The Wikimedia Foundation today filed a court brief arguing that AI training can qualify as fair use while chatbots that retrieve copyrighted material should follow publishers’ licensing requirements. The nonprofit behind Wikipedia is asking the court handling copyright lawsuits against OpenAI to distinguish between building models and supplying users with content retrieved from the web.
The friend-of-the-court filing went to the Southern District Court of New York in litigation combining 12 lawsuits against OpenAI, including the case brought by The New York Times against OpenAI and Microsoft. Wikimedia supports fair use for training large language models when the resulting models differ from the original works, arguing that such training should not require compliance with copyright licensing terms. Its position would allow some uncredited training uses of Wikipedia while requiring attribution when its material is retrieved to ground chatbot answers.
Read more: OpenAI Updates Model Spec With Five Values and Fewer Style Rules
That distinction centers on retrieval-augmented generation, or RAG, which finds and copies material online before returning that content through a chatbot. Wikimedia argues that RAG, along with newer methods serving a similar grounding function, should fall outside fair use and comply with the relevant publisher’s license. Wikipedia uses Creative Commons CC BY-SA 4.0, requiring creator credit and adaptations to carry the same license.
The New York Times filed its lawsuit in 2023, alleging that OpenAI and Microsoft infringed copyright by using its content, including for model training. OpenAI and Microsoft maintain that developing and deploying models using publicly accessible data qualifies as fair use. Wikimedia argues that an overly narrow interpretation could restrict innovation and tools useful to open source projects, while blanket protection could weaken the supply of secondary sources online.
Wikipedia contains more than 67 million articles in 300 languages, with nearly 250,000 volunteers contributing each month. Those contributors rely on sources including researched news reporting to create and improve articles. Wikimedia also says scraping its projects must respect public API rate limits or use its Enterprise API offerings, regardless of how the material is used.
The filing follows Wikimedia’s September advocacy at the United Nations General Assembly, where CEO Bernadette Meehan and its Global Advocacy team met policymakers, philanthropists and executives about protecting human knowledge amid AI development. In a separate dispute earlier this year, OpenAI accused DeepSeek of extracting knowledge from its models to train competing systems. OpenAI and Microsoft investigated and blocked accounts suspected of breaching service terms through that process.
