Wikimedia Foundation Wants Courts to Split AI Training From AI Output

Key points

• The Wikimedia Foundation, the non-profit that runs Wikipedia, filed a legal brief urging courts to protect AI training on copyrighted text while holding AI chatbots liable when they reproduce that text to users.
• Wikimedia argues that a blanket ruling against AI training would strip Wikipedia's volunteer editors of useful AI tools built on open-knowledge sources.
• A ruling that protects all AI outputs could let chatbots replace Wikipedia without compensating or crediting its contributors.
• No ruling has landed yet in the US cases Wikimedia weighed in on.

Wikipedia's parent organisation has stepped into one of the biggest legal fights in artificial intelligence: who owns the words an AI learns from, and what happens when it repeats them.

The Wikimedia Foundation filed a legal brief arguing that courts should treat AI training and AI output as two completely separate legal questions. Training, the process where a large language model (the technology behind chatbots like ChatGPT and Claude) reads vast amounts of text to learn patterns and facts, should be permitted even when that text is copyrighted. Reproducing that text back to users, verbatim or nearly so, should not.

Wikipedia is one of the largest bodies of freely licensed text on the internet, and virtually every major AI system has learned from it. Wikimedia's position isn't that AI companies owe it money for that. Its concern runs in both directions.

If courts rule that training on copyrighted work is always illegal, tools that help Wikipedia's volunteer editors, such as AI assistants that flag errors or suggest citations, could become legally impossible to build. That would hurt the open-knowledge project directly.

But if courts rule that everything an AI does with copyrighted material is protected, chatbots could answer questions by drawing on Wikipedia's content without sending traffic or credit back to the community that created it. Volunteers lose readers. Without readers, they lose their reason to keep writing.

First reported by The Verge AI, the foundation described an overly sweeping ruling in either direction as a threat to the open web.

The timing is pointed. We covered Anthropic's submission to Australian lawmakers on 9 October 2026, when the company argued that current Australian copyright law makes AI training there impossible. The same fault line runs through every jurisdiction: training versus output, and who pays for what.

Wikimedia's brief tries to steer between those two extremes. It's an unusual position for a non-profit that gives its content away for free, and it signals how seriously the open-knowledge community is taking the risk that AI could hollow out the very sources it was trained on.

The practical stakes for ordinary readers are real. When you ask a chatbot a factual question and it answers correctly, where did that answer come from, and does anyone who originally wrote it see any benefit? Wikimedia's argument is that the law should ensure the pipeline doesn't run only one way.

What happens next?

The cases Wikimedia weighed in on are still working through US courts. No ruling has landed yet, and the outcome will set a precedent that every AI company and every open-knowledge project will have to live with.

The foundation is not a plaintiff or defendant. It filed as an interested third party, meaning it's trying to shape how judges think about the question rather than win a direct dispute. That kind of brief can influence a ruling, but there's no guarantee a court adopts its framing.

Watch for how judges respond to the training-versus-output split. If courts accept it, AI companies keep broad freedom to learn from the web while facing new limits on what they reproduce. Rejection makes the fight much messier for everyone.