By our counsel Matthew Hersh:
2023 was the year of artificial intelligence. No, the robots haven’t become our overlords yet. And we are “still a long way from the science fiction version of artificial general intelligence that thinks, feels, and refuses to open the pod bay doors”, as one wag put it. But this was the year that AI broke out of the labs and into the consumer space in full form.
The phrase you need to know is “generative AI”. That’s the form of AI that is capable of generating human-like responses to inputs.
Some of the main examples of generative AI are the following:
- Large Language Models: These are the ones that can give human-like responses to text inputs. (They’re “large” because they consumed massive amounts of text in order to learn.) You probably know ChatGPT, but there are others—Meta’s LlaMA, Google’s BARD, and so on. They’re pretty amazing. Ask ChatGPT when Shakespeare was born and whether Francis Bacon wrote his plays, and you’ll get surprisingly accurate responses.* Ask it to write you a sonnet about copyright law in Shakespearean style and it will whip one up in seconds that is freakishly good, if a bit incoherent. They’re not perfect of course. When in doubt, they prioritize human-sounding responses to accurate ones. Ask for a legal opinion and it will sometimes make up caselaw, as more than one lawyer has learned the hard way. But they get better every day. Actually, every millisecond.
*The answers are “1564” and “probably not”.
- AI Art Generators: These are the programs that allow you to create images out of text input. Stable Diffusion might be the best known, but there’s also Midjourney, DALL-E, Deviantart, and countless others. They’re also capable of cool stuff. Put in “dog wearing a baseball cap eating ice cream” and you’ll get something that more or less resembles just that.
- Music and Video: There are plenty of these too. Try Boomi AI and you can, as their website proclaims, “create original songs in seconds, even if you’ve never made music before”. (And then “submit them to streaming platforms and get paid”, supposedly.) And as anyone who has ever been fooled by a “deepfake” knows, it’s easy to generate video and audio that sounds and looks just like a politician or a celebrity.
AI threatens to impact, and even reshape, our copyright jurisprudence in any of a number of ways. Here are some of the questions that the Copyright Office, Congress, and the courts—and eventually the Supreme Court—will have to address:
- Can a work made by AI be copyrighted? We know only works by humans can be copyrighted, not works by machines. But where is the line crossed? If a human does most of the creative work, but the machine just acts as a tool (think Computer Aided Design), then that’s a human product. If a human just asks a computer to create a picture of a dog eating ice cream, that’s not a human product. What if the human does some of the creating and the machine does the rest? What if the human provides detailed instructions to the computer that guides its creative efforts? Somewhere in there is the line, but we haven’t found it yet.
- When generative AI programs ingest copyrighted material for training purposes, is that fair use? There’s some latitude under the fair use doctrine to make limited amount of copies of protected work in order to create something with a wholly different purpose or character. (Read Justice Sotomayor’s majestic opinion in Warhol v. Goldsmith to know more.) But wholesale copying of entire libraries raises different issues. Expect this debate to be front and center in 2024.
- Is the output of generative AI programs also infringement? Suppose a program hoovers up books I’ve written and uses what it learns from those books to respond more accurately to user queries. Are those responses derivative works that require my license? What if the output uses a small snippet of my work, a sentence here or a phrase there? Where is that line crossed?
- Does generative AI threaten the name, image, or likeness of celebrities and politicians? If I ask for writing in the distinctive style of Isabel Allende, does the computer’s response impinge on her own right to her likeness?
I will focus on copyright in this writeup—but it won’t surprise you to learn that AI is talked about everywhere. President Biden put out an executive order on AI (and every federal agency is engaged in the issues too). Congress is holding hearing after hearing. The European Union just reached an agreement on comprehensive regulations (they’re ahead of the U.S. on this). AI is playing a role in labor negotiations (case in point: the agreement that ended the Screen Actors Guild strike has a provision that AI can’t be considered a writer). Pretty much every area of law and commerce is changed—for the better, we can all hope—by AI.
Copyright Office Action
The Copyright Office—to its great credit—has been out in front of this issue. They’re working on two distinct fronts.
One front is decision-making. More and more authors are now generating works with artificial intelligence and trying to persuade the Copyright Office to register a copyright in them. So far, the Office has been holding firm. In March, the Office put out guidance on registration of AI-generated works. The gist of the guidance was that there’s a difference between a human using AI as an assisting tool (think CAD, Instagram filters, Adobe formatting tools, and the like) and using a computer to make the actual creative decisions that otherwise would be left to humans. “[If] the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine”, the Office says, then no luck.
The Office has consistently applied this guidance. In February, it refused to register a work created by an inventor with the aid of Midjourney. In September, it refused to register another Midjourney work (this one, as it happens, had just won an art award at the Colorado State Fair—you can imagine how upset people were when they learned it was created by AI!). And earlier this month (in a decision we recently covered on our blog), the Office refused to register a piece of artwork created using an AI machine known as RAGHAV. (That one was particularly interesting because putative author had success getting the same work registered in India and Canada—but no dice here.)
The Copyright Office’s view on registration of AI-created works was significantly bolstered in August when the federal court for the District of Columbia sided with the Office over a work called A Recent Entrance to Paradise. The colorful, unusual artwork—showing a set of train tracks running through an array of unidentifiable foliage—was submitted for registration by Stephen Thaler, a prominent inventor and developer of a computer system entitled the Creativity Machine. The Copyright Office rejected the claim, and the court agreed. Humans can use computers to aid their work, the court held, but “copyright has never stretched so far as to protect works generated by new forms of technology operating absent any guiding human hand”.
The other aspect of the Office’s work has been policymaking. The Office maintains a website that is chock-full of information about its public activities—testimony to Congress, guidance letters, webinars, and other information. The Office has also undertaken a massive study on artificial intelligence which kicked off an open call for opinions on the impact of artificial intelligence and copyright law. They received a whopping ten thousand responses to the inquiry. (I asked a copyright office attorney at one point who was responsible for reviewing those responses. She said: “I have been instructed to tell you that the review is being undertaken by humans.”)
The courts are also busy making law. Over the course of the last year, a wide range of songwriters, writers, visual artists, photographers, and other creators have been filing lawsuits against the makers of artificial intelligence programs. Those lawsuits, which are now in their preliminary stages, will require the courts to eventually decide all of the key issues involving artificial intelligence—from ingestion to output to everything in between.
Here’s a look at the most prominent ones.
A. Visual Artists
- The visual artists were the first to strike in 2023, when a trio of artists filed a lawsuit against Stability AI, Midjourney Inc., and Deviantart Inc. Andersen v. Stability AI Ltd., No. 23-201 (N.D. Cal. filed Jan. 13, 2023). I covered the filing of the lawsuit here. The gist of the lawsuit was that their works were used as training images in the underlying software—scooped up and copied by the software and then fed into its algorithm. Moreover, the artists claimed, the software’s output was necessarily derivative of their works and—for good measure—violated their right of publicity to the extent that the software could “be directed to prioritize inclusion of specific artists’ works by invoking the name of the artist or artists”.The artists ran into a roadblock at the end of October, however, when Judge Orrick of the San Francisco bench—a longtime legend in IP circles—dismissed much of the claim. I covered the opinion for IP Law Daily here. The direct infringement claim against Stability AI for copying up the images was fine. But everything else in the complaint would have to be repleaded with more specificity—or dropped. The most important flaw, Judge Orrick found, was the claim that the image software’s output constituted derivative works. The problem for the artists, he noted, was none of them alleged that any of the output images were substantially similar to their originals. “The alleged infringer’s derivative work must still bear some similarity to the original work or contain the protected elements of the original work”, Judge Orrick noted.The plaintiffs have just amended their complaint, so we will see where this goes. At least on my quick review, it doesn’t seem to fix the problem with the derivative works claim, although it adds a good deal of additional details that might address Judge Orrick’s other issues.
- Photographers got into the game in February, led by Getty Images. It’s somewhat of a battle of the titans—the leading photo agency takes aim at Stability AI, perhaps the most prominent of the image-generation programs available. Getty Images v. Stability AI, No. 23-135 (D. Del. filed Feb. 3, 2023). I covered the complaint for IP Law Daily here. The complaint alleges that Stability AI has copied without permission more than 12 million photographs from Getty Images’ database, along with the associated captions and metadata, to train its software. The lawsuit also contends that some of the computer-generated renderings even contained distorted recreations of the watermark that Getty Images embeds in all of its photos, thus tarnishing the photo agency’s reputation. Stability AI has contested the Delaware court’s jurisdiction and the parties are engaged in jurisdictional discovery now.Notably, the United States is not the only battleground for these two behemoths. Earlier this month, a UK court allowed Getty’s claim against Stability AI to go forward on that side of the Atlantic.
C. Print writers
- Writers joined the mix in July when Sarah Silverman, Richard Kadrey, and Christopher Golden filed a lawsuit against Meta. Kadrey v. Meta Platforms, No. 23-03417 (N.D. Cal. filed July 7, 2023). (The case has been since been consolidated with a similar case that Paul Tremblay and Mona Awad filed against OpenAI a week earlier. Tremblay v. OpenAI, No. 23-03223 (N.D. Cal. filed June 28, 2023) The writers claimed that Meta’s LLaMA infringed their works by direct infringement (that is, copying them to ingest into its database) and indirectly (by outputting works that were impermissible derivatives of their originals). I wrote about this case for the IP Law Daily hereBut the lawsuit has run into immediate trouble. In November, Judge Chhabria (following the path already taken by his fellow jurist Orrick in the Anderson case) wrote a concise opinion dismissing all of the claims except the direct infringement claims. The derivative infringement claim, the court noted, was “nonsensical” because the authors failed to allege any actual similarity between their works and the program output. “To the extent that they are not contending LLaMa spits out actual copies of their protected works”, the court observed, “they would need to prove that the outputs (or portions of the outputs) are similar enough to the plaintiffs’ books to be infringing derivative works”. They failed to do that.Silverman and company are not done yet. On December 11, they filed an amended complaint that joins a host of other authors to the cause, including Michael Chabon, Ta-Nehisi Coates, Junot Díaz, Andrew Sean Greer, and half a dozen others. But it also appears that they have given up the ghost on the dismissed derivative works claims—the new complaint includes only a direct infringement claim.
- The next big move was made by the Authors Guild, the main trade organization for working writers. The Guild, along with a star-studded list of authors including John Grisham, David Baldacci, Jodi Picault, among others, sued OpenAI for infringement in September. Authors Guild v, OpenAI, No. 23-8392 (S.D.N.Y. filed Sep. 19, 2023). I wrote about the lawsuit for the IP Law Daily here. Like the original Silverman et al. lawsuit, the Authors Guild lawsuit alleges infringement on both ends: by ingesting the authors’ work without a license, and by spitting out works that are derivative of the originals. Unlike the Silverman suit, though, the complaint seems to contain a very detailed set of allegations to shore up the derivative works claim. The case has been assigned to Sidney Stein, a legend on the New York bench, and is still in preliminary stages.
- Former Arkansas Governor Mike Huckabee, along with several other non-fiction authors, filed a lawsuit in October against Meta and Microsoft, also taking aim at the LLaMA program. Huckabee v. Meta Platforms, Inc., No. 23-9152 (S.D.N.Y. filed October 17, 2023). I wrote about the lawsuit for the IP Law Daily here. The parties have been wrangling about jurisdiction and venue with no substantive movement.
- The newest lawsuit to hit in this area is Sancton v. OpenAI, Inc. (S.D.N.Y filed November 27, 2023). Writer Julian Sancton (you might know him from Madhouse at the End of the Earth: The Belgica’s Journey into the Dark Antarctic Night) filed a class action lawsuit against OpenAI and Microsoft. The lawsuit claims that the material fed into ChatGPT was “scraped indiscriminately from the internet and included a massive quantity of pirated and copyrighted material”, including Madhouse at the End of the Earth. (Interesting side note: How did Sancton know that his own work had been incorporated into the ChatGPT training database? He asked ChatGPT—and it told him.)
- Songwriters finally got into the game in October, when the major music publishing companies brought a lawsuit against the company that owns the AI chat program Claude. Concord Music Group, Inc. v. Anthropic PBC, No. 23-1092 (N.D. Tenn. filed October 18, 2023). The issue here is lyrics. The music publishers own or manage the music and lyrics to millions of songs (on behalf of the songwriters who hire them). Claude, which is run by a company called Anthropic, happily spits out the lyrics to those songs when it is asked to. Briefing is currently underway on a motion to dismiss.
Last but not least—or given the state of political affairs these days, perhaps indeed least—is Congress. It’s hard to expect a lot out of the legislature these days. But Congress is nonetheless doing its thing, holding hearings, etc.
The main action this year has been in the Senate Judiciary Committee’s Subcommittee on Intellectual Property. That committee held a two part hearing (I covered the second one for IP Law Daily here) over the summer. The biggest issues on tap are the following:
- Should the Copyright Act be amended to expressly address artificial intelligence? For example, should Congress create an express “safe harbor” of fair use in order to ensure that AI developers can advance technology without fear of lawsuit, as tech companies generally want? Or on the other side, as some creators advocate, should Congress tighten the law to make clear that the ingestion of creative works to train generative AI programs is never permitted?
- Should Congress enact a federal right of publicity law? One of the biggest concerns of creators is that, in addition to having their work used to train AI programs, the programs are also able to mimic their writing style, singing style, or image. But to the extent that individuals have the right to control their “name, likeness, and image”—or NIL, as the phrase is usually used—it is only through a patchwork of state laws that offer differing levels of protection from jurisdiction to jurisdiction.
- What kind of transparency requirements should be imposed on AI models? For example, should the developers of every generative AI program be obliged to make public all of the works that were used to train that program? And what about outputs? Should every AI-generated product be expressly labeled as such? (Think: deep fakes.)
- These issues are going to keep Congress busy for a while. Don’t expect answers soon.
Ending (Deep) Thoughts
Hang on tight. 2024 will be a wild ride.