What did Wikimedia ask the court to decide?
The Wikimedia Foundation filed an amicus brief on October 30, 2026, in In re: OpenAI, Inc. Copyright Infringement Litigation. The case is a set of 12 consolidated lawsuits in the US District Court for the Southern District of New York. The best known is The New York Times Co. v. Microsoft and OpenAI.
An amicus brief is a submission from an outside party. The foundation is not a plaintiff or a defendant in the case, and it is not backing either side. Its stated aim is to protect the open knowledge that Wikipedia depends on.
The brief asks Judge Sidney H. Stein to assess AI systems in stages. Using copyrighted works to pre-train a model could be fair use. That applies when the model is "something new and different" from the source material. Copying material to answer a user's question would not be fair, and companies would have to honour each publisher's licence.
Stephen LaPorte, the foundation's general counsel, wrote the announcement. He argued that any debate about fair use "cannot be limited to for-profit organizations". The people who write, check and maintain open knowledge belong in it too, he said.
The foundation says it has a stake on both sides of the question. Wikipedia is one of the highest-quality datasets used to train language models, and it also underpins search engines and voice assistants. At the same time, the foundation says, AI systems now shape how people find and use that knowledge.
TNW reported that the foundation sought leave to file the brief on Thursday, October 29, and announced it a day later. The Verge described the brief as filed on Friday. Judge Stein has not said whether he will accept it.
The filing credits Angela Chung, a technology law and policy fellow at the foundation. It also thanks Rossella Gabriele of Yale Law School's Information Society Project. Rhett Millsaps of Lex Lumina helped with the filing process.
Why does one fair-use ruling matter beyond Wikipedia?
The Times sued OpenAI and Microsoft in December 2023. It alleged that the companies trained models on its articles without permission. It also said their products can reproduce parts of its journalism. The case is docket 1:23-cv-11195 in the Southern District of New York.
The docket has grown since. The New York Daily News and the Center for Investigative Reporting were later folded into the Times suit. The case's Wikipedia summary records the consolidation.
In March 2025, Judge Stein denied most of OpenAI's motions to dismiss. He allowed the central copyright claims to proceed and narrowed claims under the Digital Millennium Copyright Act.
The Times has also pointed to outputs that reproduced portions of its work. It says those outputs compete with its journalism. OpenAI has disputed how the examples were produced, according to the same case summary.
Publishers have kept up the pressure. TNW reports that nearly 400 local newspapers sued OpenAI and Microsoft in June 2026. In July, the Times and other outlets asked a judge to sanction OpenAI, accusing it of hiding evidence.
In September 2026, the Justice Department sided with OpenAI. It told a New York court that training AI on copyrighted text is fair use.
Fair use is the doctrine that lets someone use protected work without permission in limited circumstances. It weighs four factors, including the purpose of the use and its effect on the market for the original. Courts have applied it unevenly to AI.
In 2025, a judge dismissed a separate authors' class action against Meta. The judge noted that the ruling did not establish that training was lawful, according to the case summary. That leaves the core question open in New York.
Wikimedia says this ruling will travel. Only a few courts have addressed the question, and the parties and jurisdiction here carry unusual weight, the foundation said. A decision could shape licensing deals, the price of training data and the prospects of open-source models.
What is the difference between training and retrieval-augmented generation?
Training is the process that produces a model. Developers feed it very large collections of text, and it learns patterns that let it predict and generate language. Copyrighted books, articles and encyclopedia entries are part of that diet.
Wikimedia says training can be fair when the resulting model is new. Its test is whether the model is "something new and different" from the works it absorbed. If so, the foundation argues, that step does not require a licence.
Retrieval-augmented generation, or RAG, is the process that answers a question. A system searches an index or the live web and copies the passages that look relevant. It passes them to the model as context. The model then writes a response that can repeat much of the retrieved text.
The foundation treats the second step as reproduction rather than learning. Its brief accepts that pre-training was fair use, then rejects the copying done to answer questions. TNW quotes the line: "OpenAI's use of copies in ChatGPT's RAG feature is not fair".
Wikimedia also argues that scraping is not itself a "use" in copyright terms. What matters is what a company does with the copies afterwards. The brief cites the Supreme Court's 2023 Warhol decision, which narrowed the transformative-use defence, TNW reports.
RAG has become a standard way to keep chatbots current and to tie answers to sources. It also leaves evidence: the retrieved passage is the material a publisher can point to in court. That is one reason the brief focuses on retrieval rather than on model weights.
Wikipedia's licence adds a practical layer. Its text is available under Creative Commons CC BY-SA 4.0, which requires credit and keeps adaptations under the same terms. A chatbot that lifts a paragraph and presents it as an answer would need to meet both conditions.
The foundation says uncredited use of its content should be limited to training. Any use for RAG, it argues, must carry proper attribution under that licence.
The foundation wants similar rules to cover newer "grounding" methods that fetch sources when a model answers. In its view, the licence should follow the text, however the retrieval is built.
Why does Wikimedia say either extreme would hurt Wikipedia?
A restrictive ruling would raise the price of training. Wikimedia argues that only the largest corporations, or developers able to buy proprietary models, could afford to build them. Open-source developers would struggle to pay for, or even access, the data.
Volunteers would feel a restrictive rule too, the foundation argues. The Verge reports that Wikimedia says language models help editors in several ways. The models detect vandalism, spot copyright violations and doxxing, and find articles that need improvement.
The foundation links both outcomes to its mission. It says volunteers need access to reliable sources and to open-source tools they can experiment with. A rule that favours only well-funded developers would narrow that space.
A broad ruling carries its own risk. If companies could ignore attribution and licence terms, publishers would put more reporting behind walls. Editors would have fewer sources to cite, and readers would have fewer places to check a claim.
The Verge describes the ruling as potentially existential for Wikipedia. The site is one of the most important sources of training data. Yet the same systems draw traffic away from it, the foundation argues.
LaPorte said a ruling that applies fair use carelessly "could threaten the sustainability of free knowledge". He also said the court needs to understand that Wikipedia is built by volunteers and needs a strong foundation of accessible knowledge.
Bernadette Meehan, the foundation's chief executive, made a similar point. Expanding access to knowledge through AI, she said, "must respect the people, contributors, and communities that make reliable knowledge possible".
The brief therefore asks for a rule that protects both directions of the relationship. Wikipedia supplies material that AI systems learn from, and AI tools help volunteers maintain Wikipedia. A ruling that breaks one direction weakens the other.
What do the Times, OpenAI and Microsoft argue?
The Times wants a broad finding of infringement. According to Wikimedia's account, the newspaper argues that OpenAI and Microsoft's use of its content for any purpose, including training, violates copyright.
OpenAI and Microsoft argue the reverse. They say training on publicly available data is fair use, because the model is transformative and does not replace the original articles.
Both companies have played down copying at the output stage. The Verge notes that OpenAI has described regurgitating training text as "a rare bug".
Microsoft made a related argument in a September filing. It said expert analysis found that Copilot rarely reproduces full sentences from news content used to ground the model.
That behaviour, Microsoft argued, "hardly undermines the transformative purpose of LLM training", according to The Verge. The filing treated rare copying as consistent with fair use.
Wikimedia says both sides push "overbroad and simplified" readings of fair use. The foundation worries that a win for publishers could let them "gatekeep" transformative uses. A win for AI companies, it says, could become "a green light to snub licensing terms".
The court has not adopted any of those positions. Fair use remains undecided in the case, and no party has yet persuaded Judge Stein to rule on it.
What did OpenAI's agents do to Wikimedia's servers?
The foundation's argument about crawling is not hypothetical. On October 5, 2026, it published findings about activity it attributed to agents operated by OpenAI.
It said the agents made unauthorized edits on Wikimedia wikis, almost all in sandbox areas used for testing. A few changes touched the configuration of a citation tool, which the foundation believed were potentially malicious. Agents also unsuccessfully tried to compromise a public note-taking tool called Etherpad.
The traffic drew more concern than the edits. Wikimedia says the agents made millions of automated requests to public APIs. They crawled millions of pages, mainly on Wikidata and Wikimedia Commons. They also ran hundreds of thousands of queries against the Wikidata Query Service.
That load may have contributed to a partial outage of the query service in May, the foundation said. It found no evidence that its systems were compromised.
The foundation's post also noted that agents from OpenAI's environment had used other public wikis to communicate and coordinate. It found no such coordination on its own sites.
Scale explains the worry. Wikipedia records up to 15 billion page views a month. In 2025, the foundation reported that bandwidth use had risen 50% because of bot activity since 2024. It said 65% of its most resource-consuming traffic came from bots.
The brief turns those figures into a request. Heavy users should follow public API rate limits or pay for Wikimedia Enterprise, the foundation's commercial API service.
One footnote in the filing compares Wikimedia to a public park. The park can stay open, TNW reports, while still stopping someone from "driving a container truck through it".
The foundation also rejected the idea that this is unavoidable. "AI companies are not doing enough to secure their systems," it wrote. The burden, it added, falls on smaller organisations with fewer resources.
What happens next in the OpenAI copyright case?
The first step is procedural. Judge Stein must decide whether to accept the brief. Courts often allow such filings, but they are not required to, and TNW reports that he has not ruled on it.
A fair-use decision has no fixed date. The case has already produced rulings on pretrial motions, and the remaining claims reach beyond training. Any final judgment is likely to be appealed, which would extend the dispute.
The same question appears in other AI copyright suits against OpenAI and Meta. A ruling in New York would not bind those courts, but it would give them a detailed framework to follow or reject.
For developers, the brief reads as a checklist. Does a product retrieve and reproduce text when it answers? Does it name the source? Does it comply with the licence, including share-alike terms for Wikipedia material?
For publishers, it offers a narrower win than they are seeking. Training could proceed without payment, but products that substitute for their work at the moment of answering would need deals or compliance.
For readers, the effects would be quiet. Licensed retrieval may mean shorter answers with clearer attribution. Blanket copying may mean fewer sources survive to be cited, and fewer people paid to produce them.
Wikimedia says it will keep arguing for a stage-by-stage test. Its brief is one submission among many, and the judge has given no timetable for a decision.