Should tech companies have free access to copyrighted books and articles for training their AI models? Two judges recently nudged us toward an answer.
More than 40 lawsuits have been filed against AI companies since 2022. The specifics vary, but they generally seek to hold these companies accountable for stealing millions of copyrighted works to develop their technology. (The Atlantic is involved in one such lawsuit, against the AI firm Cohere.) Late last month, there were rulings on two of these cases, first in a lawsuit against Anthropic and, two days later, in one against Meta. Both of the cases were brought by book authors who alleged that AI companies had trained large language models using authors’ work without consent or compensation.
In each case, the judges decided that the tech companies were engaged in “fair use” when they trained their models with authors’ books. Both judges said that the use of these books was “transformative”—that training an LLM resulted in a fundamentally different product that does not directly compete with those books. (Fair use also protects the display of quotations from books for purposes of discussion or criticism.)
At first glance, this seems like a substantial blow against authors and publishers, who worry that chatbots threaten their business, both because of the technology’s ability to summarize their work and its ability to produce competing work that might eat into their market. (When reached for comment, Anthropic and Meta told me they were happy with the rulings.) A number of news outlets portrayed the rulings as a victory for the tech companies. Wired described the two outcomes as “landmark” and “blockbuster.”
But in fact, the judgments are not straightforward. Each is specific to the particular details of each case, and they do not resolve the question of whether AI training is fair use in general. On certain key points, the two judges disagreed with each other—so thoroughly, in fact, that one legal scholar observed that the judges had “totally different conceptual frames for the problem.” It’s worth understanding these rulings, because AI training remains a monumental and unresolved issue—one that could define how the most powerful tech companies are able to operate in the future, and whether writing and publishing remain viable professions.
So, is it open season on books now? Can anyone pirate whatever they want to train for-profit chatbots? Not necessarily.
When preparing to train its LLM, Anthropic downloaded a number of “pirate libraries,” collections comprising more than 7 million stolen books, all of which the company decided to keep indefinitely. Although the judge in this case ruled that the training itself was fair use, he also ruled that keeping such a “central library” was not, and for this, the company will likely face a trial that determines whether it is liable for potentially billions of dollars in damages. In the case against Meta, the judge also ruled that the training was fair use, but Meta may face further litigation for allegedly helping distribute pirated books in the process of downloading—a typical feature of BitTorrent, the file-sharing protocol that the company used for this effort. (Meta has said it “took precautions” to avoid doing so.)
Piracy is not the only relevant issue in these lawsuits. In their case against Anthropic, the authors argued that AI will cause a proliferation of machine-generated titles that compete with their books. Indeed, Amazon is already flooded with AI-generated books, some of which bear real authors’ names, creating market confusion and potentially stealing revenue from writers. But in his opinion on the Anthropic case, Judge William Alsup said that copyright law should not protect authors from competition. “Authors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works,” he wrote.
In his ruling on the Meta case, Judge Vince Chhabria disagreed. He wrote that Alsup had used an “inapt analogy” and was “blowing off the most important factor in the fair use analysis.” Because anyone can use a chatbot to bypass the process of learning to write well, he argued, AI “has the potential to exponentially multiply creative expression in a way that teaching individual people does not.” In light of this, he wrote, “it’s hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars” while damaging the market for authors’ work.
To determine whether training is fair use, Chhabria said that we need to look at the details. For instance, famous authors might have less of a claim than up-and-coming authors. “While AI-generated books probably wouldn’t have much of an effect on the market for the works of Agatha Christie, they could very well prevent the next Agatha Christie from getting noticed or selling enough books to keep writing,” he wrote. Thus, in Chhabria’s opinion, some plaintiffs will win cases against AI companies, but they will need to show that the market for their particular books has been damaged. Because the plaintiffs in the case against Meta didn’t do this, Chhabria ruled against them.
In addition to these two disagreements is the problem that nobody—including AI developers themselves—fully understands how LLMs work. For example, both judges seemed to underestimate the potential for AI to directly quote copyrighted material to users. Their fair-use analysis was based on the LLMs’ inputs—the text used to train the programs—rather than outputs that might be infringing. Research on AI models such as Claude, Llama, GPT-4, and Google’s Gemini has shown that, on average, 8 to 15 percent of chatbots’ responses in normal conversation are copied directly from the web, and in some cases responses are 100 percent copied. The more text an LLM has “memorized,” the more it can potentially copy and paste from its training sources without anyone realizing it’s happening. OpenAI has characterized this as a “rare bug,” and Anthropic, in another case, has argued that “Claude does not use its training texts as a database from which preexisting outputs are selected in response to user prompts.”
But research in this area is still in its early stages. A study published this spring showed that Llama can reproduce much more of its training text than was previously thought, including near-exact copies of books such as Harry Potter and the Sorcerer’s Stone and 1984.
That study was co-authored by Mark Lemley, one of the most widely read legal scholars on AI and copyright, and a longtime supporter of the idea that AI training is fair use. In fact, Lemley was part of Meta’s defense team for its case, but he quit earlier this year, criticizing in a LinkedIn post about “Mark Zuckerberg and Facebook’s descent into toxic masculinity and Neo-Nazi madness.” (Meta did not respond to my question about this post.) Lemley was surprised by the results of the study, and told me that it “complicates the legal landscape in various ways for the defendants” in AI copyright cases. “I think it ought still to be a fair use,” he told me, referring to training, but we can’t entirely accept “the story that the defendants have been telling” about LLMs.
For some models trained using copyrighted books, he told me, “you could make an argument that the model itself has a copy of some of these books in it,” and AI companies will need to explain to the courts how that copy is also fair use, in addition to the copies made in the course of researching and training their model.
As more is learned about how LLMs memorize their training text, we could see more lawsuits from authors whose books, with the right prompting, can be fully reproduced by LLMs. Recent research shows that widely read authors, including J. K. Rowling, George R. R. Martin, and Dan Brown may be in this category. Unfortunately, this kind of research is expensive and requires expertise that is rare outside of AI companies. And the tech industry has little incentive to support or publish such studies.
The two recent rulings are best viewed as first steps toward a more nuanced conversation about what responsible AI development could look like. The purpose of copyright is not simply to reward authors for writing but to create a culture that produces important works of art, literature, and research. AI companies claim that their software is creative, but AI can only remix the work it’s been trained with. Nothing in its architecture makes it capable of doing anything more. At best, it summarizes. Some writers and artists have used generative AI to interesting effect, but such experiments arguably have been insignificant next to the torrent of slop that is already drowning out human voices on the internet. There is even evidence that AI can make us less creative; it may therefore prevent the kinds of thinking needed for cultural progress.
The goal of fair use is to balance a system of incentives so that the kind of work our culture needs is rewarded. A world in which AI training is broadly fair use is likely a culture with less human writing in it. Whether that is the kind of culture we should have is a fundamental question the judges in the other AI cases may need to confront.
The post What Two Judicial Rulings Mean for the Future of Generative AI appeared first on The Atlantic.
