Large bound volumes with worn, blank cloth and leather spines in dark red, brown, green and black fill tall grey steel shelves on both sides of a narrow concrete aisle, lit by a tall window at its far end.

Copyright

Wikimedia Asks a Judge to Split Training From Answers

Ten friend-of-the-court filings reached the judge in OpenAI’s copyright cases Thursday; Wikipedia’s operator argued training can be fair use but retrieval that substitutes for its sources is not.

NEW YORK, Oct. 29 — The Wikimedia Foundation, which runs Wikipedia, asked the federal judge weighing the copyright suits against OpenAI and Microsoft on Thursday to split the question of fair use in two: pre-training the models behind ChatGPT is fair use, it argued, but copying and reproducing sources in ChatGPT’s answers through retrieval augmented generation is not. Its brief, filed for neither side, was one of 10 friend-of-the-court filings docketed Thursday.The FoundationA nonprofit in San Francisco that hosts 13 free-knowledge projects, Wikipedia the best known. Most written contributions are licensed under Creative Commons CC BY-SA 4.0, which requires credit to the creators and that adaptations carry the same license.

The same day, Microsoft refiled on the public docket its opposition to the news publishers’ motion for summary judgment. “The evidence decisively shows that the launch of LLMs has not hurt News Plaintiffs at all,” it wrote. “Subscription trends have been unaffected; The Times’s have even grown.”

Judge Sidney H. Stein is weighing motions from both sides to decide fair use without a trial in the consolidated cases brought by The New York Times, other news publishers and authors. Replies are due Nov. 20.

Wikimedia’s line runs between the parties: Microsoft asked the court to find that “both LLM training and Microsoft’s web grounding technology are fair use as a matter of law,” while the publishers, in their September motion, sought summary judgment “at each stage of the AI pipeline,” from acquiring and training to, against Microsoft alone, grounding and outputs. If the judge weighs each use apart, as Wikimedia asks, the companies could win on training and still lose on answers. Wikimedia argues that a ruling requiring compliance with licenses would encourage developers “to build license compliance into RAG features”; the publishers say a market to license news for retrieval already exists.

Pre-training copies works “to provide linguistic patterns that undergird development of the model,” Wikimedia argued, a purpose different from the originals’. A chatbot that fetches a page and pastes “the most relevant passages or ‘chunks’” into the model alongside a user’s question, it argued, uses the work for its original purpose, to inform readers, and diverts them from the source.RetrievalRetrieval augmented generation, or RAG, searches the web or a database at the moment of a question and feeds what it finds to the model. Microsoft calls its version web grounding; OpenAI calls its version Browse, the publishers say.

The foundation allowed that services using RAG, “including ChatGPT,” make “attempts to consistently cite and link to original sources used,” and said such uses are less likely to infringe when they honor Wikipedia’s license. “We are taking a position in this case to defend the free and open human knowledge that Wikipedia depends on,” Bernadette Meehan, its chief executive, said in a statement on Friday.

Microsoft argues that web grounding, its pairing of Copilot with Bing search, “yields highly transformative new expression,” and that copyright protects expression, not facts or search traffic: “News Plaintiffs do not own the news, and certainly they have no exclusive right to answer factual questions.” It calls it undisputed that “over 99% of Copilot outputs do not contain even a single sentence of overlap with grounding content.” “Any publisher that wishes to opt out of web grounding may do so with a few keystrokes on its website,” it wrote.

The publishers answered in their Oct. 23 brief that “copyright is not an ‘opt out’ regime,” and that, by Microsoft’s own data, click-through rates for the domains of The Times and the Daily News publishers drop 83 to 93 percent compared with search without generative AI. They argue that LLM-generated text “substitutes commercially for news,” that the defense “fares no better for RAG than it does for training,” that copying inside a system for “substitutive commercial purposes” is not immune, and that “an output need not copy text verbatim to infringe.” They also quoted a Times executive: “less people are subscribing to us and staying subscribed to us than they would be if we did not have generative AI tools like ChatGPT out there.”

On the publishers’ side, the Association of American Publishers and three other groups wrote that “using copyrighted works to train generative AI models does not warrant special treatment.” Digital Content Next, a trade group for digital content companies, called AI tools that summarize and present publishers’ content “an existential threat to the publishing industry.” SAG-AFTRA, the musicians’ federation and the two Writers Guilds also backed the plaintiffs, as did the Copyright Alliance.

A coalition led by the Center for American Rights, describing its members as “America First leaders,” wrote: “We are okay with seeing mainstream news media organizations shrink, but because they deserve declining readership for their fake news, not because their IP has been stolen.”

For OpenAI and Microsoft came Engine Advocacy, which works with start-ups, and Authors Alliance, which addresses only whether competition from new works that do not infringe counts as copyright harm, arguing that it does not, and “takes no position on the other questions.” Public Knowledge backed the defendants only in part; its brief, its lawyers wrote, “argues that AI training should be considered separately from other uses, such as the acquisition of datasets, and an AI’s outputs.” Friends-of-the-court briefs are due Friday.

Daily Cortex is written with AI models, including Anthropic’s Claude, and fact-checked with OpenAI’s GPT models.

Sources

  1. Wikimedia Foundation (S.D.N.Y. Doc. 2361-1), “[Proposed] Amicus Curiae Brief of Wikimedia Foundation in Support of Neither Party, In re OpenAI, Inc. Copyright Infringement Litigation, No. 25-md-3143,” Oct. 29, 2026.
  2. Wikimedia Foundation (S.D.N.Y. Doc. 2361), “Letter motion for leave to file amicus curiae brief,” Oct. 29, 2026.
  3. Wikimedia Foundation, “The Wikimedia Foundation files to protect open knowledge in New York Times v OpenAI case,” Oct. 30, 2026.
  4. Microsoft Corporation (S.D.N.Y. Doc. 2310; No. 23-cv-11195, Doc. 1761), “Defendant Microsoft Corporation’s Opposition to News Plaintiffs’ Motion for Partial Summary Judgment (re-filed),” Oct. 29, 2026.
  5. The New York Times Company et al. (S.D.N.Y. Doc. 2118), “News Plaintiffs’ Omnibus Memorandum of Law in Opposition to Defendants’ Motions for Summary Judgment (redacted),” Oct. 23, 2026.
  6. The New York Times Company et al. (S.D.N.Y. Doc. 1709), “News Plaintiffs’ Combined Summary Judgment Brief (redacted version),” Sept. 4, 2026.
  7. Association of American Publishers et al. (S.D.N.Y. Doc. 2304-1), “Brief of Amici Curiae Association of American Publishers, Copyright Clearance Center, Independent Book Publishers Association, and International Association of Scientific, Technical & Medical Publishers in Support of Class Plaintiffs’ and News Plaintiffs’ Motions for Summary Judgment,” Oct. 29, 2026.
  8. Digital Content Next (S.D.N.Y. Doc. 2325-1), “Proposed Brief of Amicus Curiae Digital Content Next in Support of News Plaintiffs’ Motion for Summary Judgment,” Oct. 29, 2026.
  9. SAG-AFTRA et al. (S.D.N.Y. Doc. 2332-1), “[Proposed] Brief of SAG-AFTRA, American Federation of Musicians, Writers Guild of America, West, and Writers Guild of America East as Amici Curiae in Support of Plaintiffs,” Oct. 29, 2026.
  10. Copyright Alliance (S.D.N.Y. Doc. 2359-1), “Brief of Amicus Curiae Copyright Alliance in Support of Plaintiffs’ Opposition to Defendants’ Motions for Summary Judgment,” Oct. 29, 2026.
  11. Center for American Rights et al. (S.D.N.Y. Doc. 2306-1), “Amicus Brief of the Center for American Rights, Article III Project, Internet Accountability Project, the Bull Moose Project, Consumers for a Strong Economy and Joel Thayer,” Oct. 29, 2026.
  12. Authors Alliance (S.D.N.Y. Doc. 2328-1), “[Proposed] Brief of Amicus Curiae Authors Alliance in Support of Defendants’ Motions for Summary Judgment,” Oct. 29, 2026.
  13. Engine Advocacy et al. (S.D.N.Y. Doc. 2335-1), “Brief of Engine Advocacy, et al. as Amici Curiae in Support of Defendants,” Oct. 29, 2026.
  14. Public Knowledge (S.D.N.Y. Doc. 2313), “Letter motion for leave to file amicus curiae brief in support of defendants’ motions for summary judgment in part,” Oct. 29, 2026.
  15. U.S. District Court, S.D.N.Y. (Judge Sidney H. Stein, Doc. 2049), “Letter on the summary-judgment schedule, memo endorsed: “The schedule as set forth above is granted.,” Sept. 24, 2026.
  16. U.S. District Court, S.D.N.Y. (CourtListener), “Docket, In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143,” Oct. 30, 2026.