Two judges in one courthouse split the law on AI training

Two days apart, two theories of the same factor

On June 23, 2025, Judge William Alsup held that using books to train a large language model was "quintessentially transformative," and threw out the authors' central theory of harm. Their worry that a trained model would flood the market with competing books was, he wrote, "no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works," and "not the kind of competitive or creative displacement that concerns the Copyright Act."¹ Two days later, in the same courthouse, Judge Vince Chhabria took the opposite view of the same question. Market dilution, the claim that a flood of machine-made books would sap demand for human-written ones without any one output copying any one book, was "the most important factor in the fair use analysis" and "the potentially winning argument" that these particular plaintiffs had failed to make.² He named Alsup's schoolchildren analogy and called it "inapt," "not a basis for blowing off the most important factor."²

This is not two outcomes. It is two incompatible rules about whether a specific kind of market harm counts, entered forty-eight hours apart by two judges of the Northern District of California. Four months earlier, in Delaware, Judge Stephanos Bibas had rejected a fair-use defense outright, for reasons that reach neither of them.³ The district courts that have reached the merits of training-data fair use disagree less case by case than axis by axis. Three fault lines run through the rulings: whether a court makes transformative purpose or market substitution decisive, whether the trained system is generative or not, and whether the corpus was bought or pirated. No court of appeals has yet resolved any of them.

The split runs through factor four

Bibas gave Thomson Reuters both of the factors that matter most. He held Ross's use non-transformative because it served the same purpose as Westlaw, "to make it easier to develop a competing legal research tool," and he found the fourth factor for Thomson Reuters on a market that did not yet exist: "the effect on a potential market for AI training data is enough," whether or not Thomson Reuters had entered it.³ Factors one and four, he wrote, "weigh most heavily."³

Alsup and Chhabria both gave the developer the first factor and agreed the training was transformative. Alsup called it "spectacularly so"; Chhabria called Meta's use "highly transformative."¹² They split on the fourth. Alsup treated the dilution theory as a category the Copyright Act does not recognize, and rejected the lost-licensing theory in the same stroke: a market for training licenses "is not one the Copyright Act entitles Authors to exploit."¹ Chhabria treated dilution as not only cognizable but usually decisive. "It seems likely that market dilution will often cause plaintiffs to decisively win the fourth factor — and thus win the fair use question overall — in cases like this," he wrote.² Meta won anyway, because its plaintiffs "made the wrong arguments and failed to develop a record in support of the right one."²

The developers' losing scenario, at its strongest, runs like this. A model trained on an author's work can generate an unbounded supply of near-substitutes at almost no cost, displacing that author's market as surely as verbatim copying would, without ever reproducing a sentence. Chhabria found that argument strong enough to predict developers "will often" lose on it. Alsup found it outside the statute. The facts narrow the gap a little: Ross was a direct Westlaw competitor copying Westlaw's own curated product, while Meta trained a general model on books it did not compete with. But the stated rules still point in opposite directions on whether a training-license market counts, and neither opinion pretends otherwise.²

For a rightsholder, the split sets the burden. Before a judge who follows Chhabria, dilution is the winning theory, but it has to be built with evidence, the survey and expert testimony the Kadrey plaintiffs never produced. Before a judge who follows Alsup, the same evidence is beside the point.

The generative line that saved Anthropic and sank Ross

Ross turned on a distinction the other cases barely had to discuss. "It is undisputed that Ross's AI is not generative AI," Bibas wrote, meaning AI "that writes new content itself."³ Ross's tool answered a legal query by returning existing judicial opinions rather than composing new text, and it had been trained on Westlaw headnotes, the short summaries Westlaw writes for each decision. Bibas cabined his holding to that fact: "Because the AI landscape is changing rapidly, I note for readers that only non-generative AI is before me today."³

Alsup took him at his word and used the line to distinguish Ross. The Ross court, he noted, "twice emphasized" that what was trained "was not generative AI," and "that was not transformative."¹ Training a model to generate new text was a different use, closer to a 2014 decision that found fair use when a company copied lawyers' briefs to build a searchable database of them.¹ Chhabria's Llama analysis ran the same way: the model was "highly transformative" because it was built to generate diverse new text, a purpose the books themselves did not serve.²

So the generative fact does opposite work depending on the corpus, clearing the first factor for Anthropic and Meta and sinking Ross. That matters for the appeal now pending, because a ruling that trumpets its own limits is hard to extend. If the Third Circuit affirms on Bibas's reasoning, it affirms a holding that on its face governs only non-generative systems, and the generative cases keep whatever protection the first factor gave them. For a company building a non-generative tool that repackages a rival's curated data, Ross is already a loss. For a company training a generative model, Ross is a case about someone else.

Where the books came from splits the same two judges

Alsup split Anthropic's conduct in three. Training on books was fair use. Converting purchased print books to digital was fair use for a narrower reason, since Anthropic destroyed the print copy and kept the file as a one-for-one replacement. But downloading and permanently keeping more than seven million pirated books from shadow libraries to build "a central library of all the books in the world" was not fair use, and on that piece "every factor points against fair use."¹ "The person who copies the textbook from a pirate site has infringed already, full stop," he wrote, and buying a copy afterward "will not absolve it of liability for the theft."¹ That conduct went to trial with damages on the table, willfulness included.¹

Chhabria refused to treat shadow-library sourcing as an automatic loss. To say Meta's downloading "was 'piracy' and thus cannot be fair use begs the question," he wrote, and the download had to be judged "in light of its ultimate, highly transformative purpose," a conclusion he marked as "contra Bartz."² Same sourcing method, roughly the same shadow libraries, opposite rule on whether it independently defeats the defense. Part of the distance is posture: Anthropic retained its pirated files permanently and the authors pressed the point, while Meta's plaintiffs under-argued it. The stated legal rules conflict all the same.

The stakes on this axis are the most concrete in the litigation. Anthropic's training holding was a clean win, but its piracy exposure was large enough that the case never reached appeal on the merits. The parties settled the pirated-library claims for a reported $1.5 billion, and the training holding was never tested above the district court.⁴ For a developer, the lesson is that how the corpus was obtained can carry more financial risk than whether training is fair use, and a favorable ruling on the second question does not touch the first.

The Times case has the fact the others lacked

One fact unites the three completed cases and separates them from the fourth. In none of them was any infringing output proven. Alsup was explicit that "Authors do not allege that any LLM output provided to users infringed," and that "if the outputs seen by users had been infringing, Authors would have a different case."¹ Chhabria's record was the same: even under adversarial prompting, no Llama model would reproduce more than fifty words of a plaintiff's book, so "Llama cannot currently be used to read or otherwise meaningfully access the plaintiffs' books."²

The New York Times case is the different case Alsup described. Its core allegation is that GPT-4 and ChatGPT reproduced Times articles at length to users, the regurgitation the California records lacked. The case survived a motion to dismiss, with the direct and contributory infringement claims allowed to go forward.⁵ It is now the lead matter in a consolidated proceeding, In re: OpenAI, before Judge Sidney Stein in the Southern District of New York, and the plaintiffs have pushed discovery into tens of millions of ChatGPT conversation logs.⁶ In July 2026 they moved to sanction OpenAI over deleted output data, an accusation OpenAI disputes and the court has not ruled on.⁷

That posture is why the California holdings do not decide the Times case. Both rested on the absence of infringing output. The Times pleads its presence. Whether it can prove that models reproduced protected expression, and that the reproduction cost it a market, is a separate question the record does not yet answer. For a developer whose model can be induced to emit its training text, the output seam is where the favorable California reasoning stops. For a rightsholder, it is the one theory the completed cases left open.

What is settled, and what an appeals court still has to decide

What is settled is narrow and almost entirely district-level. A non-generative tool that copies a competitor's curated data to build a rival product is not protected: that is Ross, and it is on appeal.³⁸ Training a generative model on lawfully obtained books is fair use in the two courts that have ruled, and building the corpus out of pirated files is not: that is Bartz.¹ A generative model can be highly transformative under the first factor even where the plaintiffs lose the case they should have won: that is Kadrey, which formally binds only its thirteen named plaintiffs.² None of it binds beyond the deciding court.

What is pending is the part that would turn any of this into law. The Third Circuit heard oral argument in Thomson Reuters v. Ross on June 11, 2026, and has not yet decided.⁸ The Bartz settlement went to a final fairness hearing in May 2026 and, as reported, still awaits a final-approval order, so its training holding stays unappealed.⁴ The factor-four split is not going up quickly either. On July 8, 2026, Chhabria denied the Kadrey plaintiffs' Section 1292(b) motion to certify the fair-use ruling for interlocutory appeal, choosing to send the question up later on a final judgment, "as part of a tidy package."⁹¹⁰ The distribution claims are not set for a dispositive hearing until February 2027.¹¹

What can be predicted comes with a probability and a reason, or it stays out.

It is tempting to read the four cases as a line running from a pro-rightsholder pole at Ross to a pro-developer pole at Bartz and Kadrey. The opinions resist it. Kadrey sits in the developer's win column and reads as the sharpest argument against developers yet written, a roadmap for the plaintiff who builds the dilution record Meta's opponents did not. Bartz split the defense down the middle and produced the largest copyright settlement on record. The honest description is a fracture, not a trend: three axes decided differently by trial judges who cannot bind each other, waiting on appeals courts that have said nothing. Until they speak, the same training run can be fair use in one courthouse and a billion-dollar liability in the next.