The people who built the internet's biggest commons just told the courts how to save it from both the copyright maximalists and the AI absolutists.
The Summary
- The Wikimedia Foundation filed an amicus brief in the NYT v. OpenAI lawsuit, arguing courts should protect LLM training on copyrighted work while restricting verbatim regurgitation to users
- Wikipedia's position: training is transformative use (protected), but chatbots spitting out near-copies of source material is not
- If courts get this wrong in either direction, they could either kill AI research tools or destroy the open web's fair use foundation
The Signal
Wikimedia just did something rare in 2026: they picked a nuanced position on AI and copyright when everyone else is screaming from opposite corners. Their amicus brief argues that training LLMs on copyrighted content should be legal, but serving up chunks of that content to users should not be. This matters because Wikipedia runs on the same fair use doctrine that's now under siege.
The foundation's concern is existential. If courts rule that any use of copyrighted material in AI training is infringement, Wikipedia loses access to the same AI tools its volunteer editors need. We're talking summarization bots, citation checkers, vandalism detection, translation assistants. All of those rely on models trained on copyrighted text. Kill that training and you don't just hurt OpenAI, you gut the tooling that keeps 6.8 million Wikipedia articles maintained by volunteers.
"An overly black-and-white ruling could either deprive Wikipedians of useful AI-based tools or endanger the open web."
But Wikimedia also sees the other cliff edge. If courts give AI companies a blank check to train on anything AND serve back near-verbatim copies, you've just legalized the replacement of the open web with a walled garden of chatbots. Why link to Wikipedia when ChatGPT can summarize the article without attribution? Why maintain a commons when the for-profit summarizer becomes the default interface?
Key tension in the brief:
- Training on copyrighted work = transformative, should be protected fair use
- Regurgitating copyrighted work = derivative output, not protected
- Courts need to split this difference or break something important
This is the Web4 dilemma in legal form. The agent economy runs on models trained on human knowledge. But if those agents just extract and repackage that knowledge without attribution or compensation, you've built an intelligence layer that cannibalizes its own training data. Wikipedia gets this because they've been dealing with the same problem for two decades: how do you build collective intelligence infrastructure that doesn't get strip-mined by whoever has the biggest servers?
The Implication
Watch how this brief lands with the court. If Wikimedia's position becomes precedent, you get a workable framework: train on everything, serve back nothing verbatim. That's a world where AI research continues but chatbots can't just regurgitate the New York Times without consequences. It also means agent builders need to think harder about attribution and how their outputs relate to training sources.
For anyone building in the agent space, this is your heads-up: the law is about to draw lines between acceptable and unacceptable uses of copyrighted training data. Design your agents to transform, not transcribe. The courts are coming.