The Wikimedia Foundation wants the judge in The New York Times’ copyright fight with OpenAI to split the question in two. Training an AI model on copyrighted work is likely fair use, it argues. A chatbot that copies websites to answer a user’s question is likely not.

The foundation runs Wikipedia. On Thursday, it asked the US District Court for the Southern District of New York for leave to file an amicus, or friend-of-the-court, brief. It announced the filing on Friday. The case, In re: OpenAI, Inc. Judge Sidney H. Stein will decide whether using copyrighted content to build large language models is fair use. That US legal doctrine allows some use of protected work without permission. He has not ruled on that, or on whether to accept the brief.

“We are taking a position in this case to defend the free and open human knowledge that Wikipedia depends on,” said Wikimedia Foundation chief executive Bernadette Meehan.

Training yes, chatbot answers no The proposed Wikimedia amicus brief supports neither party. It says both sides push “overbroad and simplified” readings of fair use. The Times alleges that using its content for any purpose, including training, breaks copyright law. OpenAI and Microsoft say training on publicly available data is fair use, according to Wikimedia’s summary.

Wikimedia asks the court to judge each use separately, citing the Supreme Court’s 2023 Warhol ruling. Scraping a website is not itself a “use”, it argues. Only what a company does with the copies should count.

Training a general-purpose model is “likely to be fair”, the brief says. The model is something new, and it does not replace the works it learned from.

Retrieval-augmented generation (RAG) is different. A chatbot using RAG copies text from sites such as Wikipedia or The New York Times when a user asks a question. It feeds the most relevant passages to the model and often reproduces much of the original in its answer, the brief says.

“While the pre-training of ChatGPT’s general LLM was fair use, OpenAI’s use of copies in ChatGPT’s RAG feature is not fair,” Wikimedia wrote in the brief.

Companies using RAG should follow each publisher’s licence terms, it argues. The same goes for newer techniques that serve a similar “grounding” purpose.

Credit and crawlers For Wikipedia, that licence is Creative Commons CC BY-SA 4.0. It requires credit to the original authors, and adaptations must stay under the same licence. When chatbots use Wikipedia content without following those terms, people are steered away from contributing, Wikimedia told the court.

Scraping also strains its servers, the filing says. That pushes Wikimedia to add technical safeguards, which bad actors often get around. A public park can stay free and open while stopping someone from “driving a container truck through it”, one footnote says.

Wikimedia wants heavy users to follow its public rate limits or pay for its Enterprise service. This month, it said rogue OpenAI agents had edited its wikis without approval.

“Any conversation around fair use and copyright in the age of AI cannot be limited to for-profit organizations: it must include stakeholders across the entire online information ecosystem,” said Stephen LaPorte, the foundation’s general counsel.

The wider fight Publishers have kept up the pressure on OpenAI. In June, nearly 400 local newspapers sued OpenAI and Microsoft. In July, the Times and other outlets asked a judge to sanction OpenAI, accusing it of hiding evidence. In September, the Justice Department sided with OpenAI, telling a New York court that AI training is fair use.

Wikipedia has more than 67 million articles in 300 languages, according to the foundation. Volunteers write it using reliable sources, including news articles.

“Expanding access to knowledge through AI is a powerful goal, but it must respect the people, contributors, and communities that make reliable knowledge possible,” Meehan said.