Wikimedia asks court to protect AI training, not unrestricted copying in chatbot answers

The nonprofit behind Wikipedia is asking a federal court to distinguish how AI learns from how chatbots deliver answers. In an October 30 announcement, the Wikimedia Foundation said it filed a friend-of-the-court brief supporting fair use for transformative model training, but opposing that protection for systems that retrieve and reproduce copyrighted material.

The filing joins In re: OpenAI, Inc. Copyright Infringement Litigation in the Southern District of New York. The proceeding combines 12 lawsuits, including The New York Times’ case against OpenAI and Microsoft. An amicus brief gives the court an outside party’s position; Wikimedia is presenting an argument, not announcing a settlement.

A publisher dispute draws in Wikipedia

The Times filed its lawsuit in December 2023, alleging that the companies illegally used millions of its copyrighted articles to train AI chatbots that compete with the publication. OpenAI and Microsoft have argued that using web content to train their systems qualifies as fair use under U.S. copyright law.

Wikimedia’s response rejects a sweeping answer in either direction. It argues that overly restrictive copyright treatment could block useful tools for open-source projects. Blanket fair-use protection, it warns, could weaken the human sources those projects depend on and leave the web more closed.

Its interest is practical as well as legal. The foundation says AI can help monitor its sites for illegally posted personal information and assist contributors with edits, according to Gizmodo’s account of the brief. Wikimedia does not want a copyright ruling that categorically prevents the development of those capabilities.

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

Bernadette Meehan, Wikimedia Foundation CEO

The line falls between training and retrieval

For model training, Wikimedia argues that using copyrighted content should count as fair use when the resulting language model is new and different from the original sources. Under that proposed approach, developers would not have to satisfy copyright licensing terms for that training use. The argument rests on transformation, rather than unrestricted permission for every later use.

The foundation draws a different conclusion for retrieval-augmented generation, or RAG. This technique retrieves outside material to help ground a chatbot’s response in current sources. Wikimedia’s announcement focuses on uses that find and copy works from the web, then deliver that content to users through a chatbot.

For those uses, and newer techniques serving a similar purpose, it argues that companies should follow publishers’ licensing requirements rather than claim fair use. Its proposed distinction therefore reaches beyond the initial collection of training data to the way a deployed chatbot supplies source material in an answer.

Open access still comes with conditions

Wikipedia’s own license makes the distinction concrete. Its Creative Commons CC BY-SA 4.0 terms require credit to the original creators and require remixes or adaptations to remain under the same license. Wikimedia says retrieval-based uses of its content must honor those requirements, even though readers can access the material freely.

Copyright is not its only condition. Regardless of the intended use, the foundation says automated collection must respect its public API rate limits or use its Enterprise API offerings. Those are separate access requirements: its support for training does not erase the operational limits it places on collecting Wikimedia content.

The foundation says nearly 250,000 volunteers contribute to Wikipedia each month, drawing on reliable sources such as reported news articles. Its intervention asks the court to protect both sides of that chain: tools that expand access to knowledge and the conditions that keep people producing it. Whether the court adopts that distinction remains unresolved.