Now we know that we won’t have any new decisions on fair use in AI training from other district court judges until mid-2026 at the earliest, let’s take stock of where things stand.
3 Decisions on Fair Use in AI Training: 2 decisions for fair use and 1 against
It’s fair (no pun intended) to say the state of the law is nuanced, if not conflicted. We created this visual to summarize the current state of fair use law.
We call it the Fair Use Triangle:
Not fair use: 1 decision
Thomson Reuters v. ROSS Intelligence: Let’s start at the bottom of the Fair Use Triangle: In Thomson Reuters v. ROSS Intelligence, Judge Bibas reversed his 2023 ruling (that found a potential transformative purpose in training an AI model) and ultimately held that ROSS’s training an AI model that could produce a legal research tool that would compete with the plaintiff’s research tool was not a fair use. That was so even though ROSS’s legal search engine did not produce any infringing outputs, but rather, produced direct quotes from uncopyrightable judicial opinions. The case is now on interlocutory appeal in the Third Circuit.*
Fair use: 2 decisions
Several months later, Judge Alsup in Bartz v. Anthropic and Judge Chhabria in Kadrey v. Meta, in the very same week, ruled that training an LLM model was a highly transformative fair use in the respective copyright cases against Anthropic and Meta. But, as we explained in the post below, both opinions were not complete wins for the defendants and came with other parts of the decisions that were unfavorable to the defendants. These parts are depicted above in the red octagons, resembling stop signs.
Bartz v. Anthropic: Judge Alsup ruled that Anthropic’s training of its AI model that can respond to human queries in non-infringing ways was a highly transformative fair use. Indeed, Judge Alsup said it was “exceedingly transformative,” even “spectacularly so.” It was “among the most transformative [technologies] many of us will see in our lifetimes.” But Judge Alsup also held that Anthropic’s initial acquisition of copies of books from shadow libraries was for the purpose of building a general research library at Anthropic (in part because some copies were not used in any training) and such library building was not a fair use. Judge Alsup later clarified that his ruling on library building did not preclude Anthropic’s re-assertion of fair use at trial. (The case eventually settled and Judge Alsup preliminarily approved the $1.5 billion class settlement, the largest in U.S. copyright history.)
Kadrey v. Meta: A few days later, Judge Chhabria also held that Meta’s training of its model with copyrighted books was a highly transformative fair use. Judge Chhabria’s fair use analysis is similar to Judge Alsup’s in many aspects of the four factors.
However, Judge Chhabria differed from Judge Alsup in two key respects. First, Judge Chhabria ruled that Meta’s initial acquisition of book copies from shadow libraries was a part of the same use to train its models because the ultimate purpose of the acquisition was to use the copies to train.
Second, in dicta, Judge Chhabria also stated that, where the plaintiffs showed with evidence that they suffered a new type of harm called copyright “market dilution” from non-infringing works that other people created with AI in the same genres of works used in the training materials, he expected that such training “in most cases” would be illegal and not fair use. However, the problem in Kadrey was that the plaintiffs failed even to advance this new theory of copyright dilution and did not submit meaningful evidence to substantiate that any of the plaintiffs actually suffered any dilution at all on summary judgment. (By contrast, Judge Alsup rejected the book authors’ argument for recognizing a new theory of market dilution based on similar genres.)
DOWNLOAD A PDF OF THE FAIR USE TRIANGLE WITH LINKS TO EACH DECISION
(*Full disclosure: I joined an amici brief of copyright law professors arguing that this ruling in Thomson Reuters was wrong.)
Related Stories