[House Hearing, 118 Congress]
[From the U.S. Government Publishing Office]
ARTIFICIAL INTELLIGENCE AND INTELLECTUAL PROPERTY: PART I_
INTEROPERABILITY
OF AI AND COPYRIGHT LAW
=======================================================================
HEARING
BEFORE THE
SUBCOMMITTEE ON COURTS, INTELLECTUAL PROPERTY, AND THE INTERNET
OF THE
COMMITTEE ON THE JUDICIARY
U.S. HOUSE OF REPRESENTATIVES
ONE HUNDRED EIGHTEENTH CONGRESS
FIRST SESSION
__________
WEDNESDAY, MAY 17, 2023
__________
Serial No. 118-20
__________
Printed for the use of the Committee on the Judiciary
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Available via: http://judiciary.house.gov
______
U.S. GOVERNMENT PUBLISHING OFFICE
52-318 WASHINGTON : 2023
COMMITTEE ON THE JUDICIARY
JIM JORDAN, Ohio, Chair
DARRELL ISSA, California JERROLD NADLER, New York, Ranking
KEN BUCK, Colorado Member
MATT GAETZ, Florida ZOE LOFGREN, California
MIKE JOHNSON, Louisiana SHEILA JACKSON LEE, Texas
ANDY BIGGS, Arizona STEVE COHEN, Tennessee
TOM McCLINTOCK, California HENRY C. ``HANK'' JOHNSON, Jr.,
TOM TIFFANY, Wisconsin Georgia
THOMAS MASSIE, Kentucky ADAM SCHIFF, California
CHIP ROY, Texas DAVID N. CICILLINE, Rhode Island
DAN BISHOP, North Carolina ERIC SWALWELL, California
VICTORIA SPARTZ, Indiana TED LIEU, California
SCOTT FITZGERALD, Wisconsin PRAMILA JAYAPAL, Washington
CLIFF BENTZ, Oregon J. LUIS CORREA, California
BEN CLINE, Virginia MARY GAY SCANLON, Pennsylvania
LANCE GOODEN, Texas JOE NEGUSE, Colorado
JEFF VAN DREW, New Jersey LUCY McBATH, Georgia
TROY NEHLS, Texas MADELEINE DEAN, Pennsylvania
BARRY MOORE, Alabama VERONICA ESCOBAR, Texas
KEVIN KILEY, California DEBORAH ROSS, North Carolina
HARRIET HAGEMAN, Wyoming CORI BUSH, Missouri
NATHANIEL MORAN, Texas GLENN IVEY, Maryland
LAUREL LEE, Florida
WESLEY HUNT, Texas
RUSSELL FRY, South Carolina
------
SUBCOMMITTEE ON COURTS, INTELLECTUAL PROPERTY, AND
THE INTERNET
DARRELL ISSA, California, Chair
THOMAS MASSIE, Kentucky HENRY C. ``HANK'' JOHNSON, Jr.,
SCOTT FITZGERALD, Wisconsin Georgia, Ranking Member
CLIFF BENTZ, Oregon TED LIEU, California
BEN CLINE, Virginia JOE NEGUSE, Colorado
LANCE GOODEN, Texas DEBORAH ROSS, North Carolina
KEVIN KILEY, California ADAM SCHIFF, California
NATHANIEL MORAN, Texas ZOE LOFGREN, California
LAUREL LEE, Florida MADELEINE DEAN, Pennsylvania
RUSSELL FRY, South Carolina GLENN IVEY, Maryland
CHRISTOPHER HIXON, Majority Staff Director
AMY RUTKIN, Minority Staff Director & Chief of Staff
C O N T E N T S
----------
Wednesday, May 17, 2023
Page
OPENING STATEMENTS
The Honorable Darrell Issa, Chair of the Subcommittee on Courts,
Intellectual Property, and the Internet from the State of
California..................................................... 1
The Honorable Henry C. ``Hank'' Johnson, Ranking Member of the
Subcommittee on Courts, Intellectual Property, and the Internet
from the State of Georgia...................................... 3
The Honorable Jerrold Nadler, Ranking Member of the Committee on
the Judiciary from the State of New York....................... 4
WITNESSES
Sy Damle, Partner, Latham & Watkins LLP
Oral Testimony................................................. 7
Prepared Testimony............................................. 9
Chris Callison-Burch, Associate Professor, Computer and
Information Science, University of Pennsylvania
Oral Testimony................................................. 25
Prepared Testimony............................................. 27
Ashley Irwin, President, Society of Composers and Lyricists
Oral Testimony................................................. 51
Prepared Testimony............................................. 53
Dan Navarro, Grammy-nominated songwriter, singer, recording
artist, and voice actor
Oral Testimony................................................. 58
Prepared Testimony............................................. 60
Jeffrey Sedlik, President & CEO, PLUS Coalition
Oral Testimony................................................. 62
Prepared Testimony............................................. 64
LETTERS, STATEMENTS, ETC. SUBMITTED FOR THE HEARING
All materials submitted by the Subcommittee on Courts,
Intellectual Property, and the Internet, for the record are
listed below................................................... 99
Materials submitted by the Honorable Darrell Issa, Chair of the
Subcommittee on Courts, Intellectual Property, and the Internet
from the State of California, for the record
An article entitled, ``Will AI Value Human Creators? Now's
the Time to Decide the Future of Our Culture (Guest
Column),'' May 16, 2023, Billboard
A sample of AI entitled, ``Make the argument that AI-
generated material should not be copyrightable,'' AI-
generated
A sample of AI entitled, ``Make the argument that AI-
generated material should be copyrightable,'' AI-
generated
APPENDIX
Materials submitted by the Honorable Darrell Issa, Chair of the
Subcommittee on Courts, Intellectual Property, and the Internet
from the State of California, for the record
A paper entitled, ``Position Paper--Artificial
Intelligence,'' Copyright Alliance
Statement by the Honorable Darrell Issa
A summary from Stability.AI
QUESTIONS AND RESPONSES FOR THE RECORD
Responses to questions from the Honorable Darrell Issa, Chair of
the Subcommittee on Courts, Intellectual Property, and the
Internet from the State of California, for the record
Response from Stability.AI
A summary from Stability.AI
ARTIFICIAL INTELLIGENCE AND INTELLECTUAL PROPERTY: PART I--
INTEROPERABILITY OF AI AND
COPYRIGHT LAW
----------
Wednesday, May 17, 2023
House of Representatives
Subcommittee on Courts, Intellectual Property, and
the Internet
Committee on the Judiciary
Washington, DC
The Committee met, pursuant to notice, at 10:04 a.m., in
Room 2141, Rayburn House Office Building, Hon. Darrell Issa
[Chair of the Subcommittee] presiding.
Members present: Representatives Issa, Fitzgerald, Bentz,
Gooden, Cline, Kiley, Moran, Lee, Fry, Johnson of Georgia,
Nadler, Lieu, Ross, Schiff, Lofgren, Dean, and Ivey.
Also present: Representative Jackson Lee.
Mr. Issa. The Committee will come to order. I want to
welcome our guests and welcome the Members to what is
undoubtedly going to be the first of many AI hearings.
My staff, in preparation for this, knows one of my favorite
jokes on artificial intelligence, which I was telling long
before it is becoming a direct issue for this Committee, and
that is that artificial intelligence is what Members have when
their staff prepares their opening statements.
So, yes, today we stand at the intersection of two rapidly
evolving domains, generative artificial intelligence and
existing copyright law that must adapt to it. The advent of
generative AI technologies has sparked a profound
transformation in the creation, distribution, and consumption
of a new form of creative work. As we embark in our legal
journey along with the administration and their regulatory
powers, it is vital that we explore the complex relationship
between generative AI and copyright law, recognizing both the
challenges and opportunities that lie ahead.
Generative AI holds immense potential for innovative and
artistic expression. It empowers creators to explore new
frontiers fueling their imagination with AI-generated content.
Yet, with the power that comes with it comes responsibility,
responsibility for negative activity on the web and of
copyright laws being trampled. Copyright laws were designed to
protect intellectual property. Copyright laws also exist under
our Constitution specifically to encourage and then reward
creation. It is that encouragement that, in fact, creates the
right of ownership, not the right of ownership having some core
item.
In this hearing, we aim to strike a delicate balance. Let
me rephrase that in spite of what it says here. In this
hearing, we recognize there is a delicate balance, and we would
hope that our panel today represents both a diversity of views
and a possible collaboration.
We must consider though that there will be individuals on
each side who will not want to move. There will be companies
which would include some of the household words, Microsoft,
Meta, which is not yet a household word, but Facebook is, and
in fact, Google, that are on both sides of this issue being
both massive creators of their own copyright and massive users
of others.
It is our duty to adapt, refine copyright laws to
accommodate the transformative potential of generative AI while
safeguarding interests of existing creators and of the right of
the society as a whole to benefit. Generative AI presents both
challenges and opportunities for creative works and copyright
holders. It requires thoughtful consideration and ongoing
discussion with stakeholders to strike a balance between
protecting intellectual property rights, encouraging
creativity, and fostering innovative content, AI generated
included.
Today, all of us here understand that generative AI is
nascent. As we receive briefings, we hear about first
generation, second generation, third generation, now fourth
generation. By the way, the education of this young child is
profound from something that was hard to understand to
something that said the wrong thing, to something that now is
so useable that we might often forget to fact check its output
because it seems so good.
We must, first and foremost, address properly the concerns
surrounding unauthorized use of copyrighted material while also
recognizing the potential of generative AI can only be achieved
with massive amounts of data, far more than is available
outside of Copyright. By embracing a forward-thinking approach,
we can establish guidelines to promote responsible and ethical
practices in the realm of generative AI.
As we embark on our legal journey, let us approach it with
a spirit of collaboration as those who were on the call the
other day clearly were doing. Let us find common ground seeking
solutions that promote the flourishing of both creative
expression and intellectual property protection. The stakes
couldn't be higher, and the outcome will shape the future
landscape of art, technology, and copyright.
Today, let us navigate this uncharted territory to ensure
that generative AI and copyright laws foster in the future a
fairly harmonious and vibrant ecosystem for generations to
come.
In closing, I might say that we clearly could have
substituted patent or other areas of innovation and use just as
easily and we would have been having substantially the same
hearing, perhaps in some cases with different witnesses, and in
that sense, I think it is important for us all to look at this.
I will close by saying that it has come to my attention
that for once, Congress may not do either of the things we are
known for which is nothing at all or overreact. That is not
true around the world. Spain is moving forward with what I
think might be a very restrictive interpretation. Japan
believes, apparently, that all information that goes into the
teaching is, in fact, free of any copyright restriction in its
use. I do not believe that today's discussion will take us down
either road. I believe that we will measure carefully and find
middle ground that respects existing copyright law while
allowing the future of generative AI to flourish. With that, I
yield back and recognize the Ranking Member for his opening
statement.
Mr. Johnson of Georgia. I thank the Chair for holding this
hearing. I would assure those who are listening that Congress
is getting more and more information about artificial
intelligence and discussing the ramifications and all
enterprises of human behavior. I share your prediction that we
will act, and we will do so in a reasonable way. I don't know
if we will do the Japanese hard-liner position on inputs
though. I want to thank you for holding this hearing.
Artificial intelligence is a promising technology with the
potential to revolutionize a range of industries, but with that
promise also comes potential peril, especially to the creators
of art and other copyrightable works. It is critical that we
think through the many ways that AI will change our lives and
whether our existing laws are up to the challenge. I am glad
that we are beginning that process today.
Among the many questions that we must confront as
artificial intelligence takes on a larger role in our society,
is how our copyright system should view AI, both in terms of
how we could treat copyrighted works that are used to train an
AI model, in other words, the inputs and whether the new work
that is generated by AI, in other words the outputs, should be
eligible for copyright protection itself. Just as AI machine
learning is a complex process, so too is our inquiry today as
there may not be simple answers to any of these questions.
First, let's consider the inputs used to train AI systems.
A typical generative AI system is fed vast amounts of human-
authored work on which to train. This may include written word,
visual art, and music. The model then processes this
information with the help of various algorithms to detect
patterns and probabilities. A foundational principle of
copyright law generally requires users of copyrighted works to
obtain the permission of the copyright owner. While much of
input into generative AI systems tends to consist of works that
are protected by copyright, these works are typically used
without seeking consent or a license.
Some argue that this constitutes fair use and indeed this
question is being litigated in the courts as we speak. I am
hard pressed to understand how a system that rests almost
entirely on the works of others and can be commercialized or
used to develop commercial products owes nothing, not even
notice, to the owners of the works it uses to power its system.
Even if we determine that AI systems must seek permission
to use copyrighted works, that only leads to more questions.
For example, what sort of licensing system should be required?
What would represent fair compensation for these works?
Other questions arise as well, such as what degree of
transparency should be built into AI models, and how can we
ensure that proper credit is attributed to copyrighted works?
On the other side of the equation is we must consider the
status of the works, the outputs that are generated by the AI
model. Should a purely AI generated work be eligible for
copyright protection if it does not otherwise meet the
definition of infringement? What if there was significant human
creative involvement in generating the AI work? What about
works that may be of a hybrid manner, consisting of both human
authorship and elements that are purely AI? Whether or not an
AI-generated work is eligible to be copyrighted, such works
will compete and indeed are already competing in the
marketplace against human-authored works.
No examination of AI is complete without considering the
impact that AI works will have on human creators. How do we
balance the need for innovation with the need to protect human
creators? None of these questions have simple answers and that
is why I am pleased that we are joined by such a distinguished
panel of witnesses to help guide our Subcommittee's inquiry
into important copyright issues raised by the use of artificial
intelligence.
I am interested to hear your perspectives on these
questions, as well as whether our existing copyright framework,
including recent guidance issued by the U.S. Copyright Office,
is sufficient to address these questions or whether legislation
is needed.
I thank the Chair for holding this important hearing again
and I yield back the balance of my time.
Mr. Issa. I thank the gentleman. In lieu of the Chair's
opening statement, in addition to the opening statement, I read
which was a hybrid of GPT and my staff, I also will put a pure
opening statement that was done through regenerative AI into
the record and without objection it will be placed in the
record.
I now recognize the Ranking Member of the Full Committee,
Mr. Nadler.
Mr. Nadler. I thank the gentleman, thank you, Mr. Chair,
and thank you for holding this important hearing.
Yesterday, some of us were privileged to attend a briefing
by Sam Altman on generative AI. Thirty years ago, as the
Internet was in its nascent stage, we were forced to grapple
with a series of questions about how that then emerging
technology would change our society and whether our laws were
sufficient to address these new challenges. We are still
wrestling with many of those issues today.
Now, a new technology, artificial intelligence, has emerged
that has a similar potential to transform many aspects of our
lives. Once again, as our technology evolves, we must determine
if our laws must evolve as well.
Although AI gives rise to a myriad of issues, today we are
focusing on the intersection between artificial intelligence
and copyright. Even within this limited spirit of the law, I
find myself with far more questions than answers as we begin
this inquiry.
At its core, the fundamental question we must ask is how
can we promote innovation and further development of generative
artificial intelligence models while also protecting the rights
of creators whose works are the engine that fuels these models
and must compete with AI generated work in the marketplace?
This is not an easy balance to strike, and many answers just
lead to more questions.
For example, a typical AI model trains on vast quantities
on data analyzing the underlying text, images, or music as it
learns to make predictions based on that information to
generate new works. Most of the information which these
generative AI systems rely, is found in copyrighted works, but
typically, these works are used without seeking the permission
of the copyright holder. Worse yet, there is little
transparency as to which copyright works are used, thus
preventing creators from asserting their rights.
In fact, AI systems are generating new content based on
their work, content that may 1 day directly compete with their
work, but without their knowledge or consent and certainly
without compensation. This is a troubling development, but the
solutions are far from clear. What is the proper way to license
these works and to monitor compliance? What sort of
transparency and accountability should be built into these
systems? It is time to begin considering these important
issues.
As we wrestle with these questions about how to treat the
inputs into a generative AI system, there are equally thorny
questions about the outputs. How should the copyright laws
treat works that are generated by an AI model? The Copyright
Office recently offered guidance on that question. According to
the guidance, only works of, ``human authorship are eligible
for copyright protection.'' It goes on to explain that it is
necessarily a fact-specific inquiry to determine where the line
is between human work that is merely assisted by AI technology
and works under which the traditional elements of authorship
are actually conceived and executed by the AI model itself.
To add further complexity, in some cases, certain aspects
of a single work may be eligible to protection while others are
not. We must consider whether the Copyright Office has the
tools and resources it needs to evaluate AI-related
applications and to enforce existing rules.
Finally, there are important questions about the impact
that AI-generated works will have on the market for human-
offered works, works that may have served as the foundation for
generating the AI work in the first place.
While we work to promote innovation in the AI realm, we
must also ensure that artificial works do not displace human
creators or threaten the ability of the creative class to
thrive. Already creators like the members of the Writers Guild
of America are uniting to speak up for their own rights and to
ensure they are not replaced by AI systems.
All these issues lead to a final set of questions. What, if
anything, should Congress do to ensure that we strike the right
balance between protecting creators and promoting innovation?
Does our existing copyright framework up to the task of
navigating these complex issues? Should we wait and see how AI
technology evolves before taking any action or let these issues
play out in the courts? There are no easy answers and I
appreciate the opportunity to be thinking through these
important questions with our distinguished panel of witnesses.
I thank the Chair for holding this hearing and I yield back
the balance of my time.
Mr. Issa. I thank the gentleman. The gentleman yields back.
I now ask unanimous consent that an article titled, ``Will
AI Value Human Creators? Now's the Time to Decide the Future of
Our Culture,'' to put it into the record. I now ask that this
be placed in the record.
I additionally ask and I will distribute copies for
unanimous consent that ChatGPT's answers for when we asked it
to argue for regulation and when we asked it to argue against
regulation be both put in the record. I will say that it is a
convincing argument in both directions.
With that, I now have the honor of introducing our
distinguished panel.
Mr. Sy Damle is a partner in copyright practice at Latham &
Watkins. He is a former software engineer and specializes in
technology matters including those involving computer systems
and networks and artificial intelligence. He is previously
General Counsel at the Copyright Office. For all those reasons
and more is why you are here. Welcome.
Mr. Callison-Burch is an Associate Professor of Computer
and Information Science at the University of Pennsylvania. He
is also a Visiting Research Scientist at the Allen Institute
for Artificial Intelligence and welcome.
Mr. Ashley Irwin, Mr. Irwin is President of the Society of
Composers and Lyricists and an Emmy Award-winning Music
Director, Conductor, Composer, Arranger, and Producer who has
scored over 30 featured films, 300 hours of television, and
over 3,000 commercials. I suspect the commercials probably made
you the most as that sometimes goes.
Mr. Dan Navarro, Mr. Navarro is a Grammy-nominated
songwriter, singer, recording artist, and voice actor. He was
formerly in the duo Lowen and Navarro, released 13 albums and
performed over 1,500--this was done by my staff, not by AI,
gigs, and is currently a solo artist who tours constantly, but
has blessed us with a day off the road with no compensation and
for that we are very appreciative.
Mr. Jeffrey Sedlik is the President and CEO of PLUS
Coalition, which is a global, nonprofit initiative to make it
easy and fast to search for, find, communicate, and understand
information about photographs and visual artwork. Mr. Sedlik is
also a Member of the Joint Committee on Ethics in AI and a
professional photographer.
I want to welcome all our witnesses, recognizing that this
is a Judiciary Committee, I am going to ask you to all rise and
take the oath.
Raise your right hand for the camera. Do you solemnly swear
or affirm under penalty of perjury that the testimony you will
give will be the truth, the whole truth, and correct to the
best of your knowledge so help you God? Thank you.
Please be seated. Let the record indicate that all
witnesses answered in the affirmative.
As you may know from watching CSPAN, your entire statement
will be placed in the record, along with, without objection,
any extraneous material you see fit now or in the next five
days to include so that as we go through this, if you realize
there is something that you should have said, would have said,
or wanted to say you will be allowed to supplement.
With that, Mr. Damle, you are first up for five minutes.
Because we have a lot of people up on the dais that want to ask
you questions, I will ask that each of you try to limit right
to five minutes or less.
The gentleman is recognized.
STATEMENT OF SY DAMLE
Mr. Damle. Chair Issa, Ranking Member Johnson, and Members
of the Subcommittee, thank you for inviting me to participate
in today's hearing.
I am a former software engineer, former General Counsel of
the U.S. Copyright Office, and currently a partner at the law
firm Latham & Watkins, LLP. I am here testifying solely in my
personal capacity and not on behalf of my law firm, any of the
firm's clients, or the U.S. Copyright Office.
My testimony today will focus on the copyright implications
of AI training, but first I would like to put those issues into
context. The AI tools of the present and near future will
impact almost every aspect of the human experience. It will
improve our science and our medicine. It will make our military
more effective. It will make our businesses more efficient and
productive. It will enable anyone to more fully unlock their
creative potential. In short, AI has the potential to transform
our economy and improve our society as a whole. That outcome is
not guaranteed.
The way we regulate AI will directly determine whether the
United States will continue to lead the world in AI
development, or whether another country will take up that
mantle. In considering whether to impose intellectual property-
based restrictions on AI innovation, Congress should carefully
evaluate whether those restrictions will hamper the development
of AI here in the United States.
At the same time, artists, writers, and other creators have
expressed genuine concern that the rapid development of AI will
displace human authors. Policy makers should take those
concerns seriously. By taking a step back, it is important to
appreciate that every new technological development has caused
similar fears and in hindsight, we can see that those fears
have not come to fruition. To just take one example, when
photography was invented in the mid-1800's, one prominent
critic dismissed the medium as, ``the refuge of every would-be
painter too ill-endowed or too lazy to complete his studies.''
He predicted that photography would ``corrupt art altogether.''
Now, I am sure Mr. Sedlik can tell us how much he disagrees
with that sentiment as an accomplished photographer. Of course,
society embraced the camera as a creative tool and photography
blossomed as an art form that deepened rather than diminished
in the field of human creativity.
There is no reason to believe that generative AI is any
different. Like the camera, or the many creative tools adopted
since, generative AI will be an engine of human creativity, not
a replacement for it.
So, with that context in mind, I want to make two points
today.
First, Copyright's well established fair use doctrine is
the best way to balance the competing interest in the AI space.
For the reasons I explain in detail in my written testimony,
the training of AI models will generally fall within the
established bounds of their use.
While some AI models may very well exceed those bounds, our
courts are well equipped to handle those situations.
The concerns that some have raised that AI models can
replicate artistic styles are completely understandable. The
solution does not lie in copyright law which does not and has
never granted monopolies over artistic or musical style.
Second, some groups have proposed a collective licensing
regime for AI training data. Such a regime would eliminate fair
use in this area, replacing it with a rigid assumption that AI
training is infringing. I believe that would be a mistake.
Plus, if Congress were nevertheless interested in setting up a
collective licensing regime, it should be aware of some of the
serious practical challenges it will face. I detail those
challenges in my written testimony, but fundamentally, they are
rooted in the fact that successfully training an AI model
requires using many billions of pieces of content. That is many
orders of magnitude larger than the number of works covered by
any similar scheme in the history of American law.
In short, Congress has already adopted a copyright right
that is technology neutral and flexible enough to balance the
need for a dynamic domestic AI industry with the right co-
creators. I look forward to answering your questions today.
[The prepared statement of Mr. Damle follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Issa. The gentleman yields back early.
Mr. Burch, Professor Burch.
STATEMENT OF CHRIS CALLISON-BURCH
Mr. Callison-Burch. Chair Issa, Ranking Member Johnson,
distinguished Members of the Subcommittee, thank you for the
chance to testify on this important topic.
My name is Chris Callison-Burch. I'm a professor at the
University of Pennsylvania, a visiting researcher at the the
Allen Institute for Artificial Intelligence, and the Deputy
Chair of the Board of Advisors of the Johns Hopkins University
Human Language Technology Center of Excellence.
Generative AI had its breakthrough moment in November of
last year with the release of OpenAI's ChatGPT. This brought my
field of research into the public eye and generated a huge
amount of enthusiasm. I had access to OpenAI's large language
models about a 1\1/2\ years before the public. Despite having
worked in this field for over 20 years, I was shocked by its
capabilities. My first encounter with it pitched me into a
career existential crisis.
The technology had seemingly solved many of the problems
that I was researching, that it could translate texts from
Russian into English, it could write coherent summaries of long
documents, and then answer questions about them. I wondered
whether there was any room left for academic research in light
of the fact that these large language models required google-
sized data centers to train.
So, I asked myself should I just drop out of computer
science and become a poet? Of course, the next week I
downloaded 15,000 poems from the Internet and trained the
system to write much better poetry than I ever could. I've
subsequently calmed down and I do not think that my job is at
imminent risk of being replaced by ChatGPT, but I understand
that many other people are experiencing the same sense of panic
that I had.
Artists and writers are worried about their work being
devalued. I worry that careers like a paralegal might go the
way of a lamplighter. I think that at its core what we're
talking about today goes far beyond copyright. It's about the
value of work.
This is a truly transformative technology that will shape
many aspects of our lives. I hope that it is for the better. I
optimistically believe that AI will enable us to be more
productive workers and to allow more people to realize their
creative visions.
In my testimony today I hope to offer my expertise in the
technical aspect of generative AI, and I promise to explain it
in a way that doesn't require a Ph.D. in computer science,
answers to any questions that you have about the potential for
legislation impacting on innovation in this field, and advocacy
for retaining fair use for the purposes of training generative
AI systems.
In my written testimony I've provided an overview of how
these systems work. I'm happy to explain during the hearing
today how they do or to have a one-on-one meetings with you or
your staff at a later date.
To briefly summarize the points that I want to highlight
from my written testimony, generative AI is trained on huge
amounts of data. Large language models are now trained on
roughly one trillion words. Image generators are trained on
hundreds of millions of images. Much, or even most of that data
consists of copyrighted works that have been gathered by
automatically crawling the web.
It's important to remember that from these copyrighted
works AI systems learn. This learning process is called
pretraining, which is the ``P'' in GPT. Pretraining AI systems
is different than how we teach our children to learn, but the
effect is similar. AI systems learn how to use language. They
learn facts about the world. They learn ideas and opinions.
They learn visual concepts. They even learn some rudimentary
common-sense reasoning skills.
This pretraining happens on copyrighted data which is then
set aside as models are fine-tuned to perform more specific
tasks. For instance, a large language model can be fine-tuned
on a much smaller purpose-built set of data to become an
intelligent tutor or a computer vision system can be fine-tuned
to detect cancerous growths in mammograms. These systems could
not be as easily adapted to these specialized tasks without the
general knowledge that they acquire from the copyrighted data
that they're pretrained on.
I believe like Sy that pretraining these systems squarely
falls within fair use and that Internet era court precedents
likely established this as the case, although as the Ranking
Member mentioned, this is currently being litigated in the
courts.
I do believe that the output of generative AI systems can
infringe on copyright and it's worth Congress considering
legislation to better shape copyright to govern things like
copyrightable characters and possibly to extend copyright to
cover things like right-of-publicity.
I look forward to discussing this topic with you today.
[The prepared statement of Mr. Callison-Burch follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Issa. Thank you.
Mr. Irwin?
STATEMENT OF ASHLEY IRWIN
Mr. Irwin. Chair Issa, Ranking Member Johnson,
distinguished Members of this Subcommittee, thank you for the
opportunity to testify today on the important issues involving
artificial intelligence and the impact on the creative
industries.
My name is Ashley Irwin, and I am currently serving my
sixth term as President of the Society of Composers and
Lyricists. I'm here today to advocate for my members who are
already feeling the negative impact of generative artificial
intelligence on their craft and its potential threat to their
profession.
First, some background on me and on the SCL: In my 40-plus
years as a composer, arranger, and such I have written music
for over 1,000 hours of film and television and more than 3,000
commercials. Since 1990, I've provided musical compositions and
arrangements for 23 Academy Awards shows and been part of
several Emmy-winning teams. I've had the pleasure of
collaborating on many films with Oscar winners such as Clint
Eastwood and Bill Conti. Over the years my arrangements have
been performed on numerous occasions for Presidents Clinton,
Bush, Obama, and one of my chorale orchestrations was performed
during President Ronald Reagan's State funeral service.
The SCL background: The Society of Composers and Lyricists
is the premier organization for music creators working in all
forms of visual media. With chapters in Los Angeles, New York,
and Nashville the SCL operates as the primary voice for over
3,000 members who work as creators of scores and songs for
film, television, video games, and theater.
While the SCL advocates for our members on several
different fronts, the issue that is consumed the majority of
our time recently has been generative AI. The rapid
introduction of generative AI systems is seen as an existential
threat to the livelihood and continuance of our creative
professions unless immediate steps are taken on legal,
interpretive, and economic fronts to address these emerging
issues.
I want to be very clear: My goal is raising these issues
pertain to the rights of writers and creators is not to block
AI research and usage. We're simply advocating for the creation
of a policy framework that ensures generative AI is developed
and utilized responsibly, ethically, and with respect for human
creators and copyright, so that the creative arts that are the
real engine of generative AI can continue to flourish.
The SCL believes that AI companies in their generative
model should adhere to the fundamental ``Three Cs'': Consent,
credit, and compensation. Consent by creators for the first use
of their works in generative AI media; credit wherever
audiovisual creators' works are used; and compensation at fair
market rates for the ingestion of any portion of human
creators' copyrighted works by AI generative machines and the
subsequent output of new derivative works.
I'd like to highlight three challenged posed to music
creators by generative AI, potential solutions to which I've
offered in my written testimony.
Issue 1: Generative AI has been equipped using copyright-
protected human-authored works and programmed to mimic those
works without consent, compensation, or credit.
Issue 2: Copyright information, metadata, has been removed
during the ingestion process of these models.
Issue 3: The market will be diluted due to AI-generated
works. As a result, copyright protection should not be granted
to AI-generated works.
In closing, I thought I should address why America's
success is important to me. As you can no doubt tell by my
accent I'm not originally from the United States. I came here
from Australia as a young man because I wanted to be a part of
the vibrant culture that is the U.S. entertainment industry.
However, the rise of generative AI poses a threat to this
unique American art form.
If we do not protect and nurture our human creators, we
risk losing one of our greatest exports and its profound
influence. It's essential to prioritize policies and
regulations that safeguard the intellectual property and
copyright of creators and preserve the diverse and dynamic U.S.
cultural landscape.
Protection of creators is not now, nor has ever been in
conflict with technological development. Our Founding Fathers
recognized that. The only place the term ``right'' is mentioned
in the Constitution is with regard to intellectual property.
Specifically, the rights granted to authors and inventors in
Article 1, Section 8, Clause 8, are to promote the progress of
science and useful arts. Respect for copyright and development
of AI should go hand in hand. That way all humanity can
benefit.
I believe this Committee has the power, authority, and
motivation to lead that charge and I look forward to working
with every one of you to achieve that common goal.
[The prepared statement of Mr. Irwin follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Issa. Thank you.
Mr. Navarro?
STATEMENT OF DAN NAVARRO
Mr. Navarro. Good morning, Chair Issa, Ranking Member
Johnson, and esteemed Members of the Subcommittee. Thank you
for the opportunity to appear before you today.
My name is Dan Navarro and I've been a songwriter,
recording artist, session singer, voice actor, and music
activist for over four decades. Throughout my career I've
frequently been asked how did you come up with that song? While
I often feel like saying it just came to me, the truth is all
my work reflects a lifetime of personal emotions, rich
experiences, and even shattered dreams. In other words, it's
complicated.
Sometimes it's a moment. The song that saved and sustained
my career, ``We Believe,'' recorded by Pat Benatar 40 years
ago, came to me right as I was giving up a career in music and
a long-time collaboration with my best friend was decaying. We
decided to give it one more shot and I started with the end of
the song and soon we were trading lyrics back and forth. It
worked for us in one way or another. Two estranged friends
found a space to connect and a song that people have enjoyed
for the last 40 years was born in 90 minutes. That human
alchemy can't be fully explained, but it's the heart of music
creation.
For generative AI the answer to the question where did that
come from, is in many ways much simpler. These machines have no
emotions or experiences or dreams of their own to draw from.
All they have are millions and millions of imported songs and
lyrics, most copyrighted, hoovered off the Internet without
permission. Training AI to mimic professional performers or
generate new works based on millions of copies of published
songs and recordings presents a host of legal implications,
from copyright infringements, to violation of rights of
publicity and trademark, to name, voice, and likeness abuses.
It's a long-term threat to music itself. By marginalizing
and ultimately abandoning the fundamental human spark and music
creation we are inviting a future that sees fakes as real and
that debases our art and culture with soulless brown food
product mediocrity.
Does anyone thing a computer-generated song can give you
goose bumps, or comfort, or become a theme song to the loves of
our lives? That's our song. How can AI give a goose bump if it
can't get a goose bump? That's human.
To fight for human creativity, I was proud to help launch
the Human Artistry Campaign in March and I'm here today as a
supporter of that initiative. Now, boasting more than 100
organizations globally representing all kinds of arts and
creativity it is the global center of gravity advocating for
the rights of creators in the age of AI.
This coalition believes AI is exciting and promising but
can never replace human artistry and soul. It's based on seven
core principles that I support wholeheartedly:
(1) LTechnology has long empowered human expression, and AI
will be no different. Musicians will use this technology to do
great things.
(2) LHuman-created works will remain essential in our
lives. At the heart of the connection between the artist and
the audience are shared lived experiences only humans can
relate to and convey.
(3) LThe use of copyrighted works for AI purposes and the
use of voices and likenesses of professional performers
requires permission. Like all predecessor technologies AI must
be subject to authorization and free market licensing from all
rights holders and creators.
(4) LGovernments should not create new copyright or other
IP exemptions that allow AI developers to exploit creators
without permission and compensation. People looking to make a
fast buck from technological change routinely as policymakers
to pick winners and losers. Here that would be devastating.
(5) LCopyright should only protect the unique value of
human intellectual creativity. The copyright clause of the
Constitution exists to incentivize humans to create. Machines
don't need incentives.
(6) LTrustworthiness and transparency are essential to the
success of AI and the protection of creators. Without
transparent AI we will have no idea whether the inputs AI
systems were trained on were licensed leaving us no way to
enforce our rights.
(7) LCreators must have a seat at the table, not just
developers. Our creativity, our rights, and our livelihoods are
at stake.
If AI is allowed to take away the ability of authors and
artists to control and make a living from their art, we will
lose all authenticity in our expression. We'll lose culture
itself. The next decision by the courts and Congress in this
area will decide our cultural future and it's your
responsibility to make sure the cultural promise of reward for
human genius remains viable.
Guided by the principles of the Human Artistry Campaign we
can look forward to the real emotions, experiences, and dreams
of future generations of creators, perhaps facilitated by AI,
but never silenced by it.
I thank you and I look forward to answering your questions.
[The prepared statement of Mr. Navarro follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Issa. Thank you.
Mr. Sedlik?
STATEMENT OF JEFFREY SEDLIK
Mr. Sedlik. Chair Issa, Ranking Member Johnson,
distinguished Members, thank you for the opportunity to testify
here today.
I've been a professional photographer and film maker for 37
years. I'm a Professor at the Art Center College of Design, the
former President of the American Photographic Artists, and the
current President of the PLUS Coalition.
As mentioned by Chair Issa, PLUS is a global nonprofit
organization focused exclusively on simplifying the
identification of visual works. PLUS is currently developing a
global nonprofit visual registry and in cooperation with the
IPTC updating our widely adopted metadata standards to allow
artists to declare AI-related permissions and prohibitions in
their image files. You can learn more about that a plus.org.
As a professional visual artist, I make my living by
creating, and most importantly licensing my works. My ability
to create new works, sustain my business, and support my
family, depends directly on my exclusive rights to reproduce,
distribute, publicly display, and adapt my original works.
These fundamental rights are the core of my business providing
a strong incentive to create new works. In fact, they're the
only way that I can afford to create new works.
Unfortunately, many AI developers and platforms have built
their businesses by exploiting billions of visual works without
authorization from or compensation to copyright owners. The
theft and exploitation visual works by AI platforms displace a
common long-standing practice, in which, creators and their
agents offer and sell artist reference licenses permitting the
use of works for reference in creating new derivative works.
AI ingestion falls squarely within this reference license
category. In fact, stock photo agencies routinely sell
reference licenses to AI platforms permitting the use of visual
works for AI ingestion. A market clearly exists for these
licenses. The unlicensed ingestion of photographs by AI systems
usurps that market and forces human creators to compete with
machine-made derivatives of their own original works.
Many AI platforms are trained on copies of creative works
scraped from websites that display those works without the
knowledge or permission of the copyright owners. I've found
thousands of unlicensed copies of my works in open data bases
used by AI systems to support image ingestion and generation.
It is clear that many generative AI platforms were founded on
copyright infringement.
In defense of their actions AI developers attempt to apply
blanket clearings of fair use, counter to the spirit and letter
of the Copyright Act. Fair use is not a right. It's an
affirmative defense requiring a fact-specific inquiry on a
case-by-case basis.
AI developers claim that AI-generated works are not
substantially similar to source works and thus can't be
infringements, but this ignores the fact that the exclusive
reproduction right is a stand-alone right under Section 106 of
the Copyright Act. The creation and use of copies for AI
ingestion purposes is copyright infringement on a massive
scale.
AI developers further suggest that a photographer's use of
cameras is the creative equivalent of drafting prompts for
generative AI. They minimize the creative process in
photography and attempt to frame photographers as mere button
pushers. This is a false equivalency.
We photographers are visual storytellers. Our creative
decisions are guided by our life history and our unique
combination of training, experience, personality, aesthetic
sensibilities, dreams, memories, research, and other factors.
When creating our works, we anticipate and respond dynamically
to the subject matter and shooting environment. We exercise
control over the visual rendition of the scene. We decide which
elements to include and exclude and where to place those
elements within the frame. We determine how to juxtapose
people, objects, and other compositional elements for a desired
creative effect. We control the placement and interplay of
color, tone, texture, contrast, light, and shade. We control
the perspective, distortion, depth of field, and selective
focus to guide the viewer's eye through the image. We select
the precise moment at which to create the ultimate photograph.
This substantial creative human expression is not the
equivalent of submitting a text prompt instructing a machine to
generate a work.
Copyright law affords protection only to human expression.
The output of an image by machines in response to prompts is
and should remain ineligible for copyright protection. Policies
must must not favor machines over human creators. AI developers
must be required not only to obtain advance permission to
ingest and exploit creative works, but to compensate creators
whether directly or through collective licensing organization
such as the American Society for Collective Rights Licensing.
AI technologies must be developed and used in a manner that
is responsible, respectful, and ethical upholding the
underlying goals and purposes of our copyright system.
I appreciate the opportunity to testify, and I look forward
to answering your question.
[The prepared statement of Mr. Sedlik follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Issa. Thank you. I thank all our witnesses.
I am going to forego my questioning until the end or near
the end, and so I am going to go to Mr. Fitzgerald, somebody
who understands what it is to own copyright in his former life.
The gentleman is recognized for five minutes.
Mr. Fitzgerald. Thank you, Mr. Chair.
Mr. Damle, the Supreme Court has held since an 1884 case,
Borrow-Giles Lithographic Co. v. Sarony, that photographs can
be entitled to copyright protection where the photographer
makes decisions regarding creative elements of the work.
You were quoted in your testimony as saying,
. . . society embraced the camera as a creative tool and
photography blossomed as an art, as an art form that deepened
rather than diminished this whole field of creativity.
There is no reason to believe AI is any different, I think was
your point. The Copyright Office has disputed this comparison,
instead comparing the generative AI to a client to hires an
artist to create something because users do not exercise
ultimate creative control.
What is kind of your response or your thoughts to copyright
comparison and do you believe that Copyright was correct in
denying copyright for lack of the ultimate creative control as
they did in their decision in ``Zarya of the Dawn''?
Mr. Damle. Congressman, thank you. Thank you for that
question. So, the way I would start by framing this issue is
the point I was trying to make in my testimony is that
generative AI can be a tool that humans use to enhance their
creative output. So, the Copyright--what the Copyright Office
has said is where that tool is doing all of the work of
creative output, then that's a situation where we don't need
the economic incentive that the Constitution has an incentive
for creating that output.
I think that's sort of one end of the spectrum of the
question. There's going to be a big gray area where there's
going to be basically human and AI together creating output. I
think that's going to be a very common situation going forward.
In that situation, I would say where the human is exercising
some control over the AI and its output, where there is a sort
of iterative process between what the generative AI produces
and what the human produces, then that is a situation where you
should have copyrightable output, that the output of that
process should be copyrightable to some degree.
It's no different I think than a camera, than a
photographer's interaction with a camera. They're adjusting the
settings on the camera. They're choosing the framing of the
image. They're employing a lot of choice. Yes, the camera is
the one that's actually recording the image, but the human has
control over that process. So, by the same token where a human
has that level of control over the generative AI process, as I
think will happen in many cases, then the output of that should
be copyrightable in the same way that a photograph is
copyrightable.
Mr. Fitzgerald. So, let me just reiterate that. Do you
think Copyright got this correct in the way that they had come
up with their determination?
Mr. Damle. I think in the particular instance and the
particular things that they said, I think, they got it right,
but they're addressing a very just sort of extreme example
where the human is not actually providing anything other than a
simple prompt to the generative AI system. Then they decided in
that instance the output is not copyrightable. I think that's
going to be very different than the mine-run of cases using
generative AI. The mine-run of cases are going to involve much
more involvement of human authorship than just providing a
simple one-sentence prompt.
Mr. Fitzgerald. Yes, so just because I just have 1 minute
left actually, so what other situations do you think could
emerge that Copyright would then have to deal with where you do
find this mix of artistry along with what AI has capabilities
of doing that could somehow put us in a place where it could be
undetermined who actually is the creator?
Mr. Damle. So, those are going to have to be determined on
a case-by-case basis. The office, actually, in its guidance
provided some examples, a few examples, but some examples of
where there would be sort of a creative--enough creative human
authorship. For instance, where you provide a prompt, it
generates an image, and then you adjust that image. The human
actually takes that image as a starting point and then adds
more creative authorship to that image. That's one example.
Another example that you might see is where you start with
an image that you have--I'm using an image, ones as an example,
but you start with an image, you feed it into the generative
AI, and the generative AI helps you make changes to that image
in certain ways, in the same way that you might use Photoshop.
In those circumstances you're going to have situations where
the output is copyrightable. Then there may be questions about
how much of it is copyrightable, how much of it was generated
by the AI, but those can be dealt with on a fact-by-fact basis.
Mr. Fitzgerald. Thank you very much, and I would yield
back.
Mr. Issa. I thank the gentleman.
We now recognize the Ranking Member of the Subcommittee,
Mr. Johnson, for five minutes.
Mr. Johnson of Georgia. Thank you, Mr. Chair.
The other day I was in my car cruising and there was an
interview with Smokey Robinson on the radio, and Smokey
Robinson was saying that his big hit ``Cruisin'' took in total
about five years to come to the final product.
I would like to ask you, Mr. Irwin, Mr. Navarro, and Mr.
Sedlik--your works can be scraped from the Internet in a matter
of seconds for AI ingestion. Can you explain what goes into
creating a single work of art and how long that process might
take? Starting with you, Mr. Irwin.
Mr. Irwin. Sure. The answer is how long is a piece of
string, to be honest. It's always different. When we're working
in audio-visual space, be it film, television or something like
that we're commissioned to work, you have deadlines, and you
have to come up with the goods by a certain time and deliver
them.
When working as a songwriter more like Dan does, you have a
little more freedom and you can work on a lyric, you can work
on a song, you can collaborate, you can have a half-written
song, and then bring someone else in. Everything takes time.
Nothing is the push of a button. There is no push-button music
up to this point, that's any good anyway. Most of it is just
not usable.
Mr. Johnson of Georgia. Mr. Navarro? Thank you.
Mr. Navarro. I'd like to affirm what Mr. Irwin says. I made
reference to my big hit song that took 90 minutes to write. The
emotions that led to my contributions and the contributions of
my late partner took three years to percolate, to run through
my emotions and my system be expressed, be wept, be verbalized,
and be internalized before they came out in an evening blast
when everything was ready, it didn't take 90 minutes. It took
several years. I've had songs take seven years; I've had songs
take seven days. It varies depending on what it takes to sit
and look at something and go it's done. Now, it's done.
Mr. Johnson of Georgia. Thank you.
Mr. Sedlik?
Mr. Sedlik. Thank you. There are many genres of photo-
graphers, each of whom works in a different way, including
within the genres. Personally, I'm very controlling in my
photography. Every aspect of every photograph that I take is
something that is planned and controlled, whereas another
photographer might be more spontaneous.
In my process I use free association first. It will take
weeks, months, even more than a year. I use free association,
come up with ideas, make thumbnail sketches, make iterative
sketches, do tests, do planning, set the whole thing up, test,
and then during the process I control it, whereas a
photojournalist might make all those decisions in an instant.
It's their lifetime of experience and all their capabilities
that lets them accomplish the same thing, thousands of
decisions in a second.
Mr. Johnson of Georgia. Let me ask you this, Mr. Sedlik:
Can you compare the process of creating works, your works
through--or can you compare that to works created through
generative AI?
Mr. Sedlik. Yes. So, the same goes for AI. There's the full
spectrum--
Mr. Johnson of Georgia. If you could tell me also how it
feels as an artist to have your work used without your
permission to serve as the basis for a product that might then
compete against your product.
Mr. Sedlik. Yes.
Mr. Johnson of Georgia. If you could leave some time for
Mr. Navarro and Mr. Irwin to respond to that question also.
Fifty-eight seconds. Go ahead.
Mr. Sedlik. OK. So, I'll answer your second question first.
So, photographers are used to the advancement of technology.
We've been using technology for 179 years and we're usually
first in using it. So, we anticipate that technology is going
to continue to develop, that new opportunities for creation are
going advance, that it's going to be easier to create great
works, and we accept that. That's not our concern here. Our
concern is that our works are being used without our permission
and without any compensation. The creative process--
Mr. Johnson of Georgia. Thank you. Let me go to Mr.
Navarro. Appreciate it.
Mr. Navarro. I feel the same way. I think AI is a tool that
I, myself, could use. My partner in my duo passed away. I could
maybe impersonate his voice and put out a brand-new loan and
borrow record. Permission, credit compensation.
Mr. Johnson of Georgia. Thank you for that.
Mr. Irwin?
Mr. Irwin. We've been using elements of AI as tools for--
I've certainly been using them for almost 30 years, but always
with the control of what the output is. Quite often, create the
initial piece of music that you're going to work with. Then,
you process it in a particular way, and you make the decisions
of what is the final product going to be. Some of the options
you are given are not usable. It's a matter of taste, and only
humans have that kind of discerning taste.
Mr. Johnson of Georgia. Thank you.
Mr. Chair, I appreciate the indulgence.
Mr. Callison-Burch, if I had time, I would be asking you
about the value of work that you would see as so critical for
us to think about.
Thank you.
Mr. Issa. I thank the gentleman.
We will make sure we get that answer.
Mr. Issa. The gentleman from Oregon, Mr. Bentz.
Mr. Bentz. Thank you, Mr. Chair.
I thank all of you for being here today.
So, my question is--I'm not sure which one of you to ask--
but the issue is whether our technology is such that we can
actually do what I think many of you want us to. I'm just
thinking of the Led Zeppelin situation that took years to sort
out whether or not, quote, ``the descending chord sequence had
been used for centuries,'' and whether or not it had been
stolen. Of course, although it was decided after five, six,
seven, or eight years of litigation, that Led Zeppelin had, for
various reasons, not usurped it.
My question to you--and I'm going to ask it of you, Mr.
Navarro--if you think that we have the technology available to
sort these things out after the fact. If not, what would you be
suggesting that we do to try to prevent it before it happens? I
just want to know, given the nature of music--and I'm a very
bad musician, but I know what a guitar is--how in the world do
you sort this out? Do we have the technology available?
Mr. Navarro. I'm not sure we're qualified to determine how
it's sorted out. I know that I, as a creator, if I get too
close to something, I pull back. If I get a little too close to
Ray Charles, I pull back. If I get too close to John Lennon, I
pull back. If I don't do it right, there are legal remedies.
Just ask George Harrison or Robin Thicke.
Mr. Bentz. Mr. Damle, I'm not sure I pronounced your name
correctly, but could you address the question?
Mr. Damle. Thank you, Congressman. Yes, you got it exactly
right.
So, I would point to the sort of cases that you're talking
about, like the Led Zeppelin case, like the Blurred Lines case,
as example of instances where copyright did not stay within its
proper bounds, where courts decided that borrowing somebody's
musical style counted as copyright infringement.
I think when those decisions came down, it really threw the
music industry into chaos. I'm a music lawyer, in addition to
being a technology lawyer. I know that the music industry
really grappled with those decisions. You had artists saying
that they were afraid to create because they were worried that
they were, inadvertently perhaps, borrowing the style from
somebody else. It got to the point that even the rights owner
groups started criticizing those decisions, saying they went
too far; they did not allow artists to express themselves as
freely as they should be able to.
Thankfully, I think we've seen a return back to those core
principles of the Copyright Act, which is ideas/styles should
be able to be used by all. So, that's what we've seen in these
recent decisions involving Ed Sheeran--a return back to that
principle of being faithful to the Copyright Act, faithful to
the Constitution's mandate for what the Copyright Act is
supposed to do.
So, that would be my response to that issue.
Mr. Bentz. Thank you for your answers.
With that, Mr. Chair, I yield the rest of my time to you.
Mr. Issa. I thank the gentleman for yielding.
I'll continue along somewhat an earlier line. Mr. Navarro,
we talked about credit; we talked about permission, and we
talked about compensation--or you did.
Credit would seem to be one that Congress could mandate
that the data base input could be searchable. So, you would
know that your work or your name, or something, was in the data
base. Fairly easy, probably doable. It's a credit index of
billions, or trillions in some cases, but that would be--the
output might be a little more complex, and we'll talk about
that later.
The compensation is a question I want you to opine on, and
maybe both ends of the question. If there are 10 billion or 10
trillion inputs--and let's just assume for a moment that there
is a billion copyrighted. We all know what it's like to get
that big check from Spotify for the hundred or a thousand times
you were played, and it comes out in pennies. OK. What is the
division of a billion pieces of music, and how would it, in
fact, assuming that this was part of the output, how would you
actually quantify it? Because we have to put a number on it at
some point.
Mr. Navarro. It's a difficult thing to do. I believe in
free-market negotiations with regards to this. I know that can
be cumbersome. I don't believe in compulsory licenses,
especially as regards to this particular issue.
The compulsory licenses I'm used to, whether it's at
SiriusXM or in the use of a song that I've already recorded,
which is a compulsory license for someone else to use, benefits
me directly. When my stuff is part of a large number of stuff
that's scraped, it supersedes my work. It doesn't even just
compete with it. That's a footrace. It supersedes it. In that
context, I believe in free market.
Mr. Issa. We will, undoubtedly, be asking that question of
others in another forum.
With that, we go to the Chair of the Full Committee, Mr.
Nadler.
Mr. Nadler. Thank you, Mr. Chair.
Mr. Irwin, according to reports, OpenAI, the company behind
ChatGPT, is valued at $29 billion. Stability AI, the company
behind Stable Diffusion, was valued at around $1 billion late
last year. For an area of comparison, can you give us an idea
of how much the average composer's salary would be?
Mr. Irwin. That's very difficult to say. Because the way
composers--first, we are not governed by any kind of collective
bargaining. We're not a union or anything like that. So, we're
very much independent contractors.
It's easier to talk about ranges for television shows. An
hour of television, a feature film, those sort of budgets are
easier to quantify, and then, of course, it comes into the
experience of whether you're an entry-level person or a very
experienced, high-level person. All of those change the rates.
Then, of course, on the back end, where we get our royalty
streams, which is where a lot of us make our money, the
performing royalties that are collected by the societies, like
ASCAP, BMI, and SESAC, they are dependent completely on the
number of performances of those works that include your music.
That's the same for audiovisual and for streaming, and
everything.
So, it's very hard to say. Some years, it's like--it's
almost like a farmer. Some years you have a bumper crop; the
next year, you might not see much at all.
Mr. Nadler. I see.
Mr. Sedlik, what about the average photographer? The same?
What about the average photographer? The average salary?
Mr. Sedlik. I don't have statistics on that, Representative
Nadler. However, many photographers make as little as $20,000-
$40,000 a year, and amounts greater than that outside of
certain spaces can be unusual.
Mr. Nadler. OK. Thank you.
Mr. Irwin, can you talk a little more about what threats
generative AI poses to composers and songwriters? What actions
do you think we should take to ensure that your work and
artistry is protected? Actually, you talked about the threats.
Can you talk about--
Mr. Irwin. Yes, I can talk--do you want to talk about the
threat. or do you want to talk about the solution?
Mr. Nadler. I think you've talked about the threats. What--
Mr. Irwin. Yes. Some of it's in my written testimony.
One of the things that the music industry has done,
particularly, well in this area is collective licensing. They
have, as I just mentioned, for ASCAP, BMI, SESAC, they have a
way to monitor performances. For mechanical royalties, which
are the physical royalties, you have SoundExchange. Then,
there's streaming royalties as well through the Music Licensing
Collective.
There are reciprocal organizations set up all over the
world.
Mr. Nadler. Well, how would we apply that to OpenAI?
Mr. Irwin. Well, music is probably more easily applied to
music than it possibly is to some of these other art forms.
Because every piece of music that's registered has a
registration number, a work number, and a recording number.
Every artist has their own number. I have what's called an IPI
number; Dan has one. So, they're already in the system. You
don't need to reinvent them. You just need to make sure they
are attributed to those works as they're being used, or logged,
or however you want to do it.
Mr. Nadler. How do you determine what works are used in
generative AI?
Mr. Irwin. Well, that's something that will need to be
determined. I mean, I write music. I'm not a technologist.
Mr. Nadler. OK. Maybe, Professor Callison-Burch,
transparency in AI training models is a concern, but the
transparency to the end user that the media they are viewing
was AI-created is critical as well.
I see that in your testimony you note that parts of it were
composed using ChatGPT. Should some sort of disclosure like
that be required? Currently, how are individuals alerted to the
fact that they're hearing or viewing an AI-created work, if at
all?
Mr. Callison-Burch. Thank you for the question.
So, disclosure of AI-generated works I think is valuable. I
think, especially with our potential for societal harms through
generating deepfakes or works that could be used to influence
elections by mocking up instances like Trump being arrested in
New York, Assad being generated by Midjourney. Certainly, a
disclosure--
Mr. Nadler. Or generating a fake speech by me?
Mr. Callison-Burch. I think that this is an important issue
that touches on output of generative AI systems, and that is
where I think that regulation is deserved.
Mr. Nadler. How are individuals alerted to the fact that
they're hearing or viewing an AI-created work, if at all?
Mr. Callison-Burch. I'm sorry, repeat the question, please?
Mr. Nadler. How are, currently, how are individuals alerted
to the fact they are hearing or viewing an AI-created work, if
at all?
Mr. Callison-Burch. There is a variety of technological
devices that our field is innovating to mark AI-generated
works, similar to a watermark on a stock photography site. This
is not an established industrywide practice, but it is
something that our field has been discussing. At the moment,
it's up to the user of the AI system who's generating it to
disclose to people who they're transmitting that image to that
it was generated.
Mr. Nadler. Thank you. My time is well expired, and I thank
the Chair for his indulgence.
Mr. Issa. Mr. Chair, you gave me the similar indulgence
when you sat in this Chair. So, I'm only returning the favor. I
thank you.
We now go to the gentleman from Virginia, Mr. Cline.
Mr. Cline. Thank you, Mr. Chair.
I want to followup on that line of questioning about
existing technologies that are there to identify digital works
and enhance copyright protection.
Mr. Sedlik, digital watermarks, tags, metadata, what
challenges are you aware of in the use of these types of
technologies for works that are used in training AI?
Mr. Sedlik. We have mature, very capable technologies to
identify visual works, including image recognition and what's
called steganography, which is burying signals in the image to
identify it.
Creators use embedded metadata to pass information into
their images, so that, as their images are distributed, their
images can be identified. A big problem for us is that all that
information is stripped out by the social media platforms and
other platforms when images are used. We would very much
benefit from a change, an improvement to the law to make it
illegal to remove embedded rights metadata, even if it's not
done for the purpose of infringement. Right now, under Section
1202, it's only illegal if it's done intentionally for the
purpose of inducing, enabling, concealing, or facilitating
infringement.
Mr. Cline. Mr. Irwin, what about music?
Mr. Irwin. Sorry?
Mr. Cline. Do you have any comments on that?
Mr. Irwin. Yes. There is watermarking and fingerprinting
used in music technology over and above the identifies that I
talked about. There's technology that allows, currently allows,
music to be identified even within a program that has dialog
and sound effects over the top of it. You can still identify
that music.
So, the technology is there for this to be done. It's just
a matter of having the will and sitting with the organizations
who are doing it and getting this discussion going. I have no
doubt that there's a way to track this stuff directly.
Mr. Cline. Professor, you've commented on that and talked
about the trending toward use of this technology by industry.
Is that something that's happening too slowly? Is there
something that needs involvement from government? What do you
think?
Mr. Callison-Burch. So, I think at the core of the problem
with an idea like a compulsory license--and Mr. Issa's
suggestion of what is the value of one over a trillion, when
you think about the volume of this work--and I can say that,
definitively, that the value of one over a trillion is going to
be vanishingly small.
So, another key that is a practical consideration here is,
unlike compulsory licenses, which are based on performance--so,
the MMA, where Spotify plays music, Taylor Swift gets more
money than some random person in the catalog because her songs
are played more--there's no equivalent here for generative AI.
It's hard, it's impossible to understand how much of a system's
output is due to Stephen King versus a random Reddit poster
who's written a paragraph in the collection.
So, I think that the lack of that performance is key to one
of the tricky things about establishing a compulsory license
here.
Mr. Cline. Mr. Irwin, some have suggested that imposing IP-
related obligations on the AI developers would hobble
development because of the inability to feasibly use
copyrighted content for training purposes. Would you be willing
to license your works for training purposes? Do you think other
creators would also be willing?
Mr. Irwin. If I was being compensated for them, absolutely.
Yes. I don't have a problem with the technology at all. I have
a problem with the stealing of the material.
Mr. Cline. Mr. Navarro, you indicated your willingness to
use AI and--
Mr. Navarro. A similar answer. The ability to approve or
disapprove of a particular use is why I oppose compulsory
licenses. I don't oppose blanket licenses, which might make--
streamline a process. I still have the ability to say yes or no
and opt out.
Mr. Cline. Now, in your testimony, you advocated against
granting AI any special IP exemptions. Depending on how some
lawsuits turn out, courts may decide that AI doesn't need
special exemptions in training AI, that copyrighted music is
not infringement under current law. In that event, would you
advocate for changing the law to make training AI with
copyrighted works a type of infringement?
Mr. Navarro. I'm not sure that I am qualified to answer
that. It's a very technical question. I do believe that, as
we're looking at guardrails, guardrails used to be made of
wood. Then, they started becoming made of steel. Now, as cars
get faster and more powerful, maybe they need to be made of
titanium. As these technologies progress, and as its scope
increases, our guardrails need to be modified and improved.
Mr. Cline. Thank you. I yield back.
Mr. Issa. The guardrails of the future will be software-
driven.
With that, we go to the gentleman from California, my
colleague, Mr. Lieu.
Mr. Lieu. Thank you, Chair Issa and Ranking Member Johnson,
for holding this important hearing.
As a recovering computer science major, I am enthralled
with AI, and I believe it has, and will continue to,
revolutionalize society. It can also cause us harm and it
creates all sorts of unanswered questions.
So, I'd like to walk through some examples, so that the
American public and I can better understand how artificial
intelligence interacts with copyright.
I'd like to ask Mr. Damle this example. Let's say I create
a generative AI Internet application related to music. I do it
for commercial purposes. To train it, I scrape the entire
Internet of all songs, including every one of Taylor Swift's
copyrighted songs without her permission. Your view is I
wouldn't have to compensate her in any way, is that right?
Mr. Damle. So, I think it would very much depend on the
particular way, in which, you trained the AI models. Not all AI
models are constructed the same way. Some are constructed in
ways that might very well exceed the bounds of fair use.
So, it's inevitably going to be a fact-by-fact--a case-by-
case determination of whether a particular model is going to
exceed those bounds or not.
Mr. Lieu. Let's just use ChatGPT's model.
Mr. Damle. So, in an instance where what you have done is
extract unprotectable facts from any copyrighted work--so,
stepping aside and generalizing this point to any kind of
work--if what you've done is extract unprotectable facts from
those works, and then, used those facts to generate a new work,
then, under well-established principles of copyright law, that
is not infringement.
Mr. Lieu. I'm not talking about generating any, just about
training the model.
Mr. Damle. Just on the training side, if that's all that
happens, then, under a long line of cases that I've laid out in
my written testimony, that is fair use--
Mr. Lieu. To train a model, you need to actually download
the Taylor Swift songs?
Mr. Damle. That's correct. That's correct.
Mr. Lieu. You view that as fair use?
Mr. Damle. That would be fair use. The premise of any fair
use case is going to be--
Mr. Lieu. So, Internet applications, like YouTube, pay a
licensing fee to Taylor Swift when they download her songs.
What is the difference?
Mr. Damle. Well, the difference would be in those
instances, what they're doing is they're taking the work, and
then, they're taking that work and they're streaming it to end
users. So, that's a--that's a public performance of her work.
Mr. Lieu. So, let's say ChatGPT lets you just put out
Taylor Swift lyrics. What's the difference?
Mr. Damle. I think that might be an issue that exceeds the
bounds of fair use, where you're taking--
Mr. Lieu. Now, let's say I take my model and I generate a
new song similar to a Taylor Swift song in terms of lyrics. Is
that a copyright infringement?
Mr. Damle. That's correct, Congressman.
Mr. Lieu. It would be infringement?
Mr. Damle. It would not be infringement.
Mr. Lieu. It would not be infringement.
Mr. Damle. Because one of the basic principles of the
Copyright Act, which I discussed earlier, is replicating
somebody's style, writing a song--if I were to write a song in
the style of Taylor Swift, I would not be committing copyright
infringement. That's one of the founding precepts of copyright.
Mr. Lieu. Now, as you know, voice cannot be copyrighted.
Let's say my model also generates voice very similar to Taylor
Swift, and I had this brand-new song similar to Taylor Swift's
voice, similar to her lyrics. You believe that would not be
copyright infringement?
Mr. Damle. You would have to look at other bodies of law to
determine whether that would be illegal. As a matter of
copyright law, and just the basic principles of copyright law,
that would not be copyright infringement, which is not to say
it's not concerning for other reasons. In just looking at the
copyright law itself, that's not--
Mr. Lieu. Then, finally, in fact, you believe I could,
then, copyright this Taylor Swift-like song with voice like
Taylor Swift that I generated by scraping the Internet with
Taylor Swift copyrighted songs I didn't pay her for, right?
Mr. Damle. I don't know that would necessarily be the case.
Certainly, as we were discussing earlier, the Copyright Office
has taken the view that AI-generated works like that may not be
subject to copyright protection at all.
Mr. Lieu. OK. Thank you.
So, my remaining comment and question for Professor
Callison-Burch about disclosure--and Congressman Nadler asked
about this--many creators already use AI, right, in their
creative works? I mean, there's a whole bunch of algorithms
that make your song sound better. They don't disclose that,
right? How would you even define what kinds of AI they need to
disclose that help them with their particular creative work?
Mr. Callison-Burch. That's a great question. If you'll
indulge me, I want to answer your question to Sy as well.
So, your question about if I download Taylor Swift's songs
and I learn from it, it could be equally well posed to a
teenaged pop star who's learning how to sing. So, if that
person learned from Taylor Swift, the decision of whether or
not they're violating copyright is not at the time when they're
listening to the songs and learning to perform music. It's when
they release an album and whether that album is sufficiently
similar to, say, Taylor Swift's songs.
So, if we release, instead of Taylor's version, we release
ChatGPT's version of an album, that's infringement, but the
learning from it is not.
In terms of disclosure of use of copyrighted, of generative
AI in materials, I think there's an interesting ill-advised
guidance from the Copyright Office that works that involve
substantive use of generative AI are not copyrightable at the
moment. I believe that AI is going to be used in a
collaborative way with humans, and the human using it deserves
that copyright. Whether or not they acknowledge it, I think
depends on the use of it.
So, there was a law passed that political ads must disclose
if they're using generative AI to create images of politicians.
That seems like a very valid case to disclose. If I'm creating
a comic book, it seems less high stakes, so probably not
necessary.
Mr. Lieu. Thank you. I yield back.
Mr. Issa. I thank the gentleman.
We now go to the gentleman from Texas, Mr. Gooden.
Mr. Gooden. Thank you, Mr. Chair.
That was very interesting. I'd like to maybe keep going
along those lines.
It seems as if, as this develops, we'd like to perhaps see
more transparency in the process. Is it unrealistic to think
that the songwriters, the people, the American people, whoever,
could kind of see what goes into these sources of what has an
influence on the AI. Mr. Callison-Burch, I'll give it to you. I
don't think the average person, myself included, understands
the technology. Is it unrealistic to ask that we know if a
particular song got more influence from Taylor Swift, or
whoever, and how that process comes about?
Mr. Callison-Burch. So, I think this is a super-interesting
question that, again, involves the output of a generative AI
system, rather than the training per se. I think that there is
a valid case to be made that copyright should be reshaped to
protect against a case where I, as a user of an AI system, ask
it to generate something that mimics a particular artist.
Like this concern that artists have, I think is 100 percent
valid, that you can currently say, ``Generate a comic strip in
the style of Sarah Anderson,'' and it produces something
similar in style, but does not reproduce any of her published
works. That, to me, seems more like a right of publicity style
concern than current copyright law is addressing.
I do think it's an ethical issue that we should consider as
an industry, and I think that there should be an opt-out
mechanism for artists to explicitly exclude their work from the
vast amount of training data that we have. Again, I want to
make this distinction clear. What AI systems are learning from
their training data is more akin to facts and patterns and
statistical correlations than it is memorization or directly
lifting from copyrighted works.
Mr. Gooden. Mr. Damle, please forgive me if I've
mispronounced your last name.
Do you believe that there will be more of a push to
actually copyright some of these AI works? I feel like that
will be controversial and a tough sell for those on the other
side of your table. Could you explain the thought process
behind that?
Mr. Damle. I think there's going to a lot of hard questions
that get raised. Just as there are many ways in which to train
an AI model, there are many, many ways in which to use an AI
model in the creative process. I think we're really at the
very, very early stages of trying to figure out where you draw
the line between an AI-generated output that we don't think
deserves copyright protection under sort of the constitutional
precept of what copyright law is about and what creative output
that is assisted with in AI deserves copyright protection, and
it's necessarily going to be fact by fact.
I agree with the Copyright Office's view that the principle
here is that you need sufficient human authorship, sufficient
human input into the creative process, to warrant copyright
protection. That's almost a constitutional requirement.
Exactly how that plays out in any given case is going to
require over time, over the next few years and more, looking at
every case that comes and trying to decide, OK, is this on one
side of the line or the other?
Mr. Gooden. OK. Thank you.
Maybe it's a far-fetched analogy, but many years ago, when
I was in the statehouse, we passed a bill--I authored it--that
restricted the use of drones over people's backyards. The drone
industry was against that, and they said,
You should be able to fly a drone over anyone's house and park
it and watch them all day. A helicopter can do that. So, what's
the difference?
I said,
Well, a helicopter is operated by an individual. It can't stay
up there forever, and you know if it's there because it's loud
and big.
So, I thought of that as you were talking about your
example of the middle school child who sings like Taylor Swift,
compared to the AI, it sounds like Taylor Swift, and how
there's no difference. The difference is that's a middle school
child, and I'm not real worried about a middle school child
taking over the music industry. So, I think we have to
differentiate between the two.
I yield back.
Mr. Issa. Would the gentleman yield?
Mr. Gooden. This gentlemen?
Mr. Issa. Yes, would you yield for a moment, please?
Mr. Gooden. Oh, of course. Please.
Mr. Issa. Thank you.
At some future time, I'd like you to talk about--and
particularly, Mr. Navarro--about how you view the difference
between a cover band and AI and where the guidelines are
similar and when they would be different, and the same,
obviously, from a legal standpoint. I won't ask to have it
answered at this time, where I'm out of time.
Mr. Issa. I thank the gentleman for yielding.
We now go to the gentlelady from North Carolina, Ms. Ross.
Ms. Ross. Thank you, Mr. Chair.
Thank you to all the witnesses for being here today.
We've heard multiple perspectives today on whether training
AI on unlicensed copyrighted material should be permitted
through the fair use doctrine. The Copyright Office considers
several factors in evaluating this question of fair use.
One of those factors is the amount and how substantial the
copyrighted work was that was used. The guidance notes that in
some context using an even small amount of copyrighted work was
determined to not be fair because the selection was an
important part, or actually, the heart of the work. So, it
could be just very small, but have such an imprint.
That strikes me as a key factor in the consideration of AI
training as well. AI is built upon creative works, whether it's
art or music or writing. AI would not exist at all if it
couldn't learn from human beings. So, the work that AI learns
from seems to me to constitute the heart of AI itself.
Now, the Copyright Office also considers whether the
unlicensed use of copyrighted work would harm the existing or
future market for the work, and we've heard about that. We've
heard from the creators today about how AI is already doing
that.
Mr. Altman has been with us this week, and I'm quoting him
when he says,
When we're working on new models where, if an AI system is
using your content, or if it's using your style, you get paid
for that.
I hope that he's going to follow through on that because,
as the Chair--or Ranking Member Nadler told us, Mr. Altman has
a lot of money to pay you.
My first question is for Professor Callison-Burch. Is it a
common industry practice to keep a careful record of how and
whether copyrighted works, performances, and likenesses were
used to develop or train an AI dataset?
Mr. Callison-Burch. Thank you, Congresswoman.
Yes, it is. In fact, there was a Washington Post
interactive feature published about two weeks ago sort of
provocatively named ``the secret lists of websites that make
ChatGPT so smart.'' That was produced in collaboration with
researchers at AI2, where I'm currently taking my sabbatical,
to exactly search for which websites were included in the
training data of a very common training set, not necessarily
the one that ChatGPT uses, but the one that many people in our
industry do use.
Ms. Ross. Just as a followup, do you think that would make
it easy to devise a compensation system, since we have that
trail of what's been used?
Mr. Callison-Burch. That's a great question. So, I think
the tricky part about creating a compensation scheme is, once
again, there are a trillion words' worth of text in our
training datasets. Each author represents a vanishingly small
portion of that. We do not re-perform any of the songs or texts
that are in our corpus. It's not a performance-based
compensation scheme that would be possible. So, instead, it
would have to be something to do with the volume of each
person's contribution to that work. Again, I think if you do
the math, it will end up being everyone gets a check for two
cents, which doesn't make sense.
Ms. Ross. We have similar things in the music industry,
though. I want to get on to my--
Mr. Callison-Burch. Those are orders of magnitude smaller.
Ms. Ross. Yes, to my next question. This is for all three
of our artists and creators.
Some AI developers have suggested that, if using
copyrighted works is not deemed to be fair use, it would stop
the development of AI. I imagine that most copyright owners
would be willing to license their works to people who they
would consent to license their works to, given the opportunity.
So, if a company wanted to license your work for AI, would
you be willing to grant permission if you could agree on
reasonable compensation? Go in whatever order you would like to
go.
Mr. Irwin. Absolutely, yes, of course. We're professional
people. We get paid to do what we do. So, why wouldn't we want
to license our work? The more work we get out there, the better
it is for--it makes more incentive for everyone to create it,
if there's compensation to create it.
Ms. Ross. OK. Mr. Navarro and Mr.--
Mr. Navarro. The ``if'' is the big part of it. If we can
agree, absolutely. If we can't agree, then I retain the right
to withhold it.
Mr. Sedlik. I agree, copyright protection is automatic.
This can't be an opt-out and shouldn't require an opt-in. We
are protected by copyright, and most visual artists are willing
to license their works for various usages. I, personally, have
offered my works for AI training licensing for years, and stock
agencies increasingly do sell licenses for AI training.
Ms. Ross. Thank you, Mr. Chair. I yield back.
Mr. Issa. I thank the gentlelady.
We now go to my colleague from California, Mr. Kiley.
Mr. Kiley. Thank you, Mr. Chair, and thank you for putting
together today's hearing, which is a very important and timely
topic.
I have a very open mind on these matters. I'm extremely
sympathetic to the concerns raised by the artists and the
predicament that they're now in. At the same time, I understand
the practical difficulties with some of the ideas that are
being proposed, the potential impediments they might pose to
innovation.
I'm also wondering whether copyright law is well-suited to
the matter at hand in a lot of respects, as clearly the
Copyright Clause and the body of law around it did not
anticipate sort of a capacity for creation that is non-human in
origin.
So, just to sort of start out, I want to try to analogize
this to the process of human creation as best as possible. So,
Mr. Callison-Burch has drawn a distinction with the training
process and the output. You mentioned earlier how a child who's
learning how to sing by studying Taylor Swift, there's no
infringement there. It's the output that is going to be judged.
All of the artists here, I'm sure, could cite people, artists
who have been influences on their style, as they developed.
So, let's take another analogy. Let's say that I'm writing
an essay on Abraham Lincoln and arguing how he ranks among U.S.
Presidents. I go to the library. I check out every book I can
possibly get on Abraham Lincoln. I learn all about him, and
then, I reach an opinion and I write an essay about it. Of
course, I'll cite to sources when I'm quoting them directly or
for specific facts, but that learning process by which I formed
my opinion, I'll--this is for you, Mr. Damle. Is there any
copyright claim that is specific to that process?
Mr. Damle. I think that if, to just modify your
hypothetical slightly, if you were to go to a library and make
photocopies of books, right, that's a copy that you're making
of those books. That would be copyright infringement if you
didn't have a fair use defense. In your hypothetical, that
would be very clearly within the scope of copyright
infringement. You're making those copies not to--for the only
purpose for which to learn the facts that are being conveyed in
those copyrighted materials.
Then, yes, you're producing an output that might borrow
those facts into a new work, but that work, as long as that
work does not infringe the original work that you copied from,
that's quintessential fair use.
Mr. Kiley. So, a court that would sort of study any
copyright claim, would they look at my learning process or
would they just sort of judge the output? Maybe they would if
it's, like, about intentionality or damages claim or something.
The core copyright claim, infringement claim itself, would they
study my learning process, or would they just look at my work
product?
Mr. Damle. I think it would be all the above. These
copyright cases, having litigated a lot of these, really delve
into a lot of the facts. So, the Google Books case, for
example, looked at the way the manner, in which Google had
acquired the copies, but, then, the outcome was driven mainly
by the fact that all that Google was doing was providing a way
for you to search within those books, and not providing you the
whole book. It was providing just snippets of the book, and
said, looking at that whole process, we're going to call the
copying and retention of the copies for that end purpose to be
fair use.
Mr. Kiley. Interesting. So, one of the things that Mr.
Altman said is that he thinks we're moving toward a stage
where, really, most of the training is based upon synthetic
data that's generated by the AI itself. Maybe an analogy to
this is sort of when DeepMind was training its goal-playing AI.
At first, it studied the masters' games, then it, eventually
just learned how to succeed by playing itself. So, does that
impact the analysis?
This is for either of you, Mr. Damle or Mr. Callison-Burch.
Do you have thoughts on that?
Mr. Callison-Burch. This is a very interesting question.
So, I've done a thought experiment regarding training AI
systems on completely synthetic data. So, after having been a
panelist on the U.S. Copyright Office listening tour, I thought
about what about our training data do we need to learn the
facts about the world that we care about, and learn the
language patterns?
So, we need examples of language to learn the structure and
grammar of language. We need some anchor into how people
discuss the world to learn facts about it. There's no
obligation that this be human-written text now.
I think that, in my little armchair experiment, we may have
reached a copyright escape velocity, where, in theory, you
could have a system like ChatGPT generate a trillion words'
worth of text, which current copyright would be not
copyrighted, and then, retrain a subsequent system, where you
throw away the original one entirely. It's not derived from
copyrighted works at all, but it still is likely just as
performant as the original.
Mr. Kiley. My time is up. So, I'll just say, in closing I
kind of worry that many of the issues we're talking about here
might sort of be obsolete within a matter of years, especially
as the capabilities of these systems advance.
I do think that there will always be a desire for works of
purely human creation. I don't know. This is a bad analogy,
but, we have at the grocery organic and the nonorganic
sections. So, I think that this is an argument for making sure
that we have transparency throughout the process, as to when
you're dealing with something that was created by an AI versus
a human. That may be well into the future we could have
different markets, but I do believe there will always be a
desire for the works that are produced by, the kind of works
that are produced by the artists that are here today.
Thank you very much.
Mr. Issa. I thank the gentleman.
I might note that, after we all saw the movie ``WarGames,''
and found out that the outcome of tic-tac-toe, ultimately, is
nobody wins, we still have generals, colonels, captains,
soldiers, and some of them fighting in Ukraine as we speak.
With that, I go to the former U.S. Attorney and my
colleague from California, Mr. Schiff.
Mr. Schiff. Thank you, Mr. Chair.
I should correct the record. I was an Assistant U.S.
Attorney, so as not to think I was promoted more than I was.
Thank you.
Mr. Issa. Humility is also one of the traits I admire in
you.
Mr. Schiff. Thank you, Chair.
The explosion of everyday AI use practically overnight has
caused, caught the attention of many in the IP space, and for
good reason. As many of you know, my constituents in Hollywood
in the entertainment industry, and more broadly, in California,
are particularly impacted by the rise of AI. We eagerly look to
see how it will affect the creative industries.
A recent analysis of ChatGPT's training sources by The
Washington Post found that 11 percent of the models' input data
comes from arts and entertainment, including movies and
television, art and design, and entertainment events.
Mr. Navarro, I want to echo a remark that you submitted in
your testimony today. ``Technology has long empowered human
expression and AI will be no different.'' It's true that new
technologies have the potential to complement or augment art
and have certainly done so in the past. AI can be used to
enrich the work of those in the creative industries.
Behind all that work are human artists, people. The
copyright system is the foundation for the entire economic
marketplace that allows American artists and creatives to earn
a living and American companies to create jobs for the sake of
producing art.
Copyright law must continue to incentivize and protect this
activity in the United States and beyond. Blanket AI exemptions
do not exist in current law, nor should they, for copyright
infringement.
I wanted to ask two questions:
(1) Mr. Navarro, and that is, under what circumstances
would you be interested, or do you think other artists would be
interested, in licensing their work for AI? Or do you think
there's a broad enough concern about moving away from human-
generated music that you think artists shouldn't sell their
works to be used for AI? Is one question.
Then, another question for the broader panel, and that is,
disinformation is a grave concern with AI. It may very well
affect people in the political world, but it also will affect
people in the arts.
(2) How do you see disinformation about artists or about
their works being a danger, and how do you think that can be
addressed?
Mr. Navarro. Certainly, the area of disinformation is
underscored by the rise in the last few years of the deepfake
phenomenon, which I was at your talk at SAG-AFTRA. I'm on the
National Board of SAG-AFTRA.
This is a tremendous issue in terms of right of publicity,
but, even just putting faces on bodies that are doing things
those bodies wouldn't do--I don't think there's a single answer
to whether it should or should not be allowed. I mean, what
might be right for somebody to have their face put on a body
that, suddenly, is a war hero might be OK. To have a face put
on a body that's doing something pornographic would not be OK.
I don't mean to be indelicate in that, but these are some of
the concerns.
With regards to music, it's probably a little different. I
think that every individual has their own line of demarcation
as to what should or should not be done with their work.
Certainly, when you're dealing with the physical countenance of
somebody, that the standards and the strictures are going to be
greater.
Mr. Schiff. Thank you.
We held an open hearing in the Intelligence Committee years
ago on deekfakes, and among the most chilling observations was
that, once you see a deepfake, even if you're later persuaded
that what you saw was a fake, you never completely lose the
lingering negative impression it left with you. So, the damage
is done when you see it.
Would others like to comment on the danger of
disinformation to artists?
Mr. Sedlik. There's a very important initiative called the
C2PA. It's the Content Authenticity Initiative, and it's led by
Adobe and others. They've determined a way to track any changes
made to images, whether by AI or otherwise.
So, that, especially when it comes to photojournalistic
images, people could be confident in the providence of the
images that they're looking at. That's C2PA.
Mr. Schiff. Thank you.
Mr. Irwin. I was just going to say that, certainly, in my
own experience--and I'm sure for the other people here--you put
in a lot of hours from a very young age to try and do what you
do and become really good at it. To dismiss that in any way, I
think that's the problem that a lot of us are having, it was
Malcolm Gladwell who wrote, ``in 10,000 hours, you start to
know what you're doing.''
If the machines can do it that quickly, what is the
incentive for us to keep going? What really is it? For years
and years and years, for as long as you can remember, going
back to the court composers, the arts people have been traded
on. Oh, they'll do it because they love doing it. They love
doing it. It's true, we do love doing it.
At some point, the love doesn't feed your family, and
that's the real harm here, is there has to be a way for us to
coexist. That's all we're looking for really.
Mr. Callison-Burch. Mr. Chair, may I add a comment? Because
I'd like to highlight agreements on this issue.
Mr. Issa. Uh-hum, quickly.
Mr. Callison-Burch. I am in absolute agreement with artists
on the panel here that this is an important issue and people
should not be allowed to imitate another person through
deepfakes or through imitation. I think the right of publicity
is something that's very worth considering, as you consider
legislation on this.
Mr. Issa. I thank the gentleman.
We now go to the gentleman from South Carolina, Mr. Fry.
Mr. Fry. Thank you, Mr. Chair. Thanks for having this
hearing today.
This is a really interesting topic, right? We look at this.
I have not ever in my practice done IP work. AI is the next
wave of--or the wave of the future, if you will, what we're
going to be dealing with in this country, really, on all
spectrums.
Mr. Burch, I got a kick out of reading your bio, where your
Ph.D. students joke that, whenever they ask you anything, your
first response is, ``Did you ask ChatGPT for that?'' I think
that really kind of sums up why we're here, right?
Mr. Damle, I want to ask you something. I've been bouncing
around in committee hearings all day, so I may have missed
this. Can you identify inadequacies of our existing laws to
address copyright protection of AI?
Mr. Damle. I think that there's a lot of questions that
still need to be answered about whether AI output is
protectable by copyright and the circumstances in which it is.
This is such a new issue. The Copyright Office has put out its
guidance, but that guidance really addressed kind of one end of
the spectrum of the question, which is, where there is
virtually no or minimal human input into the creative process,
does the output qualify for copyright protection? The office
says no. That's just one end of the spectrum.
There's a whole area from there to somebody sitting with a
paint brush and paint painting on a canvas, where you're using
technology to assist you in the creative process, whether
that's autofocus on a camera, whether that's Photoshop, or,
indeed, whether that's generative AI.
I think generative AI, properly understood, is going to be
a tool for human creativity. If you talk to artists that really
have incorporated generative AI into their creative process,
they don't see it as a substitute for their own creativity.
They see it as a way to enhance their own creativity.
So, we're going to have a lot of hard questions in that
space, where there is really an iterative process between the
human author and the generative AI system to determine whether
the output of that process is copyrightable.
Mr. Fry. Do you think we need a completely new set of
rules? Does existing contract law maybe cover this? Or is it
possible for this to fit in with existing law?
Mr. Damle. I think existing law is well-suited to deal with
all the questions we've been talking about today. Congress had
the wisdom in the 1976 act, and various amendments since then,
to build a technology-neutral, flexible copyright regime. I
think it's proven time and time again that, no matter the new
technology that comes along, the laws are able to adapt to
them.
There are instances where perhaps they're not, and Congress
can step in and act in those circumstances. In general--and I
think in this space--I think our existing copyright laws are
well-suited to handle the questions that--as they arise in
this.
Mr. Fry. When I was reading the guidance for this, and
reading the CRS report that was issued kind of surrounding
this, what struck me was it reminded me of a test in law on
whether somebody was an independent contractor or an employee
of a company, right? So, the test that they look at is the
degree of control in which somebody exercises over that
individual on whether they're an employee. You can call them an
independent contractor all you want to, but if you're the one
doing the schedule, you're the one putting in all these
requirements in what they do, they're not actually an
independent contract; they're actually an employee at that
point.
So, to me, there are some similarities there that, when
we're talking about the degree in which there's litigation on,
hey, create a song that sounds like this, and there's no input,
but where do we go? Should we be looking at it in those terms,
as the degree in which we provide input to AI to the generated
product as a test?
Mr. Damle. Thank you, Congressman.
I think that's a really useful analogy. In fact, there is
part of copyright law, called the work for hire doctrine, that
really asks that exact question--the extent to which you
exercise, as the employer, control over the creative output of
your employee, or to the extent they are able to do things on
their own. That determines whether it's a work made for hire or
not.
I think that we are people that are copyright lawyers are
looking at that body of law already as a way to draw analogies
in this space--considering the AI as an employee, or is it an
independent contractor? Is it off running on its own or does
the human author actually have some control over what the AI is
outputting? So, I think it's a very, very useful analogy.
You've sort of anticipated where, where copyright lawyers
already are on this.
Mr. Damle. Thank you, Mr. Chair. With that, I yield back.
Mr. Issa. I thank the gentleman.
We now go to the gentlelady from California, Ms. Lofgren.
Ms. Lofgren. Thank you, Mr. Chair.
This has been a very useful hearing, and I appreciate the
testimony of each of these witnesses.
I believe that AI is going to upend a lot of careers. In
fact, the House of Representatives has had three bipartisan,
Congress-wide briefings on AI in the last two weeks. It's going
to upend the practice of law. It's going to upend the practice
of medicine. It will probably upend engineering.
The difference for creators, at least you have some
protection in the law, which is copyright, which is absent
other professions. The question is, how will that work to
protect creators?
I was glad that Mr. Altman in his testimony indicated that
creators should be compensated and do have rights and that
recognition. The technology is complicated, and how that is
going to work, we don't know yet.
I'll disclose that, a number of months ago, I put together
some creators with the AI people to see if we couldn't have
discussions. I was thinking, honestly, about the Music
Modernization Act, which I think was very successful in
reaching negotiations, so that creators could be compensated.
I'll express some disappointment that progress that I expected
to have been made by now has not yet been made. So, I'm hoping
that those discussions will re-energize.
Here's a question, I guess, for Mr. Burch, Professor Burch,
and maybe Mr. Sedlik, since you represent different ends of the
knowledge base here. We've got, basically, two questions.
You've got the input question, which is, basically, lines of
code that have been assembled, and then, you've got the output
question, which is how much of this is infringing.
Is it, in your judgment, even possible to reach an
agreement like the Music Modernization Act did, so that
creators can be fairly compensated?
Mr. Callison-Burch. I think that there are a number of
practical issues that make this very difficult to imagine, not
least of which is the fact that we're not performing anything
when we're outputting something from a system. It's not simply
a collage where we're combining elements of existing work. It's
genuinely been distilled into a form that's completely
different than the original. That is going to be the crux of
what makes this difficult.
I think that there might be a market for licensing images
and songs, and things like that, that companies end up
voluntarily entering into, but I don't think that the practical
implementation of such a thing will be as easy as it was for
the MMA, which I understand was already very complex.
Mr. Sedlik. I don't think that most photographers are
concerned or whining about the fact that AI is going to
affect--it's going to compete with them in terms of it's easier
to create images with AI than it is to create a photograph. It
is.
What they're concerned with is the use of their works in
that system, grinding it up and spitting it out as generated
AI, based on their works, and the fact that they're not
compensated, and it's done without their authorization.
Images are data, but they're not merely unprotected facts.
An image, my image is my depiction of a fact and applying my
creative expression to depict it. So, the copying of that,
under 106, irrespective of anything else, is an infringement.
It's copied into a system. That's infringement right there.
In terms of the output, that output may or may not resemble
my work. It may or may not infringe on my work. The input,
copying the work under 106 would be an infringement, unless,
fact-specifically, it's fair use.
Ms. Lofgren. I thank you for that. As I was listening. I
was thinking back to my prior service as a staffer to my
predecessor in office, Congressman Don Edwards, and his
partnership with Bob Kastenmeier in the 1976 Act, and how the
work that they produced has endured to this day--to protect the
creative forces.
Obviously, we need to meld technology to help that
protection. Adobe, which is located in my district, actually
did the water-
marking or their tagging. We actually used that in the January
6th Committee to prevent our material from being altered.
So, I think there's some real possibilities here. I think
we've got a long road ahead, but I'm hoping that the
discussions that are ongoing can ramp up a little bit, because
I think that's probably the most productive way to reach a
successful conclusion.
With that, Mr. Chair, I'd yield back.
Mr. Issa. I thank the gentlelady.
We now go to the gentleman from Texas, Mr. Moran.
Mr. Moran. Thank you, Mr. Chair.
I know that all of you have been sitting there for quite
some time. So, before I ask my questions, I'd like to pause for
a quick musical interlude.
[Music plays.]
I know that was just a short interlude, and all of you
would, no doubt, like to hear the rest of that song.
[Laughter.]
As you heard, it is a rendition of something by Drake and
the Weeknd. Many of you, I would suspect--maybe all of you--
believe that this is likely their true voices and, in fact, a
musical, a musical song written and produced by both the Drake
and the Weeknd. In fact, it was computer-generated using only
snippets of those artists' original voices.
Even the most ardent fans of Drake and the Weeknd--and I'll
admit, they're not my first choice; I'm more of a TobyMac and
Steven Curtis Chapman guy--but even the most ardent fans didn't
realize that this was not their voices. In fact, in just a few
short days, that song garnered over a half of million streams
on Spotify before it was, ultimately, taken down.
I'll also admit that I am profoundly blown away by the
artistic values sitting at this table, but then, also concerned
about what's going to happen to the creativity if we don't get
a hold of this artificial intelligence issue and protect the
creativity of human development.
Just like what Drake and the Weeknd need protection for,
all human creativity needs protection from what we're seeing
from the artificial intelligence community here. I'm curious if
any of you knew that story about that song. Were you all aware
that this had happened? Did you guys hear that song before you
knew that it was not actually Drake and the Weeknd? Yes, it's
amazing to me that it sounds exactly like them.
Mr. Navarro, I want to ask you a couple of questions. Mr.
Irwin mentioned three primary issues to focus on--consent,
credit, and competition--compensation. I appreciated the fact
that you mentioned those. Mr. Navarro, as a generational singer
and songwriter, why is not enough just to give credit to an
artist if AI uses your voice or prior works as a basis to
create something new? Why is credit--
Mr. Navarro. I am very familiar with this particular case.
Royalties were generated. I serve on the Unclaimed Royalties
Oversight Committee of the Mechanical Licensing Collective
under the Music Modernization Act and the U.S. Copyright
Office. Where do those royalties go? Who do they go to? They
didn't authorize the use of their voices. AI could be used to
put abhorrent words into their mouths. They didn't get their
permission. They have contractual relationships with Universal
Music. This operates in violation of it.
Many, many entities are harmed by using this without going
through whatever the proper channels are. No means no. If they
had said no, no means no.
Mr. Moran. Yes, because you mentioned earlier, even if you
have credit and compensation, without the consent it's really
nothing more than compulsory licensing. Is that true?
Mr. Navarro. Yes.
Mr. Moran. Do you consider, Mr. Navarro, derivative AI
works from your original works that sound like you to be works
in competition with you?
Mr. Navarro. I think they are works superseding me.
Mr. Moran. Mr. Damle, I want to ask you a couple questions,
because I wrote down some quotes that you had throughout the
testimony today. I couldn't follow some of the consistency in
them. I want to give you, read you back some of your quotes.
You said at one point, ``existing law is well suited for
everything we are dealing with today.'' Then at another point,
you said, ``people should not be allowed to imitate other
works.'' Then you answered the question, you said earlier, ``we
need to look at other areas of the law other than copyright,
for examples, like the Drake and the Weekend,'' the example
that I just gave you, for protection in examples like that.
So, are we perfectly suited under existing law, or do we
need additional laws to protect artists like the ones sitting
at the table?
Mr. Damle. Thank you, Congressman. Just to clarify, I am
not sure all those quotes were mine. I think only a couple
were.
Mr. Moran. All those quotes were--
Mr. Damle. So, the question of whether copyright law needs
to be changed, I think the answer to that is no. I think that
our copyright law is--I am a copyright lawyer, so I sort of
focus my testimony on copyright law. I think copyright law is
well suited, flexible enough to deal with the copyright
questions that are being raised.
Now, I acknowledge that there are concerns like with this,
with the Drake and Weekend track that are legitimate and need
to be considered. I think copyright law is really a blunt
instrument, too blunt an instrument to deal with that.
So, it may be worth looking at other areas of law outside
of copyright. Professor Callison-Burch mentioned right of
publicity. There is trademark law. There are other areas of law
that I am not expert in that may be better suited to deal with
these situations than copyright law.
Mr. Moran. OK. You may be right. Professor Burch may have
said people should not be allowed to imitate other works. He
may be the one I need to attribute that quote to. So, I stand
corrected, not my first time in Congress, won't be my last
time.
Thank you, gentlemen, for your time today. We appreciate
it. It is a complicated issue. We want to get this right. Thank
you.
Mr. Issa. I thank the gentleman.
We now go to the gentlelady from Pennsylvania, Ms. Dean.
Ms. Dean. I thank you, Chair Issa, Ranking Member Johnson,
for holding this important and very interesting hearing. It is
also very pleasant to be in the company of this talent and to
have us really getting along very, very well because of the
subject matter. So, thank you, thank you. It has been a real
pleasure for me.
I know this won't be the last of our hearings. It is really
among the first. There is so much to learn. There is so much
for me to learn. We know as lawmakers we can't wait after
widespread societal use to come back and say what did we do
right, what should we have done in advance. We have learned
that from other technologies.
I want to lay a couple of my biases on the table. I taught
for 10 years at a different Philadelphia university. I taught
writing, all different levels of writing, to students at
LaSalle University. To very much what most of you have said,
Mr. Navarro, in particular, I always taught my students know
your craft, understand the mechanics, break the rules when it
makes sense, but make sure you place your humanity in whatever
you write. That is the genius of what you create, placing
humanity in it.
My other bias is I am a copyright holder myself. My son,
Harry Cunnane, and I wrote and published a memoir of his
struggles with addiction, our family struggles with his disease
of addiction, but much more importantly the power and the hope
in recovery. We also wrote a children's version of that book.
It was made into an audio book with the extraordinary
generosity of Mr. Paul Williams writing and producing and
performing the background music for our children's book.
My son, Pat Cunnane, is a television writer and movie
screen writer. So, when some of you talk about how long some of
this stuff comes, takes to come, I keep saying to Pat, where is
the movie, it is years. Maybe we will see his name. He has done
really well. I bring those biases to the fore.
When I consider what we have seen of generative AI, two
questions seem glaringly obvious, so if you will help me with
these. The first is the creation of these models being done in
a way that respects the rights and interests of authors,
musicians, artists, content creators, to your very point,
consent, credit, and compensation? If I frame it a different
way, is anybody doing this right?
Maybe I will start here with Mr. Sedlik. Anybody doing it
right?
Mr. Sedlik. I think that I am seeing a silver lining on the
cloud. Some of that AI platforms are beginning to listen and to
adapt their systems to respect rights of authors, or at least
they are saying that this is in progress.
Ms. Dean. OK.
Mr. Sedlik. There is really two gateways here. There is the
gateway to allow works into the data bases that are used for
ingestion of images into these systems. Then there is also a
gateway in terms of the prompts that are entered and uploading
of images as image prompts, copies of our works uploaded by
others without our permission as image prompts. Those two
gateways are of concern.
Ms. Dean. Mr. Navarro and Mr. Irwin, could you offer me
your thoughts? Anybody doing this right?
Mr. Navarro. I am not aware of who is doing it right. I
don't believe that everyone is doing it wrong. I think it is so
brand new, as it is the Wild West out there.
I think of the earliest days of sampling, when people
sampled stuff for records routinely and didn't give credit,
compensation, nothing. Now, it is routine. Can't Touch This by
MC Hammer credits Rick James for Super Freak, and his estate
gets compensated. So, we are in the process of trying to get it
right.
Ms. Dean. OK. Mr. Irwin, briefly, if you don't mind, I have
one more question to ask.
Mr. Irwin. Yes, I have no knowledge of anyone doing it
right. I think there is starting to be some overtures as to,
and inquiries as to let's get together and talk about it. At
the moment, no, not to my knowledge, no.
Ms. Dean. Dr. Callison-Burch, in the time I have remaining,
I was particularly taken in your testimony by something you
said about what is, the impact on labor, on workers. You used
the expression and you suggested will paralegals go the way of
the lamplighter. What are the implications for the labor
market?
Mr. Callison-Burch. I think this is the large issue that
everyone needs to deal with and where Congress can have the
most effect. I feel that at the moment we may be on a precipice
of mass unemployment. I think the probability is very small. It
is such a dire outcome that you really need to consider some
sort of legislation like in case of emergency, start a new WPA.
I think that dealing with this as a copyright issue almost
entirely misses the point that these systems are coming. We
have in America experienced unemployment as a result of
automation before. It has largely affected blue collar work but
now has the potential to also affect white collar work as well.
Ms. Dean. Fascinating.
Mr. Chair, again, I thank you. I yield back.
Mr. Issa. I thank the gentlelady.
We now go to the very patient gentlelady, Ms. Lee of
Florida.
Ms. Lee. Thank you, Mr. Chair. Thank you to all our
witnesses who are here today. We so appreciate your time and
your testimony helping inform us about how we might embrace the
emerging technology of artificial intelligence, but also
recognize the immeasurable value of our artists and the need to
protect and balance intellectual property and copyright
protections.
I would like to return, Mr. Damle, to your testimony.
Specifically, during the questions by Congressman Moran and
Fry, you talked a bit about your perception that existing
copyright law was adequate to take on this new emerging
landscape and continue to resolve that balancing between the
property rights of artists and our new technology that we see.
One thing that concerns me is that when Congress fails to
be sufficiently clear we leave to judges the task of figuring
it out. Of course, we want judges applying the law not creating
it. We don't want to put judges in the role of being
policymakers.
So, I would like for you to elaborate a little bit more on
your conclusion that our existing statutory framework is
sufficient to take on this new challenge.
Mr. Damle. Thank you, Congresswoman. It is an excellent
question.
I think if you just look at history here fair use has
existed in the copyright law for about almost 200 years. Over
that time, it has dealt with lots of massive shifts in
technology.
In 1984, the Supreme Court applied fair use to the then-new
technology of VCRs and said, applying again 180-year-old law
that did not have any understanding of that kind of recording
technology, and made a really fact-bound, cautious, and careful
decision about whether home recording was fair use or not.
Fast forward through the era, recently the Supreme Court
handled a case involving software APIs, again, something no one
could have ever dreamed of as being a copyright issue 180 years
ago. Yet, the court, again looking at the very specific facts
of that case, determined that the reuse of software APIs was
fair use.
There have been other cases going the other way, looking at
new technologies like Napster and saying that is not fair use,
that is an exploitive use of the technology.
So, my perspective comes from looking at that long history
and how copyright has been able to manage shifts, even major
shifts in technology.
Ms. Lee. Professor Burch, what is your take on that same
question and the conundrum of not overregulating, but at the
same time not leaving it to courts to try to create policy?
Mr. Callison-Burch. I think that it is worth considering
all these issues and deciding whether or not copyright needs to
be extended. I think there is many things that my fellow
panelists are talking about that are 100 percent valid and need
to be incorporated into the ethical guidelines that AI system
developers create, including right of publicity and copyright
and characters and things like that.
I don't, and those are not currently covered sufficiently
by copyright law. They may be covered by other laws
sufficiently. I think that is where the target should be, like
what is the output of these systems and what is correct use and
incorrect use of the output of these systems.
Ms. Lee. OK. Mr. Navarro, one of the advantages about
getting to ask questions near the end is that I can bring to
you this one. Is there anything that you wanted to share with
the Committee today that you have not to this point in the
hearing been asked?
Mr. Navarro. That is a good question. Yes. First, when we
look at the ingestion of trillions of pieces of data, that is
one way of looking at it. When we look at the impersonation of
voices, that is really one piece of data. That is very
specific. It is easy to get lost in the shift between the macro
and the micro as we figure out where to go on this. So, I think
we need to take a look at both.
I think the other is that I appreciate the technology. I
appreciate the technical use of terms. My music isn't data.
Ms. Lee. Mr. Irwin, the same question to you, anything you
wish to share with us today that you haven't already testified
to.
Mr. Irwin. I am going to follow on what Dan said. I am a
little distressed that we are calling this training, to be
honest, because in my mind we train athletes or animals. We
don't train machines. We equip machines with data, as the
technology people like to call it, but as we like to call it
music.
It is very Orwellian how the tech industry manages to
change terminology on us. It is not data or content to us. It
is music. It is photographs. It is not file sharing. It is
stealing. It is very simple. They are the sort of things I
wanted to get into the record that I didn't get to say.
Ms. Lee. Thank you.
Mr. Chair, I yield back.
Mr. Issa. A brilliant move for such a new member.
With that, we go to the gentleman from Maryland for five
minutes, Mr. Ivey.
Mr. Ivey. Thank you, Mr. Chair and Mr. Ranking Member.
Ms. Lee, you stole my question there. I did want to go back
to the deepfake issue. Mr. Sedlik, I think you talked a little
bit about technology that exists that allows, that gives the
ability to separate out what is real and what is fake. You said
C2 something. I wanted to get more details on that.
Mr. Sedlik. Sure. So, a consortium came together managed I
think by Adobe or founded by Adobe and with various industry
players. They looked at, they explored methods of creating a
technology that would allow you to use software to determine
whether or not an image has been altered and what the
provenance of that image is.
That information is stored in the image so that when you
are using software to view an image you can tell whether or not
that image has been altered, for example, swapping out a head,
changing, removing something, adding something, any sort of
revision to the image, to provide that information to the
public and to people who might rely on those images. That is
C2PA.
Mr. Ivey. OK. Here is where I am going with this. So, I
guess what you are saying is that, a movie or some visual
depiction that has the code embedded in it so that distinction
could be made, then you could, you have the software to
identify it. I guess the question I would have is: What if it
is not software-generated or if it is just a visual?
Here is sort of the big question. So, for example, body
cameras worn by police, and I know this isn't a creative
content question. Body cameras have become super relevant in
court with respect to police cases and in many instances
videos. Former President Trump's video in his most recent trial
was pretty significant as well.
I was sort of curious if it is AI-generated entirely, so
you would take the subcomponents of the training piece or
whatever, and I went to law school instead of any kind of hard
science classes, so I am struggling with the technology terms.
Would you be able to tell if an image or a video is generated
entirely by an AI approach as opposed to some of it being real
or some of it not? Is there a way that you could tell if it is
authentic or not?
Mr. Sedlik. There is, the AI platforms are doing some work
on this and coming forward with technology to be able to allow
the public to discern what is AI and what is not. With respect
to other types of creations or recordings, like you mentioned
the police cameras and such, those have, those embed time code
and other information in the recording so that you can tell
whether or not there has been anything removed.
In terms of finding something or determining whether an
image is AI-generated or not, that is right now at the
experimental phase with scientists looking at it, image
scientists. I think we are going to see something in the very
near future.
Mr. Ivey. OK. Mr. Navarro, did you have some thought on
that as well or--
Mr. Navarro. From a musical standpoint, we sometimes use
our taste. I did a touring test at South by Southwest in March
where pieces were being played. I was able to guess about \3/4\
of them, but not all of them. It happened to be with very
simple forms of music that were style-based. Sometimes the
vocals were gibberish. Other times it was just, were moves made
musically that no great musician or a great producer would
allow.
It gets more difficult. This particular piece that we just
heard from the gentleman from Texas, that would have been
really, really difficult to tell. So, from that standpoint, I
think that is where one of the great dangers is, is in not
knowing the difference.
Mr. Ivey. Professor Callison-Burch, I wanted to sort of
touch base with you on this kind of issue as well. I know we
have been talking about copyright law, but from the standpoint
of protecting the community, the world, whatever from these
sorts of fabrications, which I don't know if they are illegal
at this point or not. Are there steps we need to take to
address this?
Mr. Callison-Burch. Absolutely. My Ph.D. students have done
the largest-ever study on human detectability of machine-
generated text. We have found that over time it is getting
harder and harder to detect, but that people can be trained to
detect, so I think similarly to how Dan is saying that the
latest clip of the fake Drake song is increasingly passable as
human, whereas five years ago it wouldn't have been. Like that
is a trend that we are on. I think any sort of basic media
literacy that we teach to our children should include this as
an element.
Mr. Ivey. OK. With respect to visuals? This is from the
court standpoint on some of these images. Sometimes people's--
whether you go to jail or not is determined by a visual or a
video clip that a jury might see. Are there ways to make sure
that whatever is presented or a judge making a determination
about what could be viewed is actually authenticate?
Mr. Callison-Burch. There is expert testimony through
examination of the images. I think you could likely still
detect artifacts of a machine-generated image if it is done
carelessly. I think it is going to get harder. So, it is
certainly something to keep under consideration as time
progresses.
Mr. Ivey. All right. Well, I want to thank all of you for
coming in today. I apologize for a question that is a little
off the topic, but I think important for us.
Mr. Chair, I really wanted to thank you and the Ranking
Member. I agree with some of the previous comments about this
being a refreshing change from some of the types of hearings we
have had, previously. I thank you for this.
Mr. Issa. Well, I thank the gentleman. I will welcome you
to the Subcommittee, which has historically been extremely
bipartisan. So, we are going to keep that tradition going when
it comes to, we can't pass intellectual property reform except
on a completely bipartisan basis.
Now, last and probably least, you get me. I am going to
start with this. Thank you, Mr. Ranking Member. This is an
actual portrait. You will recognize some of the characters.
They are pretty much mostly all dead. Some of them, in fact,
all of them were produced by an artist who took a number of
photographs, including Lincoln photographs, to produce this
product. I am confident that Andy Thomas did not pay for most
of the photographs.
Do you see a problem in, and I will start with Mr. Navarro,
in that being fair use, the collection of photographs, since
for the most part these gentlemen were not available to sit or
stand for this portrait?
Mr. Navarro. I think if it were me, I probably would have
made an attempt to compensate the photographers. I don't know
enough about the derivative work clauses in the Copyright Act
that can allow something like that. I happen to think that is a
wonderful painting. I am a Democrat. I still think it is a--
Mr. Issa. I wanted to bring both, because I have another
one with all the Democrats. They both hang in my office. You
are welcome to come afterwards.
Mr. Navarro. As such, my personal opinion is that
creativity has been enhanced, and communication has been
enhanced. However, were it me, I would have made an attempt to
contact the photographers.
Mr. Issa. I believe that he would have bought those
pictures if they were commercially available, whether they were
under copyright or not.
So, and I point that out because it is one of the
challenges that I face in trying to steer this Committee now
and in the future is I try to look back, as I did there, on
existing copyright implementation, existing art. Andy Thomas
clearly does his own art. He works off of, even when he is
doing a living person, he generally will come in, take
pictures, and take those pictures back to do his work. It is a
style. It works very well for him obviously. I ask that more
for all of us to opine on afterwards.
I do have one similar situation. If we believe Wikipedia as
always correct, The Beatles derived their inspiration from
among others Elvis and Chuck Berry. Elvis gave credit to his
inspiration, including none other than Dean Martin. Dean Martin
credited The Mills Brothers and Perry Como. Perry Como gave
significant credit to his inspiration to Bing Crosby.
Oddly enough, Bing and each of those people were all alive
at one point, all entitled to their copyrights at the same
time. Yet, we were able to work our way through who, what, and
where. Every one of these artists undoubtedly took sheet music
and practiced with those inspirations that they had. Every one
of them listened to those people's recordings. Undoubtedly,
every one of them performed and tried to do it at least in
training in that style.
So, the question I ask, and I will just go right down the
line starting with Sy, we have a conundrum here at a minimum.
If I am taking a class, high school, college, whatever, the
material used for that is typically paid for as copyrighted
material once to educate me. After that, my education, assuming
I am not using the script itself, my education goes forward
without further copyright. If I take a significant amount from
one of those college books that I purchased on which the
copyright was paid, I have an obligation to disclose it, and if
it is beyond a snippet to pay.
Is that really what you believe is at the root of how
computers, and I apologize to Mr. Irwin, learn or collect data?
Mr. Damle. I think the analogy is a very close one, both as
a legal matter and, as Professor Callison-Burch can talk about,
the technology. At least as I understand the technology, that
is also true as well, that the copyrighted works are being used
not to create a collage or record the copyrightable content
within them, but to simply learn statistical facts about the
works themselves.
It is a very similar process to the way humans learn. Now,
the problem with machines is that they learn much more
inefficiently than humans do. I can read three or four books
about a topic and then be conversant about that topic and maybe
even write my own article about that topic. Machines are not
that skilled. They haven't quite caught up to us yet. So, they
require, to do a similar kind of learning, they require
billions and billions of pieces of work to reach those same
learnings.
Mr. Issa. Just because they are a bad student doesn't mean
that the copyrighted material shouldn't be paid for, does it?
Mr. Damle. Now, the question of whether it should be paid
for is a different question.
Mr. Issa. You are the lawyer. The gentlemen on this end
want their work paid for even if it is used in the classroom.
That is the final part of this question, bringing it all back
together, is you two have been very good at calling this
education, training, et cetera. If the analogy of the
classroom, of teaching, of learning is there, there is also the
analogy that this copyrighted material is paid for. You didn't
go to the public library, so to speak. You went and got this
material and ingested it just as I ingested, for the most part,
my college years.
So, I'll go to Mr. Burch. I really want to get through this
for everyone because I think it is part of what brings a close
to this hearing.
Mr. Callison-Burch. Thank you. This is a great analogy. So,
again, I think it highlights the fact that the systems are
learning. They are learning facts about the world. Those facts
are not copyrightable.
I think where the material is acquired from and whether
that is fairly, fair use or not fair use is exactly the right
question. Many of my students learn by going out to the
Internet and retrieving facts from the Internet that they do
not pay for. That is what is happening here. So, I think--
Mr. Issa. You mean you haven't published a book that you
make them buy like my professors? No--
Mr. Callison-Burch. I have not published a book that I have
forced my students to buy.
Mr. Issa. Mr. Irwin.
Mr. Irwin. I think what was really interesting in your
little Wikipedia piece you read was the word inspiration. I
think inspiration is what we need to keep hold of here, because
there is a big difference between getting a book and you want
to emulate your idol. You want to emulate The Beatles, or you
want to emulate The Rolling Stones.
In my case growing up, my parents bought me sheet music.
The sheet music was Mozart or Beethoven. It was public domain.
They still had to buy the sheet music for me to learn. I
learned how Beethoven writes music, how Mozart writes music,
learned The Beatles, and I learned all these things. I learned
them.
Every time, in your case, the inspiration was because they
heard it on the radio. If they heard it on the radio, there
were royalties being paid. If they bought the sheet music,
there were royalties being paid. If they bought the record and
took it home and played it, there were royalties being paid.
That is really where we are now. Yes, they were learning. That
inspiration was paid for at every time along the way.
Mr. Issa. Mr. Navarro.
Mr. Navarro. Mr. Irwin has hit it on the head with the
notion of inspiration. I don't think we should ever litigate
against inspiration. That is how we evolve as a culture. I
think the beauty of it is that no matter how inspired you are
you are never going to get it perfect. Therein lies the
individuality of the subsequent artist is they bring their own
stamp to whatever they learned on. I learned on many singers to
develop myself as a singer.
Mr. Issa. You are not giving credit to one the way
Wikipedia did for these?
Mr. Navarro. Well, no. Well, there are a few. It branches
out. We still change things a little bit.
Also, with regards to stuff like that, I spent some years
in advertising before I became a professional songwriter and
musician. I date back from the era where if it got too close on
the radio you said celebrity voice impersonated. You had to
reveal it, that it was an impersonation lest somebody--yes,
well, Tom Waits and Bette Midler with that.
All this is to say is that inspiration should never be
stifled. I am not interested in stifling technology. We benefit
from it, guardrails, credit, consent, credit, and compensation.
Thank you.
Mr. Issa. Mr. Sedlik.
Mr. Sedlik. I agree with my colleagues. Your example of the
classroom was very insightful. One of the largest areas for
copyright licensing is educational use, the textbook industry
and in turn photography and illustrations being licensed for
educational use.
If you want to teach in the classroom about B.B. King, you
can't use my photograph of B.B. King without licensing that
photograph. If you want to teach about me or my photograph of
B.B. King, that is going to be fair use.
It is true that all artists have been influenced by others.
There is a difference between influence and inspiration or
copying or theft. All our works are made based on our lived
experience as humans and taking those works and being inspired
by multiple of them to create, for example, the portrait that
you showed.
If he relied on multiple portraits to create each
individual portrait and was inspired by them rather than
copying them, then there might not be an infringement there. If
he traced it on a light box, if he projected it on the canvas,
if he copied all the expression or the heart of the expression,
there might be infringement there. That is a case-by-case,
fact-dependent analysis.
Mr. Issa. I think one of the interesting things about Andy
Thomas is he manages to capture each of these men, both
Republicans and Democrats, better than they were, which is
something that a computer probably will take a long time to
learn to do.
In closing, there are two things that were not discussed
fully. One is the fact that even if copyright use is not
attributable to a single artist, in other words the trillion
into the works may not be effectively able to be done, it
doesn't mean that the ingestion of that material should not
somehow go to the benefit of the copyright industry as a whole.
That will be one of the things that is not a statutory remedy
today but could be.
To a certain extent it is like orphan works. You can't
necessarily get them back to somebody, but you don't get to use
them completely for free just because we can't assign them.
The last one, which was brought up here today, and I would
like all of you to opine on it, because I think it was a real
threat, but there wasn't time in this hearing to pick it up. In
patents, in copyrights, and in trademarks currently based on
some bad actors, computers can generate an infinite amount of
combinations. If they generate an infinite amount of
combinations of copyright material, they could, in fact, create
a body of copyright that could eclipse future innovation,
simply push out tens or hundreds of trillions of songs, of
variations of art and, in fact, then make a claim, a troll-like
claim that everybody else who comes up with an original piece,
there I find enough to say that you took it from me even though
you may never have seen the trillion different outputs.
It doesn't sound, it sounds far-fetched until those of you
who look at AI and look at the petabytes per minute now that
are being ingested on the Internet, and you realize that, in
fact, infinity is closer to us than we ever thought it was.
So, I would like you to opine on that, because one of the
challenges that I face is to limit copyright or patent
applications or trademark applications that are computer-
generated if, in fact, they serve only to limit human's ability
to do individual creation. I would like you all to opine on
that within your own expertise.
We will leave the record open for the next, I think five
days is the Committee rule. I will shove it in if I get it
later, as long as the Chair lets me.
So, I want to thank you. I expect that we have your
numbers. We will be calling you. You have our contact
information. I would hope that you continue to help us.
Mr. Ranking Member, do you have any closing statements?
Mr. Johnson of Georgia. I do not, other than to thank the
witnesses for your testimony today.
Mr. Issa. With that, we stand adjourned.
[Whereupon, at 12:38 p.m., the Subcommittee was adjourned.]
All materials submitted for the record by Members of the
Subcommittee on Courts, Intellectual Property, and the Internet
can be found at: https://docs.house.gov/Committee/Calendar/
ByEvent .aspx?EventID=115951.