Wikimedia asks court to separate AI training from retrieved answers

The Wikimedia Foundation has asked a US court to assess AI training separately from systems that retrieve source material to answer questions. In an October 30 statement, Wikipedia’s operator backed fair use for transformative model training while arguing that answers which reproduce material and replace visits to the original should respect content licences.

The distinction concerns two uses of the same material: building a model from a collection of works, and fetching documents after a user asks a question. The latter is commonly called retrieval-augmented generation, or RAG. Wikimedia wants the court to assess the purpose of each use, including whether retrieved output substitutes for its source.

The court-stamped record is dated October 29, a day before the Foundation’s announcement. It contains a motion seeking permission to file a proposed friend-of-the-court brief, whose cover expressly supports neither party. The request for leave is a procedural step; the document does not establish that a judge has accepted Wikimedia’s legal position.

OpenAI separately published selected excerpts from supportive outside briefs on October 30. Its page highlights arguments about the benefits of AI training for research, creators and access to information. Its selected material includes arguments from accessibility and library groups as well as independent creators.

Fair use permits certain uses of copyrighted works without permission. US copyright law directs courts to consider the purpose and character of a use, the nature of the original, the amount and significance copied, and the effect on a work’s potential market or value. Purpose and market effects are central to the competing submissions.

Training copies and answers that replace the source

Public Knowledge said its October 29 brief asks the court to separate training from questions about acquiring data and infringing outputs. It describes training as extracting information, such as word frequencies, from works. The group acknowledges risks to creators but argues that addressing those risks should preserve the protections fair use provides.

The News/Media Alliance takes the publishers’ side. On October 30, it announced a joint brief with the National Newspaper Association, the National Press Photographers Association and ProPublica. The groups argue that AI summaries and reproduced news replace visits to publishers, damaging advertising, subscriptions and a developing market for licensing news to AI companies. Their objections cover both training and real-time retrieval.

Authors Alliance focuses on competition from new works that do not copy protected expression. It sought leave to file on October 29, arguing that losing readers to such works—what the parties call market dilution—should not count as copyright harm. Its position is expressly limited: it takes no view on how OpenAI acquired training material or whether particular outputs infringe.

Wikipedia uses CC BY-SA 4.0, a licence requiring attribution and the same terms for adaptations. Creative Commons’ official explanation specifies appropriate credit, a licence link and an indication of changes. It permits commercial sharing and adaptation while barring additional restrictions on those freedoms. Applicable legal exceptions remain available. Wikimedia’s submission asks the court to distinguish uses that qualify for an exception from those that must follow licensing terms.

The consolidated litigation concerns claims by authors and news organisations against OpenAI and Microsoft. The cases were consolidated in April 2025 before Judge Sidney Stein. Public Knowledge says both sides have sought summary judgment on the facts of the case. Wikimedia’s published record asks the court to allow its proposed brief; it contains no order granting that request and supplies no date for a final ruling.