The AI Copyright Fight Has Narrowed to One Fair-Use Factor

Three district rulings, none of them the last word

Fair use is the escape valve in the Copyright Act. Section 107 lets someone copy a protected work without a license when four factors weigh, on balance, in the copier's favor: the purpose and character of the use, the nature of the copyrighted work, how much of it was taken, and the effect of the use on the market for the original. Whether training a generative model on copyrighted books and articles clears that bar reached the merits three times in 2025. The rulings do not line up.

In June, in the same San Francisco courthouse and two days apart, two judges read comparable records and came out opposite ways. Judge William Alsup held that training Anthropic's Claude models on books was "exceedingly transformative," a use that repurposes rather than substitutes, and therefore a fair use.¹ Judge Vince Chhabria granted Meta summary judgment on its own training claim, then wrote that he had done so only because the thirteen authors before him "presented no meaningful evidence on market dilution at all," and that in most cases training on copyrighted books will likely be illegal.² Four months earlier, Judge Stephanos Bibas had handed down the first of the three orders, holding that Ross Intelligence's copying of Westlaw editorial headnotes to build a competing legal-research tool was not transformative and not fair use, with the market-effect factor favoring Thomson Reuters because Ross set out to build a market substitute; Ross was found to have copied 2,243 of the 2,830 headnotes at issue.³

Neither generative ruling is the clean victory a headline could make of it. Alsup split his case in two. Training on lawfully acquired books was fair use, but the way Anthropic had stocked its library was not: it had downloaded "over seven million copies of books" from pirate sites and kept them, and "such piracy of otherwise available copies is inherently, irredeemably infringing" even when the pirated copy is then put to a transformative use.¹ Anthropic settled that surviving liability in September. The plaintiffs' motion for preliminary approval valued the fund at about $1.5 billion, roughly $3,000 for each of some 500,000 works, and called it the largest publicly reported copyright recovery in history.⁴ Anthropic's deputy general counsel said the deal would "resolve the plaintiffs' remaining legacy claims."⁵ Chhabria's win for Meta is narrower still: his order "does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful," and binds only the thirteen plaintiffs before him.² No appellate court has yet weighed in. The Ross appeal, cabined to non-generative AI, is the first to reach the Third Circuit.³ What each side would put its name on is already on the record, in briefs, testimony, and public statements, and in the two June orders that adopted opposite halves of it.

The disagreement has collapsed onto the fourth factor

What no longer divides the sides is most of the doctrine. Both generative rulings hold that training is transformative; Alsup calls it spectacularly so,¹ and Chhabria agrees the point is not seriously disputable.² The Copyright Office concedes that "various uses of copyrighted works in AI training are likely to be transformative."¹⁵ The rights-holders' own scholar grants that outputs often will not infringe, which is why she moves the fight to the inputs.¹² Strip away what everyone agrees on and one factor is left holding the entire question: the effect of the use on the market.

Inside that factor sit two disagreements of different kinds, and the cases keep fusing them. The first is legal, a dispute about what the statute protects. Does the market-effect factor reach "market dilution," the slow erosion of a work's market by a glut of machine-made substitutes, or only the direct substitution of one copy for another? Alsup says the statute "seeks to advance original works of authorship, not to protect authors against competition."¹ Chhabria says "indirect substitution is still substitution," and the Copyright Office agrees dilution is a "serious risk" the factor should count.²¹⁵ The EFF answers that policing dilution would reach uses copyright was never meant to control.⁸ Neither judge misread the record here; they read the same clause of the same statute to protect different things.

The second disagreement is a forecast, and it is where the litigation actually turns. Even Chhabria's theory needs proof: Meta won not because dilution does not count but because the plaintiffs "presented no meaningful evidence on market dilution at all."² The theory that could sink the labs has never been carried on a trial record. Whether it can be is a claim about the world, not the statute: whether the outputs of a given model measurably displace sales of the specific works it trained on, and whether a plaintiff can quantify it. Each camp's strongest voice concedes the hinge. Sag grants that if a model's ordinary outputs infringe or its access was unlawful, "there is no obvious fair use rationale to replace" non-expressive use;⁷ Alsup proved the access half by holding Anthropic's piracy unprotected, and Anthropic paid $1.5 billion to close it.¹⁴ Neither absolute survives its own side's best minds.

Two things would settle it, and both are in motion. One is an appellate ruling on whether dilution is cognizable under the fourth factor; the split across the district orders makes review close to inevitable, and the Ross appeal, though it turns on non-generative copying, is already before the Third Circuit.³ The other is a trial record in which a well-resourced plaintiff finally proves dilution of the works actually trained on. The New York Times, which has pleaded substitution directly and can fund the market economics Kadrey's authors did not, is the case that will test whether the harm can be shown rather than asserted.⁹ Until one plaintiff proves the market harm and one appeals court says whether that harm counts, the record supports no verdict wider than the one both judges reached in their own cases: it depends on the use, and on evidence no one has yet put in.