June 19, 2025 • 39 mins
Kevin Werbach interviews Dale Cendali, one of the country’s leading intellectual property (IP) attorneys, to discuss how courts are grappling with copyright questions in the age of generative AI. Over 30 lP awsuits already filed against major generative AI firms, and the outcomes may shape the future of AI as well as creative industries.
While we couldn't discuss specifics of one of the most talked-about cases, Thomson Reuters v. ROSS -- because Cendali is litigating it on behalf of Thomson Reuters -- she drew on her decades of experience in IP law to provide an engaging look at the legal battlefield and the prospects for resolution.
Cendali breaks down the legal challenges around training AI on copyrighted materials—from books to images to music—and explains why these cases are unusually complex for copyright law. She discusses the recent US Copyright Office report on Generative AI training, what counts as infringement in AU outputs, and what is sufficient human authorship for copyirght protection of AI works. While precedent offers some guidance, Cendali notes that outcomes will depend heavily on the specific facts of each case. The conversation also touches on how well courts can adapt existing copyright law to these novel technologies, and the prospects for a legislative solution.
Dale Cendali is a partner at Kirkland & Ellis, where she leads the firm’s nationwide copyright, trademark, and internet law practice. She has been named one of the 25 Icons of IP Law and one of the 100 Most Influential Lawyers in America. She also serves as an advisor to the American Law Institute’s Copyright Restatement project and sits on the Board of the International Trademark Association.
Thompson Reuters Wins Key Fair Use Fight With AI Startup
Dale Cendali - 2024 Law360 MVP
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:01):
Yeah, welcome to the Road to Accountable AI.
Happy to be here.
Thank you.
You've been in this space of intellectual property law for some time, uh involved with anumber of important influential cases.
Technology is always a part of what happens in IP, but is technology changing faster?
(00:22):
Are the issues more challenging today with AI?
don't think technology is changing faster.
It's certainly changing.
But whether it was the creation of movable type or a photocopier or the internet orworried about Y2K, there's always something new from a technological perspective.
(00:44):
And the law and thinking in business needs to adapt.
So I don't feel like it's changing faster.
I just feel like
With AI, this is a moment in time where there is the latest, bigger change.
So let's get into some of those issues with AI.
We'll start on the input side in terms of the training of models with copyrightedinformation.
(01:10):
There have been a whole slew of cases filed against uh AI developers.
Can you just first generally give us a sense of what the state of play is legally?
Sure.
According to the Copyright Alliance, over 30 lawsuits have been filed already in thisspace with regard to the input side.
(01:33):
And to me, there are a few interesting things to note.
First off, let's make sure we know what we're talking about.
When we say the input side, that's the side that focuses on the cases where the primaryclaim, there could be other claims, but is based on
that you use my copyrighted work to train your LLM, your large language model, to do AIwork.
(01:58):
And so it's concerned about that, the training to create the AI.
And to me, there's a few interesting things about the cases that have been filed so far.
First, it's interesting to me that they're groups of stakeholders.
who were filing them.
And by that I mean you have authors and writers, like in the Authors Guild case, and yousee authors and writers complaining.
(02:28):
You see news publishers, like the New York Times case and the Dow Jones case.
uh You see visual artists and photographers, like Anderson versus Stability AI and theGetty Images case.
You have music cases by people like Concord Music with musicians and lyricists and peopleconcerned about that.
(02:51):
And there have even been some suits by uh software developments like Doe versus GitHub.
So what you're seeing is people are approaching it from their own particular focus and arebringing lawsuits that are focused on kind of their angle, their.
their type of work.
(03:12):
It's not just one type of work.
The second thing I wanted to mention about these cases, which are a little different, istypically copyright cases are a particular copyright holder is suing about their work
being infringed.
And there are plenty of cases like that.
But you're also seeing a number of class action cases being filed in this space.
(03:37):
And that's a little bit more unusual for copyright cases.
And it's going to be interesting to see whether those classes are going to be certified,how they're going to be certified, and what their potential impact might be on the
litigation.
So the bottom line is there's a lot of focus on these kinds of input cases from a lot ofdifferent sources and some different procedural frameworks.
(04:07):
One of the core issues though is whether this material can be used as fair use even if itis protected under copyright.
Can you help people who are not familiar with the legal issues understand at least whatare the courts going to have to work through to make those determinations?
Right.
Well, one of the things that the courts will do is a threshold matter is does someone havea copyright?
(04:31):
And under US law, you're supposed to have a registered copyright registration or else beturned down for a copyright registration.
Then you can effectively appeal that to a court.
So you need to own a copyright to start with.
And then someone needs to copy what it is that you have, which usually the model or thefact pattern for these input cases
(04:55):
that's usually not so hard because almost by definition, these cases involve copying ofthe copyrighted material.
That's the whole point of it.
And so then the issue is, well, is that fair use or not?
And fair use under United States law is, courts are supposed to look at four factors at aminimum.
(05:17):
They can consider other things like public policy or special things.
but those four factors are the purpose and character of the use.
Like what are you using it for?
What's your purpose?
What's your justification?
Second, the second factor is what's the nature of the work?
Works that are more factual, more functional, like computer programs tend to get lessprotection than more creative works, like a novel or something that's highly expressive.
(05:47):
The third factor is
How much of the work are you using?
Are you copying all of it or just a small amount of it?
And then the last factor is the effect of the use on the market or potential market forthe work.
And in AI cases, that focuses a lot on is there a market for training data?
Was there one originally?
(06:08):
Is there one now?
How should that be factored in in light of the other policies?
And those are the kinds of things that courts will be looking at from a variousperspective.
And obviously, your litigator, in any case, you would be making an argument based on thosefactors if their use is an issue.
When I talk to people about this that are not lawyers, they struggle with the fact thatit's a balancing test.
(06:32):
There's not a bright line.
And it's not just that any one of those factors is necessarily enough by itself.
So does this mean that courts are just going to resolve a whole bunch of individual casesone way or another?
but we won't get to any clarity about the fundamental issues.
No.
And that's because we live in a world of precedent where, yes, it's one copyright statute,which also helps, but you learn from each case.
(07:03):
And recently there were two Supreme Court cases that address fair use, Google v.
Oracle, and then more recently Goldsmith v.
Warhol, both of which tried to shed light on how courts should look at the four fair usefactors.
I'm confident that as time goes on and we get more courts digging into these oh AI cases,patterns will develop, thinking will develop, and things will become more clear.
(07:34):
Although it is 100 % true that fair use is seen as very fact dependent.
And what that means is this particular justification for
the copying in one case may be different from another case.
And the effect on the market might be more directly competitive and easy to understand andharmful in one case, and it might be much more indirect in a different case.
(08:04):
And so there could be variations on the outcome depending on the facts.
And frankly, that's how law works or otherwise one side would always win and one sidewould always lose.
So on that point about precedent, you mentioned at the beginning that there's always wavesof new technology.
We had issues about intellectual property and fair use with the rise of the internet.
(08:28):
uh Google makes copies of things on the internet to put in its database.
There was a big case about Google Books where they were actually copying books and thatwas found to be fair use.
how much does that internet wave of cases predetermine what's going to happen in the AIcontext?
Well, I don't think anything predetermines everything, but you're absolutely right.
(08:51):
People look at precedent.
And you're mentioning the Google Books case.
And that was a case where absolutely similar, you could argue, to training for data.
That's where someone would say there's a parallel.
Google was copying lots of copyrighted material in its entirety, not for purposes of AI,though, but for purposes of devising a search methodology that would let people
(09:17):
find books.
And one of the things about that case, and courts have talked about this, including someof my own, have been that it helped people find books.
didn't act as a substitute for the books.
It didn't replace books.
It didn't deliver the content in the books, except perhaps for small snippets.
(09:38):
It let people find books.
And so both sides of the equation, those who are arguing
uh for fair use in the AI context and those arguing against fair use in the AI contextfind things in the case to talk about that might be more or less similar to their
(09:58):
particular uh case.
But I don't think oh most people would feel that it's not an AI case.
It's not on all fours with what the facts that people are dealing with here.
So just like
that case, just like the Google, the Oracle case, just like the Warhol case, a lot ofthese cases are going to have to be considered and the parts that are more relevant than
(10:24):
others will be the ones that will be pulled out.
And as you talked about, these cases involve various different modalities, text, images,music, and so forth.
How likely is it that the outcomes are going to be different based on the kinds of thingsthat are being at issue?
(10:44):
Well, that again is going to be interesting.
I don't know if it's going to be a big difference if it's, oh let's say it's a literarywork versus a photograph.
mean, both of those tend to be relatively expressive works.
And I don't know.
There could be reasons why the market might be different for one or the other or thejustification, the transformative purpose could be different.
(11:13):
But it's harder for me to see why that would necessarily make a huge difference given thenature of some of those works.
You could see other circumstances though, something was, especially something that's usedfor, that's very scientific, for specific scientific research, especially if that wasn't
(11:38):
substitutive, it didn't cause market harm.
that could lead to a different kind of result potentially.
But I think it would be more like the science stuff used for non-competitive purposesversus that seems more likely to be different than if it's a song versus a photograph
(12:07):
versus a text.
But I could be wrong and there could be reasons, very smart lawyers and judges we willdeal with in those cases that will point out that there are differences that may be less
obvious to us having this conversation now.
The Copyright Office has a proceeding or they did a request for comments on AI and got, Ithink, 10,000 comments and then have a series of reports that they're putting out.
(12:35):
And the most recent one was on this question of generative AI inputs.
um What was your reaction to that report?
Well, part of my reaction was just that there was a lot of reaction because the communitywas very interested in it.
People were waiting and copyright lawyers were all saying to one another, when do youthink they're going to come up with a report?
(12:59):
Because they always do a very thorough job, the copyright office and the reports.
as you know, there's three reports that they did.
And this one from May dealt with fair use.
dealt with generative AI and dealt with the inputs.
And just to kind of step back as to what the report said, as that relates to the reactionto the report, one of the things the report said was that using copyrighted works to train
(13:32):
AI models could be infringement of the right to reproduce the works.
and that you really need to go back to Warhol, that's that very use case by the SupremeCourt I mentioned, and assess whether the use was transformative and the copyright off is
noted, that's a determination, that's a matter of degree.
(13:55):
In other words, it's not just binary, it's transformative, it's definitely fair use, orit's not transformative, it's definitely not.
It could be somewhere in between.
And it said that
gee, if you were taking a work to train something and it looks like you're doing it forthe same audience as the original, that doesn't seem that transformative to me.
(14:19):
So part of the report looked like it had stuff in it that suggested that there could be ohan issue and that there could be infringement.
Same thing on the output side.
Again, but the report
consistent with what I was saying before, was very measured and had a lot of explanationsabout matter of degree and depending on the facts and how the law will evolve.
(14:52):
But what's interesting to me is the reaction to the report because the Copyright Alliance,which I alluded to earlier, which is a coalition of
various companies and people that own copyrights.
They said that they were pleased with the report and that it struck the right balancebetween not saying everything was necessarily fair use or that everything necessarily
(15:19):
wasn't fair use.
And they agreed with the importance of Warhol and certain other stakeholders like the NewYork Times made similar kind of comments.
On the other hand, the Electronic Frontier Foundation, the EFF,
know, disagreed and felt that the report, you know, more firmly placed its thumb, theythought, on the side of saying it was copyright infringement and not a fair use.
(15:45):
And similarly, the Chamber of Progress, which is a group formed by a number of differenttech companies, was concerned that the Copyright Office didn't fully appreciate the uh
constitutional incentive behind copyright of creating new works.
And the thought was that, that,
(16:05):
over broad copyright protection in their mind could stifle an invention and theadvancement of technology.
Of course, someone else could say, well, you need to have created these works to beginwith in order to train it.
what I'm, to me the interesting thing is that everybody is paying a lot of attention towhat the Copyright Office says, trying to get learning and guidance from it.
(16:33):
and are commenting on it and reacting to it from where they sit in their prism.
Absolutely.
And part of that points out that this is now a major uh policy issue, a major politicalissue, an issue with a lot of uh economic impact at stake on all sides.
(16:56):
Again, going back to history, when we had the internet raising copyright issues 30 yearsago in the US, eventually Congress passed the Digital Millennium Copyright Act to address
some of these issues.
And historically at other times we have, for example, compulsory license.
for radio and for streaming and things like that.
How likely do you think it is that Congress will ultimately resolve these issues with somekind of mechanism?
(17:20):
I can firmly say I don't know.
All kidding aside, it's a volatile world in terms of not just copyright, but in terms ofoh the legislature and what they might or might not do.
(17:40):
But I think it is helpful to go back and look at, as you said, what happened with theDigital Millennium Copyright Act because
That was intended to strike a balance between the different coalitions.
The thought was that the service providers that created this important thing, the internetthat we all have come to know and love and rely on, or mostly love in any case, they
(18:07):
needed to be protected with safe harbor so that they weren't responsible like a newspapermight be for things that users might post on their
their sites.
That's the premise behind a lot of the DMCA was to provide that type of safe harbor.
On the other hand, the trade off was that the people who own copyrights would get aneasier mechanism to have this take down procedure so that they can give a notice of take
(18:39):
down and then the ISP was supposed to take down the user generated comment in order topreserve their safe harbor.
Now,
A lot of people disagree today with provisions of the DMCA.
Some think it goes too far one way.
Some think it goes too far another way.
But the point was it was intended to be a compromise.
(19:04):
So you have to think about, know, what would that compromise be here or not?
You can also have more specific legislation if we want to go back to the internet thing inthe sense that there were uh
the Anti-Cyber-Scotting Consumer Protection Act, which was really specific to deal withfraudulent people misusing domain names and the names of folks like Wharton or something
(19:31):
like that and causing mischief.
You could see some parallels to that today.
You probably know about more specific legislation like the DeepFace Fakes Act and similarthings done at least right now.
efforts made on the state level concerned about uh especially right of publicity issues,privacy issues, things like that.
(19:58):
In other words, there could be broad legislation or there could be more narrow castinglegislation depending on what the issues may be.
And it could be that we end up with a broad legislative solution or a more narrow.
legislative solution, depending on the will of Congress and how such a bill might bewritten.
(20:26):
And you could also end up with a situation where uh the balancing, the coalition is notthere and we need to have the courts spend more time.
uh
figuring out a solution, which may not be a bad thing because as time goes on, you learnmore and you see more of the consequences.
(20:57):
And sometimes you can make a better decision when you have more information about uh thepotential impacts.
So it's too early to tell though as to whether here we're going to end up with.
em
uh a giant legislative solution or not.
(21:18):
Okay, why don't we move on to the output side?
We've been focusing predominantly on the training question, but what are the major issuesaround whether a work created with genera.ai infringes on?
So many cases that have been brought that are input cases that are complaining about thetraining are also output cases because it's kind of hard to have the outputs be similar
(21:46):
and this would be your output, you're letting users come up with things that aresubstantially similar.
That's the magic phrase for copyright infringement if the works are substantially similar.
So that the idea is that the outputs, the photographs or the song lyrics or the book orwhatever is substantially similar to my copyrighted work.
(22:11):
Sometimes the training needs to have been fed in with that material for the outputs to dothat, although not necessarily.
So the output cases uh focus much more on, hey, I'm being harmed because
Instead of coming to me, copyright holder, for my photograph, my character, whatever itis, a user is able to type in what they want and get that image and they don't have to
(22:44):
come to me for a license or hire a graphic designer or someone like that for that purpose.
And most recently, very recently, this week from when we're talking, Disney
Ununiversal filed a lawsuit against Mid Journey, which is one of the AI platforms thathelps to create uh visuals.
(23:06):
It's not a text focus, the focus is on the visuals.
And their complaint had all sorts of images of Elsa and other princesses that weregenerated using Mid Journey or Marvel superheroes and Simpsons characters and
and the like, and the concern was, hey, you're using my IP improperly.
(23:32):
And this is the first time the movie studios have really come together to file a lawsuitlike this.
it's already gotten a lot of attention in the newspaper.
So we're gonna have to see how it plays out, but that's what the input cases focus on.
With that, though, just to be clear, is that there's uh the different AI platforms varysomewhat.
(24:01):
Some of them have put in place guidelines to try to prevent the ability for a user to comeup with something that looks like a copyrighted character or reveals too much of the text
of a book or of a particular article or something like that.
Others feel that, no, that shouldn't be their obligation and they shouldn't have to dothat.
(24:29):
It's the user that put in that request.
It's not their responsibility.
People who are on the side of the platforms will often argue something called volitionalconduct.
Is it really volitional conduct?
If I'm the one, the user who's typing something in saying, me X, uh they'll say that I'mnot really responsible for that.
(24:56):
Others would say, on the other hand, well, you set it all up that they can do this by partof what you train.
So you are responsible.
So those are going to be some of things that are going to uh be dealt with in some ofthese cases.
And this issue of to what extent uh
(25:17):
there should be guardrails or not uh is something that you're seeing litigated more andmore.
For example, in one of these music cases I mentioned, the Concord case, initially therewas a preliminary injunction filed and trying to get certain guardrails in place, but they
(25:37):
reached an agreement that the platform would put in certain guardrails that would make it
harder for people to get, for example, song lyrics, in which case the court said, themotion's now mute, moo, because people have agreed to this.
And there'll be a debate, and the cases are going to try to figure out, as I said, to whatdegree is the platform responsible for what the users do.
(26:05):
And there'll be a debate as to whether there should or shouldn't be these kind ofguardrails with regard to uh limiting.
what people can ultimately see.
Is this going to be a similarly fact-specific kind of inquiry?
Because even if there are guardrails, there are lots of examples you can find online oryou can probably find yourself with some of these tools where you type in and say, give me
(26:32):
Elsa from Frozen, and it says, I can't do that.
And then you say, give me a princess who looks like this, and so forth, and you cancircumvent.
So does it inevitably have to be a very specific inquiry about what's enough?
I, as any lawyer would say, it depends, right?
(26:55):
And because again, you can understand that part of it will depend legally on the thoughtabout the, to some degree, this volitional conduct idea.
But the other thing is if ever some of these cases get to a jury, juries may have todecide credibility issues,
(27:19):
responsibility issues and issues relating to uh how similar it is because uh people coulddisagree as to whether something really is a recognizable character or not.
There's things like scenes of fair, certain tropes are available for everyone to use.
(27:44):
uh On the other hand,
It's the same coloring and same hair and same design and same eyes and is returned by aspecific choice to that.
oh That makes it a little bit harder for someone to say it isn't a copy of that image.
And then the last major scenario I just want to ask you about is uh the question aboutwhether a general AI work itself is entitled to copyright protection.
(28:10):
And as I understand it, the copyright was just said if it's, there's no human involvementat all, you just push a button, no.
But what's the line beyond that where there's enough human involvement?
This is where I think that the law is starting to develop so people are starting tounderstand how it's supposed to work.
You're absolutely right, and this was in yet a different copyright office circularguideline that they report, the three reports we were talking about that came out.
(28:40):
And it did talk about the idea that generally speaking, the concept is if the
artwork or article or whatever it is was all created by AI, then that can't becopyrightable.
But the issue is it could be copyrightable if there was sufficient human involvement.
(29:01):
And then everyone's trying to figure out, what does that mean?
Well, you go back and you look at some cases and you can start to figure it out.
There was a Thaler case which went all the way up to the DC circuit where someone reallywanted to make the
point, he was really trying to say that AI should be allowed to copyright things.
(29:22):
And the court said, no, you're admitting that there was no human involvement.
So that's not good enough.
And to be clear, the Copyright Office report that I mentioned said that the level of humaninvolvement can't just be in coming up with prompts.
It has to be actually some act of creation beyond coming up
(29:44):
Well, then you compare the Thaler case with the Zariah the Dawn uh case involving agraphic novel.
And I think that really tells you a lot about it's a good practical guidance.
Let's put it that way for those of us advising clients on this, because originally uh thisperson had created a graphic novel, Zariah the Dawn, and they got a copyright registration
(30:11):
for it.
She did.
Well, then.
uh it was revealed to the Copyright Office that the graphic novel, the visuals to it atleast, had been created by Mid Journey, that same company that we were talking about a
minute ago.
And that she wrote the text and she arranged the material uh into a graphic novel, but theactual visuals were not something that she had created.
(30:39):
In which case, they amended the Copyright Office
uh registration to say that she had a copyright in the text and she had a copyright in theselection and arrangement of the text to the pictures.
In other words, she did the creative work of putting it all together into a graphic novel.
But the actual images that were generated by many detailed text prompts asking themajority to create X, Y, and Z wasn't something that could be owned.
(31:12):
by themselves.
So I think that what you're seeing from that is at least how copyright could be used tosay if it's pure AI, no, but you can still protect a lot because there's still a lot of
human involvement and what the humans do with the material that's created by AI.
(31:38):
that seems to be getting
more clarity compared to some of these other issues.
uh Both the content owners and AI companies and those that are advocating for them tend tomake statements that these are existential issues, how these cases come out.
(32:02):
How much of a business impact is it really um on either side, whether there's infringementin these kinds of cases?
Well, it could have a huge impact depending on what the use is and depending on what thebusiness is.
Something could go to the core of what that business does.
(32:22):
And when I say that business, that could be on both sides, the core of what the AIbusiness is trying to do and the core of what the copyright creator type of business do.
So it could have a big impact on both, which is why
Sometimes when there's a lot of risk for either or both sides, you end up withsettlements.
(32:47):
And certainly you're seeing, if not settlements, but you're reading it out in the press,different licensing schemes being agreed to.
There was, think, the New York Times recently did a deal, I think, with Amazon and there'sother.
(33:07):
as I already mentioned, at least some, not settlement, but some agreement on some of theguidelines in the Concord case.
So you're seeing that there could be an incentive to try to work out a resolution in somecases that's a compromise for everyone.
(33:27):
uh But sometimes that compromise is not very happy.
Sometimes it could uh still end up killing the golden goose on either side or not becommercially uh sensible.
(33:47):
And once had a court that uh was urging me to, you know, both the parties to settle and itwasn't an AI case or anything like that.
But I said, well, Your Honor, that would be like having us uh lose money on every sale.
but make it up on volume, which if you think about, you really can't do it that way.
(34:12):
So it has to work from a dollars and cents point of view.
Lawyers do tend to be extreme and make uh dramatic arguments.
Sometimes there's a reason though to do so.
Sometimes the consequences could be dramatic for
(34:34):
both sides of the equation.
And I'm not in any way minimizing that.
But it sounds like from the tenor of the conversation that you're, am I right to saybroadly optimistic that through some combination of litigation, maybe some legislation,
these private resolution of disputes that will ultimately navigate through thesecomplexities.
(34:58):
Yes.
And why?
Well, again, what's past is prologue.
You started off by asking, doesn't technology change?
And yes, it does.
And there have been lots of new innovations and they'll probably be even more innovations.
mean, who thought that we'd be living in a world you can have this amazing computer thatyou carry around with you that's also a camera and all sorts of other kinds of things.
(35:24):
mean, the technology.
changes all the time and I'm not as uh inclined to uh run around like Chicken Little andsay the sky is falling.
But I do think that it's uh useful to remember that there are ardently well-reasoned,thoughtful lawyers and companies on both sides of this debate and
(35:55):
given facts of different cases, different oxes that have been, or wrongs that have been uhdone, you could see a situation where one size doesn't fit all in all cases as to whether
something is fair or not fair, whether it's infringement or not infringement.
(36:19):
And what I think is really important is for people to listen to one another.
to try to come up with best practices and uh a path going forward that protects thetremendous, wonderful innovation on, I'm not even gonna say both sides of this debate, I'm
(36:42):
gonna say all sides of this debate because I think it's a table with uh many sides.
That's a great place to end.
Dale, thank you so much for your time.
It's always a pleasure.
Thanks for being interested.
Take care.
Yeah, welcome to the Road to Accountable AI.
Happy to be here.
Thank you.
You've been in this space of intellectual property law for some time, uh involved with anumber of important influential cases.
Technology is always a part of what happens in IP, but is technology changing faster?
(00:22):
Are the issues more challenging today with AI?
don't think technology is changing faster.
It's certainly changing.
But whether it was the creation of movable type or a photocopier or the internet orworried about Y2K, there's always something new from a technological perspective.
(00:44):
And the law and thinking in business needs to adapt.
So I don't feel like it's changing faster.
I just feel like
With AI, this is a moment in time where there is the latest, bigger change.
So let's get into some of those issues with AI.
We'll start on the input side in terms of the training of models with copyrightedinformation.
(01:10):
There have been a whole slew of cases filed against uh AI developers.
Can you just first generally give us a sense of what the state of play is legally?
Sure.
According to the Copyright Alliance, over 30 lawsuits have been filed already in thisspace with regard to the input side.
(01:33):
And to me, there are a few interesting things to note.
First off, let's make sure we know what we're talking about.
When we say the input side, that's the side that focuses on the cases where the primaryclaim, there could be other claims, but is based on
that you use my copyrighted work to train your LLM, your large language model, to do AIwork.
(01:58):
And so it's concerned about that, the training to create the AI.
And to me, there's a few interesting things about the cases that have been filed so far.
First, it's interesting to me that they're groups of stakeholders.
who were filing them.
And by that I mean you have authors and writers, like in the Authors Guild case, and yousee authors and writers complaining.
(02:28):
You see news publishers, like the New York Times case and the Dow Jones case.
uh You see visual artists and photographers, like Anderson versus Stability AI and theGetty Images case.
You have music cases by people like Concord Music with musicians and lyricists and peopleconcerned about that.
(02:51):
And there have even been some suits by uh software developments like Doe versus GitHub.
So what you're seeing is people are approaching it from their own particular focus and arebringing lawsuits that are focused on kind of their angle, their.
their type of work.
(03:12):
It's not just one type of work.
The second thing I wanted to mention about these cases, which are a little different, istypically copyright cases are a particular copyright holder is suing about their work
being infringed.
And there are plenty of cases like that.
But you're also seeing a number of class action cases being filed in this space.
(03:37):
And that's a little bit more unusual for copyright cases.
And it's going to be interesting to see whether those classes are going to be certified,how they're going to be certified, and what their potential impact might be on the
litigation.
So the bottom line is there's a lot of focus on these kinds of input cases from a lot ofdifferent sources and some different procedural frameworks.
(04:07):
One of the core issues though is whether this material can be used as fair use even if itis protected under copyright.
Can you help people who are not familiar with the legal issues understand at least whatare the courts going to have to work through to make those determinations?
Right.
Well, one of the things that the courts will do is a threshold matter is does someone havea copyright?
(04:31):
And under US law, you're supposed to have a registered copyright registration or else beturned down for a copyright registration.
Then you can effectively appeal that to a court.
So you need to own a copyright to start with.
And then someone needs to copy what it is that you have, which usually the model or thefact pattern for these input cases
(04:55):
that's usually not so hard because almost by definition, these cases involve copying ofthe copyrighted material.
That's the whole point of it.
And so then the issue is, well, is that fair use or not?
And fair use under United States law is, courts are supposed to look at four factors at aminimum.
(05:17):
They can consider other things like public policy or special things.
but those four factors are the purpose and character of the use.
Like what are you using it for?
What's your purpose?
What's your justification?
Second, the second factor is what's the nature of the work?
Works that are more factual, more functional, like computer programs tend to get lessprotection than more creative works, like a novel or something that's highly expressive.
(05:47):
The third factor is
How much of the work are you using?
Are you copying all of it or just a small amount of it?
And then the last factor is the effect of the use on the market or potential market forthe work.
And in AI cases, that focuses a lot on is there a market for training data?
Was there one originally?
(06:08):
Is there one now?
How should that be factored in in light of the other policies?
And those are the kinds of things that courts will be looking at from a variousperspective.
And obviously, your litigator, in any case, you would be making an argument based on thosefactors if their use is an issue.
When I talk to people about this that are not lawyers, they struggle with the fact thatit's a balancing test.
(06:32):
There's not a bright line.
And it's not just that any one of those factors is necessarily enough by itself.
So does this mean that courts are just going to resolve a whole bunch of individual casesone way or another?
but we won't get to any clarity about the fundamental issues.
No.
And that's because we live in a world of precedent where, yes, it's one copyright statute,which also helps, but you learn from each case.
(07:03):
And recently there were two Supreme Court cases that address fair use, Google v.
Oracle, and then more recently Goldsmith v.
Warhol, both of which tried to shed light on how courts should look at the four fair usefactors.
I'm confident that as time goes on and we get more courts digging into these oh AI cases,patterns will develop, thinking will develop, and things will become more clear.
(07:34):
Although it is 100 % true that fair use is seen as very fact dependent.
And what that means is this particular justification for
the copying in one case may be different from another case.
And the effect on the market might be more directly competitive and easy to understand andharmful in one case, and it might be much more indirect in a different case.
(08:04):
And so there could be variations on the outcome depending on the facts.
And frankly, that's how law works or otherwise one side would always win and one sidewould always lose.
So on that point about precedent, you mentioned at the beginning that there's always wavesof new technology.
We had issues about intellectual property and fair use with the rise of the internet.
(08:28):
uh Google makes copies of things on the internet to put in its database.
There was a big case about Google Books where they were actually copying books and thatwas found to be fair use.
how much does that internet wave of cases predetermine what's going to happen in the AIcontext?
Well, I don't think anything predetermines everything, but you're absolutely right.
(08:51):
People look at precedent.
And you're mentioning the Google Books case.
And that was a case where absolutely similar, you could argue, to training for data.
That's where someone would say there's a parallel.
Google was copying lots of copyrighted material in its entirety, not for purposes of AI,though, but for purposes of devising a search methodology that would let people
(09:17):
find books.
And one of the things about that case, and courts have talked about this, including someof my own, have been that it helped people find books.
didn't act as a substitute for the books.
It didn't replace books.
It didn't deliver the content in the books, except perhaps for small snippets.
(09:38):
It let people find books.
And so both sides of the equation, those who are arguing
uh for fair use in the AI context and those arguing against fair use in the AI contextfind things in the case to talk about that might be more or less similar to their
(09:58):
particular uh case.
But I don't think oh most people would feel that it's not an AI case.
It's not on all fours with what the facts that people are dealing with here.
So just like
that case, just like the Google, the Oracle case, just like the Warhol case, a lot ofthese cases are going to have to be considered and the parts that are more relevant than
(10:24):
others will be the ones that will be pulled out.
And as you talked about, these cases involve various different modalities, text, images,music, and so forth.
How likely is it that the outcomes are going to be different based on the kinds of thingsthat are being at issue?
(10:44):
Well, that again is going to be interesting.
I don't know if it's going to be a big difference if it's, oh let's say it's a literarywork versus a photograph.
mean, both of those tend to be relatively expressive works.
And I don't know.
There could be reasons why the market might be different for one or the other or thejustification, the transformative purpose could be different.
(11:13):
But it's harder for me to see why that would necessarily make a huge difference given thenature of some of those works.
You could see other circumstances though, something was, especially something that's usedfor, that's very scientific, for specific scientific research, especially if that wasn't
(11:38):
substitutive, it didn't cause market harm.
that could lead to a different kind of result potentially.
But I think it would be more like the science stuff used for non-competitive purposesversus that seems more likely to be different than if it's a song versus a photograph
(12:07):
versus a text.
But I could be wrong and there could be reasons, very smart lawyers and judges we willdeal with in those cases that will point out that there are differences that may be less
obvious to us having this conversation now.
The Copyright Office has a proceeding or they did a request for comments on AI and got, Ithink, 10,000 comments and then have a series of reports that they're putting out.
(12:35):
And the most recent one was on this question of generative AI inputs.
um What was your reaction to that report?
Well, part of my reaction was just that there was a lot of reaction because the communitywas very interested in it.
People were waiting and copyright lawyers were all saying to one another, when do youthink they're going to come up with a report?
(12:59):
Because they always do a very thorough job, the copyright office and the reports.
as you know, there's three reports that they did.
And this one from May dealt with fair use.
dealt with generative AI and dealt with the inputs.
And just to kind of step back as to what the report said, as that relates to the reactionto the report, one of the things the report said was that using copyrighted works to train
(13:32):
AI models could be infringement of the right to reproduce the works.
and that you really need to go back to Warhol, that's that very use case by the SupremeCourt I mentioned, and assess whether the use was transformative and the copyright off is
noted, that's a determination, that's a matter of degree.
(13:55):
In other words, it's not just binary, it's transformative, it's definitely fair use, orit's not transformative, it's definitely not.
It could be somewhere in between.
And it said that
gee, if you were taking a work to train something and it looks like you're doing it forthe same audience as the original, that doesn't seem that transformative to me.
(14:19):
So part of the report looked like it had stuff in it that suggested that there could be ohan issue and that there could be infringement.
Same thing on the output side.
Again, but the report
consistent with what I was saying before, was very measured and had a lot of explanationsabout matter of degree and depending on the facts and how the law will evolve.
(14:52):
But what's interesting to me is the reaction to the report because the Copyright Alliance,which I alluded to earlier, which is a coalition of
various companies and people that own copyrights.
They said that they were pleased with the report and that it struck the right balancebetween not saying everything was necessarily fair use or that everything necessarily
(15:19):
wasn't fair use.
And they agreed with the importance of Warhol and certain other stakeholders like the NewYork Times made similar kind of comments.
On the other hand, the Electronic Frontier Foundation, the EFF,
know, disagreed and felt that the report, you know, more firmly placed its thumb, theythought, on the side of saying it was copyright infringement and not a fair use.
(15:45):
And similarly, the Chamber of Progress, which is a group formed by a number of differenttech companies, was concerned that the Copyright Office didn't fully appreciate the uh
constitutional incentive behind copyright of creating new works.
And the thought was that, that,
(16:05):
over broad copyright protection in their mind could stifle an invention and theadvancement of technology.
Of course, someone else could say, well, you need to have created these works to beginwith in order to train it.
what I'm, to me the interesting thing is that everybody is paying a lot of attention towhat the Copyright Office says, trying to get learning and guidance from it.
(16:33):
and are commenting on it and reacting to it from where they sit in their prism.
Absolutely.
And part of that points out that this is now a major uh policy issue, a major politicalissue, an issue with a lot of uh economic impact at stake on all sides.
(16:56):
Again, going back to history, when we had the internet raising copyright issues 30 yearsago in the US, eventually Congress passed the Digital Millennium Copyright Act to address
some of these issues.
And historically at other times we have, for example, compulsory license.
for radio and for streaming and things like that.
How likely do you think it is that Congress will ultimately resolve these issues with somekind of mechanism?
(17:20):
I can firmly say I don't know.
All kidding aside, it's a volatile world in terms of not just copyright, but in terms ofoh the legislature and what they might or might not do.
(17:40):
But I think it is helpful to go back and look at, as you said, what happened with theDigital Millennium Copyright Act because
That was intended to strike a balance between the different coalitions.
The thought was that the service providers that created this important thing, the internetthat we all have come to know and love and rely on, or mostly love in any case, they
(18:07):
needed to be protected with safe harbor so that they weren't responsible like a newspapermight be for things that users might post on their
their sites.
That's the premise behind a lot of the DMCA was to provide that type of safe harbor.
On the other hand, the trade off was that the people who own copyrights would get aneasier mechanism to have this take down procedure so that they can give a notice of take
(18:39):
down and then the ISP was supposed to take down the user generated comment in order topreserve their safe harbor.
Now,
A lot of people disagree today with provisions of the DMCA.
Some think it goes too far one way.
Some think it goes too far another way.
But the point was it was intended to be a compromise.
(19:04):
So you have to think about, know, what would that compromise be here or not?
You can also have more specific legislation if we want to go back to the internet thing inthe sense that there were uh
the Anti-Cyber-Scotting Consumer Protection Act, which was really specific to deal withfraudulent people misusing domain names and the names of folks like Wharton or something
(19:31):
like that and causing mischief.
You could see some parallels to that today.
You probably know about more specific legislation like the DeepFace Fakes Act and similarthings done at least right now.
efforts made on the state level concerned about uh especially right of publicity issues,privacy issues, things like that.
(19:58):
In other words, there could be broad legislation or there could be more narrow castinglegislation depending on what the issues may be.
And it could be that we end up with a broad legislative solution or a more narrow.
legislative solution, depending on the will of Congress and how such a bill might bewritten.
(20:26):
And you could also end up with a situation where uh the balancing, the coalition is notthere and we need to have the courts spend more time.
uh
figuring out a solution, which may not be a bad thing because as time goes on, you learnmore and you see more of the consequences.
(20:57):
And sometimes you can make a better decision when you have more information about uh thepotential impacts.
So it's too early to tell though as to whether here we're going to end up with.
em
uh a giant legislative solution or not.
(21:18):
Okay, why don't we move on to the output side?
We've been focusing predominantly on the training question, but what are the major issuesaround whether a work created with genera.ai infringes on?
So many cases that have been brought that are input cases that are complaining about thetraining are also output cases because it's kind of hard to have the outputs be similar
(21:46):
and this would be your output, you're letting users come up with things that aresubstantially similar.
That's the magic phrase for copyright infringement if the works are substantially similar.
So that the idea is that the outputs, the photographs or the song lyrics or the book orwhatever is substantially similar to my copyrighted work.
(22:11):
Sometimes the training needs to have been fed in with that material for the outputs to dothat, although not necessarily.
So the output cases uh focus much more on, hey, I'm being harmed because
Instead of coming to me, copyright holder, for my photograph, my character, whatever itis, a user is able to type in what they want and get that image and they don't have to
(22:44):
come to me for a license or hire a graphic designer or someone like that for that purpose.
And most recently, very recently, this week from when we're talking, Disney
Ununiversal filed a lawsuit against Mid Journey, which is one of the AI platforms thathelps to create uh visuals.
(23:06):
It's not a text focus, the focus is on the visuals.
And their complaint had all sorts of images of Elsa and other princesses that weregenerated using Mid Journey or Marvel superheroes and Simpsons characters and
and the like, and the concern was, hey, you're using my IP improperly.
(23:32):
And this is the first time the movie studios have really come together to file a lawsuitlike this.
it's already gotten a lot of attention in the newspaper.
So we're gonna have to see how it plays out, but that's what the input cases focus on.
With that, though, just to be clear, is that there's uh the different AI platforms varysomewhat.
(24:01):
Some of them have put in place guidelines to try to prevent the ability for a user to comeup with something that looks like a copyrighted character or reveals too much of the text
of a book or of a particular article or something like that.
Others feel that, no, that shouldn't be their obligation and they shouldn't have to dothat.
(24:29):
It's the user that put in that request.
It's not their responsibility.
People who are on the side of the platforms will often argue something called volitionalconduct.
Is it really volitional conduct?
If I'm the one, the user who's typing something in saying, me X, uh they'll say that I'mnot really responsible for that.
(24:56):
Others would say, on the other hand, well, you set it all up that they can do this by partof what you train.
So you are responsible.
So those are going to be some of things that are going to uh be dealt with in some ofthese cases.
And this issue of to what extent uh
(25:17):
there should be guardrails or not uh is something that you're seeing litigated more andmore.
For example, in one of these music cases I mentioned, the Concord case, initially therewas a preliminary injunction filed and trying to get certain guardrails in place, but they
(25:37):
reached an agreement that the platform would put in certain guardrails that would make it
harder for people to get, for example, song lyrics, in which case the court said, themotion's now mute, moo, because people have agreed to this.
And there'll be a debate, and the cases are going to try to figure out, as I said, to whatdegree is the platform responsible for what the users do.
(26:05):
And there'll be a debate as to whether there should or shouldn't be these kind ofguardrails with regard to uh limiting.
what people can ultimately see.
Is this going to be a similarly fact-specific kind of inquiry?
Because even if there are guardrails, there are lots of examples you can find online oryou can probably find yourself with some of these tools where you type in and say, give me
(26:32):
Elsa from Frozen, and it says, I can't do that.
And then you say, give me a princess who looks like this, and so forth, and you cancircumvent.
So does it inevitably have to be a very specific inquiry about what's enough?
I, as any lawyer would say, it depends, right?
(26:55):
And because again, you can understand that part of it will depend legally on the thoughtabout the, to some degree, this volitional conduct idea.
But the other thing is if ever some of these cases get to a jury, juries may have todecide credibility issues,
(27:19):
responsibility issues and issues relating to uh how similar it is because uh people coulddisagree as to whether something really is a recognizable character or not.
There's things like scenes of fair, certain tropes are available for everyone to use.
(27:44):
uh On the other hand,
It's the same coloring and same hair and same design and same eyes and is returned by aspecific choice to that.
oh That makes it a little bit harder for someone to say it isn't a copy of that image.
And then the last major scenario I just want to ask you about is uh the question aboutwhether a general AI work itself is entitled to copyright protection.
(28:10):
And as I understand it, the copyright was just said if it's, there's no human involvementat all, you just push a button, no.
But what's the line beyond that where there's enough human involvement?
This is where I think that the law is starting to develop so people are starting tounderstand how it's supposed to work.
You're absolutely right, and this was in yet a different copyright office circularguideline that they report, the three reports we were talking about that came out.
(28:40):
And it did talk about the idea that generally speaking, the concept is if the
artwork or article or whatever it is was all created by AI, then that can't becopyrightable.
But the issue is it could be copyrightable if there was sufficient human involvement.
(29:01):
And then everyone's trying to figure out, what does that mean?
Well, you go back and you look at some cases and you can start to figure it out.
There was a Thaler case which went all the way up to the DC circuit where someone reallywanted to make the
point, he was really trying to say that AI should be allowed to copyright things.
(29:22):
And the court said, no, you're admitting that there was no human involvement.
So that's not good enough.
And to be clear, the Copyright Office report that I mentioned said that the level of humaninvolvement can't just be in coming up with prompts.
It has to be actually some act of creation beyond coming up
(29:44):
Well, then you compare the Thaler case with the Zariah the Dawn uh case involving agraphic novel.
And I think that really tells you a lot about it's a good practical guidance.
Let's put it that way for those of us advising clients on this, because originally uh thisperson had created a graphic novel, Zariah the Dawn, and they got a copyright registration
(30:11):
for it.
She did.
Well, then.
uh it was revealed to the Copyright Office that the graphic novel, the visuals to it atleast, had been created by Mid Journey, that same company that we were talking about a
minute ago.
And that she wrote the text and she arranged the material uh into a graphic novel, but theactual visuals were not something that she had created.
(30:39):
In which case, they amended the Copyright Office
uh registration to say that she had a copyright in the text and she had a copyright in theselection and arrangement of the text to the pictures.
In other words, she did the creative work of putting it all together into a graphic novel.
But the actual images that were generated by many detailed text prompts asking themajority to create X, Y, and Z wasn't something that could be owned.
(31:12):
by themselves.
So I think that what you're seeing from that is at least how copyright could be used tosay if it's pure AI, no, but you can still protect a lot because there's still a lot of
human involvement and what the humans do with the material that's created by AI.
(31:38):
that seems to be getting
more clarity compared to some of these other issues.
uh Both the content owners and AI companies and those that are advocating for them tend tomake statements that these are existential issues, how these cases come out.
(32:02):
How much of a business impact is it really um on either side, whether there's infringementin these kinds of cases?
Well, it could have a huge impact depending on what the use is and depending on what thebusiness is.
Something could go to the core of what that business does.
(32:22):
And when I say that business, that could be on both sides, the core of what the AIbusiness is trying to do and the core of what the copyright creator type of business do.
So it could have a big impact on both, which is why
Sometimes when there's a lot of risk for either or both sides, you end up withsettlements.
(32:47):
And certainly you're seeing, if not settlements, but you're reading it out in the press,different licensing schemes being agreed to.
There was, think, the New York Times recently did a deal, I think, with Amazon and there'sother.
(33:07):
as I already mentioned, at least some, not settlement, but some agreement on some of theguidelines in the Concord case.
So you're seeing that there could be an incentive to try to work out a resolution in somecases that's a compromise for everyone.
(33:27):
uh But sometimes that compromise is not very happy.
Sometimes it could uh still end up killing the golden goose on either side or not becommercially uh sensible.
(33:47):
And once had a court that uh was urging me to, you know, both the parties to settle and itwasn't an AI case or anything like that.
But I said, well, Your Honor, that would be like having us uh lose money on every sale.
but make it up on volume, which if you think about, you really can't do it that way.
(34:12):
So it has to work from a dollars and cents point of view.
Lawyers do tend to be extreme and make uh dramatic arguments.
Sometimes there's a reason though to do so.
Sometimes the consequences could be dramatic for
(34:34):
both sides of the equation.
And I'm not in any way minimizing that.
But it sounds like from the tenor of the conversation that you're, am I right to saybroadly optimistic that through some combination of litigation, maybe some legislation,
these private resolution of disputes that will ultimately navigate through thesecomplexities.
(34:58):
Yes.
And why?
Well, again, what's past is prologue.
You started off by asking, doesn't technology change?
And yes, it does.
And there have been lots of new innovations and they'll probably be even more innovations.
mean, who thought that we'd be living in a world you can have this amazing computer thatyou carry around with you that's also a camera and all sorts of other kinds of things.
(35:24):
mean, the technology.
changes all the time and I'm not as uh inclined to uh run around like Chicken Little andsay the sky is falling.
But I do think that it's uh useful to remember that there are ardently well-reasoned,thoughtful lawyers and companies on both sides of this debate and
(35:55):
given facts of different cases, different oxes that have been, or wrongs that have been uhdone, you could see a situation where one size doesn't fit all in all cases as to whether
something is fair or not fair, whether it's infringement or not infringement.
(36:19):
And what I think is really important is for people to listen to one another.
to try to come up with best practices and uh a path going forward that protects thetremendous, wonderful innovation on, I'm not even gonna say both sides of this debate, I'm
(36:42):
gonna say all sides of this debate because I think it's a table with uh many sides.
That's a great place to end.
Dale, thank you so much for your time.
It's always a pleasure.
Thanks for being interested.
Take care.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
New Heights with Jason & Travis Kelce
Football’s funniest family duo — Jason Kelce of the Philadelphia Eagles and Travis Kelce of the Kansas City Chiefs — team up to provide next-level access to life in the league as it unfolds. The two brothers and Super Bowl champions drop weekly insights about the weekly slate of games and share their INSIDE perspectives on trending NFL news and sports headlines. They also endlessly rag on each other as brothers do, chat the latest in pop culture and welcome some very popular and well-known friends to chat with them. Check out new episodes every Wednesday. Follow New Heights on the Wondery App, YouTube or wherever you get your podcasts. You can listen to new episodes early and ad-free, and get exclusive content on Wondery+. Join Wondery+ in the Wondery App, Apple Podcasts or Spotify. And join our new membership for a unique fan experience by going to the New Heights YouTube channel now!