[Senate Hearing 118-66]
[From the U.S. Government Publishing Office]
S. Hrg. 118-66
ARTIFICIAL INTELLIGENCE AND INTELLECTUAL
PROPERTY--PART II: COPYRIGHT
=======================================================================
HEARING
BEFORE THE
SUBCOMMITTEE ON INTELLECTUAL PROPERTY
OF THE
COMMITTEE ON THE JUDICIARY
UNITED STATES SENATE
ONE HUNDRED EIGHTEENTH CONGRESS
FIRST SESSION
__________
JULY 12, 2023
__________
Serial No. J-118-25
__________
Printed for the use of the Committee on the Judiciary
[GRAPHIC NOT AVAILABLE IN TIFF FORMAT]
U.S. GOVERNMENT PUBLISHING OFFICE
53-116 PDF WASHINGTON : 2024
COMMITTEE ON THE JUDICIARY
RICHARD J. DURBIN, Illinois, Chair
DIANNE FEINSTEIN, California LINDSEY O. GRAHAM, South Carolina,
SHELDON WHITEHOUSE, Rhode Island Ranking Member
AMY KLOBUCHAR, Minnesota CHARLES E. GRASSLEY, Iowa
CHRISTOPHER A. COONS, Delaware JOHN CORNYN, Texas
RICHARD BLUMENTHAL, Connecticut MICHAEL S. LEE, Utah
MAZIE K. HIRONO, Hawaii TED CRUZ, Texas
CORY A. BOOKER, New Jersey JOSH HAWLEY, Missouri
ALEX PADILLA, California TOM COTTON, Arkansas
JON OSSOFF, Georgia JOHN KENNEDY, Louisiana
PETER WELCH, Vermont THOM TILLIS, North Carolina
MARSHA BLACKBURN, Tennessee
Joseph Zogby, Chief Counsel and Staff Director
Katherine Nikas, Republican Chief Counsel and Staff Director
Subcommittee on Intellectual Property
CHRISTOPHER A. COONS, Delaware, Chair
MAZIE K. HIRONO, Hawaii THOM TILLIS, North Carolina,
ALEX PADILLA, California Ranking Member
JON OSSOFF, Georgia JOHN CORNYN, Texas
PETER WELCH, Vermont TOM COTTON, Arkansas
MARSHA BLACKBURN, Tennessee
James Barton, Democratic Chief Counsel
Seth Williford, Republican General Counsel
C O N T E N T S
----------
JULY 12, 2023, 3:03 P.M.
STATEMENTS OF COMMITTEE MEMBERS
Page
Coons, Hon. Christopher A., a U.S. Senator from the State of
Delaware....................................................... 1
Tillis, Hon. Thom, a U.S. Senator from the State of North
Carolina....................................................... 3
WITNESSES
Witness List..................................................... 33
Brooks, Ben, head of public policy, Stability AI, San Francisco,
California..................................................... 5
prepared statement........................................... 34
Harleston, Jeffrey, general counsel and executive vice president,
business and legal affairs, Universal Music Group, Santa
Monica, California............................................. 12
prepared statement........................................... 45
Ortiz, Karla, concept artist, illustrator, and fine artist, San
Francisco, California.......................................... 10
prepared statement........................................... 51
Rao, Dana, executive vice president, general counsel, and chief
trust officer, Adobe, Inc., San Jose, California............... 7
prepared statement........................................... 63
Sag, Matthew, professor of law, artificial intelligence, machine
learning, and data science, Emory University School of Law,
Atlanta, Georgia............................................... 9
prepared statement........................................... 71
QUESTIONS
Questions submitted to Ben Brooks by Senator Tillis.............. 98
Questions submitted to Jeffrey Harleston by Senator Tillis....... 104
Questions submitted to Karla Ortiz by Senator Tillis............. 110
Questions submitted to Dana Rao by Senator Tillis................ 115
Questions submitted to Matthew Sag by Senator Tillis............. 121
ANSWERS
Responses of Ben Brooks to questions submitted by Senator Tillis. 128
Responses of Jeffrey Harleston to questions submitted by Senator
Tillis......................................................... 138
Responses of Karla Ortiz to questions submitted by Senator Tillis 151
Responses of Dana Rao to questions submitted by Senator Tillis... 161
Responses of Matthew Sag to questions submitted by Senator Tillis 171
MISCELLANEOUS SUBMISSIONS FOR THE RECORD
Submitted by Chair Coons:
Computer and Communications Industry Association (CCIA),
letter, July 12, 2023...................................... 192
Copyright Alliance, letter, July 17, 2023, and attachment.... 195
Digital Media Association (DiMA), letter, July 12, 2023...... 201
Motion Picture Association (MPA), statement.................. 203
Screen Actors Guild-American Federation of Television and
Radio Artists (SAG-AFTRA), statement....................... 211
Songwriters Guild of America (SGA), et al., letter, July 18,
2023....................................................... 214
ARTIFICIAL INTELLIGENCE
AND INTELLECTUAL
PROPERTY--PART II: COPYRIGHT
----------
WEDNESDAY, JULY 12, 2023
United States Senate,
Subcommittee on Intellectual Property,
Committee on the Judiciary,
Washington, DC.
The Subcommittee met, pursuant to notice, at 3:03 p.m., in
Room 226, Dirksen Senate Office Building, Hon. Christopher A.
Coons, Chair of the Subcommittee, presiding.
Present: Senators Coons [presiding], Klobuchar, Hirono,
Padilla, Tillis, and Blackburn.
Also present: Chair Durbin.
OPENING STATEMENT OF HON. CHRISTOPHER A. COONS,
A U.S. SENATOR FROM THE STATE OF DELAWARE
Chair Coons. This hearing will come to order. I'd like to
thank all of our witnesses for participating today, and I'd
like to especially thank my friend and colleague, Ranking
Member Thom Tillis and his staff for working with us on such a
collaborative basis to put this hearing together.
Welcome back from Vilnius. Senator Tillis was over at the
NATO summit, and I am thrilled he is able to join us, and we
are able to do this hearing today. This is our second hearing
in as many months on the intersection of artificial
intelligence and intellectual property law and policy.
You and your team have been great partners in pursuing
this. If you will indulge me for a moment, Senator, and before
I proceed with my remarks, I'd like to ask that we play just a
little clip of something to frame the challenges of this
topic--made with the permission of all the relevant rights
holders.
[Laughter.]
[Video is shown.]
Chair Coons. Thank you for your forbearance. Yes, a round
of applause is certainly welcomed.
[Applause.]
Chair Coons. My team actually produced a version of that
where it is a duet between me and Frank Sinatra, but my voice
came out so horribly flat, I didn't want to impose that on any
of you.
Senator Tillis. Thank you, Mr. Chair, for your judgment.
[Laughter.]
Chair Coons. Creating the song, ``AI, AI'' to the tune of
``New York, New York'' was great fun and I appreciate my team
that worked so hard on pulling that off, but the very existence
of it highlights a couple of the core questions around
copyright raised by generative artificial intelligence.
ChatGPT wrote the lyrics following the style of ``New York,
New York,'' although perhaps not quite as moving and inspiring
as the original words to any but IP enthusiasts. Another
generative AI tool was used to take Mr. Sinatra's recorded
songs and his voice and his phrasing and his style and set that
to music.
So, a couple of those core questions: Did ChatGPT infringe
the copyright on ``New York, New York'' when it drafted lyrics
representing its lyrical style? What about the AI tool that set
those lyrics to music? Did either that tool or did I run afoul
of Mr. Sinatra's rights by mimicking his voice?
In my case, no, because we got specific approval. Or did I
just use AI tools to enhance my own creativity? And if so,
should this newly created song be entitled to copyright
protection? These are just a few of the questions I hope we
will explore with our panel of talented and insightful
witnesses as we consider the impact of artificial intelligence
on copyright law and policy and the creative community.
As we all know, AI is rapidly developing. Generative AI
tools have brought many new people into the creative fold and
have opened new avenues for innovation. People who may not have
ever considered themselves creatives are now using these tools
to produce new works of art. Artists themselves have used AI
tools to enhance their own creativity.
Paul McCartney recently made headlines announcing that AI
helped create the very last Beatles song 50 years after the
band broke up. As I have previewed, AI creates new copyright
law issues, including whether using copyrighted content to
train AI models is copyright infringement, whether AI-generated
content should be given copyright protection, and many more.
These questions are working their way through the courts.
How the courts decide them, and the decisions we make here in
the Senate and in Congress about how to respect and reinforce
existing copyright protections and works will have significant
consequences, not just for the creative community, but our
overall competitiveness as a country. While generative AI
models are trained on copyrighted content, IP considerations
haven't been included or sufficiently considered in proposed IP
regulatory frameworks here in the U.S.
In contrast, some of our competitors recognize IP policy as
an important tool. The EU is currently planning to require AI
companies to publish the copyrighted materials used in training
models. The UK provides copyright protection for computer-
generated works. These are just some initial concerns, and I
think there are initial steps that we can take to ensure
sustained U.S. leadership on artificial intelligence.
First, it is critical to include IP considerations in any
regulatory framework for artificial intelligence and to give
our Copyright Office, in this framework, a seat at the table.
We should also consider whether changes to our copyright laws
or whole new protections like a Federal right of publicity may
be necessary to strike the right balance between creators'
rights and AI's ability to enhance innovation and creativity.
I am excited to explore these issues today. We have got a
great panel, and great partner, and great Members of the
Committee. With Senator Tillis' cooperation, we have assembled
this wonderful panel. I will introduce them shortly, but first
I will turn it over to my Ranking Member, Senator Thom Tillis.
OPENING STATEMENT OF HON. THOM TILLIS,
A U.S. SENATOR FROM THE STATE OF NORTH CAROLINA
Senator Tillis. Thank you, Mr. Chairman, and thanks for
everyone here. It is great to see the number of participants in
the audience. It is even more amazing to see an equal number
who would like to get in here, but who says AI and IP can't be
sexy?
But, you know, in all seriousness, I appreciate that we are
having another hearing on the opportunity to highlight the
importance of intellectual property when it comes to emerging
technologies, and today we are talking about AI. During our
last hearing, we really discussed the impact of AI on patent--
in a patent context, which explored ideas such as whether or
not AI can be considered an inventor.
And it cannot, and hopefully it will not in the future. But
while many of these issues we discussed in the last hearing
were perspective, the creative community is experiencing
immediate and acute challenges due to the impact generative AI
in a copyright context.
Strong, reliable, predictable IP rights are paramount to
incentivizing U.S. innovation and creativity. It is this
innovation and creativity that fuels growth of our country's
prosperity and drives enormous economic growth. In fact, core
copyright industries added $1.8 trillion of value to U.S. GDP,
accounting for almost 8 percent of the U.S. economy.
These copyright industries also employ 9.6 million American
workers. The sales of major U.S. copyright products overseas in
markets also constitute $230 billion and outpaced exports of
other major U.S. industries. Advances in generative AI have
raised new questions regarding how copyright principles such as
authorship, infringement, and fair use will apply to content
created or used by AI.
We must not only consider how our current IP laws apply to
the field of generative AI, but also what changes, if any, may
be necessary to incentivize future AI innovators and creators.
So, Chairman Coons, I am happy to have this Committee. I
will submit the remainder of my statement for the record. But
for those of you who have watched our Committee over the past
several Congresses where either Senator Coons or I were in
Ranking Member or Chairmanship, I think if anything, I hope
people understand that we are very thorough and we are very
persistent in our approach, and we are inclusive.
I have told everyone on this issue, whichever end of the
spectrum you are, if you are at the table and the work groups,
we are going to find a reasonable solution and compromises. If
you are outside of the work group process and you are just
taking shots at it, you may find yourself on the table, from my
perspective.
So, we encourage you to get to the table and make what we
are doing better. The reason why I think it is so important,
and I am glad the IP Subcommittee is leading on this in terms
of formal hearings with a focus on potentially drafting
legislation, is I think we run the risk of some in Congress who
think AI is bad--that it's a threat to the future.
I am not in that camp. I think that AI is good. It is
something that I first developed expertise in back in the late
80s, and have followed it every sense. It is a matter of
getting the regulatory construct, the intellectual property
construct, all the other underlying policies that you need when
a new, I think positive--in a positive term, disruptive
technology hits the field.
So, the reason that we need to move forward, address
potential concerns is precisely because I want the United
States to lead in innovation. And so much innovation is going
to be premised on properly exploiting the capabilities
responsibly, and that is what I hope we learn in this hearing
and subsequent hearings and work group. So, thank you all for
being here. And thank you, Mr. Chair, for having the hearing.
Chair Coons. Thank you, Senator Tillis. I am now going to
turn to our witness panel today. We welcome five witnesses to
testify about the intersection of artificial intelligence and
copyright law.
Our first witness is Mr. Ben Brooks, head of public policy
at Stability AI, a company that develops a range of AI models
that help users generate images, text, audio, and video.
Next, we have Dana Rao, executive VP, general counsel, and
chief trust officer at Adobe. I'd like to be chief trust
officer in the United States.
Senator Tillis. They don't have titles like that.
[Laughter.]
Chair Coons. Mr. Rao leads Adobe's legal security and
policy organization, including Adobe's Content Authenticity
Initiative, which promotes transparency principles around the
use of AI.
Next, we have Professor Matthew Sag, a professor of law in
artificial intelligence, machine learning, and data science at
Emory University School of Law. Professor Sag is a leading U.S.
authority on the fair use doctrine in copyright law and its
implications for researchers in text data mining, machine
learning, and AI.
Next, we will hear from Karla Ortiz, an artist, a concept
artist, illustrator, and fine artist who has worked on a
variety of well-known and widely enjoyed projects, including
``Jurassic World,'' ``Black Panther,'' ``Loki,'' and she is
most famous for designing, in my assessment at least, for
designing Doctor Strange for Marvel's first ``Doctor Strange''
film. Welcome.
Last but certainly not least, we have Jeffrey Harleston,
general counsel and executive VP of business and legal affairs
for Universal Music Group. Mr. Harleston is responsible for
overseeing business transactions, contracts, litigation for all
of Universal Music Group's worldwide operations in more than 60
countries.
After I swear all of you in, each of you will have 5
minutes to make an opening statement. We will then proceed to
questioning. Each Senator, depending on attendance,
questioning, and time, will have a first round of 5 minutes. We
may well have a second round in 5 minutes, and we may be the
only two left for a third round of 5 minutes, but we will see.
So, could all the witnesses please stand to be sworn in.
Please raise your right hand. Do you swear or affirm that the
testimony you are about to give before this Committee will be
the truth, the whole truth, and nothing but the truth, so help
you God?
[Witnesses are sworn in.]
Chair Coons. Thank you. Mr. Brooks, you may proceed with
your opening statement.
STATEMENT OF BEN BROOKS, HEAD OF PUBLIC POLICY, STABILITY AI,
SAN FRANCISCO, CALIFORNIA
Mr. Brooks. Thank you, Chair Coons and Ranking Member
Tillis, for the opportunity to testify today. AI can help to
unlock creativity, drive innovation, and open up new
opportunities for creators and entrepreneurs across the United
States. As with any groundbreaking technology, AI raises
important questions, and we recognize the depth of concern
among creators.
While we don't have all the answers, we are committed to an
open and constructive dialog, and we are actively working to
address emerging concerns through new technology standards and
good practices.
At Stability AI, our goal is to unlock humanity's potential
with AI technology. We developed a range of AI models. These
models are essentially software programs that can help a user
to create new content.
Our flagship model, Stable Diffusion, can take plain
language instructions from a user and help to produce a new
image. We are also working on research for safe language models
that can help to produce new passages of text or software code.
AI is a tool that can help to accelerate the creative process.
In our written testimony, we shared examples of how
Broadway designers, architects, photographers, and researchers
are using our models to boost their productivity, experiment
with new concepts, or even study new approaches to diagnosing
complex medical disorders. We are committed to releasing our
models openly with appropriate safeguards.
That means we share the underlying software as a public
resource. Creators, entrepreneurs, and researchers can
customize these models to develop their own AI tools, build
their own AI businesses, and find novel applications of AI that
best support their work. Importantly, open models are
transparent.
We can look under the hood to scrutinize the technology for
safety, performance, and bias. These AI models study vast
amounts of data to understand the subtle relationships between
words, ideas, and visual or textual features, much like a
person visiting an art gallery or library to learn how to draw
or how to write.
They learn the irreducible facts and structures that make
up our systems of communication. And through this process, they
develop an adaptable body of knowledge that they can then apply
to help produce new and unseen content.
In other words, compositions that did not appear in the
training data and may not have appeared anywhere else. These
models don't rely on a single work in their training data, nor
did they store their training data. But instead, they learn by
observing recurring patterns over billions of images and
trillions of words of text.
We believe that developing these models is an acceptable
and socially beneficial use of existing content that is
permitted by fair use and helps to promote the progress of
science and useful arts. Fair use and a culture of open
learning is essential to recent developments in AI. It is
essential to help make AI useful, safe, unbiased, and it is
doubtful that these groundbreaking technologies would be
possible without it.
The U.S. has established global leadership in AI thanks in
part to an adaptable and principles-based fair use doctrine
that balances creative rights with open innovation. We
acknowledge emerging concerns, and these are early days, and we
don't have all the answers, but we are actively working to
address these concerns through safe technology, standards, and
good practices.
First, we have committed to voluntary opt-outs so that
creators can choose if they don't want their online work to be
used for AI training. We have received opt-out requests for
over 160 million images to date, and we are incorporating these
into upcoming training. We are hoping to develop digital opt-
out labels as well that follow the content wherever it goes on
the internet.
Second, we are implementing features to help users and tech
platforms identify AI content. Images generated through our
platform can be digitally stacked with metadata and watermarks
to indicate if the content was generated with AI.
These signals can help ensure that users exercise
appropriate care when interacting with AI content and help tech
platforms distinguish AI content before amplifying it online.
We welcome Adobe's leadership in driving the development of
some of these open standards.
Third, we have developed layers of mitigations to make it
easier to do the right thing with AI and harder to do the wrong
thing.
Today, we filter datasets of unsafe content. We test and
evaluate our models before release. We apply ethical use
licenses, disclose known risks, filter content generated
through our computing services, and implement new techniques to
mitigate bias. As we integrate AI into the digital economy, we
believe the community will continue to value human-generated
art and perhaps value it at a premium.
Smartphones didn't destroy photography, and word processors
didn't diminish literature, despite radically transforming the
economics of creation. Instead, they gave rise to new demand
for services, new markets for content, and new creators.
We expect the same will be true of AI, and we welcome an
ongoing dialog with the creative community about the fair
deployment of these technologies. Thank you for the opportunity
to testify, and we welcome your questions.
[The prepared statement of Mr. Brooks appears as a
submission for the record.]
Chair Coons. Thank you, Mr. Brooks. Mr. Rao.
STATEMENT OF DANA RAO, EXECUTIVE VICE PRESIDENT, GENERAL
COUNSEL, AND CHIEF TRUST OFFICER, ADOBE, INC., SAN JOSE,
CALIFORNIA
Mr. Rao. Chair Coons, Ranking Member Tillis, and Members of
the Committee, thank you for the opportunity to testify here
today.
My name is Dana Rao, and I am general counsel, and, as
Senator Coons noted, chief trust officer at Adobe. I am happy
to provide you with this secret certificate you need to get
that title, if you would like, after the hearing.
Since our founding in 1982, Adobe has pioneered
transformative technologies in all types of digital creation,
from digital documents like PDF to image editing with
Photoshop. Our products allow our customers who range from
aspiring artists to wartime photojournalists, to advertisers
and more, to unleash their creativity, protect their craft,
empower their businesses in a digital world.
AI is the latest disruptive technology we have been
incorporating into our tools help creators realize their
potential. You have all seen the magic of text to image
generative AI. Type in the prompt, cat driving a 1950s
sportscar through the desert, and in seconds you will see
multiple variations of a cat on a retro road trip appear before
your eyes.
We have launched generative AI in our own tools, Adobe
Firefly, and has provided--this proved to be wildly popular
with our creative professionals and consumers alike. In my
written testimony, I explore a comprehensive framework for
responsible AI development that includes addressing
misinformation, harmful bias, creative rights, and intellectual
property.
Today, given Adobe's focus and our millions of creative
customers and our leadership in AI, I will focus on how the
United States can continue to lead the world in AI development
by both supporting the access to data that AI requires and
strengthening creator rights.
The question of data access is critical for the development
of AI because AI is only as powerful and as good as the data on
which it is trained. Like the human brain, AI learns from the
information you give it.
In the AI's case, the data it is trained on. Training on a
larger dataset can help ensure your results are more accurate
because the AI has more facts to learn from. A larger dataset
will also help the AI avoid perpetuating harmful biases in its
results by giving it a wider breadth of experiences from which
it can build its understanding of the world. More data means
better answers and fewer biases.
Given those technical realities, United States and
governments should support access to data to ensure that AI
innovation can flourish accurately and responsibly. However,
one of the most important implications of AI's need for data is
the impact on copyright and creators' rights.
There are many outstanding questions in this space,
including whether creating an AI model, which is a software
program, from a set of images, is a permitted fair use. And
whether that analysis changes if the output of that AI model
creates an image that is substantially similar to an image on
which it is trained.
These questions will certainly be addressed by courts and
perhaps Congress, and we are prepared to help assist in those
discussions. Adobe recognized the potential impact of AI on
creators and society, and we have taken several steps.
First, we trained our own generative AI tool, Adobe
Firefly, only on licensed images from our Adobe Stock
Collection, which is a stock photography collection, openly
licensed content, and works that are in the public domain where
the copyright has expired. This approach supports creators and
customers by training on a dataset that is designed to be
commercially safe.
In addition, we are advocating for other steps we can all
take to strengthen creators' rights. First, we believe creators
should be able to attach a ``Do Not Train'' tag to their work.
With industry and Government support, we can ensure AI data
crawlers will read and respect this tag, giving creators the
option to keep their data out of AI training datasets.
Second, creators using AI tools want to ensure they can
obtain copyright protection over their work in this new era of
AI-assisted digital creation. An AI output alone may not
receive copyright protection, but we believe the combination of
human expression and AI expression will and should.
Content editing tools should enable creators to obtain a
copyright by allowing them to distinguish the AI work from the
human work. In my written testimony, I discuss our open
standards-based technology content credentials, which can help
enable both of these creator protections.
Finally, even though Adobe has trained its AI on permitted
work, we understand the concern that an artist can be
economically dispossessed by an AI trained on their work that
generates arts in their style, in the Frank Sinatra example you
gave.
We believe artists should be protected against this type of
economic harm, and we propose Congress establish a new Federal
anti-impersonation right that would give artists a right to
enforce against someone intentionally attempting to impersonate
their style or likeness.
Holding people accountable who misuse AI tools is a
solution we believe goes to the heart of some of the issues our
customers have, and this new right would help address that
concern. The United States has led the world through
technological transformations in the past, and we have all
learned it is important to be proactively responsible to the
impact of these technologies.
Pairing innovation with responsible innovation will ensure
that AI ultimately becomes a transformative and true benefit to
our society. Thank you, Chair Coons, Ranking Member Tillis, and
Members of the Committee.
[The prepared statement of Mr. Rao appears as a submission
for the record.]
Chair Coons. Thank you, Mr. Rao. Professor.
STATEMENT OF MATTHEW SAG, PROFESSOR OF LAW, ARTIFICIAL
INTELLIGENCE, MACHINE LEARNING, AND DATA SCIENCE, EMORY
UNIVERSITY SCHOOL OF LAW, ATLANTA, GEORGIA
Professor Sag. Chair Coons, Ranking Member Tillis, Members
of the Subcommittee, thank you for the opportunity to testify
here today. I am a professor of law in AI, machine learning,
and data science at Emory University, where I was hired as part
of Emory's AI Humanity Initiative.
Although we are still a long way from the science fiction
version of artificial general intelligence that thinks, feels,
and refuses to open the pod bay doors, recent advances in
machine learning and artificial intelligence have captured the
public's attention and apparently lawmakers' interest.
We now have large language models, or LLMs, that can pass
the bar exam, carry on a conversation, create new music and new
visual art. Nonetheless, copyright law does not and should not
recognize computer systems as authors. Even where an AI
produces images, text, or music that is indistinguishable from
human authored works, it makes no sense to think of a machine
learning program as the author.
The Copyright Act rightly reserves copyrights for original
works of authorship. As the Supreme Court explained long ago in
the 1884 case of Burrow-Giles Lithographic, authorship entails
original, intellectual conception. An AI can't produce a work
that reflects its own original intellectual conception because
it has none.
Thus, when AI models produce content with little or no
human oversight, there is no copyright in those outputs.
However, humans using AI as tools of expression may claim
authorship if the final form of the work reflects their
original intellectual conception in sufficient detail. And I
have elaborated in my written submissions how this will depend
on the circumstances.
Training generative AI on copyrighted works is usually fair
use because it falls into the category of non-expressive use.
Courts addressing technologies such as reverse engineering,
search engines, and plagiarism detection software have held
that these non-expressive uses are fair use. These cases
reflect copyright's fundamental distinction between protectable
original expression and unprotect-able facts, ideas, and
abstractions.
Whether training an LLM is in non-expressive use depends
ultimately on the outputs of the model. If an LLM is trained
properly and operated with appropriate safeguards, its outputs
will not resemble its inputs in a way that would trigger a
copyright liability. Training such an LLM on copyrighted works
would thus be justified under current fair use principles.
It is important to understand that generative AI are not
designed to copy original expression. One of the most common
misconceptions about generative AI is the notion that the
training data is somehow copied into the model. Machine
learning models are influenced by the data. They would be
pretty useless without it. But they typically don't copy the
data in any literal sense.
So rather than thinking of an LLM as copying the training
data like a scribe in a monastery, it makes more sense to think
of it as learning from the training data like a student. If an
LLM like GPT3 is working as intended, it doesn't copy the
training data at all. The only copying that takes place is when
the training corpus is assembled and pre-processed, and that is
what you need a fair use justification for. Whether a
generative AI produces truly new content or simply conjures up
an infringing cut and paste of works in the training data
depends on how it is trained.
Accordingly, companies should adopt best practices to
reduce the risk of copyright infringement and other related
harms, and I have elaborated on some of these best practices in
my written submission. Failure to adopt best practices may
potentially undermine claims of fair use.
Generative AI does not, in my opinion, require a major
overhaul of the U.S. copyright system at this time.
If Congress is considering new legislation in relation to
AI and copyright, that legislation should be targeted at
clarifying the application of existing fair use jurisprudence,
not overhauling it.
Israel, Singapore, and South Korea have recently
incorporated fair use into their copyright statutes because
these countries recognize that the flexibility of the fair use
doctrine gives U.S. companies and U.S. researchers a
significant competitive advantage.
Several other jurisdictions, most notably Japan, the United
Kingdom, and the European Union, have specifically adopted
exemptions for text data mining that allow use of copyrighted
works as training for machine learning and other purposes.
Copyright law should encourage the developers of generative
AI to act responsibly. However, if our laws become overly
restrictive, then corporations and researchers will simply move
key aspects of technology development overseas to our
competitors.
Thank you very much.
[The prepared statement of Professor Sag appears as a
submission for the record.]
Chair Coons. Thank you, Professor. Ms. Ortiz.
STATEMENT OF KARLA ORTIZ, CONCEPT ARTIST, ILLUSTRA-
TOR, AND FINE ARTIST, SAN FRANCISCO, CALIFORNIA
Ms. Ortiz. Yes. Chairman Coons, Ranking Member Tillis, and
esteemed Members of the Committee, it is an honor to testify
before you today about AI and copyright. My name is Karla
Ortiz. I am a concept artist, illustrator, and fine artist, and
you may not know my name, but you know my work.
My paintings have shaped the worlds of blockbuster Marvel
films and TV shows, including ``Guardians of the Galaxy 3,''
``Black Panther,'' ``Loki,'' you know, but specifically, the
one I am most happiest of is that I, my work helped shape the
look of Doctor Strange in the first ``Doctor Strange'' movie.
I have to brag about that a little bit, sir. I love what I
do. I love my craft. Artists train their entire lives to be
able to bring the imaginary to life. All of us who engage in
this craft love every little bit of it. Through hard work,
support of loved ones, and dedication, I have been able to make
a good living from my craft via the entertainment industry, an
industry that thrives when artists' rights to consent, credit,
and compensation are respected.
I have never worried about my future as an artist until
now. Generative AI is unlike any other technology that has come
before. It is a technology that uniquely consumes and exploits
the hard work, creativity, and innovation of others. No other
tool is like this. What I found, when first researching AI,
horrified me.
I found that almost the entirety of my work, the work of
almost every artist I know, and the work of hundreds of
thousands of artists had been taken without our consent,
credit, or compensation. These works were stolen and used to
train for profit technologies with datasets that contain
billions of image and text data pairs.
Through my research, I learned many AI companies gather
copyrighted training data by relying on a practice called data
laundering. This is where a company outsources its data
collection to a third party under the pretext of research to
then immediately use commercially. I found these companies use
big terms like ``publicly available data'' or ``openly licensed
content'' to disguise their extensive reliance on copyrighted
works.
No matter what they are saying, these models are illegally
trained on copyrighted works. To add even more insult to
injury, I found that these for-profit companies were not only
permitting users to use our full names to generate imagery but
encouraging it. For example, Polish artist Frederic Koski had
had his name used as a prompt in AI products over 400,000
times, and those are the lower end of the estimate.
My own name, Karla Ortiz, has also been used by these
companies thousands of times. Never once did I give consent.
Never once have I gotten credit. Never once have I gotten
compensation. It should come as no surprise that major
productions are replacing artists with generative AI.
Goldman Sachs estimates that generative AI will diminish or
outright destroy approximately 300 million full-time jobs
worldwide. As Ranking Member Tillis mentioned earlier,
copyright-reliant industries alone contribute $1.8 trillion
value to the U.S. GDP, accounting for 7.76 percent of the
entire U.S. economy. This is an industry that employs 9.6
million American workers alone.
The game plan is simple, to go as fast as possible, to
create mesmerizing tales of progress, and to normalize the
exploitation of artists as quickly as possible. They hope when
we catch our breath, it will be too late to right the wrongs,
and exploiting Americans will become an accepted way of doing
things.
But that game can't succeed as we are here now, giving this
the urgency it so desperately deserves. Congress should act to
ensure what we call the 3Cs and a T: consent, credit,
compensation, and transparency.
The work of artists like myself were taken without our
consent, credit, nor compensation, and then used to compete
with us directly in our own markets--an outrageous act that
under any other context would immediately be seen as unfair,
immoral, and illegal.
Senators, there is a fundamental fairness issue here. I am
asking Congress to address this by enacting laws that require
these companies to obtain consent, give credit, pay
compensation, and be transparent. Thank you.
[The prepared statement of Ms. Ortiz appears as a
submission for the record.]
Chair Coons. Thank you, Ms. Ortiz. Last but certainly not
least, Mr. Harleston.
STATEMENT OF JEFFREY HARLESTON, GENERAL COUNSEL AND EXECUTIVE
VICE PRESIDENT, BUSINESS AND LEGAL AFFAIRS, UNIVERSAL MUSIC
GROUP, SANTA MONICA, CALIFORNIA
Mr. Harleston. Thank you, Chairman Coons, Ranking Member
Tillis, and Members of the Committee. It is an honor to be here
before you today. I am Jeff Harleston. I am the general counsel
of Universal Music Group. And what is Universal Music Group? We
are the world leader in music-based entertainment.
We are home to legendary record labels such as Motown, Def
Jam, Island, Blue Note, Capitol, just to name a few. We have a
music publishing company that signs songwriters, and we have a
music merchandizing company as well, and an audio division--an
audiovisual division that produces award-winning documentaries
based on music.
UMG identifies, develops artists across every musical
genre. I think it is fair to note that Frank Sinatra is one of
our artists, and I think based on what we didn't hear today, I
am not sure if we will be pursuing a developing artist out of
Delaware named Chris Coons, but maybe we will get back to that.
Chair Coons. I am confident you will not.
[Laughter.]
Mr. Harleston. All jokes aside, I have been at the company,
I have been honored to be with the company for 30 years, and
most of the time I have spent as a lawyer, but I have also
spent some time leading the Def Jam label and also as the
management team of Geffen Records.
So, I have been on both sides of the business. We have also
helped broker deals with digital services, platforms, social
media outlets where you, all of you can access the music that
you love. It has been my life's honor to work with countless
talented and creative artists.
Their creativity is the soundtrack to our lives, and
without the fundamentals of copyright, we might not have ever
known them. I would like to make four key points to you today.
The first, copyright artists and human creativity must be
protected. Art and human creativity are central to our
identity. Artists and creators have rights. They must be
respected. If I leave you with one message today, it is this:
AI in the service of artists and creativity can be a very, very
good thing. But AI that uses, or worse yet, appropriates the
work of these artists and creators and their creative
expression, their name, their image, their likeness, their
voice, without authorization, without consent, simply is not a
good thing.
The second point I want to make is that generative AI
raises challenging issues in the copyright space. I think you
have heard from the other panelists and they all would agree.
We are the stewards at Universal of tens of thousands, if not
hundreds of thousands, of copyrighted creative works from our
songwriters and artists, and they have entrusted us to honor,
value, and protect them.
Today they are being used to train generative AI systems
without authorization. This irresponsible AI is violative of
copyright law and completely unnecessary. There is a robust
digital marketplace today in which thousands of responsible
companies properly obtain the rights they need to operate.
There is no reason that the same rules that apply to
everyone else should not apply equally to AI companies and AI
developers.
My third point, AI can be used responsibly to enhance
artistic expression. Just like other technologies before,
artists can use AI to enhance their art. AI tools have long
been used in recording studios for drum tracks, chord
progressions, and even creating immersive audio technologies.
One of our distributed artists used a generative AI tool to
simultaneously release a single in six languages in his own
voice, on the same day. The generative AI tool extended the
artist's creative intent and expression with his consent to new
markets and fans instantly.
In this case, consent is the key. There is no reason we
can't legitimate--we can't create a legitimate AI marketplace
in the service of artists. There is a robust free market for
music sampling, synchronization licensing, and deals with new
entrants to the digital marketplace, social media companies,
and all manner of new technologies. We can do the same thing
with AI.
And my fourth and final point, to cultivate a lawful,
responsible AI marketplace, Congress needs to establish rules
that ensure creators are respected and protected.
One way to do that is to enact a Federal right of
publicity. Deepfakes and/or unauthorized recordings or visuals
of artists generated by AI can lead to consumer confusion,
unfair competition against the artist that actually was the
original creator, market dilution, and damage to the artist's
reputation, potentially irreparably harming their career. An
artist's voice is often the most valuable part of their
livelihood and public persona, and to steal it, no matter the
means, is wrong.
A Federal right of publicity would clarify and harmonize
the protections currently provided at the State level.
Visibility into AI training data is also needed. If the data on
AI training is not transparent, then the potential for a
healthy marketplace will be stymied as information on
infringing content would be largely inaccessible to the
individual creators.
And I might add, based on some of the comments I heard
earlier, it would be hard to opt out if you don't know what has
been opted in.
Finally, AI-generated content should be labeled as such. We
are committed to protecting our artists and the authenticity of
their creative works.
As you all know, consumers deserve to know exactly what
they are getting. I look forward to the discussion this
afternoon, and I thank you for the opportunity to present my
point of view. Thank you.
[The prepared statement of Mr. Harleston appears as a
submission for the record.]
Chair Coons. Thank you, Mr. Harleston. Thank you to all of
the witnesses today. We will begin our first 5-minute round.
And I am going to start by just exploring how we can respect
existing copyrighted works, copyright protections, while
continuing to safely develop and advance AI technologies. If we
run out of time, we will do a second round. My hunch is there
is at least that much interest.
Mr. Brooks, if I might just start with you. Generative AI
models like those your company creates are trained in no small
part on vast quantities of copyrighted content, on data from
copyrighted content.
Do copyright owners know if their works have been used to
train Stability's models? Is Stability paying rights holders
for that use? Why not, if not? And how would doing so impact
your business and your business model?
Mr. Brooks. Thank you, Senator. So, to the first question,
models like Stable Diffusion are trained on open datasets or
curated subsets of those datasets. So Stable Diffusion, for
example, takes a 5 billion image dataset.
We filter that for content, bias, quality, and then we use
a 2 billion image subset to train a model like Stable
Diffusion. Because it is open, you can go to a website, you can
type in the URL of an image, you can type in a name.
You can see if that work has appeared in the training
dataset. And then we are working with partners to take those
opt-out requests, and as I say, to incorporate them into our
own training and development processes.
So, we do think open datasets are important. They are one
part of how we are able to inspect AI for fairness and bias and
safety. And so that is I think to the first part.
Chair Coons. So, if I heard you right, if an artist takes
the initiative to search your training set, they might be able
to identify that a copyrighted work was used and then submit an
opt-out request. And you are in the process of facilitating
that use. But to my second question, do you pay any of the
rights holders?
Mr. Brooks. As I say, Senator, we--this is 2 billion
images, a large amount of content. A lot of it, you know, all
kinds of content. Tech language models, for example, it is not
just books, it is snippets of text from all over the internet.
As I say, to make that workable, we believe, you know, that
it is important to have that diversity, to have that scale.
That is how we make these models safe. It is how we make them
effective.
And so--and so we collected, as I say, from online data.
What I will say is that the datasets that we use, like that 5
billion image dataset I mentioned, they respect protocols like
robots.txt. So, robots.txt is a digital standard that basically
says, I want my website to be available for ancillary purposes,
such as search engine indexing.
And so, the dataset that was compiled respected that
robots.txt signal, and then on top of that, as I say, we have
the opt-out facility that we have implemented.
Chair Coons. Thank you. Mr. Rao, it is my understanding
that Adobe is taking a distinctly different approach. Your
generative AI model, Firefly, was only trained on licensed
data. Were there any downsides economically to that decision?
Is your model less robust or has it had any impact on its
performance? And how would you compare these two approaches in
terms of the incorporation of opt-out and licensed?
Mr. Rao. Thank you for the question. So, as I mentioned in
the opening remarks, that we--Firefly, our generative AI tool
was trained on our stock photography collection, which are all
licensed assets with the contributors, and that is actually the
only dataset used in the version that you can use on Firefly on
the web.
We think about the quality of this, and when we think about
the quality to your question, we have to put a lot of image
science behind that to make sure it was up to the level we
require because we didn't have the most expensive version of
that dataset. So, we had to put more computer science behind it
to make it have the higher quality we needed.
As we go forward, we are looking at whether or not there
are other areas where we need to supplement that dataset, and
for those we referred to as opening licensed content or places
where the copyright has expired.
Opening licensed to us means images that come from the
rights holders who have licensed it without restriction. So,
very similar to what we are talking about in the licensed
content, this is a place--this is also what we call
commercially safe.
Chair Coons. My sense, Mr. Brooks, is Stability is trying
to honor something like 160 million opt-out requests in
training your next model. Mr. Rao, Mr. Brooks, just this will
be my last question, and then I will turn to Senator Tillis.
Should Congress be working to ensure that creatives can opt out
of having their works used to train AI models? How would you
best do that, briefly?
Mr. Rao. So, we have this technology we refer to as content
credentials in my opening remarks, and what that does, it is a
metadata that goes along with any content. So, if you are in
Photoshop right now, you can say, I want content credentials to
be associated with this image.
As part of that, you can choose to say, I want it not to be
trained on it--a ``Do Not Train'' tag that gets associated with
the image and it goes wherever the content goes.
So, we do think the technology is there and available, and
we are talking to other companies, including Stability, about
this as an approach to honor that tag so people who are
crawling it can see the tag and choose not to train on them.
Chair Coons. Should we require that?
Mr. Rao. And I do think that there is an opportunity for
Congress to mandate the carrying of a tag like that, a
credential like that, wherever the content goes. Right now, it
is a voluntary decision to choose to do that.
Chair Coons. Should we require that?
Mr. Brooks. There is some very interesting precedent
internationally for this. The European Union has introduced
certain kinds of text and data mining exceptions. And part of
that is to say that you can use this for commercial,
noncommercial purposes.
There is an opt-out requirement, but the opt-out has to be
machine readable, as I say, as a matter of practicality, when
you are dealing with trillions of words of content, for
example, or billions of images, in this case.
The machine readability is important, and that is where
these tags become an important part of how to implement it in
practice.
Chair Coons. We will keep exploring this further. Senator
Tillis.
Senator Tillis. Thank you, Chairman. I'll have Senator
Blackburn go and then I will follow Senator Hirono.
Senator Blackburn. Excellent. Thank you, Senator Tillis.
And Mr. Chairman, thank you for the hearing today. It is so
appropriate that we have this. I am from Tennessee. We have
thousands of artists and songwriters and musicians, and we have
actors and actresses, and we have authors and publishers.
And everywhere I go, people are talking about the impact of
AI, to the positive or the negative. You know, you look at
health care, you look at logistics, you look at autos, you look
at entertainment, and there are pros and cons.
But the one point of agreement is, we've got to do
something about this so that it is going to be fair, and it is
going to be level. Mr. Harleston, I want to come to you right
off the bat because you mentioned the NIL issue, which I think
is an imperative for artists to be able to own that.
And you also mentioned the right of publicity laws, and of
course, those are State level laws. And as you rightly said, we
don't have a federally preemptive right to publicity law. And I
think the dust up--a lot of people came to realize this over
Drake and The Weeknd, and ``heart on my sleeve.''
And this is something that does have to be addressed. So,
for the record, give us about 30 seconds and then you guys, I
see your capable team behind you, you can submit something
longer in writing, if you would like, on the reason State level
publicity laws are not enough.
Mr. Harleston. In 30 seconds----
[Laughter.]
Mr. Harleston. State level publicity laws are inconsistent
from State to State. A Federal right of publicity that really
elevates right of publicity to an intellectual property is
critically important to protect----
Senator Blackburn. Okay, I am going to help you out with
this----
Mr. Harleston. Okay.
Senator Blackburn. A federally preemptive right to
publicity law would provide more of that constitutional
guarantee to her works that Ms. Ortiz has mentioned.
Mr. Harleston. Absolutely.
Senator Blackburn. All right. And----
Mr. Harleston. And we will follow up with you, Senator.
Senator Blackburn. Yes, excellent. I think something in
writing would be very helpful there. Now, I think it was very
appropriate that you had Spotify and Apple Music take down
``heart on my sleeve.'' Important to do. And talk about the
role that the streaming platforms should play. Should they be
the arbiter when it comes to dealing with this generative AI
content?
Mr. Harleston. These streaming platforms, we acknowledge
that they are in a challenging position, but certainly in some
instances when there is clear, or it is clear that the content
that has been submitted to them for distribution----
Senator Blackburn. So, a knowing and willingness standard
would be nice.
Mr. Harleston. That would be very nice, yes.
Senator Blackburn. Okay. I am helping you out there.
Mr. Harleston. You are doing great.
Senator Blackburn. Thanks for being here.
Okay, Professor Sag, want to come to you. This spring, the
Supreme Court issued a--what I thought was a very appropriate
decision in Warhol v. Goldsmith, and I was very pleased to see
them come down on the side of the artist. I filed an amicus
brief in this case arguing for strong, fair use protections for
creators. Now, we have been through this thing in the music
industry where ``fair use'' became a fairly useful way to steal
my property.
And artists don't want to go through that again. Right, Ms.
Ortiz?
[Voice off microphone.]
Senator Blackburn. It didn't work the way it was supposed
to. And I would like for you to talk for a moment, should AI,
unlicensed AI ingestion of copyrighted works might be
considered fair use when the output of AI replaces or competes
with the human-generated work.
Now, Ms. Ortiz has laid this out fairly well in her
comments and the Supreme Court has sided with the artist in
Warhol v. Goldsmith. But this fair use standard comes into play
every time we talk about our fabulous creative community and
keeping them compensated. So, the floor is yours.
Professor Sag. Senator Blackburn, commercial replacement
should not be the test. The test should be exactly what the
Supreme Court said in the Andy Warhol case.
The question is, is this significantly transformative? What
that means in relation to training AI models is, does the
output of the model bear too much resemblance to the inputs?
And that is a different question to, is it competing with the
inputs? Could it be used as a commercial substitute?
If you look at some of the old cases on reverse engineering
software, companies were allowed to crack open software, find
the secret keys to interoperability, and build new competing
products that did not contain any copyrightable expression, and
the Court said that that was fair use.
So, I think on current law, the answer is no. Potential
substitution in terms of a competing product is not the test.
The test is, are you taking an inappropriate amount of an
artist's original expression.
Senator Blackburn. Well, my time has expired. Thank you for
that. We just don't want it to become a fairly useful way to
steal an artist's product. Thank you, Mr. Chairman.
Chair Coons. Thank you, Senator Blackburn. And thank you
for the passionate engagement you have always brought to these
issues on behalf of the creative community.
[Laughter.]
Chair Coons. Senator Hirono.
Senator Hirono. Thank you, Mr. Chairman. Mr. Harleston,
whenever the idea of negotiating licenses is raised, people
express concerns about how complex it would be and how AI
platform developers could never possibly negotiate with all
rights holders. But in the music context, at least, you have a
lot of experience negotiating rights.
Could you tell us a little bit about your industry's
history of negotiating rights with digital music services and
lessons that history could teach us, for whether rights
negotiations would be possible with AI platforms?
Mr. Harleston. Thank you, Senator. As you referenced, we
have had a long history with the transition of our business
from a physical business to a digital business, and having to
encounter digital platforms that were very quickly adapted by
consumers and had lots of our content on there.
What we found was ingenuity does play a role. It is not
easy. But we were able to identify or find ways to identify our
copyrights, to work out licensing schemes that allowed the
platforms to be able to carry and distribute the music.
And in a commercial environment that was positive for them,
while at the same time allowing the artists to be properly
compensated. And this is, you know, with the--in the music
side, we have two sets of rights, which makes it even more
complicated, but we have done great work over the years to
develop systems that allow identifying not only the sound
recording, but also the underlying composition.
So, it could be done. But what it needs, it needs--what we
would need is we need help to make sure that everyone
understands that there are rights that are affected and that
the activity that is happening now is violative. And once they
understand that what they are doing is violative, that brings
them to the table so we can negotiate a deal.
Senator Hirono. I note that in your testimony you said that
consent is the key. So is your position that every artist's
work before it can be used to train AI models, that the company
that is wanting to use that information has got to get the
consent of the originator?
Mr. Harleston. In a very short answer, yes.
Senator Hirono. And you think that we are able to do this
knowing that these platforms incorporate billions and billions
of information to train their AI models?
Mr. Harleston. Understanding that, but it absolutely could
be done as these--as the digital platforms that exist today,
the licensed platforms ingest millions and millions of songs
every week. So, it is not a problem in that respect. There is
metadata that we could license. We could absolutely do that.
But there has to be an initiative on the side of the companies
to reach out.
Senator Hirono. So, Ms. Ortiz, if I--Mr. Brooks, rather,
sorry. My--what I heard you say in response to the Chairman's
question is that for all of the data that you input into your
model, you do not get the consent of the artist or originator.
Is that correct, Mr. Brooks?
Mr. Brooks. So, we, Senator, we believe that yes, if that
image data is on the internet and robots.txt says it can be
subject to aggregated data collection, and if it is not subject
to an opt-out request in our upcoming models, then certainly we
will use those images, potentially use those images if it
passes our filters.
Senator Hirono. So basically, you don't pay for the data
that you put into your--to train your model.
Mr. Brooks. For the base, the kind of initial training or
teaching of these models with those billions of images, there
is no arrangement in place.
Senator Hirono. So, you have Ms. Ortiz who says that that
is wrong. Is that correct, Ms. Ortiz?
Ms. Ortiz. One hundred percent, Senator.
Senator Hirono. So do you know if--well, I think you
mentioned that your work has been used to train AI models and
you have gotten not one cent for that use.
Ms. Ortiz. I have never been asked. I have never been
credited. I have never been compensated one penny. And that is
for the use of almost the entirety of my work, both personal
and commercial, Senator.
Senator Hirono. So, if you were to allow your works to be
used to train, you would--do you think that you would
negotiate--if there was a law that required compensation, then
that compensation negotiation should be left to you and the
entity such as Mr. Brooks'.
Ms. Ortiz. Personally, I love what I do, so I wouldn't
outsource it to an AI, but that is not a choice for me to make,
and it is all about that. It is about being able to have that
choice and artists don't have that right now.
Senator Hirono. Thank you.
Senator Tillis. Thank you, Mr. Chair. I was actually
inspired by one of the opening statements, so I went out and
generated a cat driving a 1960 Corvette with a surfboard in it.
And I produced that picture.
[Phone is held up with the screen facing the witnesses.]
Senator Tillis. Actually, it gave me four options. This one
I found the most interesting. But it raised a question that I
wanted to ask you, Mr. Brooks. If an artist looked at that and
said, that is in part developed by that 60s Corvette in South
Beach, how does that artist then go about saying--I am trying
to get an understanding of your current opt-out policy.
And one of the issues that we have had here and not
completely related, but we have a notice, a takedown notice and
stay down discussion in the past around creative works. So, I
was just trying to understand, and I think it is going to be a
lengthy answer. And then if I talk to a creative, it is going
to be a lengthy answer.
But for the record, it would be very helpful to me for your
specific platform to understand how that opt-out process works.
I think I heard right that you could embed within the works
certain things that already create an opt-out, or that that
work shouldn't be used. But I want to drill down. We don't have
time to do that now.
And in a twist of irony, I was wondering if any of the
witnesses would suggest any creative works by other
governmental bodies that we should steal and use as a baseline.
In other words, what good policy seems to be being discussed or
passed? What particularly problematic at either end of the
spectrum? Because I am sympathetic to the issues at both ends
of the spectrum on this argument.
So maybe we start with you, Professor. Are you aware of any
Western democracy states, I am not particularly interested in
what China is doing because whatever they agree to, they are
going to rip off anyway, but any best practices that we should
look out there, or bad practices, or trends that we should
avoid or be concerned with as we move forward?
Professor Sag. I think that the European Union's approach,
where they have different rules for commercial and
noncommercial use, and opt-outs have to be respected for
commercial uses of the text mining in Article 4 of the DSM has
something to recommend it.
By the same token, I would note that opt-outs do not apply
to researchers working at proper research institutions in the
EU, nor do contractual overrides, which is a position that I
can't see Congress adopting, but it is certainly something to
look at. That's--that's really it.
Senator Tillis. Anyone else briefly could add what--Ms.
Ortiz, I should also add, I have seen all your works and it has
been since 11 o'clock last night that I was talking about
``Guardians of the Galaxy'' with my colleagues as we were
coming back from Vilnius.
Ms. Ortiz. It was a really fun project to work on, Senator,
so thank you. So, what the artists community have suggested is
that models be built starting from scratch via public domain-
only works, that's work that belongs to everyone. Any expansion
upon that to be done via licensing. And there is a couple of
reasons for this.
Current opt-out measures are inefficient. For starters,
machine learning models, once they are trained on data, they
cannot forget. And machine unlearning procedures are just dead
on the water right now, and this is not according to me. I am
an artist. I have no idea on this. This is according to machine
learning experts in the field.
Second, things, safety filters like, for example, prompt,
you know, filters are so easily bypassed by users. So
unfortunately, when companies say, hey, opt out, there is no
real way to do that. But even further, what happens if someone
doesn't know how to write a robot.txt?
Like, how does a person who may not know the language, may
not know the internet, may not even know that their work is in
there, recognize that, you know, they need to opt out. This is
why my community in particular has suggested over and over, opt
in should be the key in order to base the foundations of
consent, credit, and compensation.
Senator Tillis. And Mr. Brooks, I can understand the
challenges with opt-in versus opt-out in terms of the task that
you would have ahead of you. But what is your view of the
concerns that creatives have expressed in this light and the
current opt-out process that you all have in place or
procedures which I would like to get information for, for the
record.
Mr. Brooks. Thank you, Ranking Member. Look, I will say at
the start that we do need to think through what the future of
the digital economy looks like. What do incentives look like?
How do we make these technologies a win-win for everyone
involved? These are very early days from our perspective. We
don't have all the answers, but we are working to think through
what that looks like----
Senator Tillis. I am going to stick around for a second
round, so we will get a little bit deeper into that. But I want
to defer to my colleague from California.
Chair Coons. Thank you, Senator Tillis. Senator Padilla.
Senator Padilla. Thank you, Mr. Chair. And I want to thank
the witnesses for your testimony and participation today.
Speaking of California, I can't help but observe that
California is very well represented on this panel. Not only a
point of pride for me as a Senator from California, but it is
frankly not a surprise since we are the creative and tech hub
of the Nation.
Now, generative AI tools, as we have been talking about,
present remarkable opportunities and challenges for the
creative community and our broader society. And I couldn't help
but observe that in reviewing the testimony from each of you, I
noted the common goal of seeking to leverage and develop AI
tools to complement and encourage human creativity and
artistry, while also respecting the rights and dignity of the
original creators.
So, it is a tall order, a delicate balancing act in many
ways, but that is--this seems to be the shared objective here.
So, I want to thank you again for participating in this hearing
as we are working to determine what role we play in fostering
the development of AI in a manner that is a net positive for
innovation and creativity.
My first question, and I will keep it brief because it is
sort of piggybacking on--Senator Hirono has raised it, Senator
Tillis was just trying to expand upon it, and is directed at
Mr. Brooks. This whole opt in, opt out: We can talk about what
the process is, whether it is easy, clear, or not for artists.
And, you know, I don't completely agree with you that we
are in an early stage because it is happening fast. Tell me how
it is possible--explain how it works to have a system unlearn
inputs that have already been taken, if you get this after-the-
fact opt-out from an artist. It is happening now. While you are
trying to think what it means long term, it is happening now.
So how does it work, not just process, checking a box,
filling out a form, but technically?
Mr. Brooks. Thank you, Senator. So, just in terms of the
data collection piece, I just want to make it clear that today
it is very much a kind of work in process framework. You know,
you can go to this website. You can indicate you want to opt
out. We will take those opt-out requests as they come in.
But as we were talking about before, it is important that
eventually there is a standardized kind of metadata that just
attaches to these works as they go out into the wild. And as I
said, that is what the EU is requiring, and I think there will
be a lot of standards development in that space, again, with--
in terms like Adobe and others.
In terms of what then happens, you know, as I say, we
filter that training data for a few reasons. We take out unsafe
content, we adjust for issues like bias to correct the bias.
And then in addition to that, we start to incorporate, as I
say, the opt-out requests.
Sometimes some of the models we release are retrained from
scratch with new datasets. Again, they take into account the
lessons learned through previous development, both as an
organization, as a company, and potentially technical things
that we have learned as well in that process.
Some of the models that are released are just fine-tuned
variations of the model, and so those ones may have the same
kind of basic knowledge from that original training process,
and there has just been some additional training to correct for
certain behaviors or improve performance in specific tasks.
So, in terms of, you know, the future of this space, you
know, there is a lot of work being done on unlearning in
general. You know, how do you interpret the relationship
between training and the data in training and the performance
of the model?
How do you potentially adjust to that, different ways? But
as I say, at this stage, we treat it as a process of
incorporating those opt-out requests, retraining, and then
releasing a new model trained on that new dataset.
Senator Padilla. I hear you, and I just want a level set a
little bit, not just out of the concern for the artists, but
knowing that unless you are getting one, two, three inputs
today, which may be small enough to keep your arms around, I
doubt that is the case as we are getting into the hundreds and
thousands of inputs per day to go in and relearn, unlearn, and
comply with any consent or opt-out. It gets overwhelming and
unfeasible real quick, and it is happening now.
I also wanted to follow up on a subject matter that Senator
Coons touched on earlier. We know that generative AI models
need to be fed large datasets to learn how to generate images
based on user prompts, just like Senator Tillis did. By the
way, that looked much more like Pacific Coast Highway than
South Beach.
Now, AI for--this is now talking to folks back home, can
only understand what it is taught, making it critical that for
AI companies to train their models with data that captures the
full range of the human experience, want to be inclusive and
diverse, if we are going to be accurate in representing our
users, representing the diverse backgrounds of all users.
Now, Mr. Rao, you have explained how Adobe's Firefly seeks
to avoid copyright infringement by being trained on only
licensed Adobe Stock images, openly licensed content, and
public domain content.
So how do you reconcile both? You want to be as inclusive
as possible, which means as much data input as possible, but to
avoid the copyright infringement, you are being selective in
those inputs. That diversity of input is important, I think,
for the diversity of output. So how do you reconcile?
Mr. Rao. It is definitely a tension in the system. Right?
The more data you have, the less bias you will see. So, it is
great to have more data.
But when you set the expectations that we had for ourselves
of trying to design a model that was going to be commercially
safe, we took on the challenge of saying, can we also do that
and minimize harmful bias? And the way we did that, we have an
AI ethics team. We started that 4 years ago.
And one of the key things they did when we were developing
Adobe Firefly was not only do we have the dataset and we
understand what that is, we also did a lot of testing on it. We
have a series of prompts, hundreds and hundreds and hundreds of
prompts.
We were testing against it to see what the distribution of
model is. Is there going to be a bias. If you type in
``lawyer,'' are you only going to get men--or white men, and
what does that mean, and how, then, do you change that?
And you either change it by adding more data, making it
more diverse--and so that means you have to get ethically-
sourced, more data to diversify the dataset, or you can add
filters on top of the dataset to force a distribution of what
you expect to see if you are typing in certain search terms and
make sure the bias is removed.
So, you can either do it by adding more data or you can do
it with through adding filters on top of the model itself to
ensure that you are going to get the right result.
Senator Padilla. And if you ask--if you input ``Senator,''
what comes out?
Mr. Rao. An amazingly handsome man and woman, just very
intellectual.
[Laughter.]
Senator Padilla. Men and women--colors across the spectrum.
Mr. Rao. Across the spectrum.
Senator Padilla. Thank you, Mr. Chair.
Mr. Rao. The first time we did ``lawyer,'' though, we only
had white men. And as general counsel, I was like, there should
be some people who look like me as well.
Chair Coons. Thank you, Senator Padilla. Senator Klobuchar.
Senator Klobuchar. Okay, very good. Thank you. I was glad
to be here for all your testimony and thank you for that. I
guess, I will start with you, Mr. Harleston. Approximately, and
I know you talked about this a bit with some of the other
Senators, Senator Blackburn. Approximately half the States have
laws that give individuals control over the use of their name,
image, and voice.
But in the other half of the country, someone who was
harmed by a fake recording purporting to be them has little
recourse. In your testimony, you talk about new laws and how
they could protect musicians' names, likenesses, and voices--
the right of publicity, I think you called it.
Can you talk about why creating this is important in the
face of emerging AI? And how have statutes in States that have
these protections helped artists?
Mr. Harleston. Thank you, Senator, for the question. It is
critical in this environment when we are talking about the
creative expression that the artist has made, that the right of
publicity also be extended at the Federal level.
There is inconsistency, but more importantly, the
preemptive element of it is critical. Raising it to the level
of an intellectual property is also critical. What we have
seen, and this is really in the area of deepfakes, where you
have seen, I think, Ms. Ortiz referenced how many times her
name was listed.
We are finding with our artists, particularly the ones that
are most established, that their names are, you know, daily--
hundreds and hundreds of thousands of posts with their names.
And also, there is sometimes images that are used as well.
Senator Klobuchar. Mm-hmm.
Mr. Harleston. So, it is critical to have this right to
protect the artists and their use. And if I could just say one
thing on the--I know this is not your question, but I have to
say----
Senator Klobuchar. There we go.
Mr. Harleston. Because it is killing me----
Senator Klobuchar. I will just add it to my time.
[Laughter.]
Mr. Harleston. All right, thank you, thank you. On the opt
in, opt out, there is an element beyond commerciality. And I
want to make sure everyone understands. Ms. Ortiz did reference
it, about she didn't really--she probably wouldn't want a
license to AI. And there are--we have artists that don't want a
license to streaming services. So, they are not--it is not
always about the commerciality.
Some artists just don't want their art distributed in
certain ways. And the Beatles didn't come onto streaming
platforms till about 7 or 8 years ago. That was a decision that
was very important to them. So, I want to add that into the
conversation. I know that wasn't your question, sorry.
Senator Klobuchar. Okay, very good. And so, what do you see
as the obligations of social media platforms on this?
Mr. Harleston. With respect to AI?
Senator Klobuchar. Uh-huh.
Mr. Harleston. Oh, great--fantastic question. We believe
that the social media platforms absolutely have an obligation.
I will say this, that we could help them by giving them a hook
beyond copyright in terms of being able to take down some of
the----
Senator Klobuchar. Exactly----
Mr. Harleston. Some of the deepfakes. They have challenges
with some of the platforms on this.
Senator Klobuchar. Yes. Right, exactly. And I think we are
seeing the same thing. I guess I would turn to you, Mr. Brooks.
You talked about advocating for creating ways to help people
identify AI-created content.
And when we talk about deepfakes, we are already seeing
this with political ads, and not even paid ads, just videos
that are put out there. There's one of my colleague Senator
Warren that was just a total lie that, saying that--acting like
it was her that she was saying people from one party shouldn't
be able to vote.
And we have seen it in the Republican Presidential primary.
A number of us on a bipartisan basis are working on this. I
chair the Rules Committee, so it is kind of my other hat.
Do you agree that without tools for people to determine
whether an image or video generated by AI, that that would pose
a risk to our free and fair elections, if you can't tell if the
candidate you are seeing is the candidate or not?
Mr. Brooks. Thank you, Senator. We absolutely believe that
these transparency initiatives like CAI with Adobe are a really
important part of how we make the information ecosystem more
robust.
This isn't just an AI problem or a social media problem. It
is going to require everyone, and it is going to require
accountability right across that ecosystem. But what we think
is, you know, we have in place things like metadata, things
like watermarking for content.
They are just some more of the signals that social media
platforms can use to decide whether they are going to amplify
certain content.
Senator Klobuchar. Yes, and we have got this REAL Political
Advertisement Act, with Senator Booker and Senator Bennet.
There is a version initially that was also introduced in the
House. And so, that is one solution.
But we are also going to have to look at, I would say,
banning some of this content, because even a label or a
watermark--it is not going to help the artist or the candidate
if everyone thinks it is them and it is not, and then at the
end, it says generated by AI.
Mr. Brooks. It is a great question and a really important
one, I think, Senator, because there are a few things in there.
I think there's the question of the use of likeness,
particularly for improper purposes, where you are implying that
there is some kind of endorsement or affiliation between a
particular person and a particular work or idea.
That is different, I think, to the use of the kind of free
experimentation with style and some of these other issues that
tend to get lumped together in AI outputs.
Senator Klobuchar. Mm-hmm.
Mr. Brooks. And so, in terms of these scenarios that you
are talking about, there is this kind of improper use.
You are implying that someone endorses or embraces a cause
or a work that they are not affiliated with. And there needs to
be clear rules around how like this is used in that context,
whether through right of publicity or through some of the
bespoke deepfake legislation.
Senator Klobuchar. Okay. Last, Mr. Rao, our recent study--
and I know you have worked on this democracy issue, which I
truly appreciate. A recent study by Northwestern predicted that
one third of the U.S. newspapers that existed roughly 2 decades
ago will be gone by 2025.
The bill that Senator Kennedy and I have, the Journalism
Competition and Preservation Act, would allow local news
organizations to negotiate with online platforms, including
generative AI platforms.
This bill passed through this Committee now twice. Could
you describe how Adobe approaches this issue? And in your
experience, is it possible to train sophisticated, generative
AI models without using copyrighted materials, absent consent?
Mr. Rao. Thanks for the question. Absolutely. We--our
current model that is out there is trained using the licensed
content that I had mentioned before and other content that has
no restrictions on it, and it comes from the rights holders
directly.
So, we definitely think it is possible. We have done it. It
is out there on our website, and it is also in Photoshop, and
people love it. The creative professionals are using that AI.
It makes their day easy.
It lets them start their creative work in just one click
and then they finish it in the tool. So, it has really
revolutionized how we think about things. In terms of how we
acquire datasets, and we have a group inside Adobe whose--that
is their job. Their job is to think about where do we need to
go next? Do we need to get to different media types? Do we--are
we missing some sort of subject matter for our AI to be more
accurate?
That was a question we had before. We think about that
content. Maybe there is a newspaper that you mentioned that has
the kind of content we need. We go approach that organization
and say, look, we need to license that content in to make sure
our AI is more accurate.
So, we have a team that thinks about this--sources it, and
brings it in.
Senator Klobuchar. And in the absence of that? I mean, what
impacts do you anticipate this could have on local journalism
if there are no rules of the road put in place?
Mr. Rao. Yes, I think that, you know, both on the
authenticity side and on this side, if people are able to, you
know, create images and these newspapers are not able to get,
you know, the ability to license the work they are doing, it
could certainly have a negative impact to them.
On the authenticity side, the reason why so many media
companies have joined the Content Authenticity Initiative, like
AP, Reuters, Wall Street Journal, New York Times, Washington
Post, is because they know that when they are showing images,
they need to be able to show that they are actually true.
Senator Klobuchar. Mm-hmm.
Mr. Rao. They need to be able to prove that it happened. If
people stop believing that any of these digital images they are
seeing are real, then they are going to stop consuming
newspapers. They are going to stop consuming that content
because they are not going to believe it. So, you have to give
those local newspapers a way to prove what they are showing is
true----
Senator Klobuchar. Right. Absolutely.
Mr. Rao [continuing]. So people can still consume it.
Senator Klobuchar. Of course, there is a lot less famous
newspapers, including some very small ones in my State that
just you might not mention. Right?
And so, I think that part of it, is that, you know, the Ms.
Ortizes of this story need to be able to have some kind of
power to be able to protect their content, too, because they
don't have a general counsel, and they are not going to be able
to, on their own, start some major lawsuit. And so, I think
that is how we have to think about that, too, as we look at all
of this.
Mr. Rao. And that is why I would say again that when we
designed the Firefly, we designed it that way. Right?
Senator Klobuchar. Mm-hmm.
Mr. Rao [continuing]. To be commercially safe first, right,
making sure that we built the model the way----
Senator Klobuchar. Yes, no, I am not--I am saying it sort
of rhetorically to the world and to everyone that needs to get
this done, as opposed to you, Mr. Rao.
Mr. Rao. Thank you so much.
Senator Klobuchar. All right. Appreciate it. And I thank
you, both of you, for your continual bipartisan work in taking
on this very important issue. Thanks.
Chair Coons. Thank you, Senator Klobuchar. We are going to
do a last round of questioning. We may be joined by other
colleagues, but we are also in the middle of a vote. So, my
hunch is we will resolve this in 10 to 15 minutes at the most,
if I might.
I am interested in pursuing the question of a Federal
statutory right of publicity. And to me, the core issue really
is, what is the remedy?
Often, preemption is motivated by a desire for there to be
consistency, the elevation in terms of process and access to
justice, and potential remedies that comes with a Federal right
as opposed to a State right.
But, Professor, if I could start with you. You testified
earlier in response to a question from Senator Blackburn that
commercial replacement is not the appropriate test under
current fair use law in the United States.
Should we adopt a Federal right of publicity with
commercial replacement as the test or part of the test, and how
would that play out? What other remedy might you suggest under
a new Federal right of publicity?
Professor Sag. Senator, thank you for that question,
because I was quite alarmed by some of the discourse here about
the right of publicity. I think, as well as----
Chair Coons. Regulated by discourse.
Professor Sag [continuing]. As well as thinking about
publicity rights for well-known artists, musicians, etcetera,
Congress should be thinking about the right of publicity of
ordinary people, people who are anonymous, people who have no
commercially valuable reputation.
All of us deserve to be protected from deepfakes and
synthetic reproductions of our name, image, and likeness,
regardless of whether we are a famous politician or a famous
artist or just an anonymous law professor. So, I think----
Chair Coons. How would you focus the remedy in order to
make that effective?
Professor Sag. Senator, in terms of remedy, I think that
right of publicity statutes have traditionally had injunctive
relief, usually incorporating equitable balancing tests. That
is the remedy I would go for, which would mean, the models
might have to be retrained.
Chair Coons. Injunctive relief only, not commercial?
Professor Sag. Damages, potential as well. But statutory
damages, I don't think so. Statutory damages can be quite
distorting. They tend to be a honeypot for opportunistic
lawyers, as well as genuinely aggrieved plaintiffs. So, I would
steer clear of statutory damages, but actual damage and
injunctions, absolutely.
Chair Coons. Mr. Rao, I'd be interested in your views on
what a right of publicity might potentially do. I'd also be
interested in hearing your thoughts on how we should be trying
to balance respecting copyright through this or other means,
while incentivizing investment in AI and accelerating
innovation in the United States?
Mr. Rao. Thank you for the question. So, we talked about in
our testimony similar to, but not exactly like, a right of
publicity.
We referred to as a Federal anti-impersonation right. And
the reason we thought about it from an anti-impersonation
perspective is actually some of the same questions Professor
Sag raised, which is we want to make sure Professor Sag himself
is not--does not have a deepfake made of him.
So, if you think about it as an impersonation right, that
would apply to everybody. And what we are really targeting
there is we see the economic displacement that we have been
talking about here, where an AI is trained on an artist and
creates an output that is exactly like the artist, and they are
getting displaced by that work.
And copyright may not reach them, like that has been the
question. So that is why we believe they do need this right so
they can go after these people who are impersonating work,
whether that is likeness, whether that is style.
And then the test would be something that we would work out
through 6 months of deliberation here in this body, exactly how
you would decide that. But I think that is the right approach,
because you want to focus on people who are intentionally
impersonating someone in order to make or get some commercial
benefit, and I think that will help clarify what harm we are
trying to address.
Chair Coons. Mr. Brooks, how do you think a Federal anti-
impersonation right----
Mr. Rao. By the way, that spells FAIR. Just want to make
that clear. I know how Congress loves acronyms. Yes?
Chair Coons. We are enthusiastic about acronyms. We
actually are producing a Senate-only version of ChatGPT that
only produces acronyms for bill names.
[Laughter.]
Chair Coons. Mr. Brooks, how do you think a Federal
publicity right or an anti-impersonation right, a Federal
requirement that there be opt-in only rather than opt-out would
impact the business model that you are currently representing?
Mr. Brooks. So, Chair Coons, I think the actual instrument
and the content of that instrument, I think is really
diagnostic at this stage. As I said to Senator Klobuchar, it is
important from our perspective that there are clear rules
governing the use of likeness in an improper way.
I think the important thing to stress there is that it is a
use. And to some extent we can't escape the fact that the
determination of whether it is proper or improper will depend
on the application, what the user does or does not do with that
content downstream.
And so, as I say, you know, from our perspective, the lines
in the sand between improper use of likeness, free
experimentation with style, or other kinds of good or bad use
of these tools aren't easy to draw.
They are very fact sensitive. It may be appropriate for
courts to determine that. But at a high level, as I say, I
think there is a core of things around that improper use of
likeness, especially voice likeness, that there may be some
legislative intervention there that makes sense and may have
obligations, as I say, across the supply chain, across the
ecosystem.
Chair Coons. Mr. Harleston, if I might, the Copyright
Office recently issued guidance about human authorship being
critical to any copyright protection.
Is their guidance accessible enough, relevant? Did they
strike the right balance? Should we be looking at a different
policy in terms of how broadly copyright protection should
reach when there is AI assisted creativity as opposed to AI
generated?
Mr. Harleston. I think the Copyright Office did a pretty
good job. One can debate whether an AI component in a broader
work should also be afforded some form of copyright. You know,
I think they landed in the right place, that it shouldn't. That
copyright should only be afforded to human creation.
So, for example, if you had an AI-generated song--well, if
you had a song that was created by an artist and they used--a
piece of it was generative AI, there should be a copyright in
that entire work, but the AI-generated portion would not be
protectable.
So, if someone were to actually sample it, which would lift
it out and use it in another context, it would not be subject
to copyright. I think they did a pretty good job trying to
strike that balance.
Chair Coons. In the conversation I had previously with
Professor Sag, how do you feel about the scope of potential
remedies if we were to craft an anti-impersonation statute?
Mr. Harleston. I am glad you asked me that question.
Chair Coons. I thought you'd be.
[Laughter.]
Mr. Harleston. Thank you. I think there should be a private
right of action. I think that it is--I think commerciality is,
again, not always the proper standard here.
I think that in some instances we have had artists who have
had been victim of deepfakes where the voice was appropriated,
and the lyric content was something the artist would never have
said.
And that is something that can have irreparable harm to
their career, you know, in trying to explain that it wasn't
them, because there is stuff that is really good, these--sort
of these AI-generated things are really good.
Chair Coons. Ms. Ortiz, last but not least, has, in
producing some of the interesting, engaging, powerful,
inspiring content you have generated, have you ever relied on
an AI tool to help you expand or produce some of the works you
have worked on? And what is your hope about what we might do
going forward here in Congress in response to what we have
heard from you about your concerns?
Ms. Ortiz. I am very happy you asked this question,
Senator. So, I have never really--I was curious very early on
before I knew the extent of the exploitation of artists. Very
briefly used an AI to generate references, and I didn't enjoy
it at all. I am--you know, I love every step of the process of
being an artist.
And ever since I found out, you know, last August,
September, of what actually went behind the scenes, I just--I
cannot use it. My peers refuse to use it. My industry is very
clear that we do not want to exploit each other. And again, it
is important to remember that these, you know, models basically
compete in our own market.
And this isn't something that is hypothetical. It has
happened now with our own works. And one of the things that I
would hope, you know, would be kind of addressed here is that a
lot of the solutions that have been proposed--or, you know,
basically you cannot enact them unless you know what is in the
dataset.
And for this, we need to ensure that there is clear
transparency built from ground up. Like, no offense to some of
the companies here, but if you don't know what exactly is in
the dataset, how do we know?
How does the licensor know that my work is in the dataset?
And that I feel like it is one of the starting foundations for
artists and other individuals as well to be able to gain
consent, credit, and compensation.
Chair Coons. Thank you. Mr. Tillis, before I just hand it
over to you, and then we are going to conclude, I just
appreciate all of you taking the time and effort helping
educate us. You are literally training us as we try to produce
some fidelity in our legislative work. Senator Tillis.
Senator Tillis. Yes, to me, trying to figure out what may
or may not be in the language model is a lot like taking roll
in a dark classroom. I just don't understand how you would do
it. So, you know, I can see that we have to work at it.
But I want to start, Mr. Brooks, by thanking you for being
here. I think that anyone that is watching this needs to
understand that this isn't unique to Stability AI. This is a
broader set of issues that we have to deal with, and I
appreciate the fact that you'd be willing to come here because
you should expect that some of the concerns are--that were
going to be expressed to begin with.
I have one question. The bad news for you all is that my
staff are really excited about this. These are the questions
[papers are shown to the witnesses]----
[Laughter.]
Senator Tillis [continuing]. That we are going to submit
for the record. But rather than expect you all to respond to
every one of them, you are welcome to do that, your area of
expertise, your priorities, just use that to guide you, and get
that information back for the record.
But one of the ones I won't have to ask because I will ask
it now is, a recent survey on how consumers view AI found that
most consumers, nearly 80 percent, believe the use of AI should
be explicitly disclosed.
Now, in Vilnius, I happened to stay at a hotel that is
called the Shakespeare Hotel, and every room was named after
the greats. I don't see a day 100 hundred years from now where
those rooms are going to be named after great LLMs. And the
reason for that is I think there is a natural cultural bias for
rewarding the human beings who are truly the creators and the
lifeblood of our creative community.
So, does anyone here disagree that a work that is derived
even from, let's say, licensed content, that the consumer
should know that this was created by a machine versus an
original creative work by human beings?
Anybody disagree with that, or maybe technical issues I
should look at? No different than Mr. Rao, me, I use Photoshop.
I could create Corvette cat with a skateboard or surfboard
really quickly. No different than I want that, which, as you
know, again, based on prior creative work, somehow have
disclosures. Does that make sense to you?
Mr. Rao. Yes, I think the question, and we thought about,
we think it is definitely of interest to our creative
customers, is to be able to show something human created versus
AI created. In Adobe Firefly, it all comes out saying something
is AI created.
That is on by default. So, you will always know that it is
AI created. The trick going forward, though, is we anticipate
our AI features are our most popular features in Photoshop, so
we expect going forward, most images are going to have a part
that is AI and a part that is human, and you sort of have to
start thinking about what are you disclosing when you disclose
that. Right?
The content credential we mentioned before that you could
use a ``Do Not Train'' tag on it, or you could use it for, you
know, addressing deepfakes. Also, we will record the human part
versus the AI part. So, you could think about using that as a
disclosure.
But I am not sure over time people are going to be as
interested in knowing the identity of the artist who created
the work as opposed to which part of it they did with AI, and
which----
Senator Tillis. That is fair. Professor Sag, do you have a
comment?
Professor Sag. Just to follow up on that. You also have to
think that you are not just talking visual works here. Like
take the same thing with written works. Someone uses GPT to
help smooth over their writing, refine something, explain it
more clearly.
There are some awkward line drawing questions, but the
spirit of the disclosure requirement is correct. The
implementation, I think, just will be difficult.
Senator Tillis. I agree. And Mr. Chair, I am checking the
votes. I think it is probably time for us to wrap up the
Committee. I think you could see from this just by the sheer
number of Members who came to the Subcommittee, this is an area
of interest and a priority for us.
Mr. Chair, I have decided that maybe for the next hearing,
it is going to take a little bit more tuning for me to get the
answer, but I am going to do a--with the--you know the song,
``Who Let the Dogs Out''?
[Laughter.]
Senator Tillis. I was thinking we would set that to ``Don't
steal my IP,'' and I will see if I can get that done. If you
think about it, it'd be pretty snappy.
[Laughter.]
Senator Tillis. But I will work on that for those of you--
you may have to get a bigger room if people know about that in
advance.
Chair Coons. We may end up doing this as a duet.
[Laughter.]
Senator Tillis. But again, I think this Committee has
demonstrated that we are very thoughtful, and we are very
diligent, and I, for one, could sit at that table and probably
present the interests of either side of the spectrum, which is
why I believe that we need legislative certainty.
We need to learn like data privacy, data ownership. In
Europe, they don't always get it right in the first tranche, so
we wouldn't necessarily lift something up and implement it
here, but we want to think it through and make sure it is
something that scales properly. But this is clearly an area
where I don't think anyone--they would be hard pressed to
convince me that no action is required.
And again, my bias on this Committee from the beginning,
having grown up in innovation, technological innovation, seeing
the compelling numbers about how important it is to our economy
and our culture, there is a lot of work to do.
And I am confident with the leadership of the Chair, we are
going to get work done. We look forward to your continued
engagement. Thank you.
Chair Coons. Thank you, Senator Tillis. I think it was Mr.
Brooks, I may be wrong, who early on said that other
technological developments, perhaps it was you Professor, word
processing didn't end authorship, smartphones didn't end
photography, but they impacted them. They impacted them.
And we need to closely, and with some deliberation, realign
what Federal rights and protections there are, both to deal
with things like deepfakes--some argue that Shakespeare himself
was a deepfake--to protect the rights of individuals, protect
the rights of those who earn their living by being creative, to
ensure that consumers understand what they are consuming, and
to make sure that we are aligning with other countries that
share our core values and our priority on a free market and the
rights of individuals in contrast to other countries with other
systems.
So, I am grateful to all of you for testifying today, for
taking your time and contributing to this. These are very
challenging questions. Members can submit questions for the
record for these witnesses if they were not able to attend.
Questions for the record are due by 5 p.m., one week from
today, July 19th.
Again, thank you, all. I look forward to your input as we
try and craft a good legislative solution. With that, this
hearing is adjourned.
[Whereupon, at 4:40 p.m., the hearing was adjourned.]
[Additional material submitted for the record follows.]
A P P E N D I X
Additional Material Submitted for the Record
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
[all]