Large language models and image generators are trained on vast quantities of text, images, music, and news, much of it protected by copyright, and the outputs they produce now compete in the very markets those works came from. Authors, artists, record labels, news organizations, and stock-photo companies have sued, and the cases are testing whether copying works to train AI is fair use and who is responsible when a model reproduces protected expression.
In this TalksOnLaw interview, recorded in the summer of 2024, Joel Cohen speaks with Shyamkrishna (Shyam) Balganesh, the Sol Goldman Professor of Law at Columbia Law School, a co-author of portions of Nimmer on Copyright and an adviser to the American Law Institute's Restatement of the Law, Copyright. Balganesh explains how generative models work, how copyright law applies to them, and where the law may be heading.
How the Models Work, and Where Copying Happens
Balganesh describes large language models as predictive systems, autocomplete at enormous scale, and explains how training breaks content into tokens and vectors, as well as how diffusion models learn to generate images by adding and removing noise. The key legal point, he explains, is that building these systems typically involves copying entire works onto local servers. Public availability on the internet is not permission to copy, and copies made for internal use can still implicate the reproduction right, leaving fair use as the central defense.
Fair Use: Campbell, Google Books, and Warhol
The conversation walks through the four statutory fair use factors and the precedents AI developers and rights holders invoke. Campbell v. Acuff-Rose Music made transformative purpose central but did not make it decisive. Authors Guild v. Google held that scanning millions of books to build a searchable index with limited snippets was fair use, an analogy AI companies rely on but that weakens when outputs compete with the original works. Andy Warhol Foundation v. Goldsmith made clear that a transformative label does not excuse a commercial use that serves the same purpose as the original.
The Cases
Balganesh analyzes the lawsuits that were then pending, including The New York Times v. Microsoft and OpenAI, in which the Times's complaint included examples of the model reproducing articles nearly verbatim; Getty Images' suit against Stability AI, which pointed to generated images that reproduced Getty's watermark; and the record labels' suits against the AI music services Suno and Udio. He explains why output evidence matters both to prove what was in the training data and to show market substitution, why developers' growing reticence about their training data creates proof problems, and the transparency requirements proposed in the COPIED Act, introduced in Congress in July 2024.
Outputs, Style, and Licensing
Substantial similarity, Balganesh explains, is not a mathematical threshold that engineers can code around, which makes guardrails against infringing outputs imperfect. Copyright does not protect style as such, but voice and likeness may be protected by the right of publicity, as Bette Midler's case against Ford showed. He predicts that the long-term answer will be licensing, likely through collective or aggregated arrangements because negotiating with millions of individual rights holders is impractical, and he considers how human creators and AI systems may come to work together.
What to Know Now
Courts have begun to answer the questions discussed here, with mixed results. In Bartz v. Anthropic (June 2025), Judge Alsup held that training a model on lawfully acquired books was fair use but that building a library of pirated copies was not, and the case settled for $1.5 billion, approved in 2026. Days later, in Kadrey v. Meta, Judge Chhabria ruled for Meta because the authors had not shown market harm, while suggesting that market dilution could defeat fair use on a better record. On September 29, 2026, in Thomson Reuters v. ROSS Intelligence, the Third Circuit issued the first federal appellate ruling on AI training, holding that copying Westlaw headnotes to train a non-generative legal research tool was not fair use and expressly leaving generative AI questions open. The New York Times case is now part of a consolidated litigation against OpenAI in the Southern District of New York, where summary judgment motions on fair use were filed in September 2026. In the United Kingdom, Getty largely lost its case against Stability AI in November 2025, while its U.S. case continues. Balganesh's prediction about licensing has largely borne out: Universal and Warner settled with Udio, Warner settled with Suno, with each settlement paired with licensing arrangements, and AI developers have signed numerous content licensing deals with publishers and news organizations. The U.S. Copyright Office released a report on generative AI training in May 2025 concluding that fair use must be decided case by case, and the D.C. Circuit held in Thaler v. Perlmutter (2025) that copyright requires a human author.
