Latest / CLEs You Actually Want to Hear / Oral Advocacy and Credibility at Trial: Lessons from High Stakes AI Litigation
Transcript
- 0:03Welcome to Oral Advocacy and Credibility at Trial,
- 0:07Lessons from High -Stakes AI Litigation. I'm
- 0:11Laurie Critt -Martin, a Certified Family Law
- 0:13Specialist and a former Vice Chair of the BHBA's
- 0:17Litigation Section. I'm delighted to introduce
- 0:20our speaker, Angela Denning. Angela Denning is
- 0:23a partner at Cleary Gottlieb, recognized by Chambers
- 0:26USA in both Artificial Intelligence and Intellectual
- 0:30Property Trademark. copyright, and trade secrets.
- 0:33Her practice centers on generative AI, intellectual
- 0:37property, appellate litigation, and complex commercial
- 0:41disputes for many of the world's leading companies.
- 0:44And she has argued successfully before the US
- 0:47Court of Appeals for the Ninth Circuit in numerous
- 0:49published cases. Earlier in her career, she clerked
- 0:53for Chief Magistrate Judge Gregory G. Hollows
- 0:56of the U .S. District Court for the Eastern District
- 1:00of California, and she served as an Assistant
- 1:02District Attorney in San Francisco. She now teaches
- 1:06trademark law at UC Berkeley School of Law and
- 1:09was named Copyright Practitioner of the Year
- 1:11by Managing IP's Women in Business Awards. And
- 1:15now let's get started. Well, thank you so much.
- 1:19Laurie, and thanks to everyone who is joining
- 1:21us today. It's great to be with you. As Laurie
- 1:24said, I'm Angela Dunning. I'm an IP litigation
- 1:27partner at Cleary Gottlieb. I'm based in the
- 1:29Bay Area, but I grew up in Southern California
- 1:32and so feel very at home with all of you. I'm
- 1:36going to be sharing some of the recent developments
- 1:42across the generative AI litigation landscape,
- 1:47including both with focus on copyright and trademark
- 1:51litigation and developments. There have been
- 1:55some late breaking developments as recently as
- 1:58last night, so you'll get a very timely presentation,
- 2:01I hope. Let me just see if I can share my slides
- 2:04and we'll dive right in. with this quote from
- 2:22Arthur C. Clarke that any sufficiently advanced
- 2:25technology is indistinguishable from magic. And
- 2:28for those of us who have been practicing long
- 2:30enough, we have seen an awful lot of magical
- 2:34technology come our way from the internet to
- 2:38the iPhone and just from tapes and VHS to you
- 2:48know, streaming video, it's been an incredible
- 2:51journey over the last few decades that I have
- 2:55been practicing. But I don't think we've ever
- 2:58seen anything quite as transformative and truly
- 3:03revolutionary as generative AI. So it is being
- 3:10used to create just amazing works of art. I'll
- 3:14talk a little bit about this further in the presentation,
- 3:17but this work, Teatro de Paras Basial, won first
- 3:21prize at a Colorado art contest at the Colorado
- 3:26Fair a few years ago, and is now the subject
- 3:30of a challenge as to whether this can be owned
- 3:32as a human created work. AI is being used to
- 3:38create films or parts of films and to create
- 3:42new art forms and enable art by people who have
- 3:48never made it before. I just, I love the work
- 3:51by Chelsea Jones, who is an AI artist who has
- 3:57been showing amazing pieces. And it can also
- 4:01be used to create real world items. The image
- 4:05on the left is one that I created using Dali.
- 4:12It's a litchi -inspired chair. But your ability
- 4:17to create things is limited only by your imagination.
- 4:20And then real world products that could be made
- 4:23and manufactured today are just potentially at
- 4:28fingertips of people who never maybe considered
- 4:31themselves a manufacturer or designer before.
- 4:35And of course, some of the most promising research
- 4:38and development is being conducted in the context
- 4:41of scientific discovery, the finding of new therapies
- 4:48for disease, the diagnosis of disease is being
- 4:52aided, and truly remarkable things that will
- 4:54have the capability to affect all of us are coming
- 4:57online. We are also seeing, of course, many controversies
- 5:04arise, and that stems from a potential bias in
- 5:09the way these models were trained and the outputs
- 5:12that are generated where, you know, a request,
- 5:15for instance, for an image of a 1943 German soldier
- 5:20turned up images of officers of different ethnic
- 5:24backgrounds that would not be at all true to
- 5:27life, you know, to the generation of deep fakes,
- 5:31interference with election, use in film, and
- 5:35use to potentially mimic the voices or... image
- 5:40of celebrities. So these are all real considerations
- 5:44that need to be taken into account in how we
- 5:47legislate and move forward. Of course, it has
- 5:52also led to just an extraordinary amount of litigation.
- 5:56I don't expect anybody to be able to process
- 5:58all of these cases, but at last count there were
- 6:01approximately 140 cases pending across the country.
- 6:05I am handling quite a few of those, mostly in
- 6:09the Northern District of California in San Francisco,
- 6:12where I'm based, but also in New York. And these
- 6:16cases are at various stages of resolution, and
- 6:20I'll talk about where we are. So I planned to
- 6:26cover Copyright issues, trademark and right of
- 6:30publicity issues, and then some privacy and regulatory
- 6:33concerns. And certainly we'll try to leave some
- 6:35time for questions at the end and pepper in some
- 6:39stories about how these issues have arisen and
- 6:42what it has been like to litigate them in court.
- 6:48So let's start with copyright. There's an underlying
- 6:55question before we ever get to the litigation
- 6:58over training of generative AI models or copyright
- 7:02infringement and fair use of whether we can own
- 7:09the outputs of generative AI models. This is
- 7:12one of the things that artists and businesses
- 7:18and people of all backgrounds are considering
- 7:21when trying to figure out whether they should
- 7:24be using these models and for what. So we've
- 7:29got the question of whether the outputs of these
- 7:34models are human authored or whether they are
- 7:39the product of a a machine, a mechanical process.
- 7:44And this issue really first came up all the way
- 7:47back in 1884 when the camera was still new and
- 7:51there was some question as to whether a photograph
- 7:54could be entitled to copyright protection. The
- 7:58U .S. Supreme Court held that it could in this
- 8:01case involving this photograph of Oscar Wilde.
- 8:06found that the photograph was a useful, new,
- 8:10harmonious, characteristic and graceful picture
- 8:13and that the contribution of the photographer
- 8:16was in the way he arranged the photo and chose
- 8:22the lighting and the shading and got the expression
- 8:26and picked the lens and the filters for the camera.
- 8:30And that this arrangement and this set of human
- 8:36choices about what to display was sufficient
- 8:40to demonstrate sort of the the human authorship
- 8:44and creativity necessary for copyright protection.
- 8:48Well fast forward to about at this point maybe
- 8:5412 or 13 years ago, this principle got put to
- 9:00the test in a case called Naruto the Slater that
- 9:04I handled. The story is pretty well known at
- 9:08this point and this is the monkey selfie copyright
- 9:11case, but a photographer from England flew to
- 9:14Sulawesi in Indonesia to film these crested macaques
- 9:18in the wild. They were highly endangered and
- 9:20he wanted to bring attention to their state of
- 9:25danger in the world. The story goes that upon
- 9:29arriving there he had no luck in getting them
- 9:32to pose for the camera and so instead he set
- 9:35out a lead and the macaques discovered it and
- 9:40loved the clicking sound and started taking monkey
- 9:42selfies. So he gets back to London and he sees
- 9:45the film and he's overjoyed because he figures
- 9:48he's got the next cover of National Geographic,
- 9:51but that is not what happened. Instead, he created
- 9:56a book using my client Blurb's self -publishing
- 9:59platform, and he put it out and sought copyright
- 10:02protection. But the U .S. Copyright Office rejected
- 10:06protection on the ground that he had not taken
- 10:09the photos. Instead, the monkeys had taken the
- 10:12photos because they pressed the shutter. Then
- 10:16Wikipedia published his photographs from his
- 10:19book online, making his book difficult to sell.
- 10:25And then the third shoe dropped when the monkey,
- 10:27in this case, this monkey, Naruto, filed a copyright
- 10:31infringement lawsuit in the Northern District
- 10:33of California. Now I handled the case for blurb
- 10:37and I argued the motion to dismiss in the district
- 10:40court. And then I defended the company at the
- 10:44Ninth Circuit when the monkey appealed at the
- 10:47ruling that he lacked standing to enforce the
- 10:51Copyright Act because he was not a human. And,
- 10:55you know, this, this case is the best case I
- 10:59think I'll ever have. It's not every day a monkey
- 11:01sues. If you don't, if you don't win against
- 11:04a monkey in court, you really ought to find another
- 11:07job, because that can be career limiting. But
- 11:09I am pleased to say that we did prevail, and
- 11:13we prevailed on this notion that only human authorship
- 11:17is covered by the Copyright Act. And we looked
- 11:21to the language of the Copyright Act. talks about
- 11:25widows and orphans and signatures, and although
- 11:28the Copyright Act does not define the term author,
- 11:32it plainly speaks in human terms, just as the
- 11:36Burro Giles case has and pretty much the major
- 11:40decisions in the hundred years in between. And
- 11:45so the ruling from this case was that only humans
- 11:50have standing to enforce the Copyright Act. I
- 11:54won't say it wasn't surreal, being at the Ninth
- 11:57Circuit and getting the following question. Council,
- 12:01what financial interest does Naruto have in the
- 12:05copyright? To which I answered, Your Honor, he
- 12:09has no financial interest in the copyright. He's
- 12:12a monkey. That was the witting argument there.
- 12:18But I was conscious at the time that this you
- 12:22know, circumstance, although not likely to recur,
- 12:27really had implications for output from AI models
- 12:34and whether AI outputs could be deemed to be
- 12:41human authored or not. We were careful to seek
- 12:46a ruling that human authorship is required without
- 12:50defining what that specifically meant because
- 12:55we didn't know how the technology would develop
- 12:58or how quickly and develop it. It did and much
- 13:02quickly much more quickly than I anticipated.
- 13:05I will confess. Now there are other processes
- 13:09that have been, you know, looked at in terms
- 13:11of the. subject of copyright. You have stop motion
- 13:16animation. You have processes by which there
- 13:19really isn't human control in the generation
- 13:22of the work, but we nevertheless grant protection.
- 13:26But in the context of generative AI outputs,
- 13:33the Copyright Office has taken a very narrow
- 13:36view. The Copyright Office has said that if the
- 13:41traditional elements of authorship were produced
- 13:44by a machine that is a generative AI model, the
- 13:47work lacks human authorship and the office will
- 13:50not register it. And they have denied registration
- 13:54to numerous works on this basis. And the office
- 13:59has discussed this concept of whether a work
- 14:05contains sufficient human authorship to support
- 14:07a copyright claim, notwithstanding that AI is
- 14:11used to some extent. But for the most part, the
- 14:16current guidance states that if it is AI generated,
- 14:19it is not protectable. And so in this case involving
- 14:26Stephen Thaller's creation, a recent entrance
- 14:30to paradise on his, what he called his creativity
- 14:34machine. When he sought to register this work
- 14:36with the Copyright Office, he disclaimed any
- 14:41human authorship and said that it was created
- 14:43exclusively on autonomously by his AI and the
- 14:48Copyright Office rejected registration and that
- 14:51was affirmed on appeal. to the district court
- 14:55and then to the D .C. Circuit. And finally, the
- 14:58Supreme Court denied cert. And this was the easy
- 15:01case because no human authorship was claimed.
- 15:05And he tried to claim human authorship later
- 15:08in the process, but his application had had disclaimed
- 15:11it. And on that basis, the courts considered
- 15:15that to be conclusive. And this work was not
- 15:19deemed protectable. But it's harder when you
- 15:23get to works where humans are claiming that they
- 15:26do in fact have significant human involvement
- 15:28and contribution in the creation of the work.
- 15:31In the Zarya of the Dawn work, which is created
- 15:36using mid -journey by an artist called Kishtanova,
- 15:43They created a graphic novel using mid -journey
- 15:47to create the images and then they wrote the
- 15:50text and made the arrangement of the piece. And
- 15:53the copyright office originally granted protection,
- 15:55but once it found out that the images were AI
- 15:58generated, it withdrew that protection. And after
- 16:02some back and forth, the office determined that
- 16:06the text and the arrangement were protectable,
- 16:09but not the images because they had been generated
- 16:11using mid -journey. the author could not show
- 16:14that they had exerted sufficient control over
- 16:17the output to claim human authorship. As I mentioned,
- 16:22in the case of Mr. Allen's work, he was also
- 16:27denied protection, even though the creation of
- 16:30this work involved hundreds of iterations in
- 16:34which he painstakingly directed the development
- 16:38of the work using mid -journey. And he has sued
- 16:42the copyright office to require registration
- 16:46of this work. And his action is pending in the
- 16:50district court in Colorado, where summary judgment
- 16:54has been fully briefed for some time, and we
- 16:57are awaiting a ruling. And that's important because
- 16:59the courts, rather than the copyright office,
- 17:02will have ultimate authority over whether a work
- 17:05is protectable and whether the human authorship
- 17:11requirement is satisfied. In the case of Rose
- 17:15Enigma, which you see on the bottom left of your
- 17:17screen, this is another work by Kostanova, and
- 17:21they obtained a registration for the pen drawing,
- 17:25but then used stable diffusion to create the
- 17:30richer color version of the image. The Copyright
- 17:33Office found that the color image, to the extent
- 17:37it was different from the hand drawn version
- 17:41is not protectable because that was all AI generated
- 17:46and not human generated. Now again, The work
- 17:49was supported by extensive choices that went
- 17:51into picking exactly how this would look, but
- 17:55it was nevertheless rejected. And the reality
- 17:58is you can't recreate or reproduce the color
- 18:03image without infringing the underlying registration,
- 18:07which is maybe enough to protect it. And that
- 18:11takes us to a single piece of American cheese,
- 18:13which was an application filed by some friends
- 18:16of mine, former colleagues. who helped a client
- 18:20with an application for this work, which involved
- 18:25use of an AI tool to place piece of the work
- 18:30in the image. And so although every part of this
- 18:33work is AI generated, the Copyright Office ultimately
- 18:37found that it could be protected by copyright
- 18:40and granted registration over the selection,
- 18:44coordination, and arrangement of the material
- 18:47in the work. And so the key takeaway from these
- 18:52examples is that, you know, the The human aspect
- 18:58of control, where I am making choices about what
- 19:02goes where, where I am using an AI tool to place
- 19:06certain pieces in an image, to place certain
- 19:10pieces of a text in a particular order, to potentially
- 19:14create an order code in a way that is reflective
- 19:20of human choice and control, may be sufficient
- 19:25to convert an otherwise unprotectable work into
- 19:29one that can show the necessary human authorship
- 19:33to be registered. In the case of this film, Critters,
- 19:40not every part of it can be registered, the Copyright
- 19:44Office has held. The initial 2D renderings of
- 19:49certain images generated by the AI were excluded
- 19:52from the registration. But of course, the registration
- 19:55covers the human authored script, the human modifications
- 20:00to that AI generated output, the way that the
- 20:04characters and output has been selected and arranged,
- 20:07and then the human performances in terms of voice
- 20:10that went into the film. And so for that reason,
- 20:13there are multiple layers of copyright protection
- 20:15for this film, even if some of the AI generated
- 20:18portions are not protected. The U .S. is, of
- 20:27course, one country and other countries are reaching
- 20:32different outcomes. So outside of the US, for
- 20:35instance, in China, the courts have been willing
- 20:38to find not only that AI generated outputs are
- 20:43protectable by copyright in China, but also that
- 20:47reproduction of those images is infringing. So
- 20:51this image was, one, created using an AI tool,
- 20:56the court found that it was protected. And then
- 20:59when someone tried to make this image in real
- 21:02life or to use it in advertisements, even with
- 21:06some modifications, the court found that it was
- 21:10infringing and awarded damages. We've seen some
- 21:14other examples of this. The Beijing Internet
- 21:17Court found the AI -generated image of this girl
- 21:21protectable and awarded. relief in this serious
- 21:29image was one that an artist made taking a photograph
- 21:33he had taken and putting it into an AI model
- 21:37and asking it to generate an image sort of affected
- 21:42by the Van Gogh work in the middle to create
- 21:46the combined. composite derivative image on the
- 21:51right. And the US denied registration for that,
- 21:55claiming that he didn't have control over what
- 21:57came out. But Canada and India both registered
- 22:00the copyright in that work. And there are other
- 22:04countries, so UK, Ireland, Hong Kong, India.
- 22:07New Zealand and South Africa that have recognized
- 22:11copyright in computer generated works. How that
- 22:15will be interpreted in the context of generative
- 22:17AI remains to be fully fleshed out. But we are
- 22:21aware of at least one German decision that no
- 22:24found no copyright in AI generated icons because
- 22:28the AI was used. not as a tool, but as the instrument
- 22:34to create these, and they therefore lacked that
- 22:37personal creativity. So, key takeaways here.
- 22:44Human authorship is required for copyright protection
- 22:47in the U .S., and the need for a human creator
- 22:53has been consistently held to apply. prompts
- 22:58are probably not enough. The fact that you do
- 23:01a really detailed or creative prompt is probably
- 23:04not enough to make you the author in the US.
- 23:07Even if the prompts themselves are, you know,
- 23:12complicated and creative enough that they would
- 23:14themselves qualify for protection as a textual
- 23:18work. because it's the Copyright Office reasons.
- 23:22It's the AI that is creating the work from the
- 23:24prompts and not the human that is authoring them.
- 23:29But where you can show substantial creative decision
- 23:31making, where you can claim protection for the
- 23:34arrangement and selection of the images or...
- 23:38text, where you are making significant contributions
- 23:42through iterative prompting and blending elements
- 23:45with your own original works, that is most likely
- 23:49to generate the path to protection. And it's
- 23:53important to document your process. The Copyright
- 23:57Office requires you to disclose when you have
- 24:01used AI in the creation of a work, not in any
- 24:03great detail, but if you are going to disclose
- 24:06it, then it's important to be able to document
- 24:09what you contributed, both to support a claim
- 24:12of protection in the office and also to support
- 24:15a claim of authorship and protectability if you
- 24:18need to enforce against others or demonstrate
- 24:22that your work is protectable. I'll just briefly
- 24:27note The guidance from the U .S. Patent and Trademark
- 24:31Office with respect to the patentability of inventions
- 24:35that are assisted by AI is a little more generous.
- 24:40There, so long as there is sufficient human authorship,
- 24:44inventorship such that a human qualifies as an
- 24:47inventor, then the fact that AI was used in supporting
- 24:51the invention should not mean that it is unpatentable.
- 24:57That said, use of AI to actually reduce an invention
- 25:07to practice and to come up with all of the ideas
- 25:10and actually make it may not be sufficient to
- 25:13show the human inventorship required. So let's
- 25:20talk now about Copyright infringement, and this
- 25:23is where I've spent most of my time since the
- 25:25first of the copyright class action lawsuits
- 25:28was filed in the U .S. in January of 2023. Copyright
- 25:33infringement is the exercise of an exclusive
- 25:37right without permission, and those rights are
- 25:40enumerated. They are reproduction, that's to
- 25:42make copies, to prepare derivative works, to
- 25:45distribute the works, to perform or display them.
- 25:48And infringement requires substantial similarity
- 25:51of protected expression. So just because I'm
- 25:54inspired by a work doesn't mean that it is infringing.
- 25:58It actually has to have substantial similarity
- 26:02in the parts that copyright law protects. And
- 26:05that's an important qualifier because copyright
- 26:08law doesn't cover everything in a work. It only
- 26:11covers the original expression. It doesn't cover
- 26:18ideas, it doesn't cover facts or styles or content
- 26:22in the public domain. It really only protects
- 26:25the way in which your ideas or facts are specifically
- 26:28expressed and only the expression that is original
- 26:31to you. And so separating those things out and
- 26:35figuring out if there's substantial similarity
- 26:37between what is protected makes work for us copyright
- 26:42lawyers. Importantly, making a fair use is not
- 26:47an infringement. So how do we assess fair use?
- 26:53Well, there are four factors written into the
- 26:56fair use statute, which is section 107 of the
- 26:59Copyright Act. And the courts look to the purpose
- 27:03and character of the use first and foremost.
- 27:06And you may have seen this described as whether
- 27:10the work is transformative. Does it serve a purpose
- 27:14and function different from the original or is
- 27:18it basically serving the same purpose for which
- 27:21the work was created? We also look to the nature
- 27:24of the copyrighted work with whether the work
- 27:31was published or not and whether it is sort of
- 27:34creative and expressive as opposed to, you know,
- 27:37strictly mechanical or technical. We looked at
- 27:41how much was used in the context of training
- 27:44an AI model. Whole works are typically used and
- 27:50lots and lots of them, millions or billions of
- 27:53works are copied for purposes of converting the
- 27:59information in those works into what we call
- 28:02tokens, words or pieces of words that are then
- 28:05in turn turned into mathematical representations
- 28:09and then assessed for their relation to other
- 28:14words in a particular work and across the entire
- 28:18massive set of training data. The final factor
- 28:25is whether the use will cause substitute of harm
- 28:28in the market for the original. That is, if I
- 28:30can get you know, the author's protected expression
- 28:33from the infringer instead of the author, then
- 28:37that may cause harm to the author's market that
- 28:42would weigh against a finding of fair use. But
- 28:45context really matters. So this is an image that
- 28:49I created using mid -journey. And you can bet
- 28:54that if I tried to emblazon this, this is an
- 28:58early version of Mid Journey before they got
- 29:01lettering down pat. But if I had tried to emblazon
- 29:05this on a t -shirt, I probably should have expected
- 29:08Coca -Cola to send me a cease and desist letter
- 29:11with respect to their trademark rights. And Disney
- 29:14may have had something to say about use of a
- 29:17depiction of Snow White. the evil witch. But
- 29:22if instead I produced this image for purposes
- 29:26of illustrating an article I wanted to write
- 29:29about how sugar soda has become today's modern
- 29:34poison apple, then there would be a strong argument
- 29:39that this is a fair use serving a completely
- 29:42different purpose that doesn't undermine the
- 29:45market for the original and is used to convey
- 29:49a completely different message. So it can't be
- 29:53evaluated in a vacuum. It really has to be evaluated
- 29:56in the context of the actual use to which it's
- 29:59being put. And that's the core lesson that came
- 30:02out of the Andy Warhol case that the US Supreme
- 30:05Court decided a few years ago. We look at the
- 30:08exact use at issue and we don't conflate them
- 30:11and we apply the factors to assess whether those
- 30:14uses are fair. District courts in the US have
- 30:21made a determination thus far as to whether training
- 30:24a large language model on copyrighted text, specifically
- 30:28lots and lots of copyrighted books, without permission,
- 30:32constitutes fair use. And both in the Barts v.
- 30:35Anthropic case and in the Khadri v. Mehta case,
- 30:38the courts found that it was fair use. The courts
- 30:44both looked at factor one and found that that
- 30:47weighed in favor of fair use. These are both
- 30:51in the Northern District of California. In the
- 30:54Bart's case, Judge Alsop found that the use of
- 30:59books for training was highly transformative.
- 31:02He said spectacularly so among the most transformative
- 31:05technology many of us will see in our lifetimes.
- 31:09And Judge Chabria found the same in the Kadri
- 31:11V. Meda case that I handled with a number of
- 31:16excellent teammates. But in both looked at you
- 31:24know, individual uses. In the case of Judge Alsup
- 31:28in the Bart's philanthropic case, he found the
- 31:30use for training highly transformative. He also
- 31:33found the taking of books and digitizing them
- 31:37for training to be a format change that was nevertheless
- 31:40transformative for purposes of creating this
- 31:42new model that could make new things. However,
- 31:45he found that it was a separate use to take copies
- 31:49from what he called pirate websites to create
- 31:53a general purpose library that may or may not
- 31:56be used for training. And he refused to find
- 32:01that summary judgment on fair use grounds could
- 32:03be granted as to that particular use. So focusing
- 32:09on the use for training in general, both courts
- 32:13found that the use of creative works weighed
- 32:15slightly against fair use, but that the third
- 32:19factor was basically neutral or weighed slightly
- 32:22against because although the entire books were
- 32:26used, that was necessary for training. And then
- 32:30on factor four, both courts sort of reached the
- 32:34conclusion that factor four weighed for fair
- 32:35use, but they got there in very different ways.
- 32:39Judge Alsop in the Bart's case. found, as did
- 32:43Judge Chabria, that there is no right to a licensing
- 32:47market for AI training. The book authors in both
- 32:50cases argued that they were entitled to a license
- 32:53to use their works for training. And the court
- 32:56said no. For transformative use, like a parody
- 32:59or making something new, you have no right to
- 33:02monopolize that market or require a fee. They
- 33:07also the same with respect to any claim that
- 33:13there could be direct substitution and outputs.
- 33:15In both cases, there was no allegation or facts
- 33:19to establish that the outputs of the model looked
- 33:23like the plaintiff's books. However, they diverged
- 33:29on What would constitute market harm? Judge Alsop
- 33:34in the Barnes case held that it would not satisfy
- 33:38the market harm factor to show that the Claude
- 33:43model is capable of outputting books like the
- 33:48plaintiffs. He said that's not the kind of harm
- 33:51that copyright law recognizes. Copyright law
- 33:55recognizes only harm where the the accused work
- 34:01is substantially similar to the one that is thought
- 34:05to be protected. And so if I can generate a mystery
- 34:09novel using Claude, but that mystery novel doesn't
- 34:12actually have a substantially similar expression
- 34:15in it to one of the mystery novels on which it
- 34:18trained, that is not a harm that's going to affect
- 34:20the factor four analysis. Judge Chabria, on the
- 34:23other hand, found that it could. He announced
- 34:28a new theory of market harm, which he referred
- 34:32to as market dilution, premised on the notion
- 34:35that so many mystery books could potentially
- 34:38be generated using an AI model that it could
- 34:42make it hard for the authors whose works were
- 34:45trained to find a market and that they could
- 34:51be harmed. But he found that the plaintiffs had
- 34:57not come forward with any evidence to support
- 34:59this theory. They hadn't shown that in fact they
- 35:03were suffering any harm or that any such works
- 35:06were being created using Metta's models. And
- 35:09so for that reason, the court in Kadri found
- 35:14fair use at summary judgment and gave judgment
- 35:18to Metta. In the Bart's case, the court left
- 35:22open this claim based on the downloading of pirated
- 35:27content, then granted class certification of
- 35:32hundreds of thousands of book authors. And that
- 35:35ultimately led to a settlement agreement that
- 35:38was approved on a class -wide basis for about
- 35:41$1 .5 billion, which translates to several thousand
- 35:45dollars per author and publisher. With respect
- 35:50to the Khadri case, the court left for trial
- 35:53a claim that META had procured some of the training
- 35:58material via torrent and that this could have
- 36:02involved... distribution of some of those works
- 36:06via the torrent network. Now that is a disputed
- 36:09fact that is still being litigated, but importantly,
- 36:12in the Kadri case, the court found that the fact
- 36:15that Meta may have sourced some of its training
- 36:17data from these websites did not move the needle,
- 36:21did not change the fair use analysis. Unlike
- 36:23in the Anthropic case, the source didn't matter.
- 36:27It was what Meta did with it, and what Meta did
- 36:29with it was transformative. Now, the late breaking
- 36:33news I mentioned is that last night, the DOJ
- 36:37filed a statement of position in the pending
- 36:40case against open AI, which is in the Southern
- 36:43District of New York and combines both book authors
- 36:45and claims by the New York Times. And in that
- 36:48case, the government has come out with a full
- 36:52-throated position in favor of fair use at training.
- 36:56And the government has has basically taken the
- 37:01position, the training of models and is important
- 37:05for our national security, that it is important
- 37:09national interest that we promote the progress
- 37:12of science and the useful arts by enabling this
- 37:15technology, which so many use to create new things,
- 37:19that it would be unfair and harm competition
- 37:22to require a licensing of training material because
- 37:27that would vest in only the biggest, richest
- 37:30companies, the ability to create these models
- 37:32and compete in those markets. And that this highly
- 37:38transformative technology should be allowed that
- 37:42the market dilution theory advanced in the Kadri
- 37:47summary judgment decision was, quote, deeply
- 37:50flawed, unquote, because it was looking at general
- 37:55competition for books and articles, as opposed
- 37:58to the kind of harm that copyright law is concerned
- 38:02with, which is substitution in protected expression.
- 38:06And so that case is proceeding to summary judgment.
- 38:09The opening papers are going to be filed in just
- 38:11a couple days, and the full briefing should be
- 38:14concluded sometime in October. We have cases
- 38:19across the country that are going to be dealing
- 38:20with these issues. But thus far, the direction
- 38:23of travel is toward fair use in training. I'll
- 38:28just mention very briefly, mindful of time, that
- 38:31we had another huge copyright decision out of
- 38:36the US Supreme Court this year in Cox. And this
- 38:41case dealt with a question of whether Cox Communications
- 38:45and Internet service provider could be liable
- 38:49for contributory copyright infringement for failing
- 38:53to kick a bunch of infringing infringers off
- 38:57its platform. And after a winding set of lower
- 39:04court decisions, the court ultimately held in
- 39:07March that Cox could not be contributory liable
- 39:11for the user's piracy. That the fact that it
- 39:15allegedly had knowledge of the infringement through
- 39:18takedown notices was not enough. Instead, to
- 39:22be contributory liable for the infringement of
- 39:25someone else, you must have intended that the
- 39:29service you provided to that infringer be used
- 39:32for infringement. An intent can be shown in only
- 39:35two ways, according to the court. First, if the
- 39:40defendant induced the infringement through active
- 39:42encouragement, come use our website for infringement.
- 39:45Think Napster or Grokster back some time ago.
- 39:51Alternatively, if the platform or service is
- 39:55specifically tailored to infringement, that is,
- 39:57it is an infringement machine that is not. capable
- 40:00of substantial or commercially significant non
- 40:03infringing uses, then that may be sufficient
- 40:06to show the intent needed for contributory infringement.
- 40:10Otherwise, the law imposes liability on those
- 40:16who actually do the infringing, and it is more
- 40:20difficult under this current standard, which
- 40:22is being applied across the AI cases. To impose
- 40:29contributory infringement on on one party for
- 40:35conduct committed by another. Touching on trademark
- 40:41and write a publicity in the context of AI cases.
- 40:45So I represent mid journey in connection with
- 40:49the. the first copyright class action filed back
- 40:53in January of 2023. This is Anderson versus Stability.
- 40:58And so Stability AI and Mid Journey, as well
- 41:01as Deviant Art, an art website, and Runway AI,
- 41:06a video generation company, have all been sued
- 41:10in this class action for copyright infringement
- 41:14in training. But there is also a claim against
- 41:18mid -journey in this case, for vicarious trade
- 41:22dress infringement. And this claim stemmed from
- 41:26an earlier version of the complaint, which had
- 41:29attempted to assert copyright protection over
- 41:32art styles. And the court dismissed that claim,
- 41:35finding again that copyright law does not protect
- 41:38style or ideas or facts, only expression, the
- 41:42specific expression. Style is an idea. And so
- 41:47when the plaintiffs amended their complaint,
- 41:49they added a claim for vicarious trade dress
- 41:52infringement, basically asserting that their
- 41:54art styles constitute their trade dress and that
- 41:58it is being violated when people make images
- 42:02on mid -journey in those plaintiffs' styles.
- 42:07This is, as far as I know, the only case in the
- 42:11country in which this claim is being asserted.
- 42:16And I can share the arguments that have been
- 42:20made in the case, which is, first, there is no
- 42:24protection for art styles. This is basically
- 42:26a copyright case dressed up as a trademark case
- 42:30claim. And the US Supreme Court in Daystar warned
- 42:36against trying to obtain through trademark. means
- 42:45rights that copyright does not allow. But there's
- 42:51also an issue with how the trade dress has been
- 42:54identified, because in the case of Sarah Anderson
- 42:57at the top, her style allegedly consists of simple
- 43:03cartoony and often strictly in black and white.
- 43:06But that It's very similar to one of her co plaintiffs,
- 43:10Julia Kay, whose images claim style is these
- 43:14three panel black and white comics loosely inked
- 43:17with a thin pen, fixed with pen. and both reflect
- 43:23styles that have been used by many, many, many
- 43:25artists and cartoonists for a very, very long
- 43:29time. If one artist could own that as a style
- 43:33and assert trade dress protection for it, it
- 43:36would effectively box up and preclude artists
- 43:39from building on styles that should be free for
- 43:42all to use. and harm the very artists who are
- 43:47asserting this trade dress protection if they
- 43:50are exposing themselves to potential claim every
- 43:54time they generate a new comic. Similar, Greg
- 43:57Rakowski is one of the named plaintiffs. He's
- 43:59well known for his fantasy themes, but he's not
- 44:03the only one. He didn't originate dragon paintings.
- 44:08Many, many others do dragon paintings, including
- 44:10another of the co -plaintiffs. And so again,
- 44:14the notion that we can walk up through trade
- 44:16dress rights, art styles that we think everyone
- 44:21should be free to build on really risks potentially
- 44:26damaging. the ability for art movements to arise
- 44:32and for artists to take styles and ideas forward.
- 44:38We also have challenged whether plaintiffs can
- 44:41show secondary meaning, that is, whether they
- 44:43can show that these trade dress elements serve
- 44:46as a designation of source for them specifically.
- 44:49That's an element that they need to prove and
- 44:52we think will be difficult. And then there's
- 44:55a question about functionality. None of these
- 44:57trade dress claimed trade dresses are registered
- 45:01under the with the US Patent and Trademark Office.
- 45:06And so plaintiffs have the burden to show that.
- 45:08the aesthetic is not functional. But in the context
- 45:12of an artwork, there's an argument to be made
- 45:14that how it looks is its function, and functional
- 45:19trade dress is not protectable by law. And so
- 45:22these arguments are being made and played out
- 45:26in the cases as we speak. We also have a couple
- 45:33of other claims, cases involving Trademark writes
- 45:38the New York Times claims, for instance, that
- 45:40when chat GPT generates outputs that include.
- 45:46segments of New York Times articles or articles
- 45:50that are attributed to the New York Times that
- 45:51are hallucinated, its trademark is being misused
- 45:55to designate as a source of works that are not
- 45:58New York Times material in violation of its trademark
- 46:01rights. Similarly, Getty Images has sued Stability
- 46:05AI for allowing the generation of of outputs
- 46:11that show the Getty images watermark. This case
- 46:15has taken a long time to wind its way through
- 46:18to get to the right court and but it has been
- 46:21allowed to proceed. A similar case brought in
- 46:24the UK found that this use of the watermark did
- 46:30constitute trademark infringement, but the. instance
- 46:34of it were so minimal and the issue had been
- 46:38resolved in subsequent versions of Stability's
- 46:40model such that there was no demonstrated harm
- 46:43and so it dismissed the claim. Right of publicity
- 46:49is another big area and one that I think affects
- 46:52a lot of folks in Hollywood and Beverly Hills
- 46:58for sure, but it's. There's no federal right
- 47:02of publicity. Each state has its own laws, and
- 47:07that makes it hard to keep track of. But essentially,
- 47:12the right of publicity extends to use of one's
- 47:15name, voice, signature, photograph, or likeness
- 47:19without consent in advertising, selling, or promoting
- 47:22products or services or political messages. It
- 47:29does not apply when we're talking about news
- 47:31or public affairs articles can't political campaigns.
- 47:36It's highly fact -dependent and there can be
- 47:41significant damages and adjunctive relief when
- 47:44implicated. Of course, we see right of publicity
- 47:47issues arising in the context of generative AI
- 47:49because it can be used to generate images of
- 47:53celebrities or other folks and if those are used
- 47:56for a purpose that violates the law. they can
- 48:00give rise to a lawsuit. There have been some
- 48:02additional statutes filed recently. There's a
- 48:06civil action against anyone who creates and intentionally
- 48:10shares digitized sexually explicit material using
- 48:13AI. There's also a new elections code section
- 48:17that prohibits distributing materially deceptive
- 48:20audio or visual media of a candidate within 60
- 48:23days of an election. And so you can bet people
- 48:27are going to be monitoring for this in connection
- 48:29with the upcoming elections in November, and
- 48:32we're just about at that 60 -day mark. We've
- 48:36got Tennessee's Elvis Act. Why do we care about
- 48:39that in California? Well, it's one of the broadest
- 48:43laws in the country, and it was passed primarily
- 48:46to protect country music stars, but it extends
- 48:51specifically to voice. It also extends past death,
- 48:56and it makes one liable, at least in Tennessee,
- 49:01for public use of an individual's voice or likeness
- 49:04without their knowledge if it was generated by
- 49:08AI. I think there is some real issue about how...
- 49:13the extent to which Tennessee can enforce this
- 49:16law on uses made outside of the state. And there's
- 49:20also a question I have about whether use of AI
- 49:24to generate a digital voice that may happen to
- 49:30sound like but not actually be influenced or
- 49:33trained using the voice of a celebrity qualifies.
- 49:37Lots of us have voices that may sound similar,
- 49:39and we haven't seen litigation of that issue
- 49:41yet. We've also got the various laws in New York,
- 49:47as well as in California, that pertain to digital
- 49:51replicas. And really, at a high level, prohibit
- 49:59the use of AI to generate what would be perceived
- 50:03as a performance by a person or celebrity without
- 50:07their consent. and it lasts for the life of the
- 50:13performer plus 40 years. We also have laws in
- 50:17New York and California that basically render
- 50:20unenforceable contracts pertaining to the creation
- 50:24of digital replicas, where the digital replica
- 50:29is inconsistent with the fundamental character
- 50:32of the photograph or performance in the original.
- 50:37California requires that the person be represented
- 50:42in the action by either council or a union. for
- 50:50the contract to be enforceable. And then we've
- 50:55got the No Fakes Act, which has been sitting
- 50:56in Congress for some time, but would define and
- 51:03narrow the ways in which one can use a digital
- 51:07replica that fundamentally alters another person's
- 51:11performance or depiction. Finally, some privacy
- 51:18and regulatory concerns. If your company is contemplating
- 51:24using user data, consumer data to fine tune a
- 51:29model or train a model, you should be aware that
- 51:34there are some notices that are required. If
- 51:41your terms of service provide that you may use
- 51:45user data to develop or improve products and
- 51:48services, that might not be enough. The FTC has
- 51:52indicated that you should be telling your consumers
- 51:56explicitly that you may use their content to
- 52:00feed or train algorithms. So here's a couple
- 52:05of examples. The one on the top is Google. The
- 52:08one on the top is OpenAI. But these terms were
- 52:12definitely updated to account for the fact that
- 52:17their models may be trained on data provided
- 52:22by consumers using their various apps and services.
- 52:29And then there are antitrust concerns to be aware
- 52:32of. Certainly, there's been a lot of focus around
- 52:37the country, including in California and New
- 52:38York, on use of algorithms to collude as to price,
- 52:47to set prices. Multiple companies are using the
- 52:52same AI tool to figure out pricing, even if they're
- 52:56not directly talking to one another about it.
- 52:58There is some question about whether that could
- 53:00be deemed collusion for purposes of the antitrust
- 53:02laws, and it's something to be cautious about
- 53:06and certainly well informed on. So that is a
- 53:11lot of content to take in at once, and I'll just
- 53:15leave you with this talk before we open it up
- 53:17to questions. stop my share before the speaker
- 53:21biography, which you'll get in the slides. You
- 53:25know, I've been litigating for cutting edge technology
- 53:31companies and around cutting edge issues for
- 53:34coming up on 26 years now. And I've got no special
- 53:40capability or background. I started college as
- 53:43a math major. I finished as a poet. I love logic
- 53:46puzzles and I love writing. And I've had the
- 53:51opportunity to try lots of cases to argue lots
- 53:55of fantastic appeals, primarily in the Ninth
- 53:58Circuit. I think it's in terms of the takeaways
- 54:02for those who may be younger in the practice
- 54:06or coming up. The key takeaway is to sort of
- 54:09be unafraid. I think you surround yourself with
- 54:13smart people who can help you understand the
- 54:16technology and the industry and the business
- 54:19of the client that you're representing so that
- 54:22you can tell what is ultimately a compelling
- 54:24story about it. I was very privileged to have
- 54:29a significant hand in the. the strategy and framing
- 54:34of the summary judgment brief that won in the
- 54:37Meta case and working on the case for Mid Journey
- 54:40and for other developers and users and artists.
- 54:44I think you got to dive in and roll up your sleeves
- 54:48and figure it out because nobody is more capable
- 54:54of doing that than you. I think it's just a matter
- 54:57of figuring out what's important and putting
- 55:03the time in to sort of understand it and distill
- 55:06it and turn it into something that everybody
- 55:11can understand. And through this work, I have
- 55:15been able to speak to folks in Hollywood. I've
- 55:20been able to speak to studios, to filmmakers
- 55:24in Cannes. you know, Ted AI, and it's just been
- 55:31such a privilege to take something I'm passionate
- 55:35about and have become passionate about, the creation
- 55:38of new works and the voice and the capability
- 55:42to create that these tools give us all, and to
- 55:45become, in my small way, a spokesperson for that
- 55:51technology on behalf of everyone. while at the
- 55:54same time making sure that we are thinking about
- 55:57how this could affect people, the real people
- 56:00using the tools, the real people whose jobs may
- 56:03be affected, and sort of bringing that human
- 56:05element to the analysis as well. So it's been
- 56:09my pleasure to be with you and share some of
- 56:11the work that is going on and the work that I'm
- 56:13doing and happy to answer questions. Already
- 56:22so it doesn't look like we are beginning. We
- 56:23have 2 questions so far. The 1st, 1 was should
- 56:27anthropic in your opinion have waited until eventually
- 56:31SCOTUS review before settling. Well, I can't,
- 56:34I can't take a position on that specifically.
- 56:38There was no telling if or when the DOJ would
- 56:41weigh in and faced with. class certification
- 56:47and a potential jury trial. There are lots of
- 56:51considerations that go into whether you want
- 56:52to settle or litigate, but certainly further
- 56:57lawsuits have been filed against Anthropic by
- 56:59those who have opted out of the settlement, and
- 57:01it is the only settlement to date. So it's another
- 57:05data point in the way that these matters can
- 57:08be resolved. Looks like I have another one. If
- 57:15an AI creator generates a brilliant visual composition
- 57:17that fails the human authorship test, a traditional
- 57:20artist can legally copy it canvas for canvas
- 57:23and copy their painting of it. Given this reality,
- 57:26is AI law inadvertently creating a predatory
- 57:28ecosystem where traditional artists can legally
- 57:30exploit the free labor of AI prompt engineers
- 57:33while leaving the original prompt engineer with
- 57:35no legal recourse? Yeah, so I don't know how
- 57:39the Copyright Office would deal with a claim
- 57:43for a work that was plainly a derivative of an
- 57:51AI -generated work. They may say the AI -generated
- 57:55work is in the public domain, but the copyright
- 58:00protection should potentially extend only to
- 58:03the original contribution of the painter. But
- 58:07if you assume that they could, in fact, obtain
- 58:13a registration for that work, then I do think
- 58:16that there is some tension there. Look, we're
- 58:19asking traditional artists to, you know, to compete
- 58:24with a technology that they may or may not want
- 58:27to use. And my own view is that not everything
- 58:32that comes out of AI when you say, you know,
- 58:34give me a picture of a cat with a hat on, should
- 58:38be protected. On the other hand, AI, like any
- 58:43other tool, in the hands of either a budding
- 58:49artist or a really successful one, it's just
- 58:52that. It's a tool. It's still human creation.
- 58:56And where the human is the, you know, Causing
- 58:59nothing is coming out of generative AI at this
- 59:02point for the most part without a human actually
- 59:04directing that and so I think there's got to
- 59:07be a path to protection Otherwise, we're disincentivizing
- 59:11use of technologies that could open up, you know
- 59:14incredible new ways to tell stories and create
- 59:17art and text, which is ultimately in everyone's
- 59:22benefit. That's the whole purpose of copyright
- 59:24is to promote the progress of science and useful
- 59:27arts and make sure there's a proliferation of
- 59:31protected new expression for us all to learn
- 59:33from and enjoy. Yeah, it looks like we have 2
- 59:39more. 1 of them is an opinion. It's what do you
- 59:42think the delay and resolution of the deep fakes
- 59:44litigation is? Yeah, so there there are some
- 59:49deep fake cases, but there's no deep fakes. The
- 59:52deep fakes act hasn't been passed and I, I, I
- 59:56think Congress. Has it, you know, there's a lot
- 1:00:02of. inertia there and maybe people trying to
- 1:00:07get consensus on what that law should be is what
- 1:00:09is keeping that from moving forward, even though
- 1:00:13I don't know that there's a whole lot of disagreement
- 1:00:17that deep fakes are a problem and should be addressed.
- 1:00:22And then I see the last question here, is the
- 1:00:24monkey case a bright line on the human creation
- 1:00:26issue? I mean, again, it's a bright line in the
- 1:00:30sense that it holds, like most other cases up
- 1:00:35to that point, but more explicitly, that human
- 1:00:39creation is required. But what human creation
- 1:00:42means, and the extent of the necessary human
- 1:00:45creation to warrant protection, is going to be
- 1:00:48answered by the courts in the context of specific
- 1:00:50works, like Teotra do Brasil by Jason Allen and
- 1:00:56the Colorado District Court that's currently
- 1:01:00considering that and it's a look for that decision.
- 1:01:02I think that's going to be hugely important and
- 1:01:05we will see how the courts fall in line as other
- 1:01:09challenges to rejections of copyright come out
- 1:01:12of the Copyright Office. Alrighty, this concludes
- 1:01:19today's presentation. Thank you so much Angela
- 1:01:21for taking the time to educate our audience today
- 1:01:24and thank you so much to everyone for joining
- 1:01:26us. Have a great day everybody.