Law professor Matthew Sag files amicus brief arguing AI training is transformative but shadow-library sourcing needs scrutiny

What happened

Law professor Matthew Sag filed an amicus brief on 29 October 2026 in In re: OpenAI, Inc. Copyright Infringement Litigation, the consolidated litigation in the U.S. District Court for the Southern District of New York brought against OpenAI and Microsoft by The New York Times, other newspapers, The Intercept, the Center for Investigative Reporting, Ziff Davis and a proposed class of book authors led by the Authors Guild. Judge Sidney H. Stein is considering cross-motions for summary judgment on whether training large language models on copyrighted books and news articles is fair use - which Sag notes would be the first ruling on that question in the Second Circuit and the first anywhere to address news publishers' claims. The brief formally supports neither party.

The core argument

Sag argues that LLM training is highly transformative because its purpose is to learn statistical patterns and relationships rather than to reproduce protected expression for a new audience, and that what matters for fair use is what models ordinarily output under realistic conditions, not theoretical memorization that researchers can extract through adversarial prompting. He also argues that 'market dilution' - generalized competition from AI-generated content - is not a cognizable fair-use harm, and that comprehensive rights clearance for internet-scale training data is practically impossible, since the cost of identifying rights holders and administering payments would routinely exceed the value of individual permissions.

Where he says the court must still decide

The brief identifies two judgment calls that turn on evidence and could go against the defendants: whether using extracted information to create a direct substitute for a particular source counts as indirect expressive substitution under the fourth fair-use factor, and whether OpenAI's acquisition of training material from shadow libraries bypassed an available market for lawful access. On the latter point, Sag argues the court should reject the approach in Bartz v. Anthropic of treating acquisition as a separate use, at least when it is part of an integrated training process. His conclusion is that the plaintiffs' core case is weak, but the defendants are not entitled to a clean sweep until those factual questions are assessed.