The Wikimedia Foundation has asked a New York court to distinguish ChatGPT’s model training from its use of retrieved works, saying the issue may affect access to knowledge and publishers’ control over their content.
The proposed amicus brief, dated Oct. 29, was attached to a motion seeking permission to file it in In re: OpenAI, Inc. Copyright Infringement Litigation. Wikimedia announced it Oct. 30 and said it was not taking either party’s side. An amicus brief allows a nonparty to explain how it believes a court ruling could affect people beyond those in the case.
The filing applies the four-factor fair-use test to ChatGPT’s general large language model, arguing that pretraining it is likely fair use. Wikimedia treats ChatGPT’s retrieval-augmented generation (RAG) separately, saying this use is likely not fair and should comply with applicable licenses. The argument covers the two ChatGPT uses discussed in the proposed filing and does not set a general standard for other models.
Wikipedia has more than 67 million articles across 300 languages, and Wikimedia says its volunteer contributors could lose useful tools if courts limit AI training too broadly. The foundation also warned that publishers might restrict access to copyrighted works that AI services include in RAG responses without following license terms.
In a 2023 response to the U.S. Copyright Office (USCO), Wikimedia called for AI fair-use questions to be assessed case by case.
The filing enters a U.S. debate shaped by decisions on separate fact patterns. EconoTimes’ account of June 2025 rulings described how two district courts assessed different AI training practices. Neither decision set a nationwide standard for all such practices.
