Wikimedia asks courts to protect AI training but require licences for regurgitation

The Wikimedia Foundation, the nonprofit that runs Wikipedia, has filed an amicus brief in the group of copyright lawsuits that publishers including The New York Times brought against OpenAI and Microsoft. It supports neither party. It asks the court to draw a line. Training large language models on copyrighted work should count as fair use, the foundation argues, but chatbots that reproduce that work for users in place of the original should need a licence.

The foundation filed the brief on Friday, 30 October 2026, according to The Verge. The filing matters because Wikipedia is on both sides of the dispute. Its articles are some of the most heavily used AI training data on the web, and its editors depend more and more on AI tools. A ruling that goes fully one way, Wikimedia says, would hurt it either way.

Neither side's position

The brief says Wikimedia "is concerned that the parties' positions in this case push for overbroad and simplified applications of the fair use doctrine." A clear win for publishers could let copyright holders "gatekeep" even innovative and transformative uses of their work. A clear win for AI companies could act as "a green light to snub licensing terms and fragment the flow of knowledge between human authors and audiences."

The parties are far apart. The Times alleges that OpenAI infringed its copyright by scraping its website to build training datasets. OpenAI and Microsoft say their use of LLMs is fair use. OpenAI has called regurgitation of copyrighted work "a rare bug." In a September filing, Microsoft said expert analysis found that Copilot rarely reproduces full sentences from the news content used to ground the model. Microsoft argued that this "hardly undermines the transformative purpose of LLM training."

The proposed line

Wikimedia's compromise has three parts, as described in the brief and an accompanying blog post. First, scraping websites should not count as copyright infringement. Second, pre-training LLMs on copyrighted works without a licence should be fair use when the models provide "something new and different from the original sources." Third, when a model retrieves copies of material in response to a user's query and delivers it "while supplanting readers' engagement with those original works," that should require a licence.

The reasoning behind the split is substitution. Training turns source material into something else. Retrieving a passage and handing it back to the user replaces the original instead of transforming it, and so in Wikimedia's view it does not qualify as fair use. In those cases AI companies should follow the licence terms. For Wikipedia, whose text is released under a Creative Commons licence, that means crediting the source. The licence also requires that "any remix or adaptations" stay under the same licence.

Why Wikipedia has a stake

The brief says Wikimedia sites "have become one of the most critical sources of data for all artificial intelligence and machine learning training." According to the foundation, the systems built on that data also draw visitors away from Wikipedia, which weakens the volunteer communities that write and maintain it. Wikimedia argues that services like ChatGPT reproduce Creative Commons text without following the licence's terms.

The foundation also says LLMs are useful to Wikipedia. It lists helping editors fight vandalism, detecting copyright violations and doxxing, and finding articles that need improvement. A ruling that made all training legally risky could put those tools out of reach, especially open-source ones.

"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 impacting access to information for everyone in both cases," Wikimedia Foundation general counsel Stephen LaPorte said in a statement. 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."

What it means for work

Under Wikimedia's line, building a model would stay open to many developers, including small and open-source projects that cannot afford to license training data at scale. The licensing cost would shift to the point where a product answers a user by passing on someone else's text. Retrieval-based features such as chatbot answers grounded in news articles or encyclopedia entries would need attribution or a licence. That could send some credit and traffic back to the writers, editors and volunteers who produced the material. For Wikipedia editors, the arrangement would keep AI tools for vandalism patrol and copyright checks legal while giving the foundation a claim against outputs that replace its articles. What remains open: the court has not ruled, and neither party has signalled it will accept the distinction. Nor is it settled how a court would tell a transformed output from a substitute one in practice.