The Wikimedia Foundation has asked a federal court in New York to split the biggest copyright question in artificial intelligence in two. In a filing made on 30 October 2026, the nonprofit behind Wikipedia argued that training a large language model on copyrighted work can be fair use, while a chatbot that retrieves and reproduces that work in its answers should respect the source’s licence. The position backs neither side outright, and it lands in the case most likely to shape how AI companies treat other people’s work.
What Wikimedia filed
The filing is an amicus brief, a friend-of-the-court submission from an organisation that is not a party to the case. It was filed with the Southern District Court of New York in a consolidated proceeding, In re: OpenAI, Inc. Copyright Infringement Litigation, which brings together 12 lawsuits against OpenAI.
The best known of those suits is The New York Times Co. v Microsoft and OpenAI, filed in 2023. The central question in that case is whether using copyrighted content to develop large language models counts as fair use, the doctrine that allows some uses of copyrighted material without the owner’s permission.
The Times argues that OpenAI and Microsoft used its content without permission, including for training, in violation of copyright law. OpenAI and Microsoft argue that training and operating models on publicly available data is technological development that falls within fair use. Wikimedia’s chief executive, Bernadette Meehan, framed the filing as a defence of the knowledge commons itself.
She said expanding access to knowledge through AI is a powerful goal, but it must respect the people, contributors and communities that make reliable knowledge possible. Finding that balance, she said, is a shared responsibility for the entire information ecosystem.
Training is not the same as answering
The brief’s core argument is that the court should judge each use of copyrighted material on its own terms, rather than treating every AI use as the same thing. Training, in Wikimedia’s account, is a process in which a model learns patterns from large collections of material.
The resulting model is something new and different from the sources it learned from, and that transformation is what can make the use fair. For training, the Foundation argues that using copyrighted works should be considered fair when the model that results is new and different from the original sources.
In that case, it says, there would be no requirement to meet licensing terms. That part of the argument leans toward the AI companies, and it is deliberate: Wikimedia depends on open-source tools and on volunteers being able to build with new technology. Retrieval is different, and the brief treats it as different in kind.
Retrieval-augmented generation, known as RAG, is a technique in which a chatbot searches the web, finds current material and uses it to ground an answer. When a chatbot fetches a news article or an encyclopaedia entry and serves that material back to a user, the answer can stand in for the original work.
For RAG and newer techniques that serve a similar grounding purpose, Wikimedia argues that the use should not be considered fair, and companies should respect each content platform or publisher’s licensing requirements. For Wikipedia itself, that licence is Creative Commons CC BY-SA 4.0, which requires credit to the original creators and requires adaptations to remain under the same licence.
The distinction matters because the two uses have different effects on the people who create the material. A trained model does not hand the reader the original article. The Foundation’s point is that substitution, not learning, is where the creator loses the reader and the publisher loses the visit.
A grounded answer sometimes does, or comes close enough that the reader never visits the source. Wikimedia’s general counsel, Stephen LaPorte, warned that a ruling which ignores that difference could threaten the sustainability of free knowledge and stifle open-source development.
Why Wikimedia says both extremes fail
Wikimedia describes both sides’ readings of fair use as overbroad and simplified, and it spells out what it fears from each. A ruling that interprets fair use too restrictively could make it harder to access publicly available content that is critical to developing open-source AI. It could also inhibit the many ways people gather and share public knowledge online.
The practical consequence, in the Foundation’s argument, is concentration. If training is not fair use, AI development would likely become possible only for the largest corporations that can afford to pay for content, or the most well-funded developers that can buy proprietary models.
Many open-source developers would be unable to afford or even access the technology, and volunteers would lose tools that help them build Wikipedia. A blanket ruling the other way carries the opposite risk. If AI companies gain a general right to avoid existing copyright rules and licence requirements, publishers would have a strong incentive to lock more information behind barriers.
That would leave fewer sources available to the volunteers who build Wikipedia, and less visibility for publishers that depend on readership to survive. The Foundation’s own stake in the case is unusually direct. Wikipedia holds more than 67 million articles across 300 languages, built and improved by nearly 250,000 volunteer editors every month.
That work relies on trustworthy sources, including journalism and published research, and Wikipedia is itself one of the largest and highest-quality datasets used to train language models. That double role is what gives the brief its force, and also its tension.
Wikipedia is both a source that AI systems learn from and a project that depends on the open web staying open. A ruling that disrupts either side of that relationship, the Foundation argues, would change how everyone finds and uses information, not just how one company builds one product.
What this changes
If a court adopts Wikimedia’s split, the pressure would fall hardest not on model training but on the answers users actually see. Search-style AI products, news summaries and assistants that fetch live pages would need to treat retrieved material as licensed content, with attribution and licence terms attached.
That is a heavier obligation than most grounded answers carry today, where a link or a citation is often treated as enough. The argument also reframes the debate about who pays for the raw material of AI. Training has usually been discussed as a single question: did the company have the right to learn from this text?
Wikimedia’s answer is that the question should be asked twice, once for the learning and once for the reuse. A company could win the first argument and still lose the second, which would push the industry toward licensing deals for retrieval even if training itself stays free. That split would suit some companies better than others.
Firms whose products lean heavily on retrieval, such as AI search and answer engines, would face the most immediate redesign, while firms that mainly sell a trained model would feel less direct pressure. It would also give publishers a clearer negotiating target: not the training run, which is hard to price and harder to police,
but the answer that competes with their own pages. There is a limit to what this filing can do on its own. An amicus brief does not bind the court, and the judge has not yet decided the fair-use question in this case.
The court could adopt Wikimedia’s distinction, reject it in favour of either side, or rule narrowly on the facts of the Times case without setting a general principle for retrieval at all. The wider copyright fight shows why a middle position is hard to hold.
In a separate case, music publishers have pressed Anthropic over lyrics used in training, with the company urging a judge to rule that training is fair use while publishers argue market harm. That dispute, like the Times case, turns on whether learning from a work is the same as competing with it, and courts have not yet settled the answer.
Why It Matters
The ruling in this case will be highly influential because so few cases have addressed AI and fair use so far, and because the parties include the most prominent newspaper and the most prominent AI company of the moment. Whatever the court decides about training will be cited in every later dispute,
and whatever it says about retrieval will shape how chatbots are allowed to answer questions. For readers, the stakes are practical rather than abstract. If grounded answers must respect licences, AI responses should carry clearer credit to the people who wrote the underlying work, and publishers should have a stronger claim when an answer replaces their article.
If they do not, the open web’s bargain, free reading in exchange for attention and attribution, becomes harder to sustain. For Europe, the case is a signal rather than a rule. A United States ruling does not change European law directly, but it will shape how AI companies and publishers negotiate worldwide,
and it arrives as Europe’s own AI rules take effect. The contrast between court-led rules in the United States and legislation-led rules in Europe is itself becoming part of the story, as our guide to the EU AI Act explains. Wikimedia also used the filing to raise a quieter, operational demand that affects every AI company using its projects.
Regardless of whether material is gathered for training or for answers, the Foundation says scraping must be done sustainably, within its public API rate limits or through its Enterprise API offerings. That is a reminder that the commons has running costs, even when its content is free to read.
Background
The Times filed its lawsuit in December 2023, alleging that OpenAI and Microsoft used millions of its articles without permission to train ChatGPT. The case was later consolidated with other copyright suits against OpenAI, and the court is now considering whether that use qualifies as fair use, a decision that could settle the question without a full trial.
0 Comments