I watched an edit roll back on a Wikipedia page the instant a bot-cited paragraph appeared. You could feel the volunteer base hold its breath—this was no longer an abstract legal fight. The Wikimedia Foundation had just filed a brief that could decide whether the web stays open or tilts toward paywalled AI monopolies.
I’m going to walk you through what’s at stake, what the Foundation is actually asking the courts to do, and why you should care even if you never edit a wiki.
At Wikipedia’s 25th anniversary the Foundation signed commercial deals with major AI companies
That move wasn’t a PR stunt. It was a signal. Amazon, Meta, Mistral, Perplexity, and Microsoft agreed to pay for licensed access to Wikipedia content as LLMs got thirsty for high-quality data. I see this as Wikimedia trying to convert cultural capital into financial oxygen.
Wikipedia stands as a lighthouse in a storm—a public dataset that companies want to mine, and a volunteer-run ecosystem that depends on people feeling their contributions matter. The Foundation’s new amicus brief to the New York Times v. OpenAI/Microsoft case frames a bargain: let companies mine the public web for training, but stop them from reproducing expressive content in a way that kills incentives for creators.
Can AI legally train on copyrighted content?
The brief argues training should be treated differently from reproduction. You can feed an LLM publicly available pages for pretraining—the phase where models learn patterns. But when a chatbot produces a near-verbatim answer or stitches together a news article without attribution, that’s the moment the law should weigh in.
OpenAI and Microsoft have leaned on fair use defenses, while the New York Times says its copyrighted archives were used to build commercial competitors. Wikimedia rejects both extremes: it labels the AI companies’ defense too broad, and the publishers’ demand for absolute control equally impractical.
The New York Times sued in December 2023 claiming millions of articles were used without permission
The suit foregrounded a legal test for an era of opaque models. You can’t open an LLM and point to a single neuron and say “there’s the stolen sentence.” That opacity worsens the chance courts will reach a sweeping ruling that either hands AI firms carte blanche or hamstrings model research.
LLMs are being treated as a sealed vault whose hinges only engineers can hear, and the law is being asked to pick a key. Wikimedia urges the court to split the problem: allow unrestricted access for learning structure and general knowledge, but police the model’s outputs—especially when mechanisms such as retrieval-augmented generation (RAG) let models reproduce expressive sources in responses.
What does Wikimedia want from the courts?
The Foundation’s roughly 5,800-word amicus asks for a middle path. Its summary: fair use shouldn’t become a free pass to ignore licensing, nor should it be read so narrowly that it blocks legitimate LLM capabilities like moderation or editing assistance.
Wikimedia’s lawyers wrote that “fair use should not be interpreted in a manner that would gatekeep all uses of copied works under proprietary terms to the detriment of efficient innovation,” while warning courts not to let fair use become a green light for blanket copying that fragments incentives for human contributors.
The Foundation already uses LLMs for site maintenance and signed partnerships that recognize value exchange
LLMs help spot doxxing, suggest edits, and surface vandalism—concrete benefits the brief highlights. At the same time, Wikimedia has been monetizing its dataset: partnerships with Microsoft and others give paid access to the site’s content for model training.
This dual posture—both pragmatic and protective—maps onto broader industry moves. Publishers split between suing (the New York Times) and partnering (The Atlantic’s deal with OpenAI, which includes linking back to source articles and privileged product access). Wikimedia is trying to broker a market where public knowledge remains free, and companies pay to use it in ways that don’t hollow out the supply of human-created content.
How will this affect Wikipedia?
A blanket ruling in favor of tech giants could make volunteer contributions collapse. If models can scrape and regurgitate articles without consequence, why would editors spend hours polishing entries? Wikimedia warns that an unwary ruling could threaten “the sustainability of free knowledge.”
Stephen LaPorte, Wikimedia’s general counsel, told the court that Wikipedia depends on a “strong foundation of open and accessible human knowledge” to thrive. You can see why the nonprofit is pushing to codify protections for outputs—even as it admits that some model capabilities should not be blocked by an all-or-nothing fair-use reading.
On the legal timeline: courts, opacity, and a slow-moving fight
We’re early in a decades-long puzzle. The social-media era tangled courts for years; AI presents new knots. Lawsuits multiply, technical transparency lags, and judges must decide whether to protect a commons or to bless a new toll road for data.
I think the Foundation’s brief is as much an attempt to preserve bargaining power as it is legal reasoning. You and I are beneficiaries of the status quo: a sprawling, mostly free web where anyone can add knowledge. If that system starts to erode, the public loses more than free access; it loses the incentive structure that made Wikipedia possible.
I’ll watch the filings and the tech moves closely, and you should too—because this case will shape whether the internet remains a commons or becomes a data toll road for a few platforms, and that decision will touch everything from search to Wikipedia edits to your daily newsfeed?