Fair use benefits everyone

OpenAI published excerpts from amicus briefs filed in support of its position in The New York Times' copyright lawsuit, arguing that training AI models on copyrighted works is fair use. The filers it highlights include library and disability groups, the Authors Alliance, startup advocates, law professors, independent creators, a Pulitzer Prize-winning journalist, and the United States government.

This is one of the defining legal fights of the AI era. How courts treat training data will shape who can build models, what they cost, and whether publishers get paid. The breadth of support OpenAI can point to matters, especially the federal government's brief, which argues that the Times' position would severely hamper progress.

Let's be clear about what this page is. It's advocacy from a party to the lawsuit, selecting the friendliest passages from briefs filed on its side. OpenAI's own framing is sharp: it accuses the Times of chasing an undeserved payday. That's a litigation posture, not a neutral summary, and readers should weigh it that way. The full filings are on the public court docket, and the Times and its allies have their own arguments that don't appear here.

That said, some of these arguments are serious and we find them compelling. The library and accessibility coalition makes a point that often gets lost: restricting training doesn't only hit big labs, it hits medical researchers, educators, people with disabilities, and anyone relying on open-source models built by others. Public Knowledge ties AI training to a line of fair-use cases that protected image search, Google Books, and accessible digital books for readers with print disabilities. If training is the next step in that lineage, a sweeping loss for AI developers could entrench the biggest players, who can afford licenses, while freezing out everyone else.

The most important legal fight here may be over market dilution. The Authors Alliance asks the court to reject the idea that aggregate competition from noninfringing AI outputs counts as market harm under copyright's fourth factor. Their argument: copyright protects against substitution for your protected expression, not against new work competing for the same audience. Doctrinally, that's a strong point. Practically, it's also the part that should worry anyone who makes a living writing. If AI output floods the same market and erodes the value of original reporting, and the law says that doesn't count, newsrooms bear real economic damage with no remedy. Being right on doctrine doesn't make that harm go away.

I'd also flag the Foundation for American Innovation's suggestion that national-security value should weigh in the fair-use analysis. That's an expansive idea, and a risky precedent. Copyright analysis that bends toward whatever is strategically useful to the state is not a principle we'd want applied to every future technology.

Our read: the case for training as transformative use is strong, and we think AI that learns broadly from public knowledge is worth protecting. But the honest version of OpenAI's argument would admit that journalism is under real economic pressure from these systems, and that winning in court doesn't settle whether the people producing the underlying work can survive. The likelier lasting outcome is a legal win for training paired with growing pressure for licensing deals and output-side limits on regurgitation.

Fair use under 17 U.S.C. § 107 weighs four factors, including the purpose of the use and its effect on the market for the original work. Courts in the Second Circuit, where Google Books was decided, have treated highly transformative uses favorably. Several of the excerpted briefs argue that training produces intermediate copies that are not similar to model outputs and do not substitute for the original works.