Authors begin suing over training data

Comedian and author Sarah Silverman, along with novelists Richard Kadrey and Christopher Golden, filed copyright infringement suits against OpenAI and Meta in the US District Court for the Northern District of California, alleging that ChatGPT and Meta’s LLaMA had been trained on their books without permission or compensation. The filings were among the first author-led actions against AI developers over training data and opened a category of litigation that would grow substantially over the following two years.

The complaints’ central evidentiary claim was that the models could produce detailed, accurate summaries of the plaintiffs’ books — including Silverman’s memoir “The Bedwetter” — when prompted, which the plaintiffs argued was only possible if the underlying text had been ingested during training. That inference, rather than proof of verbatim reproduction, was the plaintiffs’ main basis for alleging infringement; the suits sought class-action status on behalf of authors more broadly and unspecified damages, including in some claims the destruction of the trained models.

The suits arrived as one of the earliest entries in what became a long run of copyright litigation against AI developers: further author class actions followed against OpenAI specifically, image-generator suits from Getty and visual artists, and later, The New York Times’ suit against OpenAI and Microsoft, which added detailed exhibits of near-verbatim regurgitation that the Silverman complaints had not. Courts handling the various author suits would go on to reach mixed and sometimes conflicting rulings on whether training itself was infringing, whether output resembling a work was required, and whether fair use covered the practice — leaving the underlying legal question unresolved well beyond this filing and making it a marker of the dispute’s start rather than any part of its resolution.