Wikimedia tells court in NYT v. OpenAI that training may be fair use but RAG copying is a different matter

The nonprofit behind Wikipedia has entered the biggest copyright fight over generative AI in the United States. The Wikimedia Foundation asked the federal court in Manhattan on October 29 for permission to file an amicus brief, and publicized the filing the next day, according to The Next Web.

Where the case stands

The brief targets In re: OpenAI, Inc. Copyright Infringement Litigation, which groups 12 lawsuits against OpenAI, including the suit The New York Times filed in 2023 against OpenAI and Microsoft. Judge Sidney H. Stein of the Southern District of New York must decide whether building large language models on copyrighted material counts as fair use. As of October 30, he had ruled neither on that question nor on whether to accept Wikimedia's brief. Engadget, citing AI Lawsuit Tracker, reports that a summary judgment ruling is expected soon.

Training and retrieval, treated apart

Wikimedia sides with neither party. It argues that both push readings of fair use that are too broad, and asks the court to examine each use on its own terms, drawing on the Supreme Court's 2023 Warhol decision. In its view, scraping a site is not itself the relevant use; what a company does with the copies is.

On that basis, the brief says training a general-purpose model is likely fair use, because the result is new and does not stand in for the works it learned from. Retrieval-augmented generation (RAG), where a chatbot pulls passages from sources such as Wikipedia or the Times and often repeats much of them in its answer, is treated differently: Wikimedia argues that this use is likely not fair, and that OpenAI's RAG feature in ChatGPT falls on the wrong side of the line, The Next Web reports. Engadget's summary frames the objection around answers that reproduce articles without attribution.

Wikipedia's stake

Wikipedia's text is published under a CC BY-SA licence that requires credit. The Foundation argues that chatbots ignoring those terms discourage volunteers from contributing, and that heavy scraping strains its servers; it wants large users to respect its public rate limits or pay for its Enterprise service. General counsel Stephen LaPorte warned that a careless ruling "could threaten the sustainability of free knowledge," as quoted by The A.V. Club.

In September, the US Justice Department had already told a New York court that AI training is fair use, siding with OpenAI. Neither OpenAI nor the Times had publicly responded to Wikimedia's filing in the reports reviewed.