Wikimedia Foundation asks court to separate AI training from text output

On October 30, 2026, the Wikimedia Foundation (the nonprofit organization that operates Wikipedia) filed an amicus brief (a written opinion from a party not involved in the case) in the publishers' lawsuits against OpenAI and Microsoft. This is reported by The Verge. The lawsuits combine claims by The New York Times and other publishers against companies that trained language models on their materials.

The Foundation did not take either side. The document states that the parties' positions lead to an "overly broad and simplified" application of the fair use doctrine (an exception to copyright for limited purposes). This differs from the typical logic of the dispute, where one side defends training entirely and the other demands it be prohibited. Wikimedia proposes separating two actions: training large language models (LLMs) on protected works and outputting those works to the chatbot user. The Foundation asks that the first be recognized as permissible and the second not.

In practice, the court's decision will determine whether Wikipedians retain access to AI-based tools. If the court permits training but prohibits verbatim output, developers will have to build filters and change the output architecture. If the court takes an extreme position in either direction, the Foundation warns of a risk to the open web: either the tools will become unavailable, or the protection of texts will weaken. The note does not provide specific mechanisms that Wikimedia proposes to the court.

What exactly the Foundation considers verbatim output and where the boundary between paraphrase and reproduction lies is not specified in the material. There is no data on how the court will take this document into account.