AI Training Breaks Copyright's Transformative Use Test

For three decades, U.S. copyright law has centered fair use decisions on a single question: Is the new work transformative enough to justify unlicensed copying? That framework, crystallized in Judge Pierre Leval's influential 1990 article and the Supreme Court's 1994 Campbell v. Acuff-Rose decision, worked because transformative uses rarely competed with the original works they drew from. Parodies don't substitute for the songs they mock. Legal databases don't replace the briefs they index.

Generative AI shatters that assumption. In a ProMarket analysis, legal scholar Shishene Jing argues that large language models represent both the most transformative use of copyrighted material courts have ever seen and the use most capable of displacing the original works' markets.

The transformation-substitution paradox

The problem emerged clearly in recent litigation. In the 2025 case Kadrey v. Meta, a judge noted that AI training is "simultaneously one of the most dramatically transformative uses courts have ever evaluated and one of the most plausibly market-displacing." An LLM doesn't reproduce an author's prose verbatim—it creates statistical abstractions that generate genuinely new text. Yet a model trained on a novelist's corpus can produce competing novels that vie for the same readers and publishing contracts.

This paradox has split federal courts. Judge William Alsup called Anthropic's book-training "exceedingly transformative" in Bartz v. Anthropic and ruled in favor of the AI company. Judge Stephanos Bibas, examining similar conduct in Thomson Reuters v. Ross Intelligence, found no fair use because of direct market substitution. Both judges applied the same four-factor test mandated by Section 107 of the Copyright Act, yet reached opposite conclusions.

Two competing frameworks

The divergence reflects a deeper tension in copyright scholarship. Wendy Gordon's 1982 "Fair Use as Market Failure" framework argues fair use should exist only where transaction costs prevent licensing markets from forming—essentially, where no market transaction would occur anyway. Leval's transformativeness standard, by contrast, justifies fair use based on whether the new work adds meaning or purpose, largely independent of market effects.

Legal practice has favored Leval's approach because most pre-AI transformative uses couldn't economically substitute for their sources. A teacher photocopying pages for an unanticipated class doesn't compete with textbook sales. Google's book snippet view doesn't replace buying the book. The correlation between transformation and non-substitution held—until generative AI broke it.

Why it matters

The Supreme Court's 2023 Andy Warhol Foundation v. Goldsmith decision already signaled discomfort with treating transformativeness as an all-purpose defense, emphasizing that commercial uses sharing the same purpose as the original deserve less protection. AI litigation is now forcing courts to choose: Should fair use prioritize aesthetic judgment about what makes a work novel, or economic analysis of whether it competes in the same market?

Jing argues the fourth statutory factor—market harm—offers a more administrable standard than the first factor's vague transformativeness inquiry. Market displacement is an empirical question courts and economists can answer with evidence: Does a functioning or emerging licensing market exist? Does the new use compete for the same audience? Federal judges are better equipped to evaluate market evidence than to render aesthetic judgments about creative novelty.

A market-focused framework wouldn't eliminate fair use protections for parody, criticism, or scholarship—none of those genuinely substitute for the works they reference. What it would screen out is the category AI has created: uses that are creatively transformative and commercially substitutive simultaneously.

The details were first reported by Shishene Jing in ProMarket, published by the University of Chicago's Stigler Center.