This year could be a critical one for court decisions that redefine how U.S. copyright law applies to generative AI.
More than 70 infringement lawsuits have been filed by copyright owners against AI companies, centered on whether training large language models (LLMs) on copyrighted material constitutes fair use. Major media and entertainment companies including The New York Times and Disney have launched lawsuits against OpenAI and Midjourney, respectively, and Anthropic reached a $1.5 billion settlement in the Bartz case, the largest known copyright settlement in history, after facing liability for downloading millions of pirated works used for training. Courts, however, have sent mixed signals on what constitutes illegal use of copyrighted material by AI.
Rebecca Tushnet, the Stanton professor of the first amendment at Harvard Law School, is a leading expert on copyright and trademark law. She has written extensively on fair use, Federal Trade Commission (FTC) compliance, and the legal architecture governing creative and commercial speech.
Despite her skepticism about generative AI, Tushnet thinks copyright law is poorly suited to regulate it—and that changing the laws could have unintended consequences for creators and future innovations. This interview has been edited for length and clarity.
1. What is “fair use” in copyright law, and how are AI companies using it to their legal advantage?
The basic idea of fair use is that some unauthorized uses of copyrighted works are beneficial for society: parody, satire, and academic commentary, [for example].
“Big data” uses are also often fair because they generate new insights and have different purposes from the point of any given individual work—for example, you can use Google Books to track the rise of a concept across the decades. Some AI training is that kind of use—by analyzing hundreds of thousands of different works, the output can be something different and useful.
2. In which scenarios could training an AI model lead to violation of current copyright laws, or not?
If the training produces infringing copies as outputs every time it’s used, that’s clearly not fair; if a model never produces infringing copies as outputs when it’s used, then that’s a case for fair use.
If it sometimes produces infringing copies, then the fair use case probably depends on whether the infringement is directly caused by the training (which could happen if the infringing copies pop out readily in response to innocuous prompts) or by a prompter deliberately trying to get an infringing result. In the latter case, it’s probably not right to blame the AI model. Two district court cases ruled this way last year, and I think that’s the right result.
3. A big question in these cases is the potential replacement of existing products: an AI-generated news summary undercutting a newspaper subscription, for example. What legal weight does this carry?
Some people have been pushing the argument that this makes AI training unfair. But that’s not how copyright has ever thought about copyright harm. The rise of record players wasn’t unfair even though it put the musicians who used to play in restaurants out of work.
[In the case of AI], if a search on Google Books answers the question of when FDR was born, and a searcher doesn’t need to buy the book from which that fact is taken, that’s not the kind of substitution that copyright cares about, because copyright doesn’t protect facts. Competition among non-infringing alternatives is good, not bad—it means that audiences have more choices.
4. How do companies like OpenAI store their models’ source materials, and what does that mean in terms of copyright law?
As far as I’m aware, there’s no evidence that training models contain copies (or substantial parts) of most works on which they’re trained; there is evidence that some models can be prompted to generate pretty good copies of a few works like the first Harry Potter book or George Orwell’s 1984.
If a model can be easily prompted to generate a decent copy, then current law will likely consider the model to contain a “copy” of that work. But that will have to be evaluated on a case-by-case basis, because that seems likely to be true only of a handful of widely available works.
This argument is also probably less important than it seems because everyone agrees that the training process involves making and using digital copies, which is going to require fair use or some other justification, even if the resulting model isn’t a copy of the training data.
5. Do you think the law should change so that AI companies have to pay licensing fees to creators for training models on copyrighted materials?
I find myself in an odd position: I generally don’t think that most generative AI is good for society, but copyright law is not the right way to regulate it. Fair use under U.S. law is deliberately flexible and is generally handling AI issues well so far.
Ruling against the companies [on fair use grounds] could set a new precedent by recognizing a new kind of “training market” would threaten both existing fair uses and the next, unknown innovations. This is because copyright owners can always assert that they want to get licensing fees for any use they want to control, including commentary and criticism.
Rejecting fair use would transfer money to Thompson Reuters, Warner, and other big copyright owners, but—like every increase in copyright rights in the past few decades—will not do much for artists. Even if the big copyright owners give one-time payouts to current authors, their future contracts will not provide for additional payments for AI training or will offer the kind of penny-a-year payments that musicians complain about from Spotify now. As Cory Doctorow and Rebecca Giblin point out, the problem that artists face is not lack of rights, it’s lack of market power.
