AI on Trial: The US Copyright Fights

Since we only have 25 minutes, I'm going to blast through a bunch of information here. Hopefully leave a little bit of time to chat at the end. I will say this is entirely focused on US copyright law. We could talk about Australian or European or Chinese if you want, afterwards, maybe during the break. But this will be only about the US, just because I think for obvious reasons there is a lot going on there. As a professor, I always like to start with a little bit of a history lesson.

Chapter one is the mechanical music. How many people here know who John Philip Sousa is? Raise your hand. He's sort of fallen off the cultural landscape. But for anyone who's familiar with the history of music, John Philip Sousa was an American composer who wrote many famous marches. He even invented the sousaphone, and he was very popular in the first part of the 1900s. At the turn of the century, he wrote this essay about the menace of mechanical music, talking in this kind of colourful language about sweeping across the country with the speed of a transient fashion in slang or Panama hats — comes now the mechanical device to sing for us a song or play for us a piano in substitute for human skill, intelligence and soul.

Sound familiar? What he was really talking about was the gramophone or the phonograph, as well as the player piano. If anyone knows the old-timey westerns, you have these pianos and the piano roll has all the notes in it. His complaint — he had a bunch of different complaints, and by the way, I can make these slides available if people want to look at them later — was that this is canned music. There's something about live music which is very soulful and human, and this is very canned music.

That composers won't get paid, there'll be no remuneration for them. And my favourite one is that children will suffer because they will be put to sleep by phonographs and not the singing of their parents, and so this will harm them. And again, when we have fun with this and think, oh, this is 1906, you can see actually that the themes and the concerns about what mechanical inventions pose to society, including the dangers to children, have been around for a very long time, in various moments. And so the question is, how are we dealing with it in this current moment?

For those who are curious — and again, I can talk more about this — what happened was there was a lawsuit by some musical composers against the piano player companies, not so much the phonographs that came later, but the real decision that went up to the Supreme Court in the United States in 1908 said that making piano rolls was a copy of the music. And the Supreme Court said, naturally, no, we're going to hold this not a copy, because only humans can understand sheet music. Humans can't understand these perforated rolls.

Those are for machines. That's a different thing. And they said it's hard even for us to distinguish that from a music box that you rotate. It just feels different to us. So we're going to say it's not covered by copyright. In response, the US Congress then jumped in and in the 1909 Copyright Act actually created an expansive definition of copy, which included mechanical copies, and at the same time created a compulsory license — actually called it a mechanical license. That said that you can make player piano and phonograph copies, but you must pay the composers.

There was a set of statutory rates. So just to say, some of these questions have come up historically a long time ago. There have been many iterations since then and I won't go into them, but they're not unfamiliar historically, so I'm not going to do too much on this. I'm sure everyone here has seen many of these stories about how AI is challenging our current dynamic around copyright and creativity. This particular artist, Greg Rutkowski, did a bunch of paintings that were sort of Dungeons and Dragons fantasy-style paintings, and they ended up becoming quite popular for training and also outputs in AI systems.

And he actually spoke out and said, this is a problem. We've seen this with, in the UK, musicians — they're complaining and protesting with a silent album. And we've even seen the rise of these bot farms. In fact, a bunch of artists in the US right now who are folk artists put up these videos on YouTube or TikTok or Instagram. Someone — it's a little unclear who — has downloaded them all, trained various AI systems to create music in their style, voice, likeness, uploaded those through this sort of chain of distributors to Spotify and YouTube to sell or run ads against, and then actually even gone so far as to send copyright takedown notices to the original artists, saying that they've copied the AI music, which has problems in and of itself.

Lots of problems. Some of these might be copyright problems. Some of these might be ways the problems we can solve. There are other ways, but there's no question that there are problems and issues that we have to deal with. So what's going on in the US, and why don't we have an obvious solution to it yet? Many of you may know this, but in the US, the Copyright Act actually comes from the Constitution along with patent law, where Congress is supposed to promote the progress of science and useful arts by securing to authors and inventors exclusive rights.

So this comes from our originating document. And there are the five exclusive rights that are similar in many ways to what Australia has — you can't copy, you can't distribute, you can't display, you can't create derivative works and you can't perform things. When it comes to exceptions to this, though — the sort of things you are allowed to do even without permission, without a license — the main one that's come up in these AI cases is fair use. And one of the most important things in fair use in this preamble is it talks about examples such as criticism, comment, news reporting, teaching, scholarship or research.

And in many of the cases, which I'll gloss over for time, this has been bundled together by the courts into various concepts, which we'll talk about, such as transformativeness. You may have heard this before — creating something new, creating something different, having a different take, having a different perspective, something that isn't just a substitute for the thing you copied, but you're creating something new and different, invaluable to society. So is what is happening inside an AI system fair use or not? As many of you know — and especially if you're here, you've probably talked about this — there are many different things

that systems are doing with copyrighted works. But of course, the primary thing that people are fighting over in the US is training. There are outputs and there are other things going on, but training has been the main focus. And again, I don't have time to go into the entire fair use test here. But basically you have these four factors that courts look at, there in the law. Purpose and character, [inaudible] which is when you're taking someone else's copyright work, what are you trying to do with it?

Is it a good thing or is it a bad thing? Is it commercial or non-commercial? Are you basically trying to substitute your product for theirs, or are you trying to do something new? As I mentioned, sort of transformative. The nature of the work, which doesn't really matter in many cases, but sometimes does, in the sense that if it's a more creative work, you get a little more protection. If it's more factual or scientific, you get less. How much of the thing did you take? Did you take all of it?

Did you have to take all of it? And then finally, how much of a market impact did you have on the market that the copyright owner has? If the copyright owner is not making any money in an area, it's less of a problem, more likely to be fair use. If you're directly competing head to head, it's less likely to be fair use. Again, I'm happy to show these slides, but what a lot of folks, when we talk about this here in the US, don't spend a lot of time on — I just want to flag it — is that there are 30-plus years of cases actually looking at a number of different issues where computers and copyright works came together and fair use came up, even as far as, well, actually pre-1992. But the big cases for me, starting in 1992, with reverse engineering computer games in order to make competing computer games, or reverse engineering a computer console in order to make an emulator on a PC.

And many of these cases started to formulate the law that is leading to the AI conversation right now. And most of these cases said it's fair use if you're innovating in some way. There are pluses and minuses, and we can talk to the cases again when we had more time. But really what they said is if you're not just copying it and then reproducing the exact same thing, but you're doing something useful, especially in a computer context, like creating interoperability or creating a whole new market for computer games, or allowing people who bought legitimate computer games for a console to play them on their PC, to shift formats, or to do something new in a different context, even if the content stays the same.

But there's a sort of technological innovation. This gets more complicated as technology and content mix, but some of these cases were quite important, informative, and you see them coming up again in these AI fair use lawsuits. Some of the other cases, again, we won’t go through them all. But image search was a big one. So for example, could you scrape the internet for images and then create a search engine used image. [inaudible] So, image search right there. You're making full 100% copies of images, but you're doing it for what the courts ended up calling an information location purpose.

In other words, search. You're not trying to show people images necessarily to replace the original image. You're trying to use images like a thousand words, and say, oh, that's easier to find something. And because search became easier with images, the courts found in many of those cases. There's also an anti-plagiarism case, which is one of my favourites — the iParadigms case, where a very clever plaintiff lawyer sued on behalf of all the children whose essays had been uploaded to the anti-plagiarism database, saying that they all were denied royalties for and getting caught cheating was an unfair.

Anyway, that got rejected, though, for obvious reasons you might have guessed from the equities, but also because the court said, look, anti-plagiarism software is really good for society. We should stop cheating. And this is how you do it technologically, at least at the time. You need a giant database of all these essays in order to sort of do this comparative analysis. And we don't need to pay all these children to allow this to happen. I'm just going to jump to this Authors Guild versus Google case, because this is the one that most of the people who talk about US copyright and fair use are pointing to, because it's also very recent.

So this was the case where Google worked with libraries to get ten-plus million books, physical books, from them, scan them and make book search. And then they got sued by both the publishers and the Authors Guild as a class action suit. And again, here it was found to be fair use because of two things. One, Google was doing it basically to improve search, though specifically for books, but also for general search. And the second is that you couldn't actually reconstruct the books very easily. In other words, you get these little snippets and things, but the outputs of the system did not directly compete with the inputs.

Even though the Authors Guild and publishers made an argument — hey, we would happily license these to you. You should pay us $100 million because it's valuable to you, Google. And the court said, no, you guys are in the publishing business. This is not really what you do. Google does a search thing over here. They're sort of apples and oranges, basically. So where does that leave us here? As many of you may know, the two big decisions that we actually have right now are from this case, in a Meta case.

There's Bartz versus Anthropic and Kadrey versus Meta. We're waiting on a couple of others. There's another decision called Thomson Reuters about legal databases, which is going to come out. I'll talk over 150 cases. So there's going to be a lot of decisions that come out. But what we know so far, at least in the two decisions from Anthropic and Meta, is that both courts found fair use for very different reasons with very different takes. Now, many of you may have also heard about the Anthropic settlement, and that does not wave away this fair use ruling.

I can talk about the settlement if people want, but the settlement was very much specifically for the fact that Anthropic had gotten its books — this is a books lawsuit — from two different sources. One was from an illegal website, and that's what the settlement is about. They downloaded hundreds of millions of books in the end, but really only 500,000 that mattered to the court. And the $1.5 billion settlement was about those 500,000 books that were illegally downloaded. And they also went and bought books and then scanned them and destroyed the physical copies, actually, and then used those for training.

And the court actually said, that's fine, that's fair use. That is not what the settlement's about. So just so you understand, there were two different sources. The source that led to the settlement was the illegal downloads. The buying, scanning, destroying and then using for training — the court had no problem. So looking at the case and looking at the Anthropic case and putting them together, the first factor, purpose and character of use — what are we doing here? Basically both courts said, look, we're creating something highly transformative here.

If you thought image search was transformative, if you thought book search was transformative, my God, these LLMs can do amazing things. Now, whether you agree with that or not, [inaudible] both judges were persuaded that this technology is advancing society in a bunch of different ways. It is commercial, they said, but that doesn't overcome the benefits of society. A lot of things have to be commercial. Basically they said, some industry, you wouldn't be able to afford to do this necessarily unless they were making money. There were a lot of arguments about style, which I'll talk about in a second.

But in most contexts in the United States, style is not copyrightable. It has to be a very specific work. It has to be a very specific way you wrote, drew, painted, something else, made. Music can't just be a genre or style. There was memorisation, but both courts found that Meta and Anthropic had put in sufficient guardrails to prevent outputs that were verbatim. There are exceptions to the rules here and there, but really, both courts were fairly impressed with the amount of work the companies had done to prevent people from regurgitating the works.

So memorisation was kind of put away. So factor one, which ends up being very important, ended up favouring both parties. Factor two, nature of the work, was sort of neutral. The courts just basically said [inaudible] you kind of need everything. You need creative works, non-fiction works, factual works. We're not really going to pay attention to that factor. Three, the amount — they said it was reasonably necessary to take everything. Again, this sort of comes from the Google Books line of cases. Because there was an argument in the court — both judges bought it — which said more is better, more data, more types of writing, more types.

This is mostly books. More examples gives the models better outputs, whether you disagree with that or not. That's what the courts found there. And then perhaps this is the most controversial and potentially interesting for the future, which is that in the Anthropic case, the court really said, look, I don't see the market harm here. You're not directly substituting. In other words, no one can produce Harry Potter from Claude. So whether Claude was trained on Harry Potter, if you want to read Harry Potter, you're going to go get the book from somewhere else.

You're not going to get it from Claude. It didn't bother as a licensing market, because really it said, look, these publishers — yes, in theory, they can license, just like in the Google case, but they're not AI companies. These publishing companies make books and they do their book things, and maybe they license for a movie or a sequel or something like that, but they're not making AI systems. It's a different kind of thing. And then finally there was this question of, well, what if your AI model produces so many books that it actually just floods the market and dilutes the market — say, there's so many Harry Potter rip-offs that nobody actually cares about Harry Potter anymore.

And the Anthropic judge said no, but the Meta judge said yes. The judge said, actually, in this case, the plaintiffs' lawyers didn't put forward any evidence of dilution, but he sort of said that if there had been evidence that, in fact, AI was diluting the market for copyrighted works, that he might have flipped the other way and gone against fair use. So he sort of blamed the plaintiff's lawyers for a bad job. Now, where does that leave it? We have these two decisions. The Anthropic case looks like it's settled, so that decision is there — it's not going to get appealed.

The case is facing a bunch of trials and a bunch of other issues and may eventually get appealed. But we have 125 active lawsuits in the US alone. By the way, this all comes from ChatGPT is Eating the World, which is by Ed Lee, who's a law professor in the US who runs a great blog website all about this. This is his graphic, his latest graphic, which I'm borrowing and giving attribution to, with his permission. But just so many lawsuits to keep track of. So we're going to get a lot more decisions.

This is another way to look at it. These are all the US lawsuits. It's overwhelming. And he's tracking it all, and I really appreciate that. So we're going to get a lot of decisions. Many of them are being consolidated, so we'll get one decision for like 50 or 100 cases, things like that. This is globally as well. So lots of lawsuits happening all over the world. We might get other answers. There it is. As we all talked about yesterday, there is lots of AI regulation. There are enforcement decisions happening.

The US is obviously one place that's important but not the only place. And so I'm going to end on this and just sort of open maybe conversation, questions from here, if we have a few minutes. But so what has been the US congressional response? One of the things you need to understand, if you don't already know this, is in the US copyright law is almost entirely — I would say, some would say entirely — federal law. In other words, states can do very little, if anything, with copyright law.

So it's really left up to the US Congress. And as any of you following the US Congress know, very little is possible right now, if anything, especially here. But one of the more recent examples, just to give you an example, is this Creator Act. And the Creator Act does try to give a special copyright right for style, in fact, interestingly enough. And that for visual artists' distinctive signature styles. And that they would be able to sort of send takedown notices and sue over it. Now, will this pass?

I doubt it. If it did, how enforceable would it be? Would it create like 10 to 20 more years of litigation? Almost certainly. Are there fair use defences to this? Probably, depending on how it's interpreted. So we are seeing some attempts on the congressional side, but I don't know that we're seeing any better answers. And I'll stop with five minutes to go — my prediction is that you are going to just see a lot of waiting and watching. Of course, I still think the courts are going to get there first.

I don't think you're going to see much legislative action. And I know that several of the big cases against OpenAI are going to what's called summary judgment, where the judge makes the decision about whether the case goes one way or the other, goes to trial. That's going to happen in September of this year. So I suspect in the next six months, maybe 12 months, we're probably going to get, I'm guessing, at least 4 or 5 other decisions. We'll see if they agree with the Anthropic and the Meta case or not, or if they go a different way.

And we'll see if it eventually gets all the way up to the Supreme Court. And then who the hell knows what goes on there. I'm going to end there. Thank you very much. Happy to talk more and take questions or whatever.

Greg Sadler

Thank you so much. While some questions come in, I might just ask one myself, which is about executive power. If there's this sort of menagerie of hot cases coming out, is there any power or capacity or scope for a Trump administration to say something or to wade in, and does it have any influence?

Jason Schultz

It has soft power. Congress really passes the copyright laws, but he's tried a million things, so who knows, maybe he will do an executive order on copyright. He has gone on record, for the most part, saying that he actually likes that the AI companies have fair use. So he's been sort of vocally supportive of that. But he also flip-flops things left and right, so who knows. I will say this also — as we've seen in many of the safety contexts recently, the supposed eyes are in the White House and others in the White House have heavy influence.

So we'll see what happens there. But given that most of them are investors in these companies, I wouldn't necessarily see them siding with copyright owners anytime soon.

Audience question

What would your advice be to creators that are maybe just starting out now and want to create sort of a style and not have it just be cloned somewhere?

Jason Schultz

Yeah. Look, this is a very difficult question. Also depends where you are. Again, I refer to Kim and others who teach Australian copyright law. But my sense is, though, if you're concerned about the US, these companies tend to scrape what they can find. So I actually do think that some of the efforts to use robots.txt or text or other signals, or to not have things be as publicly available on the internet, have had some effect. That, of course, presents your problem, which is how do you get seen or heard?

So I don't know that you're going to stop yourself from ending up in a data set at some point. I think the best thing you can do is try to build the direct-to-customer relationships and direct-to-listener or audience relationships, as people have had to do for a long time. I will say some AI companies are opening up more licensing options. Can always look at that. But I think it is going to be another way that creators just have to ride out in some way. It's going to be tough.

Sorry.

Audience question

One of the key arguments against Australia doing a copyright deal that I have heard is that it would invalidate the fair use provisions under factor four, the market. If Australia creates a deal here, we're creating a market and that invalidates the fair use provisions. How do you think that's a fair and accurate representation of the current case law?

Jason Schultz

So you mean if it really does a deal it would, in the US, create a market — if the tech companies do a deal with Australian creators?

Audience question

Yeah, it's demonstrating that there is a market for the works.

Jason Schultz

Yeah. No. And fortunately that doesn't work at all. They've tried this in the US for a very long time. Very similar to the Google books. Actually, just in the interest of time — there's a longer story to this — but the image search case Perfect 10, which was a porn company, actually, of course, there were all these lawsuits where porn companies were pursuing Google. Everyone else tried to create a market for thumbnails of porn images with some mobile company, say, see, there's a market, people can download thumbnail porn.

And again, the courts can smell when a deal is being made to try to manipulate the law, and they've almost always rejected it. It has to be where an honest market has actually existed, not in response to manipulating the fourth factor, but because it really does exist. And so then courts will go, oh, no, okay, this has been here for a while. The test is traditional, reasonable and predictable markets that emerge essentially naturally.

Greg Sadler

Okay. We've got time for one more question, if someone wants to. Yeah, this is a good point. So I guess the contention here would be, well, a market did evolve naturally in Australia and AI companies did come here and they licensed the work, defective licenses or individually. But that would be a real market emerging naturally. It would be a real exchange of value for money from the participants.

Jason Schultz

Right. So we'll see what happens. So this New York Times case is getting a sign of September, has some of these arguments in it. What I suspect, though, is that there have always been some people who've struck deals, but the AI companies are scraping a bunch of other stuff. In other words, it's not — you start deals with specific industries. You can make arguments for completeness, quality, all kinds of things. But they'll say, yes, but the market for generally scraping the internet is a different market than a bespoke market for a particular industry or company providing its IP.

Or they'll just say it was just to get rid of litigation. We just struck the deal to [inaudible] get ahead of it. Anyway, I can go into more detail, but again, courts have been somewhat allergic to this idea that if some copyright owners license for technology purposes, it means everyone has to. When companies started training before that licensing market came around — in other words, the training that's been happening for a long time, the licensing market develops later — courts go, well, the training started first.

It's not as if the technology is encroaching on a licensing market that already existed. We'll see. Maybe a court or two will be persuaded, but it would be unusual in the US.

Greg Sadler

We might have to wrap it there so we can make it to the closing panel. But thank you. Thanks everyone.