Wikimedia brief: LLM training can be fair use, but chatbot output should need a license

Wikimedia stakes out a middle position

The Wikimedia Foundation, the nonprofit that operates Wikipedia, has filed an amicus brief in the cluster of copyright lawsuits brought by publishers including The New York Times against OpenAI and Microsoft, according to The Verge. The organisation is backing neither side. Instead, it is asking courts to separate two activities that are often lumped together in the AI copyright debate: training large language models on copyrighted material, which it wants protected as fair use in many circumstances, and chatbots handing that material back to readers, which it argues should require a licence.

The brief, filed on a Friday according to The Verge, warns that the positions taken by both the publishers and the AI companies push for what Wikimedia calls overbroad and simplified applications of fair use doctrine. A clean win for publishers, the filing suggests, could let rights holders "gatekeep" genuinely transformative uses of their work. A clean win for the AI companies, by contrast, could amount to permission to ignore licensing terms and fragment the flow of knowledge between authors and audiences.

Why Wikipedia is caught in the middle

Wikipedia sits awkwardly on both sides of this fight. According to the filing reported by The Verge, Wikimedia's sites have become some of the most important sources of training data for AI and machine learning systems. Yet those same systems divert visitors away from Wikipedia, weakening the volunteer communities that build and maintain it.

Wikimedia also claims a direct copyright grievance: services such as ChatGPT reproduce text that Wikipedia licenses freely under Creative Commons, but do so without following the terms of that licence. At the same time, the models are genuinely useful to the encyclopedia. The filing notes that they help editors fight vandalism, detect copyright violations and doxxing, and identify articles that need improvement.

"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, adding that Wikipedia depends on a strong foundation of open and accessible human knowledge.

Where the parties stand

The Times alleges that OpenAI infringed its copyright by scraping its website to build training datasets. OpenAI and Microsoft both contend that training large language models constitutes fair use. OpenAI has previously characterised the regurgitation of training material as a rare bug rather than intended behaviour. Microsoft, in a September filing cited by The Verge, said expert analysis found that Copilot rarely reproduces full sentences from news content used to ground the model, which it argued "hardly undermines the transformative purpose of LLM training."

The test Wikimedia proposes

Wikimedia's suggested compromise has two halves. First, scraping websites should not itself be treated as copyright infringement, and pre-training language models on copyrighted works without a licence should count as fair use when the resulting systems provide "something new and different from the original sources."

Second, when a model retrieves copies of material in response to user queries and delivers them "while supplanting readers' engagement with those original works," a licence should be required. The distinction, as Wikimedia frames it, is transformation versus substitution: output that merely replaces the original work does not transform it enough to qualify for fair use.

For Wikipedia specifically, compliance would mean crediting the original source under its Creative Commons licence and keeping any remixes or adaptations under that same licence. "Expanding access to knowledge through AI is a powerful goal, but it must respect the people, contributors, and communities that make reliable knowledge possible," Wikimedia CEO Bernadette Meehan said in a statement.

Why it matters

These cases are widely expected to shape how fair use applies to AI, and Wikimedia's brief offers a template that splits the difference between the litigants: protect training, police output. If courts adopt something like it, AI developers get legal clarity for model building while publishers and open projects like Wikipedia keep a lever for attribution and licensing on the answer side. If instead the rulings come down bluntly in either direction, the consequences could be lopsided: either publishers gain veto power over transformative tools, or AI companies gain cover to bypass licences entirely, including the open licences that underpin the free knowledge ecosystem. Because Wikipedia is simultaneously a top training-data source, a victim of traffic loss and a user of the very models at issue, its argument that nuance is existential carries unusual weight.