mitigate
potential risks from AI--can help protect consumers from harm and
encourage companies to responsibly develop and use AI.\2\ However, it
is important to note that many facets of trustworthy AI cannot easily
be translated into objective, concrete metrics and technical standards
alone are not a silver bullet for trustworthy AI.
---------------------------------------------------------------------------
\2\ ``Hickenlooper, Blackburn Call on Tech Companies to Lead
Responsible AI Use,'' press release, Apr 19, 2023, https://
www.hickenlooper.senate.gov/press_releases/hickenlooper-blackburn
-call-on-tech-companies-to-lead-responsible-ai-use/.
---------------------------------------------------------------------------
For instance, ensuring AI systems are robust and secure is one
important element of creating trustworthy AI that protects consumers,
and yes, one can employ audits to check how prevalent algorithmic
errors are. There are various types of error-analysis techniques to
check for algorithmic error, including manual review, variance analysis
(which involves analyzing discrepancies between actual and planned
behavior), and bias analysis (which provides quantitative estimates of
when, where, and why systematic errors occur, as well as the scope of
these errors).
However, other facets of trustworthy AI, such as ensuring these
systems are fair or unbiased, are subjective and cannot be reduced to
fixed functions.\3\ To see why, consider two e-commerce platforms that
use AI algorithms to recommend products to their users. One platform
employs an AI system with an objective function to recommend products
solely based on customer preferences and purchase history, aiming to
provide personalized recommendations without taking into account the
price of the products. The other platform uses an AI system with an
objective function that considers both customer preferences and product
prices, trying to recommend products that not only match user
preferences but also fall within the user's budget. Assume both AI
systems are designed to be error-free. Even if both AI systems are
functioning perfectly, they may have different suggestions for
consumers from different socioeconomic backgrounds. Defining which
system is more ``fair'' in this context can be complex, as fairness
might involve considerations of affordability, accessibility, and equal
opportunity to access desirable products.
---------------------------------------------------------------------------
\3\ Rediet Abebe, Jon Kleinberg, & S. Matthew Weinberg, ``Subsidy
Allocations in the Presence of Income Shocks,'' Proceedings of the AAAI
Conference on Artificial Intelligence (2020): 34(05), 7032-7039,
https://doi.org/10.1609/aaai.v34i05.6188.
---------------------------------------------------------------------------
This example demonstrates that achieving fairness in consumer
product recommendations can be multifaceted and context-specific.
Fairness may not have a one-size-fits-all definition. Rather than
pursuing technical standards alone, policymakers should be pursuing the
principle of algorithmic accountability. As the Center for Data
Innovation explains in its 2018 report ``How Policymakers Can Foster
Algorithmic Accountability,'' this principle states that an algorithmic
system should employ a variety of controls to ensure the operator can
verify algorithms work in accordance with its intentions and identify
and rectify harmful outcomes.\4\ When an algorithm causes harm,
regulators should use the principle of algorithmic accountability to
evaluate whether the operator can demonstrate that, in deploying the
algorithm, the operator was not acting with intent to harm or with
negligence, and to determine if an operator acted responsibly in its
efforts to minimize harms from the use of its algorithm. This
assessment should guide their determination of whether, and to what
degree, the algorithm's operator should be sanctioned.
---------------------------------------------------------------------------
\4\ Joshua New and Daniel Castro, ``How Policymakers Can Foster
Algorithmic Accountability'' (Center for Data Innovation, May 2018),
http://www2.datainnovation.org/2018-algorithmic-accountability.pdf.
---------------------------------------------------------------------------
Regulators should use a sliding scale of enforcement actions
against companies that cause harm through their use of algorithms, with
unintentional and harmless actions eliciting little or no penalty while
intentional and harmful actions are punished more severely.
Defining algorithmic accountability in this way also gives
operators an incentive to protect consumers from harm and the
flexibility to manage their regulatory risk exposure without hampering
their ability to innovate. This approach would effectively guard
against algorithms producing harmful outcomes, without subjecting the
public-and private-sector organizations that use the algorithms to
overly burdensome regulations that limit the benefits algorithms can
offer.
2. Because increasing AI transparency can make some systems less
accurate, policymakers should fund research to better
understand this tradeoff and evaluate policies for transparency
against the impact on system accuracy.
One of the core tenets of transparent AI people cite is
explainability. Explainable AI systems are those that can articulate
the rationale for a given result to a query. Explanations can help
users make sense of the output of algorithms. Explanations may be
useful in certain contexts, such as to discover how an algorithm works.
Explanations can reveal whether an algorithmic model correctly makes
decisions based on reasonable criteria rather than random artifacts
from the training data or small perturbations in the input data.\5\
---------------------------------------------------------------------------
\5\ ``AI Foundational Research--Explainability'', NIST, https://
www.nist.gov/topics/artificial
intelligence/ai-foundational-research-explainability.
---------------------------------------------------------------------------
However, it is well-documented that there is often a trade-off
between explainability and accuracy. As a 2020 paper led by NIST
researcher explains ``typically, there is a tradeoff between AI/ML
accuracy and explainability: the most accurate methods, such as
convolutional neural nets (CNNs), provide no explanations, while more
understandable methods, such as rule-based systems, tend to be less
accurate.'' \6\
---------------------------------------------------------------------------
\6\ D. Richard Kuhn et al, ``Combinatorial Methods for Explainable
AI,'' October 2020, Preprint: 9th International Workshop on
Combinatorial Testing (IWCT 20), https://csrc.nist.gov/CSRC/media/
Projects/automated-combinatorial-testing-for-software/documents/xai-
iwct-short-preprint.pdf
---------------------------------------------------------------------------
Policymakers should seek to understand the extent to which this is
true for applications that impact consumers and how they can implement
policies for increased transparency in a way that does not harm system
accuracy. A 2022 paper called ``The Non-linear Nature of the Cost of
Comprehensibility,'' published in the Journal of Big Data notes that
``while there has been a lot of talk about this trade-off, there is no
systematic study that assesses to what extent it exists, how often it
occurs, and for what types of datasets.'' \7\ It could be the case that
high-risk consumer-facing AI applications are more likely to become
less accurate if they were made to be more explainable, or it might
not. More research would help answer this question. Furthermore, if
policymakers want to increase transparency for certain high-risk
scenarios, they should also fund research into methods that might limit
the impact on system accuracy.
---------------------------------------------------------------------------
\7\ Sofie Goethals et al, ``The Non-linear Nature of the Cost of
Comprehensibility,'' March 7, 2022, Journal of Big Data, https://
journalofbigdata.springeropen.com/articles/10.1186/s40537-022-00579-2.
---------------------------------------------------------------------------
3. Policymakers should hold AI systems to the same standard as human
decisions, which are not always transparent.
Policymakers should be careful of holding AI systems to a higher
standard than they do for humans or other technologies and products on
the market. This is a mistake the European Commission is making with
its AI Act. The EU's original proposal contains impractical
requirements such as ``error-free'' data sets and impossible
interpretability requirements that human minds are not held to when
making analogous decisions.\8\ Policymakers should recognize that no
technology is risk-free; the risk for AI systems should be comparable
to what the government allows for other products on the market.
---------------------------------------------------------------------------
\8\ Patrick Grady and Kir Nuthi, ``The EU Should Learn From How the
UK Regulates AI to Stay Competitive,'' (Center for Data Innovation,
April 2023), https://datainnovation.org/2023/04/the-eu-should-learn-
from-how-the-uk-regulates-ai-to-stay-competitive/.
---------------------------------------------------------------------------
More broadly, targeting only high-risk decision-making with AI,
rather than all high-risk decision-making, is counterproductive. If a
certain decision carries a high risk of harming consumers it should
make no difference whether an algorithm or a person makes that
decision. For example, if it is harmful to deceive consumers by
creating fake reviews, enforcement action should be proportional,
regardless of whether a human or an AI system was used to create them.
To hold algorithmic decisions to a higher standard than human decisions
implies that automated decisions are inherently less trustworthy or
more dangerous than human ones, which is not the case. This would only
serve to stigmatize and discourage AI use, which would reduce its
beneficial social and economic impact.
4. Policymakers should direct NIST to support work on content
provenance mechanisms, which are techniques that help users
establish the origin and source of content (both AI-generated
and human-generated), rather than create policies that simply
require systems to disclose when output is AI-generated.
Some policymakers advocate for policies mandating that generative
AI systems, such as those used in customer service, social media, or
educational tools, must include notices in their output, informing
users that they are interacting with an AI system rather than a human.
However, mandatory disclosure requirements may not always be
practical or desirable. Many AI applications aim to replicate human
capabilities, whether by crafting human-like e-mails or simulating
lifelike customer service interactions. In such cases, labeling content
as AI-generated could undermine the very purpose for which consumers
use these systems.
Instead, policymakers should support content provenance mechanisms.
Content provenance mechanisms are techniques used to trace and
establish the origin or source of digital content, whether it's text,
images, videos, or any other form of data. For example, one technique
is to embed secure metadata within digital files to provide information
about the author, creation date, location, and other relevant details.
Metadata helps users trace the origin of the content they interact
with, whether it's AI-generated, human-made, or a hybrid of both. This
approach provides transparency without mandating a disclosure that
might compromise the utility of AI systems. It also addresses concerns
about the proliferation of misinformation on social networks by
allowing users to verify the source of the content they encounter.
The private sector is already conducting work in developing tools
and research for content provenance, such as the industry-led Coalition
for Content Provenance and Authenticity (C2PA), an initiative that is
developing standards and technologies for verifying the authenticity
and provenance of digital media content to combat misinformation,
establish trust, and increase transparency. The subcommittee should
encourage and fund NIST to support and bolster such work.
In conclusion, we appreciate the opportunity to provide our
insights on enhancing AI transparency for consumers. Transparency can
play a valuable role in achieving algorithmic accountability for some
applications and we encourage the subcommittee to support research into
when and how this mechanism can be used to support greater use of AI
for consumers.
______
Prepared Statement of Jennifer Huddleston, Research Fellow,
Cato Institute
Chair Hickenlooper, Ranking Member Blackburn, and distinguished
members of the Consumer Protection, Product Safety, and Data Security
Subcommittee.
My name is Jennifer Huddleston, and I am a technology policy
research fellow at the Cato Institute. My research focuses primarily on
the intersection of law and technology, including issues related to the
governance of emerging technologies such as artificial intelligence
(AI). I welcome the opportunity to submit a statement regarding the
potential effects of government requirements around transparency in
artificial technology.
In this statement for the record, I will focus on two key points:
AI is a general-purpose technology that is already
frequently in use by the average consumer and a wide array of
industries. Overly broad regulations are likely to have
unintended consequences and impact numerous existing products.
The U.S. policy towards AI should remain flexible and
responsive to specific harms or risks and seek to continue the
light touch approach that has encouraged innovative activity
and its accompanying benefits. Consumers and innovators, not
government, are ultimately the best at deciding what
applications of a new technology are the most beneficial to
consumers.
The most basic definition of AI is a computer or robot that can
perform tasks associated with human intelligence and discernment. Most
of us have been encountering AI much longer than we realize in tools
like autocorrect, autocomplete, chatbots, and translation software.
While generative AI has garnered recent attention, the definitions of
AI typically found in most policy proposals would impact far more than
the use of ChatGPT.
The result is a proposal to require transparency that a product
uses AI, which would mandate disclosures on many already common
products like search engines, spellchecks, and voice assistants on
smart phones without any underlying change to the product. Not only
would this require compliance costs for these existing products, but it
also impacts the value of the underlying transparency. If nearly every
product requires a warning that it uses AI at some level in its
processes, then such a transparency requirement becomes nearly
meaningless to the consumer who is fatigued from seeing the same
constant disclosure. Furthermore, such general labels fail to provide
meaningful information that consumers can understand and act upon if
desired. The government should not be in the business of being a user
interface designer as the best ways to communicate such information
will vary from product to product and use case to use case.
Because AI is a broad general-purpose technology, one-size-fits-all
regulations are likely a poor fit. Building on the success of past
light touch approaches that refrained from a precautionary approach to
technologies including the internet, policymakers should resist the
urge to engage in top-down rulemaking that attempts to predict every
best-and worst-case scenario and accidentally limits the use of
technology. Industrial policy that seeks to direct technological
development in only one way may miss the creative uses that
entrepreneurs seeking to respond to consumer demands would naturally
find. For this reason, policymakers should ensure regulations are
carefully targeted at specific harms or applications that would be
certain or highly likely to be catastrophic or irreversible instead of
broad general-purpose regulations.
Instead of a top-down approach, government should also consider the
ways that it can work with innovators and consumers to resolve concerns
that may require a degree of certainty, but for which static regulation
is likely to be problematic. This should include both looking at the
possibility of removing potential barriers as well as supporting the
development of industry best practices as appropriate. If necessary,
these best practices could be more formalized to address concerns
around specific harms or create legal certainty. As I discussed in
comments to the NTIA in June 2023:
``Soft law tools--such as multi-stakeholder working groups to
develop best practices--may help identify appropriate limits on
certain applications while also providing continued flexibility
during periods of rapid development. These soft law tools also
support the interactions of various interest groups and do not
presume that a regulatory outcome is needed. In other cases,
they may identify areas where deregulation is needed to remove
outdated law or where hard law is needed to respond to harms or
create legal certainty around practices. This also provides
opportunities for innovators to learn of regulators and
society's concerns and provide solutions that may alleviate the
sense of unease while still encouraging beneficial and flexible
uses of a technology.
Soft law tools and best practices can be formalized in a way
that provides opportunities for transparency and information
sharing. In creating voluntary standards, best practices and
other soft law tools can also address areas where direct harm
is less clear, but a variety of concerns exist.'' \1\
---------------------------------------------------------------------------
\1\ https://www.cato.org/public-comments/public-comment-regarding-
ai-accountability-policy
Innovation is often disruptive and can bring with it a sense of
unease and even fear. While such uncertainty and fear is currently seen
with regards to AI, previous advances in information technologies and
media have had similar concerns around many issues, including trust of
information.\2\ Societal norms often develop around the appropriate
uses of technology and allow for a more flexible and responsive
approach to concerns than a static law or regulation. Policymakers
should build on the success of past light touch regulatory approaches
that have made the U.S. a leader in technological innovation and
narrowly tailor necessary interventions to respond to otherwise
unaddressed harms and specific uses or applications of the technology.
---------------------------------------------------------------------------
\2\ https://www.techdirt.com/2019/02/28/deception-trust-deep-look-
deep-fakes/
---------------------------------------------------------------------------
______
Response to Written Questions Submitted by Hon. Brian Schatz to
Victoria Espinel
Question 1. Do people have the right to know when they are viewing
AI-generated content or when they are interacting with an AI-chatbot?
Answer. Transparency is an important part of building trust in AI
systems, and it can be realized in a range of different ways. BSA
believes that companies that use AI systems to interact with
individuals should inform those individuals that they are interacting
with an AI system (such as a chatbot). BSA also supports the
development of technical measures that can help users identify when an
image is authentic.
Question 2. What are the dangers if individuals do not have a clear
understanding that they are viewing or interacting with AI-generated
content?
Answer. This is an important issue because deceptive content, such
as deepfakes, can be indistinguishable from real content--with
significant consequences. For example, deepfakes can lead people to
believe that individuals have said or done things that they did not say
or do. This can be especially harmful in scenarios, such as political
campaigns, where it can spur confusion and misinformation. One concrete
step we can take is to help consumers understand when the content they
are seeing is real, which is critical to confronting disinformation.
For example, the Content Authentication Initiative (CAI) and the
Coalition for Content Provenance and Authenticity (C2PA) have been
doing important work in this area to develop open standards that can
create a stamp of authenticity.
Question 3. Do people have the right to know what model created the
AI-generated content they are viewing or interacting with? Should
generative AI systems affix information to AI-generated content
reflecting that the content is generated by AI? Should that information
be permanently affixed to the extent doing so is technically feasible?
Answer. Promoting transparency in AI systems is an important goal,
but the type of information that is relevant to consumers will vary
greatly depending on the type of AI system the consumer is interacting
with and the purpose of that AI system. In many circumstances,
identifying the model creator may not meaningfully further consumer
transparency, because consumers may be more focused on other
information, such as a disclosure that the consumer is interacting with
an AI chatbot. Consumers should also know when an image or video is
authentic, and there are helpful industry standards being developed to
address this concern. If a watermark has been affixed to an authentic
image, it should not later be removed.
Question 4. What role should platforms play in detecting AI-
generated content on their sites and in informing consumers when
content is AI-generated content?
Answer. BSA represents enterprise software companies that provide
business-to-business (B2B) products and services that help their
business customers utilize AI.
In this B2B capacity, BSA member companies generally do not
interact directly with individual consumers (unlike consumer-facing
platforms). While the most suitable way to achieve the goals of
transparency with consumers with respect to AI-generated content will
vary depending on how a business uses AI and how it interacts with
individual consumers, many consumer-facing companies may be able to
leverage open standards that help consumers trace the origin of
different forms of media, such as the standards being developed by CAI
and C2PA.
______
Response to Written Question Submitted by Hon. Ben Ray Lujan to
Victoria Espinel
Question. How can a sector-or use-case specific approach improve AI
transparency?
Answer. AI is used in different ways by different sizes of
companies that operate in different industries. AI policies--including
policies that promote transparency--should reflect that these different
types of uses of AI create different risks. A use-specific approach is
helpful for promoting transparency because different types of
information will be relevant to individual consumers depending on the
type of AI system being used and the purpose of the system.
Regulations should account for these different uses, focusing on
those that pose the highest risk to individuals--those that determine
eligibility for housing, education, employment, credit, healthcare, or
insurance. It is important to note that many uses of AI, such as
optimizing manufacturing or supply chain logistics, will not be
consumer-facing and, therefore, applying consumer-focused transparency
requirements may not be appropriate.
______
Response to Written Questions Submitted by Hon. John Hickenlooper to
Victoria Espinel
International AI Regulation:
Question 1. The international community is closely watching what
guardrails the United States establishes on high-risk AI systems. At
the same time, the international community is moving forward with their
own versions of legislation or frameworks to promote responsible use
and development of AI systems. For example:
The European Union is moving forward on passing their AI Act
The United Kingdom is hosting an inaugural AI Summit this
November focused on AI safety
The United Nations and OECD are also moving forward with AI
safety principles
Could you discuss the importance of the U.S. demonstrating
international leadership on AI and building international consensus on
our principles with our international allies?
Answer. The United States has a unique opportunity to lead the
international conversation on the responsible use and development of AI
because it is the home of leading AI innovators, many of which are BSA
members. The United States should foster an environment that enables AI
innovation to continue to thrive while also establishing protections
that address its risks. The United States has recognized the enormous
benefits of working with other countries on AI governance issues and is
already participating in a range of global efforts including through
the G7 Hiroshima AI project and the Organisation for Economic Co-
Operation and Development (OECD). The U.S. voice in those efforts will
be stronger if the United States adopts national AI legislation that
creates clear guardrails for how companies develop and deploy high-risk
AI systems. Other jurisdictions, such as the EU, are moving forward
with their approaches. Now is the time for Congress to act.
Federal Capabilities:
Question 2. Several Federal agencies and offices have launched
initiatives to advance safe and trustworthy AI, including the
Department of Commerce, Department of Energy, and the Federal Trade
Commission. In your view, which Federal agencies are equipped to
implement aspects of AI policy? What strengths do these agencies offer
to support trustworthy AI development or deployment, respectively?
Answer. Because companies in all industries are adopting AI, the
use of AI technologies will implicate areas covered by a range of
agencies, e.g., the Food and Drug Administration, the Federal Trade
Commission, the Department of Justice, the Consumer Financial
Protection Bureau, the Department of Commerce, and the National Highway
Traffic Safety Administration. Each agency has a role to play in
providing domain-specific expertise on issues falling within its
jurisdiction. It is also important to ensure that Federal agencies
coordinate with each other in their approach to AI, to promote
consistency in the Federal government's overall AI policy. For example,
the National Institute of Standards and Technology (NIST) can play a
role in supporting risk management efforts that leverage its work in
developing the AI Risk Management Framework (NIST AI RMF).
Consumer Notice:
Question 3. Generative AI systems are becoming more prevalent in
the form of customer support chatbots and personal assistants. Should
consumers be informed whether they are interacting with a generative AI
system? What is the most effective way to plainly and easily notify
consumers of this interaction?
Answer. Transparency is an important part of building trust in AI
systems. One way to promote transparency is for companies that use
customer support chatbots and personal assistants to inform consumers
that they are interacting with an AI chatbot. In practical terms, an AI
chatbot could inform the consumer that it is powered by AI in response
to the first query a consumer submits.
Review Boards:
Question 4. Some companies have established Internal Boards to
assess and determine whether an AI system is developed with safety,
risk mitigation, and ethics in mind. The majority of Internal Boards
consist of personnel within the company, and may not always include
independent experts. In general, what is the typical review process for
an AI system conducted by engineers or an Internal Board within a
member company? If engineers or Internal Board members draw different
conclusions after an AI system review, how are disputes resolved among
these parties?
Answer. Companies adopt varied approaches to AI governance,
including the adoption of Internal Boards or other bodies that oversee
the development and use of AI-related products and services. For
example, companies may establish a process for reviewing AI systems
that entails an initial review by designated personnel, and then a
subsequent and more detailed review for systems identified as having a
higher risk profile. That subsequent review may be undertaken by an
internal review board or other body charged with oversight of the
company's development and use of AI-related products and services. That
body is typically empowered to identify risk mitigations for engineers
to implement. BSA has promoted the use of risk management programs that
build on the NIST AI RMF. For example, companies can apply the NIST AI
RMF's ``govern'' function to establish policies, processes, procedures,
and practices across their organization to map, measure, and manage AI
risks.
______
Response to Written Questions Submitted by Hon. Peter Welch to
Victoria Espinel
Question 1. The European Union's (EU) Artificial Intelligence (AI)
Act is set to become the world's first comprehensive legal framework
for AI. Although EU regulations in AI and privacy only apply within the
EU market, U.S. Internet companies that operate internationally tend to
apply EU regulations everywhere they operate, which is known as the
``Brussels Effect.''
a. As Congress considers comprehensive AI legislation, what lessons
can be learned from the enactment and implementation of the EU's AI
Act, particularly in the realm of consumer protection and transparency?
Answer. Trilogue negotiations on the EU AI Act are ongoing, so the
text of the legislation is not final. But one important aspect of the
EU AI Act is its risk-based approach to AI regulation. The EU AI Act
focuses on uses of AI that create specific concerns and attempts to
address those concerns from a risk management perspective. This
approach ensures that regulations focus on high-risk uses of AI, such
as credit scoring, and does not impose overly burdensome requirements
on uses of AI that create low risks to individuals. We have encouraged
the EU to maintain this risk-based approach, which we believe is
important to any AI regulation.
The transparency measures in the EU Council's version of the EU AI
Act are also instructive. The EU AI Act ensures that developers of
high-risk AI systems provide deployers with information regarding the
capabilities and limitations of the system, including any known or
foreseeable circumstances that may lead to risks to health, safety, or
fundamental rights, and human oversight measures. This ensures that key
information about an AI system is provided by companies developing the
AI system to the companies using that system.
In addition, the EU AI Act requires developers to design AI systems
to inform individuals that they are interacting with AI systems. In
addition, users of systems that generate deepfakes must disclose that
the content has been artificially generated, with exceptions for
criminal law enforcement and artistic content. We recognize the
importance of these issues and generally agree that consumers should be
told when they are interacting with AI chatbots, and support efforts to
help consumers know when an image is authentic.
b. What impact do you expect EU's AI policy to have on
technological innovation in the U.S.?
Answer. It is not yet clear what impact the EU's AI policy will
have on technological innovation in the United States because the text
of the EU AI Act is still being finalized. The original proposal would
establish a risk-based approach to regulating AI, and BSA believes that
a risk-based approach is the right way for policymakers to address AI.
Focusing obligations on high-risk uses of AI ensures that low-risk uses
of AI are not subject to overly burdensome requirements that may impede
the ability of companies to continue to innovate, while creating
important guardrails on high-risk uses of AI technologies. The full
impact of the legislation will also be determined by how other
outstanding issues are resolved, including how responsibilities are
allocated along the AI value chain.
______
Response to Written Questions Submitted by Hon. Ted Cruz to
Victoria Espinel
Question 1. Impact Assessments. In your testimony, you stated that
AI legislation should require companies to ``conduct annual impact
assessments for high-risk uses of AI.''
You note that impact assessments are ``already widely used in a
range of other fields.'' However, the frequency of their use does not
equate to good policy. For example, when I hear ``impact assessment,''
I am reminded of the National Environmental Policy Act (NEPA)--a policy
that has received criticism from both sides of the aisle for imposing
crushing requirements on American businesses. While initially concise,
NEPA assessments have ballooned to over 600 pages on average, with
appendices alone averaging over 1,000 pages.\1\ As you can imagine, the
time investment to produce a document like this is significant,
averaging 4.5 years and, in some instances, stretching beyond 17
years.\2\ This protracted and resource-intensive process often derails
or delays vital public projects and increases costs substantially.\3\
---------------------------------------------------------------------------
\1\ Council on Environmental Quality, ``Length of Environmental
Impact Statements (2013-2017),'' Executive Office of the President,
July 22, 2019. https://ceq.doe.gov/docs/nepa-practice/
CEQ_EIS_Length_Report_2019-7-22.pdf.
\2\ Eli Dourado, ``Why are we so slow today?,'' The Center for
Growth and Opportunity, Mar. 12, 2020. https://www.thecgo.org/
benchmark/why-are-we-so-slow-today.
\3\ Ibid.
---------------------------------------------------------------------------
Given how rapidly AI is advancing and how revolutionary it promises
to be, we do not have years to waste on bureaucratic processes that are
ineffective, subjective, and hinder the adoption of important
innovation.
a. Recognizing that many of the ideas for AI impact assessments
mirror NEPA requirements, how would you ensure that they are less
expensive and time-consuming than NEPA? Please provide concrete
evidence of the individual impact of mandatory impact assessments on
large, medium, and small businesses.
Answer. Impact assessments are an important accountability tool and
should be designed to focus on a key set of issues. Clearly focusing an
impact assessment on a core set of information can reduce the burden on
companies obligated to perform them. For example, an impact assessment
for developers of an AI system would generally focus on the intended
use of the AI system, the known limitations of the system, the known
likely and specific high risks that could occur and the steps taken to
mitigate those risks, an overview of the type of data used to train the
system, and a summary of how the system was evaluated prior to sale.
Legislation can create a clear and focused set of requirements for
impact assessments, which will help maximize their use as an
accountability tool.
Notably, impact assessments may also be leveraged across multiple
jurisdictions. For example, ten states have privacy laws that will
require privacy impact assessments for certain activities. In many
cases, a company operating across state lines will be able to invest
its resources in creating a process for conducting impact assessments
that can comply with requirements across these jurisdictions.
In NEPA, mitigating potential negative impacts often involves
changing the way a particular action is undertaken to minimize
environmental harm. In AI, mitigation could involve changing an
algorithm, employing differential privacy techniques, or even
reconsidering the deployment of a particular AI system.
b. Please list all companies you represent that support impact
assessments being required by the government. Do these companies
support conducting them ``annually''?
Answer. BSA supports legislation that requires annual impact
assessments for companies that develop or deploy high-risk AI. A list
of BSA member companies is available at bsa.org/membership.
c. Do you view internally-conducted, voluntary impact assessments
as valuable? Why or why not?
Answer. Yes. Impact assessments are conducted by internal corporate
teams and are important accountability tools because they are
instrumental in helping companies identify and mitigate risks
associated with an AI system. Many state privacy laws, including the
recently-enacted privacy law in Texas, require companies to undertake
internally-conducted impact assessments when they engage in certain
activities. Impact assessments are designed to identify risks
associated with those activities and to drive changes by the internal
product teams responsible for developing a product or service, so that
those teams can mitigate the identified risks. When impact assessments
are conducted internally, they can also take into account information
that may be difficult to disclose to third parties without raising
trade secret and privacy concerns.
Although voluntary impact assessments enhance the accountability of
the companies that perform them, BSA supports legislation that would
require impact assessments for all companies that develop or deploy
high-risk AI, which we believe will help build trust and confidence in
adoption of AI technologies.
Question 2. Regulation. In your testimony, you said that ``Congress
should not wait to enact legislation that creates new obligations for
companies that develop and use AI in high-risk ways.'' As I noted in my
opening statement submitted for the record, the U.S. already has laws
and regulations that protect consumers from harms, such as the Civil
Rights Act of 1964 and the Fair Credit Reporting Act. The use of any AI
system would still be subject to these laws.
a. Given the many existing regulations that we can enforce for AI
systems across all sectors, why do you think the time is now to create
new regulations?
Answer. We recognize that existing laws apply to AI technologies,
and BSA has encouraged agencies to undertake a review of existing laws
to identify potential gaps in those authorities. At the same time, BSA
supports legislation that would require companies that develop or
deploy high-risk AI systems to conduct impact assessments. Those
assessments will help companies identify and mitigate risks that arise
from an AI system, including the potential that the system may violate
an existing law or may have a significant impact on a consumer's
privacy. The United States should have a role in how AI rules are
developed globally, and the window to help shape those conversations is
closing.
b. Was this your opinion prior to the generative AI boom with
ChatGPT at the end of last year, and if not, can you please describe
how it has changed?
Answer. In 2021, BSA began calling for legislation that requires
companies that develop or deploy high-risk AI systems to conduct impact
assessments. We continue to believe that this type of legislation is
critical to identifying and mitigating risks of AI systems and
ultimately building trust in the technology.
______
Response to Written Questions Submitted by Hon. Amy Klobuchar to
Dr. Ramayya Krishnan
Artificial Intelligence (AI) and Elections: Artificial intelligence
(AI) has the potential to completely upend campaigns and elections
unless there are rules of the road in place. I introduced bipartisan
legislation with Senators Hawley, Coons, and Collins to prohibit the
use of deceptive AI-generated audio, images, or video in political ads
because some of this content needs to be banned altogether-and I also
introduced the REAL Political Ads Act with Senators Booker and Bennet,
which is led in the House by Rep. Yvette Clarke (D-NY), to require
disclaimers on political ads containing AI-generated images or video so
voters are aware when this technology is used in our elections.
Question 1. Do you agree that without giving people information to
determine whether an image or video was created by AI, that generative
AI poses a risk to our free and fair elections?
Answer. We need to give citizens information that they can use to
determine whether an image or video was produced synthetically via AI.
So content creators need to label content and provide detection tools
that can process such labels to alert citizens about synthetic media
use in images and video. Note that the proposal is not taking a
position on the content that was created but rather informing the
citizen about content provenance (how it was created and where it came
from).
Question 2. As AI makes spreading deceptive content cheaper and
easier, what steps do you think should be taken to protect our
democracy?
Answer. First, we need to protect our election systems by getting
stakeholders to agree to a code of conduct and charge and resource a
lead agency (e.g., Federal election commission) to enforce this code of
conduct. Second, we need to educate citizens (the Block Center for
Technology and Society (https://www.cmu.edu/block-center/) at Carnegie
Mellon has created a voters guide to help educate citizens. Please see
https://www.cmu.edu/block-center/responsible-ai/genai-voter-guide.html)
and harness both Federal government department resources as well as
resources from academia and civil society to assist with public digital
literacy education and disinformation awareness in the run up to the
2024 U.S. elections. Third, we need content labeling standards and
detection tools. Example of a content labeling standard is synthid from
Google. The C2PA (https://c2pa.org) alliance is promoting a content
provenance standard. However, there is no existing adopted standard for
closed and open source models as well as for devices (cameras and video
units) that create the content. Having a content labeling standard will
be a major step forward.
We also need to monitor and use technology when content labels are
not available. The state of the art of deep fake identification in the
absence of labels is constantly improving. The following papers provide
a state of the art review:
https://www.nature.com/articles/s41598-023-34629-3
https://arxiv.org/ftp/arxiv/papers/2103/2103.00484.pdf
https://par.nsf.gov/biblio/10356295
https://www.pnas.org/doi/full/10.1073/pnas.2110013119
Scams, Fraud, and Phishing: According to experts it only takes
three seconds of audio to clone a voice. That means with just a small
clip of audio taken from social media or another source, scammers can
produce an almost perfect voice match using easily accessible
artificial intelligence (AI) programs.
Question 3. What responsibilities do developers and deployers of AI
systems have to prevent their technologies from being used to defraud
consumers?
Answer. We need a robust synthetic media code of conduct. Hand in
hand with this code of conduct is technology to establish content
provenance. Audio watermarking technology can mitigate fraud or scams
as long as content labeling and detection requirements are in place.
The U.S. Government should invest in research that will continually
improve digital content provenance and watermarking and as well in
methods to detect synthetic media enabled fraud and scams in order to
stay ahead of scammers. An enforceable code of conduct with lead
agencies resourced and charged with enforcement (e.g., Federal election
commission, CPFB, FTC) combined with technical tools will provide the
incentives and norms for all developers and deployers to create a more
trusted digital system.
Impact of AI on the Workforce: As we look to the future, it is
critical to be able to measure the impact artificial intelligence (AI)
will have on wages and economic opportunity. We have heard both
positive and negative predictions about the impact AI will have on our
economy. I am concerned we do not have enough data to inform effective
policymaking.
Question 4. What types of data about the economic impact of AI do
you think is important to collect to inform policy making?
Answer. The statistical agencies (BLS, BEA, Census) are great
assets. They are comprehensive in that the data they collect span rural
and urban geography. The data are also representative of the U.S.
population. However, the agencies lack high frequency data. Fast moving
phenomena and disruptions require policy makers to have situational
awareness.
Commercial data acquirers and providers (linkedin, workday and
lightcast to name a few) have high frequency data but the data is
neither comprehensive nor representative. A 2020 national academies
report on the consumer food distribution system made very similar
observations (see https://www.ncbi.nlm.nih.gov/books/NBK561996/) and
called for the appropriate combination of government and private data
sources to support specific policy maker use cases, An upcoming
National Academies Report co-chaired by Block Center of Technology and
Society leader Prof. Tom Mitchell (see https://
www.nationalacademies.org/our-work/automation-and-the-us-workforce-an-
update) will address the question you have posed in detail. In addition
to supporting the recommendations in these reports, we need continued
investments in research into how to make creative and innovative use of
our increasingly digital world to measure economic and societal
indicators relevant to policy making.
______
Response to Written Questions Submitted by Hon. Brian Schatz to
Dr. Ramayya Krishnan
Question 1. Do people have the right to know when they are viewing
AI-generated content or when they are interacting with an AI-chatbot?
Answer. Yes, I believe they do need to be informed that they are
interacting with synthetic media generated by an AI. I called for this
during my testimony to the committee.
Question 2. What are the dangers if individuals do not have a clear
understanding that they are viewing or interacting with AI-generated
content?
Answer. Our democracy is based on having access to trusted sources
of information and beliefs founded on attributing what one sees or
hears to the individual making these remarks. If a citizen is
deliberately misled to believe something about a candidate, this will
further erode trust in our election system. While I have illustrated
the problem in the content of elections, commerce could be affected as
well. Example: purchasing digital content that sounds like it was
created or produced by a reputed star (e.g., songs like Drake--please
see https://tinyurl.com/45aba3uu).
Question 3. Do people have the right to know what model created the
AI-generated content they are viewing or interacting with? Should
generative AI systems affix information to AI-generated content
reflecting that the content is generated by AI? Should that information
be permanently affixed to the extent doing so is technically feasible?
Answer. We need a content labeling and detection system. I believe
this should go hand in hand with a synthetic media code of conduct
which is enforceable. Helping a citizen to be reliably informed that
the content she is viewing (images/video) is synthetic or not will be a
big step forward. Further capacity to inspect the content provenance
(e.g., which model/device created it and its modification history)
would be relevant for auditors. The https://c2pa.org/ is a industry-led
coalition seeking to promote content provenance standards. Their
solution calls for this information to be affixed to digital content.
Question 4. What role should platforms play in detecting AI-
generated content on their sites and in informing consumers when
content is AI-generated content?
Answer. If content standards existed, it would be straightforward
for platforms to use the detection tool offered by the content labeling
technology to inform consumers on the platform about whether the
content they are interacting with is synthetic media generated by an
AI. In the absence of such a standard, classifying content as AI
generated or not is not perfect and errors of omission and commission
can be committed. Given that detection is not perfect, reliance on some
combination of technology tools and crowdsourced complaints, and use of
oversight boards will likely be the response of platforms. They do not
want or should not be seen as censoring speech. For example, platforms
like Meta have oversight boards who provide advice to platforms such as
in the case of doctored video of Pres. Biden. In that case, Meta stated
``. . . it will remove a video if it ``has been edited or synthesized .
. . in ways that are not apparent to an average person, and would
likely mislead an average person to believe a subject of the video said
words that they did not say.''
Question 5. The AI Labeling Act, which I introduced with Senator
Kennedy, would require generative AI systems to add visible labels and
metadata information identifying content as AI-generated. Please
explain the importance of both consumer-facing labels as well as
embedded labels and associated detection tools. Should these measures
be required for all AI-generated content?
Answer. Requiring content labeling and detection tools will be an
important step forward. I think this needs to go hand in hand with a
synthetic media code of conduct which is enforceable by a lead agency
with the resources and powers to do so (e.g., FEC in the case of
elections). There are technical solutions such as synthid from Google
and the C2PA (https://c2pa.org) alliance that is promoting a content
provenance standard.
However, there is no existing standard for closed and open source
models as well as for devices (cameras and video units) that create the
content in the first place. While there will always be some open source
alternatives that do not comply with these requirements/standards,
having all the major open source and closed source players voluntarily
adopt these requirements will be a major step forward.
______
Response to Written Questions Submitted by Hon. Ben Ray Lujan to
Dr. Ramayya Krishnan
Question 1. The explosion of generative AI has serious implications
for the integrity of our information ecosystems. Should Congress
require technology companies to label any AI-generated content with
disclosures and metadata that includes information on how the content
was created?
Answer. Requiring content labeling and detection tools will be an
important step forward. I think this needs to go hand in hand with a
synthetic media code of conduct which is enforceable by a lead agency
with the resources and powers to do so (e.g., FEC in the case of
elections). There are technical solutions such as synthid from Google
and the C2PA (https://c2pa.org) alliance that is promoting a content
provenance standard.
However, there is no existing standard for closed and open source
models as well as for devices (cameras and video units) that create the
content in the first place. While there will always be some open source
alternatives that do not comply with these requirements/standards,
having all the major open source and closed source players voluntarily
adopt these requirements will be a major step forward.
Question 2. Transparency is critical not only to protecting
consumers, but also protecting artists whose work may have been used--
often without permission--to train AI models. How can standardized
documentation help consumers and oversight bodies such as the U.S.
Copyright Office understand potential risks from AI systems?
Answer. I have called for ``nutrition labels'' to document what
data are used in training an AI model and what rights the model
developer has to the data that is being used for this purpose. While
licensing offers a path forward to both creators of content as well as
to model developers, the first step is having model developers declare
and document what data they used and the sources of these data. At
present, model cards or system cards provide high level statements of
data that are used but there is no standard template that provides the
level of detail required A recent proposal for a foundation model
transparency index is a useful device to monitor how well model
developers are doing on these dimensions (see https://arxiv.org/pdf/
2310.12941.pdf).
Question 3. How can a coordinated U.S. government approach to AI
transparency ensure U.S. leadership in responsible AI?
Answer. I am attaching a detailed policy memo on accountable AI
from the Block Center for Technology and Society (please see https://
www.cmu.edu/block-center/responsible-ai/cmu_blockcenter_rai-
memo_final.pdf) in response to your question.
______
Response to Written Questions Submitted by Hon. John Hickenlooper to
Dr. Ramayya Krishnan
Federal Capabilities:
Question 1. Several Federal agencies and offices have launched
initiatives to advance safe and trustworthy AI, including the
Department of Commerce, Department of Energy, and the Federal Trade
Commission. In your view, which Federal agencies are equipped to
implement aspects of AI policy? What strengths do these agencies offer
to support trustworthy AI development or deployment, respectively?
Answer. The NIST is a center of excellence of AI risk management.
They have led the development of the AI RIsk Management Framework which
provides via a feature called profiles a way to set acceptable risk
thresholds for use cases. The AI RMF has wide support in industry. So
NIST (which is in Commerce) is well positioned in partnership with
agencies such as the Department of Energy and Department of Defense to
create measurement, test and evaluation approaches and mechanisms to
operationalize AI development and deployment. These agencies have deep,
multi-disciplinary strength in AI technology, socio-technical systems,
security, privacy, harm analysis, etc.). The FTC could monitor the
competitiveness of the U.S. market for AI technology and application to
ensure it remains vibrant and vital.
Consumer Notice:
Question 2. Generative AI systems are becoming more prevalent in
the form of customer support chatbots and personal assistants. Should
consumers be informed whether they are interacting with a generative AI
system? What is the most effective way to plainly and easily notify
consumers of this interaction?
Answer. We should inform consumers that they are interacting with a
generative AI vs. a human in a customer support type of application.
This follows from our values about not fielding deceptive systems. It
is difficult to prescribe the best way to notify the user. It will
depend on context and the specific approaches are areas of active
research.
More generally, there is research literature on humans and ai in
decision making contexts. With predictive AI, researchers have studied
``algorithm aversion'' (humans rejecting use of AI) and recent work has
looked at how to minimize this type of aversion (please see https://
dl.acm.org/doi/fullHtml/10.1145/3544548.3581253). In contrast to
predictive AI, generative AI appears to be perceived differently (see
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4453958).
Addressing Vulnerabilities:
Question 3. Many AI developers test their systems prior to release
for flaws and vulnerabilities. One common method is ``red-teaming'',
where experts manually look for problematic behavior that needs to be
fixed. What are some limitations of ``red-teaming'' and how will this
practice need to expand and evolve to ensure AI systems remain safe
over time?
Answer. ``Red teaming'' in the context of AI systems is not a well
defined term. Recent examples appear to focus on ``prompt hacking'' to
produce bad behavior from the model as the sole focus of red teaming.
This is too narrow. Borrowing from its roots in cybersecurity, what is
required is AI testing done prior to deployment keeping in mind the
needs of both AI safety (reducing harm caused by reliability errors,
misuse or malfunction) and AI security (defending systems against
malicious actors (e.g., data poisoning attacks)). It is important that
the scope of this effort include impact analysis of the use of the AI
system (the system may include human and AI components) A memo on
accountable AI from the CMU Block Center for Technoloigy and Society is
available at https://www.cmu.edu/block-center/responsible-ai/
cmu_blockcenter_rai-memo_final.pdf) and a report on red teaming from
the Frontier Model Forum is a useful resource (https://
www.frontiermodelforum.org/uploads/2023/10/FMF-AI-Red-Teaming.pdf).
Post deployment there is need for an institutional infrastructure
(pleasee the call for a trust infrastructure in my senate testimony at
https://www.commerce.senate.gov/services/files/96B6B41C-9335-43
AF-9DB1-1231AF66C493) much like there is with the computer emergency
response team (CERT) for information security. We need a CERT for AI
which will catalog ai failures, incidents and vulnerabilities and
conduct forensic analysis of these to inform model developers who will
develop model updates to address these failures to ensure safer model
deployments. Creating safe AI ecosystems will require additional
research and infrastructure to support AI safety and security through
the AI lifecycle.
______
Response to Written Questions Submitted by Hon. Peter Welch to
Dr. Ramayya Krishnan
Question 1. Earlier this year, Senator Bennet and I reintroduced
the Digital Platform Commission Act, first-ever legislation to create
an expert Federal agency to provide comprehensive regulation of digital
platforms to protect consumers, promote competition, and defend the
public interest. We updated the text to incorporate the regulation of
AI and social media, which includes the mandate, jurisdiction, and
tools to develop and enforce rules for a sector that has gone virtually
unregulated.
a. Similar to the government's strategy towards the regulation of
social media and data privacy, the proposed framework for AI
regulations incorporates a fragmented approach that includes multiple
different Federal agencies. Do you believe that a fragmented approach
is the most effective way to regulate AI?
Answer. There are two requirements. The first is to govern AI in
specific use cases and we have agencies with oversight authority with
existing laws (e.g., EEOC, CPFB etc.). These agencies have to be
resourced with expertise and AI capabilities to do their job. The
second requirement is to coordinate AI standards globally and serve as
a single point of contact to work with the EU and other governments and
their AI regulatory agencies. This requires a new agency/commission.
This commission will need to have both technology and policy expertise
that spans multiple use cases. So I think we need a federated approach
that provides for AI expertise and capability in existing agencies with
oversight over consequential use cases while standing up a new agency/
commission to coordinate globally AI (tech) policy.
b. How could the Federal government and U.S. technology companies
benefit from the creation of a single government commission focused on
the regulation of AI and digital platforms, such as what has been
proposed in my & Senator Bennet's Digital Platform Commission Act?
Answer. The principal advantage is the creation of a center of
excellence in AI technology and AI policy with a deep understanding of
the AI oversight needed in critical use cases (e.g., in lending,
hiring, health care etc.). Tech companies and the government will look
to this agency/commission as the source of expertise and information on
AI policy within the U.S. and Globally, It could also play an important
role in harmonizing AI policy with our key trading partners.
As noted above, this will require a clear scoping of mission and
responsibility as well as close partnerships (data sharing,
appointments) between the new AI agency/commission and existing
agencies and commissions that will apply existing law to govern AI in
use cases.
______
Response to Written Question Submitted by Hon. Ted Cruz to
Dr. Ramayya Krishnan
Question. Labelling. In your testimony, you said that ``Congress
should require closed-source and open-source models to actually create
this watermarking label and a detection tool to go with this label.''
a. For the companies that are already looking into watermarking and
detection methods, how often do existing, state-of-the-art tools
incorrectly identify authentic images as AI-generated?
Answer. My testimony called for model developers to adopt content
labeling (e.g., watermark) and to provide detection tools that will use
these labels to inform a citizen that they were viewing ai generated
content. To elaborate, this technology capability needs to go hand in
hand with a synthetic media code of conduct which is enforceable by a
lead agency with the resources and powers to do so (e.g., FEC in the
case of elections). Examples of content labeling standards is synthid
from Google. The C2PA (https://c2pa.org) alliance is promoting a
content provenance standard. However, there is no existing adopted
standard for closed and open source models as well as for devices
(cameras and video units) that create the content.. Having a content
labeling standard will be a major step forward.
We also need to monitor and use technology when content labels are
not available. The state of the art of deep fake identification in the
absence of labels is constantly improving. The following papers provide
a state of the art review:
https://www.nature.com/articles/s41598-023-34629-3
https://arxiv.org/ftp/arxiv/papers/2103/2103.00484.pdf
https://par.nsf.gov/biblio/10356295
https://www.pnas.org/doi/full/10.1073/pnas.2110013119
______
Response to Written Questions Submitted by Hon. Brian Schatz to
Sam Gregory
Question 1. Do people have the right to know when they are viewing
AI-generated content or when they are interacting with an AI-chatbot?
Answer. There are already a range of forthcoming and proposed laws,
voluntary commitments, and high-level guidance in the United States and
Europe which require foundation model developers, generative AI tool
builders, distributors, and others to notify people when they are
interacting with AI-generated content, including AI chatbots. Most
recently, in September, Senators Blumenthal (D-CT) and Hawley (R-MO),
announced a bipartisan framework for AI legislation intended to be a
`comprehensive legislative blueprint'.\1\ Among other proposals, the
framework states that `users should have a right to an affirmative
notice that they are interacting with an A.I. model or system.' \2\ In
July, the AI Labeling Act introduced in the U.S. Senate by Senators
Schatz (D-HI) and Kennedy (R-LA) mandates visible labels for image,
video, audio, multimedia and text.\3\ Also in July, seven leading AI
companies agreed to the U.S. White House voluntary commitments, which
are aimed at moving toward the safe, secure, and transparent
development of AI technology, including committing to earning people's
trust by disclosing when content is AI-generated.\4\ These voluntary
commitments include developing and deploying mechanisms that enable
people to understand if audio or visual content is AI-generated,
including robust provenance, watermarking, or both, for AI-generated
audio or visual content. Earlier this year, the AI Disclosure Act,
introduced in May by Representative Torres (D-NY), also required
generative AI output to include a disclaimer.\5\
---------------------------------------------------------------------------
\1\ Richard Blumenthal and Josh Hawley, Bipartisan Framework for
U.S. AI Act, https://www.blumenthal.senate.gov/imo/media/doc/
09072023bipartisanaiframework.pdf
\2\ ibid.
\3\ AI Labeling Act of 2023, S. 2691, 118th Cong, 1st Session,
(2023) https://www.congress.gov/118/bills/s2691/BILLS-118s2691is.pdf
\4\ The White House, Fact Sheet, Biden-Harris Administration
Secures Voluntary Commitments from Leading Artificial Intelligence
Companies to Manage the Risks Posed by AI, 21 July 2023, https://
www.whitehouse.gov/briefing-room/statements-releases/2023/07/21/fact-
sheet-biden-harris-administration-secures-voluntary-commitments-from-
leading-artificial-intelligence-companies-to-manage-the-risks-posed-by-
ai/
\5\ AI Disclosure Act of 2023, HR. 3831, 118th Cong, 1st Session,
(2023) https://www.
congress.gov/118/bills/hr3831/BILLS-118hr3831ih.pdf
---------------------------------------------------------------------------
Signatories to the European Union's Code of Practice on
Disinformation \6\, another voluntary framework, include companies such
as Abode, Google, Microsoft, and TikTok, among others. Companies that
have signed on to the voluntary EU Code of Practice on Disinformation
have agreed to a similar commitment, and the EU's Commissioner Vera
Jourova has recently called on these companies to label AI-generated
content.\7\ The European Union AI Act,\8\ which is expected to come
into force in 2024, also includes provisions that place transparency
requirements on AI systems. The AI Act is likely to require AI systems
that are used to generate or manipulate images, audio, or video content
that resembles authentic content to include disclosures that notify
people that the content was generated through automated means.
---------------------------------------------------------------------------
\6\ European Commission, 2022 Strengthened Code of Practice on
Disinformation, 16 June 2022 https://digital-strategy.ec.europa.eu/en/
library/2022-strengthened-code-practice-disinformation
\7\ Foo Yun Chee, AI-generated content should be labelled, EU
Commissioner Jourova says. Reuters, June 2023, https://www.reuters.com/
technology/ai-generated-content-should-be-labelled-eu-commissioner-
jourova-says-2023-06-05/
\8\ European Parliament, European Artificial Intelligence Act,
https://www.europarl.europa.eu/doceo/document/TA-9-2023-0236_EN.html
---------------------------------------------------------------------------
All these measures reflect a necessity consistently identified in
WITNESS' consultations and research:\9\ it is not possible to place the
burden on consumers and citizens to `spot' that they are interacting
with an AI-chatbot, or viewing or listening or otherwise consuming AI-
generated or edited content. People do have a right to know they are
viewing or otherwise consuming (for example, listening to)
substantively AI-based content, particularly when this occurs in
formats, contexts or settings where a reasonable person would not
expect AI to be used. Similarly they should have a right to know they
are engaging with an AI-chatbot, and to be regularly reminded of this.
---------------------------------------------------------------------------
\9\ For an overview of this research see: www.gen-ai.witness.org
---------------------------------------------------------------------------
However, any such measures need to be implemented recognizing the
complexities of explaining how AI is employed when its usage may be
partial or an accustomed and normalized use-for instance when dyslexic
people or non-native speakers of a language use AI to help make a piece
of text clearer, or when a photograph is edited with AI. Similarly, the
format in which this information is disclosed to the consumer should
also be accessible and interpretable for multiple audiences,
acknowledging the borderless nature of content dissemination and the
needs of a diverse consumer base. Proposals for a direct disclosure
should not require the creators of AI-based content to divulge
personally identifiable data as part of the `provenance' of AI content.
AI systems should never capture by default personally identifiable
information nor include other metadata that may be able to identify the
individual operating the artificial intelligence system and generating
the outputs, unless the user has been notified and they have opted-in
to include this information.
Question 2. What are the dangers if individuals do not have a clear
understanding that they are viewing or interacting with AI-generated
content?
Answer. As noted in my written testimony,\10\ there is a rapid
increase in the realism of content produced by generative AI. In the
case of AI systems that produce realistic text, this is of particular
importance given the potential privacy implications. For instance,
users may inadvertently reveal sensitive information. Even if that data
is not sold to a third party, the information may still be saved by the
system for improving performance or further research. Additionally,
these systems tend to `hallucinate', providing inaccurate answers with
an aura of credibility. More broadly, consumers may be deceived by
audiovisual AI content that seems realistic, and fall prey to fraud and
scams. As a result, we are seeing how the increasing volume of
synthetic content is already undermining consumers' confidence in
human-generated content.
---------------------------------------------------------------------------
\10\ Sam Gregory, The Need for Transparency in Artificial
Intelligence, Testimony Before the U.S. Senate Committee on Commerce,
Science and Transportation, Subcommittee on Consumer Protection,
Product Safety and Data Security, September 2023, https://www.gen-
ai.witness.org/wp-content/uploads/2023/09/Senate-Commerce-written-
testimony_FINAL.pdf
---------------------------------------------------------------------------
Threats from inadequate signaling when content has been generated
via AI are disproportionately experienced by those already at risk. In
WITNESS's research we see how threats from synthetic media impact those
already at risk because of their gender, sexual orientation,
profession, ethnicity or social group identity. For example, for years
now women have been a primary target by those misusing deepfake
technologies to create non-consensual sexual imagery. In our global
consultations, different stakeholders have also pointed to the threat
of the use of synthetic media to attack and undermine civil society and
independent media, and how it is misused to dismiss authentic
information with claims it is falsified. This is so-called plausible
deniability or the ``liar's dividend''.\11\ This pressure places a
strain on already under-resourced local and national newsrooms and
community leaders responsible for verifying digital content. With
hyperbolic rhetoric as well as the realities of advances in generative
AI undermining trust in content we encounter, human rights defenders,
journalists and civil society actors will be among the most impacted by
generative AI. My written statement has further examples of accidental,
misuse, supply chain and structural harms.
---------------------------------------------------------------------------
\11\ Robert Chesney and Danielle Keats Citron, Deep Fakes: A
Looming Challenge for Privacy, Democracy, and National Security, 107
California Law Review 1753, July 2018, https://papers.ssrn.com/sol3/
papers.cfm?abstract_id=3213954
Question 3. Do people have the right to know what model created the
AI-generated content they are viewing or interacting with? Should
generative AI systems affix information to AI-generated content
reflecting that the content is generated by AI? Should that information
be permanently affixed to the extent doing so is technically feasible?
Answer. In my written statement and in my answer to Question 3
(Senator Hickenlooper) I explain the different technical approaches to
provide consumers with information about the content they are viewing,
and the limitations of invisible watermarking, cryptographic signatures
and labeling.
In terms of visible disclosure, synthetic content co-exists with
non-AI generated content, and material that is a blend of the two (for
instance a video originally shot in a phone with synthetic audio in
only part of it, or a photograph with elements in painted or out
painted). For this reason, simply indicating the presence of AI may not
be that helpful for consumers without a more nuanced explanation of how
and where AI has been used. Additionally, most visible labels are
easily removable or non-maliciously withdrawn-for example a user may
cut a clip from a longer video made with AI that has a disclaimer at
the start, if a third user takes that video for further editing or
circulating, the disclaimer will not be attached to it. This also
raises questions about the accessibility and interpretability of
different formats by different audiences in a global information
ecosystem.
Concerning other solutions that affix information such as metadata,
these approaches should never capture by default personally
identifiable information nor include other sensitive data that may be
able to identify the user of an AI generator tool-unless they have
opted-in to include this information. Permanently affixing provenance
information when synthetic and non-synthetic artifacts co-exist in a
piece of media brings up more complexity than in situations when the
content is binary synthetic or not. Lessons from platform policies
around the use of `real names' \12\ tell us that many people--for
example, survivors of domestic violence--have anonymity and redaction
needs that can cause life threatening situations if not respected.
---------------------------------------------------------------------------
\12\ Jillian York and Dia Kayyali, Facebook's 'Real Name' Policy
Can Cause Real-World Harm for the LGBTQ Community, Electronic Frontier
Foundation, 16 September 2014, https://www.eff.org/deeplinks/2014/09/
facebooks-real-name-policy-can-cause-real-world-harm-lgbtq-community
Question 4. What role should platforms play in detecting AI-
generated content on their sites and in informing consumers when
content is AI-generated content?
Answer. Platforms play an important role in detecting and
disclosing AI-generated content to people using their sites. People and
companies working across the AI pipeline all have a duty to insert
safeguards and address the harms their work can bring. It is a recipe
for failure if the responsibility is left solely on end-users to
determine if the audiovisual content they are viewing is AI-generated,
as well as the larger context of the content they are consuming.
However, it should be noted that platforms are not able to
implement after-the-fact detection at scale in a reliable manner.
Generative AI detection for both audiovisual and textual content is
also subject to adversarial attacks, and it is not realistic to set an
expectation that puts the burden solely on distribution platforms to
detect AI-generated content. For detection to have a real impact, it
would require collaboration and standard setting across the AI
pipeline, including model developers. Currently there is not a
generally used interoperable standard for understanding how AI was used
in content creation, and for indicating this to individuals. A number
of company specific watermarking initiatives have been started,
providing indications of AI-generation and point-of-creation.
Similarly, the Coalition for Content Provenance and Authenticity (C2PA)
has introduced an open standard for metadata-based provenance
indicating how a piece of content is created, edited and shared.
However, there is as yet no widely used interoperable standard for
either watermarking or metadata-based provenance.
As highlighted in my answer to Question 1 (Senator Schatz), there
are already a range of forthcoming laws, voluntary commitments, and
high-level guidance in the United States and Europe which require
distributors such as social media platforms to disclose AI-generated
content on their sites. Any legislative proposals should ensure to
promote standardization of watermarking as well as other disclosure
techniques such as metadata-based provenance that facilitate better
detection and allow for interoperability across platforms.
______
Response to Written Questions Submitted by Hon. Ben Ray Lujan to
Sam Gregory
Question 1. To what extent are deepfakes and AI-generated
disinformation already being created in Spanish and other non-English
languages?
Answer. AI-generated or manipulated deepfakes and disinformation
are already prevalent in Spanish and other non-English languages, and
we believe that the number of cases and their sophistication will
continue to escalate in the short term, within and outside of the
United States.
In March, WITNESS started piloting a Deepfake Rapid Response Force
(the `Force'),\13\ connecting local journalists and fact-checkers
across the globe with leading AI-media forensics experts. The Force can
take cases of suspected audiovisual synthetic content in different
languages, and it has recently covered cases in Spanish, Arabic and
Russian. It is notable that a number of the cases that have been
received at WITNESS and escalated to the Deepfake Rapid Response
Taskforce have been of audio content for which we have a reasonable
suspicion to believe that it has been AI generated or manipulated.
Generative AI and deepfake usage has also been recently seen in other
non-English speaking political contexts, such as Slovakia,\14\ Brazil
\15\ and Ukraine,\16\ and reflects the ease of production of audio-
realistic AI-generated content.
---------------------------------------------------------------------------
\13\ For an example of this work, see: https://restofworld.org/
2023/indian-politician-leaked-audio-ai-deepfake/
\14\ Morgan Meaker, Slovakia's Election Deepfakes Show AI is a
Danger to Democracy, Wired, 3 October 2023, https://www.wired.com/
story/slovakias-election-deepfakes-show-ai-is-a-danger-to-democracy/
\15\ Gabriela Coelho and Gabriel Hirabahasi, Carmen Lucia manda
remover videos contra Lula com ``deep fake'', CNN Brasil, 28 October
2022, https://www.cnnbrasil.com.br/politica/carmen-lucia-manda-remover-
vid eos-contra-lula-com-deep-fake/
\16\ Bobby Allyn, Deepfake video of Zelenskyy could be 'tip of the
iceberg' in info war, experts warn, NPR, 16 March 2022, https://
www.npr.org/2022/03/16/1087062648/deepfake-video-zelenskyy-experts-war-
manipulation-ukraine-russia
---------------------------------------------------------------------------
In our recent consultation with stakeholders in Colombia,
participants pointed out that they were seeing an increasing number of
deepfakes and AI-generated content, especially in the context of
elections. Some of the examples highlighted were the fake news avatars
lauding an alleged tourism boost in Venezuela,\17\ or the deepfake in
favor of Mexican presidential candidate, Xochitl Galvez.\18\
---------------------------------------------------------------------------
\17\ Joe Daniels in Bogota and Madhumita Murgia, Deepfake `news'
videos ramp up misinformation in Venezuela, Financial Times 17 March
2023, https://www.ft.com/content/3a2b3d54-0954-443e-adef-073a4831cdbd
\18\ Rodrigo Soriano, Blanqueada y rejuvenecida: los videos de
Xochitl Galvez creados con inteligencia artificial desatan la polemica
en las redes, El Pais, 25 July 2023, https://elpais.com/mexico/2023-07-
25/blanqueada-y-rejuvenecida-los-videos-de-xochitl-galvez-creados-con-
inteligencia-artificial-desatan-la-polemica-en-las-redes.html
Question 2. Are technology companies investing enough resources
into making sure AI systems work equally well in non-English languages?
Answer. In WITNESS experimentation and consultations we have
corroborated what different studies \19\ have already shown: most
generative AI systems work best in English, they work well in high-
resourced languages such as Spanish, and they underperform in low-
resourced languages such as Punjabi \20\ or Tamil.\21\ For AI systems
to work well in non-English languages, companies need to invest in
balancing supply chain inequities, such as the fact that there is
simply more scrapable data to train models in English than in other
languages. Companies should also employ or work with speakers of other
languages to design and fine-tune models. At a broader scale, the use
of large language models and AI systems in content moderation is rarely
sufficiently localized to contexts outside English-speaking U.S. and
Europe \22\, nor is it reinforced with well-resourced, paid and
supported human content moderation in non-English languages-with
minority populations and vulnerable communities bearing the impact.
---------------------------------------------------------------------------
\19\ Viet Dac Lai et al., Chatgpt beyond english: Towards a
comprehensive evaluation of large language models in multilingual
learning, 2023, arXiv preprint, https://arxiv.org/abs/2304.05613
\20\ Alison Snyder, AI's language gap, Axios, 8 September 2023,
https://www.axios.com/2023/09/08/ai-language-gap-chatgpt
\21\ Andrew Deck, We tested ChatGPT in Bengali, Kurdish, and Tamil.
It failed, Rest of World,
6 September 2023, https://restofworld.org/2023/chatgpt-problems-
global-language-testing/
\22\ Gabriel Nicholas and Aliya Bhatia, Lost in Translation: Large
Language Models in Non-English Content Analysis, Center for Democracy
and Technology, 23 May 2023, https://cdt.org/insights/lost-in-
translation-large-language-models-in-non-english-content-analysis/
Question 3. How are non-English speaking communities uniquely
vulnerable to mis- and disinformation generated and spread artificial
intelligence?
Answer. As remarked in my answer to Question 2 (Senator Schatz),
threats from synthetic media and generative AI used deceptively and
maliciously will likely disproportionately impact those who are already
at risk because of their gender or gender identity, sexual orientation,
profession, ethnicity, belonging to a social group or because of their
language. This vulnerability reflects existing patterns of
misinformation, disinformation and harassment. For instance, generative
AI tools make it easier to create content that promotes hatred towards
minorities, and to produce it at scale. Additionally, as mentioned in
my answer to Question 2 (Senator Lujan), generative AI underperforms
with non-English languages in comparison to English, and there is an
investment gap in both AI-enabled and human content moderation that is
not conducted in the English language.
More importantly, it is worth noting that the general mistrust of
online content may make it easier to undermine or dismiss as `fake'
digital content from minorities like non-English speaking communities
in the US, putting a higher burden on them to prove that their content
is real or trustworthy.
______
Response to Written Questions Submitted by Hon. John Hickenlooper to
Sam Gregory
Open-Source AI Models:
Question 1. Open-source AI models could benefit consumers by
promoting transparency and enabling audits for accountability. They
could also create risks, like in a recent case where image watermarks
were removed by modifying only one line of open-source code. Mr.
Gregory, please explain both the benefits and risks of open-source AI
models. How do open-source foundation models benefit consumers and the
general public? In what ways could fine-tuned open-source models
potentially harm consumers, and how could we proactively mitigate these
risks?
Answer. Many of the communities and contexts in which I work have
experienced in the past, and are still experiencing, the impact of
technological concentration of power. They are currently observing the
patterns of AI tools developed in the Global North that produce
stereotypical images of their populations, underperform in their
languages, and do not have an appropriate understanding of content or
context for content moderation. I have also highlighted in my answer to
Question 2 (Senator Lujan) how supply chain inequities affect
disproportionately minority communities and other groups that have been
historically marginalized. From this perspective, open source models
can help mitigate these biases as they allow for fine-tuning or further
examination-however, it is worth noting that there are different levels
of openness in a model.\23\ Hence, an early ability to assess these
technologies for bias is key. In this regard, open-source models can
strengthen accountability, as they enable external researchers to more
easily investigate, uncover and challenge model biases and threats.
Broadening access to artifacts such as models and training datasets
also allows researchers and users to better understand systems, conduct
audits, mitigate potential risks, and find high value applications. For
example, research about watermarking by the University of Maryland was
conducted using OPT, an open-source language model developed and
released by Meta. According to the scientists, the availability of the
materials helped them understand the techniques' robustness. This
scenario can in turn encourage other researchers to test and improve
safety techniques.\24\
---------------------------------------------------------------------------
\23\ Irene Solaiman, The gradient of generative AI release: Methods
and considerations. Proceedings of the 2023 ACM Conference on Fairness,
Accountability, and Transparency, 2023. https://arxiv.org/abs/
2302.04844
\24\ John Kirchenbauer, et al., A watermark for large language
models, 2023, arXiv preprint, https://arxiv.org/pdf/2301.10226.pdf
---------------------------------------------------------------------------
While open-source generative AI enables the customization of models
to provide content and responses that are better suited to a range of
local contexts, they also enable users to evade content moderation.
Content controls are already inadequate for use at global scale, as
mentioned above, but completely open systems make vulnerable people
even more at risk to systems that have been fine-tuned to for instance,
generate non-consensual sexual images, children sexual abuse material,
biased outputs, and mis-and disinformation. We must not underplay these
potential harms. The increasing volume, ease and access to generative
AI and the possibilities for personalization that it brings, are likely
to magnify this problem. Similarly, if models to detect synthetic
content are open, they may also be rapidly subjected to adversarial
testing, triggering a cat and mouse game of generation and detection
that may make them less valuable.
To mitigate the risk of open models, Congress should consider
encouraging scientific and research communities that are conducting
active and robust testing. Gradient releases are particularly important
in relation to detection tools, for instance, but there are other
applications for which Congress should balance the benefits of
openness. In sum, a multilayered technical, regulatory and societal
approach will be necessary-one that involves and has the commitment of
the different actors across the synthetic media pipeline, from model
developers to deployers, to content distributors and platforms.
Since watermarking and other forms of durable disclosure that
material is AI-generated may be removed easily from open-source code,
legislators could consider how these watermark approaches could be
implemented (while retaining critical protections for PII) at earlier
stages of the AI pipeline, for example in training data or underlying
models. They could also consider how both novel and existing
legislation and regulation could be used to require the retention of
watermarking even in open source.
Consumer Notice:
Question 2. Generative AI systems are becoming more prevalent in
the form of customer support chatbots and personal assistants. Should
consumers be informed whether they are interacting with a generative AI
system? What is the most effective way to plainly and easily notify
consumers of this interaction?
Answer. A general principle of informed disclosure is critical
across generative-AI based content and interactions. Consumers should
not be assumed to be aware that they are interacting with realistic or
human-seeming AI-based content, chat-bots and systems. An appropriate
approach to disclosure would involve notification, visibly or audibly,
to the users when they begin interacting with a system, as well as
durable signals embedded in the interaction that indicate origins in an
AI-system. To-date it should be noted that there is not a standardized,
robust method to provide detectable traces in text-generated content
created by a chatbot. Congress could also consider if there are notable
limitations to models that promote an anthropomorphic connection to the
user such as adopting a particular persona (as seen in recently
launched Meta AI assistants)\25\, without regular reinforcement or
reminders that this is a persona. My answers to Questions 1 and 2
(Senator Schatz) provide examples of legislative initiatives regarding
the disclosure of generative AI content and reasons for providing
consumers with this information.
---------------------------------------------------------------------------
\25\ Alex Heath, Meta is putting AI chatbots everywhere, The Verge,
27 September 2023, https://www.theverge.com/2023/9/27/23891128/meta-ai-
assistant-characters-whatsapp-instagram-connect
---------------------------------------------------------------------------
Trusted Provenance:
Question 3. In some contexts, like public safety, we may want a
very secure label that clearly shows where a piece of content came
from. What technical, policy, and educational steps are needed to
increase the public's trust in the origin of critical information?
Answer. From a technical point of view, there are different
approaches to signal the provenance of a piece of content. In my
written statement I explain in further detail how invisible
watermarking, cryptographic signatures and labeling can help increase
the public's trust in information-and their limitations to each of
them. A robust approach intended for specific public safety usages, for
example critical government communications, would need to utilize a
combination of these approaches to ensure redundancy and to enable a
range of stakeholders (with varying capacities) to be able to confirm
the integrity of the content. This approach also would likely require a
pipeline of responsibility across foundation models, developers and
deployers of AI models, as well as the creators and distributors of
content to ensure these indications are durable and maintained.
In terms of policy steps, legislative proposals should avoid
introducing a blanket requirement for compulsory disclosure of
audiovisual AI. Hence, what constitutes `public safety' should be
clearly delineated and not used to undermine freedom of speech, deter
technology whistleblowing, or enable surveillance. Any regulatory
proposition should also unambiguously protect the right to privacy, and
prohibit the default collection of personally identifiable information.
For audiovisual content that is not produced via AI, legislators should
steer clear from legally requiring labeling, but could consider how
specific governmental contexts, for example, public safety agencies,
utilize these disclosure mechanisms systematically and transparently.
Concerning education, media literacy will need to accompany the
development of tools and legislation. It will be essential in order to
enable consumers to understand the meaning of a label and verifiable
provenance, and most importantly, comprehend how this information
provides a signal of trust but does not equate to truth.\26\ Relevant
research and experience on consumer communication around tamper-
proofing and accessible disclosure that exists in fields including
currency, medicine and food safety could be relevant here. A specific
usage in a particular public safety context of a set of provenance and
disclosure indicators could also be a useful way to introduce the
public to the concept and approach, rather than an attempt to do this
at a more broad level.
---------------------------------------------------------------------------
\26\ Raquel Vazquez Llorente, Trusting Video in the Age of
Generative AI, Common Place, 01 June 2023, https://
commonplace.knowledgefutures.org/pub/9q6dd6lg/release/2
---------------------------------------------------------------------------
______
Response to Written Questions Submitted by Hon. Peter Welch to
Sam Gregory
Question 1. AI systems are trained on massive datasets, from
readily available information online to personal data collected across
the internet. Generally, a larger training set corresponds with a more
powerful AI tool, incentivizing AI developers to collect and exploit
user data to strengthen their models. This raises significant concerns
about protecting consumers' sensitive personal data and privacy.
a. Would you support legislative efforts to require platforms to
allow users to opt-out of having their personal or service-generated
data collected to train AI models?
b. How could these requirements benefit consumers, and what
specific requirements should Congress consider as it explores providing
consumers with the power to opt-out?
c. What additional steps should Congress take to empower consumers
to protect their sensitive data from AI?
Answer. Users should have control over how their personal data, or
data generated while interacting with a service, is used in the
training of AI systems. This is consistent with a rights-based approach
to regulating AI that emphasizes rights such as privacy. I have
highlighted in my written statement that a rights-based approach to AI
regulation, as opposed to exclusively a risk-based approach, is
fundamental to reflect the human rights impact of AI. Legislative
efforts that protect users from having their data collected to train AI
models would also be aligned with existing data protection laws in the
U.S.\27\
---------------------------------------------------------------------------
\27\ For instance the Privacy Act of 1974; Health Insurance
Portability and Accountability Act; Gramm-Leach-Blily Act; Children's
Online Privacy Protection Act; and the California Consumer Privacy Act.
---------------------------------------------------------------------------
Following a rights-based approach that centers U.S. Constitutional
rights and fundamental human rights would mean to protect `by default'
these rights. Accordingly, users should have the choice to indicate
whether they want their data included in the training--instead of
having this information collected as a feature of the system, even with
the possibility to opt-out. Government testing with `nudging'
techniques also suggests that in order to protect consumers' rights, it
would be more desirable in this situation to give users the choice to
opt-in.\28\
---------------------------------------------------------------------------
\28\ On data protections and nudging, see: Acquisti, Alessandro, et
al., Nudges for privacy and security: Understanding and assisting
users' choices online. ACM Computing Surveys (CSUR) 50.3 (2017): 1-41;
Jan M. Bauer, Regitze Bergstr The European Union is moving forward on passing their AI Act
The United Kingdom is hosting an inaugural AI Summit this
November focused on AI safety
The United Nations and OECD are also moving forward with AI
safety principles
Could you discuss the importance of the U.S. demonstrating
international leadership on AI and building international consensus on
our principles with our international allies?
Answer. It is imperative for the U.S. to advance international
consensus on overarching principles to govern the development and use
of AI and to remain a global leader in this space. The U.S. can remain
a global leader by developing a vision for AI development and
deployment in the United States, which includes taking a thoughtful,
targeted approach to policymaking that is grounded in international,
consensus-based standards, while also collaborating with partners
around the world to ensure that any approaches to AI governance are
interoperable and aligned to the extent possible. The potential for
regulatory divergence is great, especially if countries undertake the
development and deployment of AI in a vacuum. The Global Partnership on
AI and the G7 Hiroshima AI Process have the potential to yield
important principles and guidelines for advanced AI systems, and the
U.S. should remain engaged in these and other multilateral efforts to
help produce a vision of responsible AI.
Federal Capabilities:
Question 1. Several Federal agencies and offices have launched
initiatives to advance safe and trustworthy AI, including the
Department of Commerce, Department of Energy, and the Federal Trade
Commission. In your view, which Federal agencies are equipped to
implement aspects of AI policy? What strengths do these agencies offer
to support trustworthy AI development or deployment, respectively?
Answer. Consistent with ITI's written testimony, before
comprehensive AI legislation is introduced that governs the development
and deployment of AI technologies to consumers and enterprises,
Congress should examine the current legal and regulatory landscape to
identify policy gaps that warrant legislation. Several Federal agencies
already possess sufficient legal authority and jurisdiction to
implement AI policy based on public input from the stakeholder
community. For example, ITI has submitted feedback to multiple requests
for public comments by the Department of Commerce, including its
different subagencies and bureaus, and the Office of Science and
Technology Policy (OSTP) in the White House, including:
NIST AI Risk Management Framework: Request for Information,
Concept Paper, First Draft, and Second Draft
NTIA Request for Comments on AI Accountability Policy
Misc.: OSTP Request for Information on National Priorities
for AI
Consumer Notice:
Question 2. Generative AI systems are becoming more prevalent in
the form of customer support chatbots and personal assistants. Should
consumers be informed whether they are interacting with a generative AI
system? What is the most effective way to plainly and easily notify
consumers of this interaction?
Answer. Yes, ITI and our members believe consumers should be made
aware through clear, concise, and conspicuous notifications when
interacting with generative AI systems that have the potential to
materially deceive or influence a user's decision-making. There is no
one-size-fits-all approach for how an AI interface can plainly and
easily notify consumers, but some examples include push notifications,
pop-up screens that grant consumers access to websites, and other
clearly written disclaimers. The approach should consider whether the
generative AI system materially impacts an individual's access to goods
and services.
Review Boards:
Question 3. Some companies have established Internal Boards to
assess and determine whether an AI system is developed with safety,
risk mitigation, and ethics in mind. The majority of Internal Boards
consist of personnel within the company, and may not always include
independent experts. In general, what is the typical review process for
an AI system conducted by engineers or an Internal Board within a
member company? If engineers or Internal Board members draw different
conclusions after an AI system review, how are disputes resolved among
these parties?
Answer. Internal ethics boards are another mechanism employed by
ITI companies to instill and foster greater trust in AI solutions
before such technologies are placed into the market. The review process
and board composition, understandably, will vary by company, but the
general intent is to organize a review process that collects diverse
perspectives from experts that play an essential role in overseeing the
development of AI technologies.
______
Response to Written Questions Submitted by Hon. Peter Welch to
Rob Strayer
Question 1. Information privacy protections have not necessarily
kept pace with emerging AI technologies, posing particular risk in the
health care space. Evidence shows that biases often exist in AI
diagnostic technology--for example, in the rate of detection of breast
cancers for Black Americans versus White Americans. These tools are
often not reliable and more importantly, they aren't subject to a
universal evaluation standard.
a. What are the biggest risks that the expanded use of AI poses in
the health care sector, particularly as it relates to patient outcomes,
diagnostic tools, and bias? What steps should the industry and Congress
take to mitigate these risks?
b. What specific privacy risks do AI tools pose for patients in the
health care space?
c. What steps should Congress consider to bolster patient privacy
protections as AI tools become more common in the health sector?
d. As Congress works to minimize potential harms, what steps should
we take to ensure we do not stifle AI innovations that would benefit
patients?
Answer. Regarding privacy protections, ITI broadly believes that
U.S. consumers, including medical patients, should be in full control
of how their personal data is collected, transmitted and stored.
Individuals should have the right to exercise control over the use of
their personal data where reasonable to the context surrounding the use
of personal data. These individual control rights, consistent with the
rights and legal obligations of other stakeholders, include the legal
requirements set forth in HIPAA that pertain to the right to access,
correct, port, delete, consent to, and object to the use of personal
data about themselves. For additional information on ITI's viewpoints
on consumer data privacy, we encourage staff to review ITI's Framework
to Advance Interoperable Rules (FAIR) on Privacy.
We take seriously concerns related to biased outcomes and/or biased
AI systems. Our members are committed to taking steps to mitigate bias,
including conducting impact and risk assessments throughout the
development and deployment lifecycle and coordinating with NIST to
identify standards and steps needed to mitigate bias. ITI supports
NIST's efforts to develop a standard aimed at identifying and
mitigating bias (SP-1270), as this work will be important to bolster
industry efforts. We encourage Congress to continue to support this
work, as measurement techniques will be especially important to
consistently identifying and managing bias. Bias can be introduced at
different points throughout the AI lifecycle, so an effective approach
to mitigating bias requires a holistic approach that considers risk of
bias throughout.
Question 2. Building a generative AI model requires intense
computational and financial resources. This barrier has resulted in Big
Tech firms--such as Google, Microsoft, and Meta--leading in the
development and release of AI products. Big Tech's dominance in data
gives these companies access to behavioral data insights that provide
them with competitive advantages compared to smaller, emerging tech
companies. These same Big Tech firms have made dozens of strategic
acquisitions and investments in foundational AI tech to gain access to
new technologies and capabilities, or potentially foreclose on rivals.
a. Do you believe AI will increase competition, or will it lead to
further consolidation within the Big Tech landscape?
b. What guardrails should Congress put in place to ensure the
development of generative AI will enhance competition within the tech
marketplace?
Answer. ITI members welcome increased competition across the AI
ecosystem, which extends to AI startups and small, medium, and large
enterprises. While ITI primarily represents large technology sector
companies, our members rely heavily on partnerships with AI startups
and utilizing small businesses as a customer base for future growth and
economic opportunities. Additionally, Congress should consider
legislation that encourages strategic partnerships between large
enterprises, AI startups, and other entities and institutions that
prioritize R&D projects.
______
Response to Written Questions Submitted by Hon. Ted Cruz to
Rob Strayer
Question 1. Regulation. As I noted in my opening remarks for the
record, the Biden Administration and some of my colleagues in Congress
have--unfortunately--embraced doomsday AI scenarios to justify expanded
Federal powers. Some of these proposals are extremely onerous:
licensing regimes, creating a new regulatory agency to police computer
code, and mandatory, NEPA-style impact assessments before an AI model
can be used.
These types of regulations could cripple innovation and put us at a
competitive disadvantage against AI advancements in places like China.
Just like with past stages of tech, the best ideas often come from
individuals and small startups. There is no question that new AI
regulation will target the ideas and intellectual property of
individuals. Smaller companies with fewer resources would face
disproportionately large compliance burdens that would be expensive,
confusing, and time-consuming.
Large companies with more resources may be better able to weather
such regulations. I'm concerned that large companies are dominating the
conversation about AI regulations. For example, no small companies
participated in Senator Schumer's first AI ``Insight Forum'' on
September 13. Similarly, we have only seen large AI players sign onto
the White House's ``voluntary'' commitments on AI regulation.
a. How are smaller businesses reacting to the recent push to impose
regulations on AI in the U.S., as well as in the EU?
Answer. ITI members welcome increased competition across the AI
ecosystem which extends to AI startups and small, medium, and large
enterprises. While ITI primarily represents large technology sector
companies, our members rely heavily on partnerships with AI startups
and utilizing small businesses as a customer base for future growth and
economic opportunities. Additionally, Congress should consider
legislation that encourages strategic partnerships between large
enterprises, AI startups, and other entities and institutions that
prioritize R&D projects. Haphazard or overly broad regulations will
negatively impact companies of all sizes, especially small businesses,
so it is imperative that Congress be equipped with the tools and
expertise needed to advance meaningful legislation that encourages
future AI innovation and investment in the United States while
mitigating real risks to consumers and businesses alike.
b. Can you please share some of their primary concerns with these
regulations and describe how regulations would affect innovation in the
AI field?
Answer. In general, regulatory approaches that leverage an overly
broad definition of AI, that do not take a risk-based approach, and
that mandate overly prescriptive requirements will stymie innovation.
Regulation that designates entire classes of technology as high-risk,
or that designates entire sectors of the economy as high-risk, misses
important nuance and may implicate many low-risk uses of the
technology. Additionally, regulation that provides market surveillance
authorities access to source code could serve to significantly impede
innovation.
For example, although negotiations continue in the European Union
on its AI Act, we remain concerned with the direction of certain
provisions, especially those concerning foundation models and/or
general purpose AI (GPAI) systems. Presently, both the European
Parliament and the European Council text seek to regulate the
development of foundation models and/or GPAI systems, applying high-
risk obligations to this class of AI technology. However, it is our
view that a risk-based approach should also apply to this type of AI
technology, and that further nuance is necessary in the definitions of
both foundation models and GPAI system to clearly differentiate between
the two. Only where a foundation model is implemented in a high-risk
use case should it be subject to the requirements of the AI Act.
Importantly, we are supportive of provisions that ensure a balanced
allocation of responsibilities across the value chain so that those who
implement a foundation model into a high-risk use case can comply.
For additional information, please review our recent blog posts on
the subject: The EU AI Act: What it Means for Global Policymaking and
Five Recommendations for the EU AI Act Trilogue Negotiations.
Question 2. China. I am deeply concerned that if we hamper
innovation in the United States, we will allow China to get ahead of
us. Xi Jinping views AI as key for China's military capabilities, and
the Chinese Communist Party (CCP) will stop at nothing to accelerate
their development and adoption. They are not holding themselves back on
questions of ethics. While we want to be responsible, we need to be
realistic. We need to accept some level of risk to ensure American
innovation thrives.
a. What happens if the United States fails to innovate more quickly
than China?
Answer. China has a well-established record of shifting the playing
field in its favor--whether it is creating conditions for technology
transfer through forced partnerships with Chinese companies;
establishing ambiguous and intrusive security review regimes; or
circumventing U.S. export controls laws, these unfair practices not
only create an unfair economic advantage, but may also, especially in
the case of AI innovation, pose a national security risk. Regardless of
whether China plays by the rules or not, China will continue to invest
and improve in AI technological development, innovation, and growth.
If the United States were to cede leadership in AI innovation to
China, and by extension the CCP, China will undoubtedly leverage its
state-owned enterprises and technology-driven initiatives, such as the
Belt and Road Initiative, to deploy AI technology across China's global
sphere of influence, which could establish a standard of practice
abroad that weakens fundamental human rights and exacerbates pervasive
mass-surveillance practices. Another example, according to the
Department of Homeland Security's 2024 Threat Assessment, warns that
China will likely deploy novel technologies such as AI-developed
malware to bolster the effectiveness of cyber operations and enable
larger scale attacks that threaten U.S. critical infrastructure.
[all]