Fair use is a four-factor test, and the biggest AI copyright case in the world is being argued as if it were a coin flip. On Friday the Wikimedia Foundation filed an amicus brief in The New York Times' suit against OpenAI and Microsoft, arguing that both sides have built their case on a simplification.
The brief declines to pick a winner. Publishers, it says, are pushing for an application of fair use so narrow that rights holders could "gatekeep" even transformative and innovative uses of their work. The AI companies, meanwhile, want a reading so wide that it amounts to "a green light to snub licensing terms and fragment the flow of knowledge between human authors and audiences."
Wikimedia's own interest is anything but neutral. Its sites, it says, "have become one of the most critical sources of data for all artificial intelligence and machine learning training" — and the same systems siphon away the readers who sustain the volunteers writing them. That is the awkward position of a commons: enormously valuable as training data, structurally unable to charge for it, and dependent on traffic and goodwill that the models erode.
The second, less obvious argument is that the models are useful to Wikipedia too. Editors use LLMs to fight vandalism, detect copyright violations and doxxing, and find articles that need work. A ruling that makes model training illegal by default therefore damages the encyclopedia's own tooling. That is why the foundation wants the line drawn not at training but at output: learning from a corpus is one thing, reproducing it to a user is another.
Then there is a sharper allegation: services like ChatGPT, Wikimedia says, reproduce text that is licensed freely under Creative Commons without adhering to the terms of that licence. That is a licensing argument rather than a fair-use one, and it is the kind that survives even a generous fair-use ruling.
"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," said Stephen LaPorte, the foundation's general counsel.
Why it matters beyond Wikipedia: the case will set the default for every open model and every small developer. The brief warns that denying fair use for training would leave open-source developers unable to afford or access AI technology — a cost the brief argues would fall on the ecosystem that produces the next generation of models, not just on the defendants.
What is interesting about the filing is the position it takes. The middle here is not a compromise between two equally valid extremes; it is the only line that keeps a commons and a market alive at once. Whether the court buys that is another matter — courts prefer clean tests to calibrated ones. The brief is a bet that this one will accept a two-part rule: ingest freely, output carefully.
