Wikimedia asks the court to split AI fair use: training yes, regurgitation no

The filing, and why this case matters

The Wikimedia Foundation, the nonprofit that hosts Wikipedia, filed an amicus brief on October 30 with the Southern District Court of New York in In re: OpenAI, Inc. Copyright Infringement Litigation, a set of 12 combined lawsuits whose best-known member is The New York Times Co. v Microsoft and OpenAI. The court will rule on whether AI companies' use of copyrighted content to develop large language models is fair use.

Wikimedia is not a party to the case, and it framed the brief as neither a defence of OpenAI nor of the publishers. It argues that the ruling will be unusually influential because the jurisdiction and the parties are significant, and because only a few courts have addressed the question so far. The brief asks the court to protect what Wikimedia calls the open knowledge ecosystem when it decides.

A middle ground with two different rules

The substance of the brief is a split between two uses of copyrighted text. For model training, Wikimedia argues that using copyrighted content should be considered fair when the model is something new and different from the original sources, in which case no licensing terms would be required. That covers pre-training, the stage where a model learns from a large corpus rather than reproducing it.

For retrieval-augmented generation and similar grounding techniques, Wikimedia takes the opposite view. Those systems find and copy news or other works and deliver the same content back to users through a chat interface. Wikimedia argues that this should not be fair use, that companies offering it should respect each platform's or publisher's licensing requirements, and that for Wikipedia the relevant licence is Creative Commons CC BY-SA 4.0, which requires credit and that adaptations stay under the same licence.

Why Wikipedia is taking a position at all

Wikipedia describes itself as one of the highest-quality datasets used to train language models and the backbone of information for chatbots, search engines and voice assistants: more than 67 million articles across 300 languages, written and improved by nearly 250,000 volunteers a month. Those same systems, Wikimedia says, siphon away visitors and undermine the communities that maintain the site.

The foundation also says the traffic is not the only thing at stake. It argues that language models are useful to Wikipedia's editors, who can use them to fight vandalism, detect copyright violations and doxxing, and look for articles that need improvement. That is the argument for treating training as fair use: the tools built on the corpus can help maintain the corpus. Chief executive Bernadette Meehan said expanding access to knowledge through AI is a powerful goal, but it must respect the people, contributors and communities that make reliable knowledge possible.

Its general counsel, Stephen LaPorte, framed the risk from both directions: a ruling that does not apply fair use principles carefully could threaten the sustainability of free knowledge and stifle the development of open source technologies, ultimately affecting access to information either way. He said the court needs to understand that Wikipedia is created through a collaborative process that requires a strong foundation of open and accessible human knowledge to thrive.

The parties' positions, and the two extremes

The New York Times sued in 2023, alleging that OpenAI and Microsoft infringed its copyright by using its content, including for model training. OpenAI and Microsoft contend that training and deploying models on publicly available data is fair use. OpenAI has described regurgitating work as a rare bug, and Microsoft argued in a September filing that expert analysis found Copilot rarely reproduces full sentences from news content used to ground the model.

Wikimedia's concern is that both sides push for overbroad and simplified applications of the doctrine. A nuance-free victory for publishers, it says, could let copyright holders gatekeep even innovative and transformative uses of their work; a victory for the AI companies could be a green light to snub licensing terms and fragment the flow of knowledge between authors and audiences. The brief also asks that scraping be done sustainably, in line with Wikimedia's public API rate limits or through its Enterprise API.

What a ruling could change, and what to watch

The practical effect depends on how the court writes its opinion. A narrow ruling on the Times' claims would settle one case; a broad ruling on fair use for training would set the terms for every model developer and every publisher negotiating a licence. Wikimedia's proposal is an attempt to make the second outcome less binary, and it will be cited by both sides whatever the court decides.

Three things are worth tracking: whether the court adopts a use-specific test rather than a blanket one, how it treats retrieval and grounding as distinct from training, and whether licensing requirements are treated as enforceable when a chatbot answer substitutes for the original. For background, this site has reported on the Department of Justice inquiry into Nvidia and Hugging Face, on Proton's argument about AI and the web, and on the privacy case around Meta's Muse marketplace. The brief does not predict the outcome; it argues for a rule that would keep both training and attribution possible.