Ross ruling exposes the limits of AI’s fair-use defense

The narrow case behind the ruling

The court found that Ross had effectively copied Thomson Reuters’ “editorial annotations,” short summaries of legal issues drawn from Westlaw, for its legal research system. A federal judge in Delaware had said last year that Ross used the annotations to make it easier to build a competing tool, making the use non-transformative.

That finding matters, but the case is not a direct ruling on generative AI training. Ross’s system was not generative, the company no longer exists, and damages have not been decided. The Third Circuit’s decision came on an interim appeal, not after a final resolution of the entire dispute.

Ross shut down in 2021, shortly after the lawsuit was filed, citing growing financial difficulties caused by the litigation. The parties involved have said they will appeal to the Supreme Court.

A warning, not a general rule

Fair use depends on factors including whether a work is transformed, how much of it is used, and whether copying affects the market for the original. AI companies rely on the doctrine as a defense against claims that they used copyrighted material without permission to train models.

Dozens of lawsuits have been brought by writers, comedians, music industry figures, major newspapers and Encyclopaedia Britannica. The Ross decision is the first of its kind, according to Reuters, but how far it will shape those cases remains uncertain.

Aaron Moss of Copyright Lately said even some in the AI industry had treated the ruling as a win. Adam Eisgrau, senior director for AI at industry group Chamber of Progress, wrote on X that the decision indirectly supported the view that training generative AI on deeply transformed material to create a general-purpose model with significant public benefit could qualify as fair use.

“Judging by the victory celebrations, you’d think everyone had read a different decision,” Moss wrote. “But they hadn’t.”

The bigger fight is still ahead

Moss argues that Ross is a poor vehicle for a sweeping ruling on generative AI training. The case concerned a non-generative product and a particular kind of copying, while the question of damages remains open. There are other cases, he says, that are better suited to determine the future of generative AI training.

One of them is a separate lawsuit by Concord Music Group, which accuses Anthropic of unlawfully using thousands of copyrighted song lyrics to train Claude. Hearings are scheduled for later this month.

I think the most important signal in the Ross case is its limit: courts can reject a training-related fair-use defense where the record shows close copying, without settling whether broad model training is fair use. That leaves AI companies with no general rule to rely on—and copyright owners with more cases in which to test the boundaries.