The Wikimedia Foundation wants courts to distinguish between training an AI model to do something new and using a chatbot to reproduce the sources it retrieves.
In a 30 October announcement, Wikipedia’s nonprofit operator said it had filed a friend-of-the-court brief in US copyright litigation involving OpenAI and Microsoft. Its public explanation argues that transformative training may qualify as fair use, while AI services that retrieve and reproduce copyrighted material should respect the sources’ licensing terms.
That distinction matters for Wikipedia, whose article text is generally available for reuse under an open licence with conditions. It also matters for AI users trying to understand why a model learning from a source and a chatbot delivering that source’s content raise different questions.
An amicus, or friend-of-the-court brief, offers an outside participant’s arguments. This article explains the Foundation’s public statements; the underlying filing, its exact filing date and whether the court accepted the submission could not be independently verified.
Why Wikimedia treats training and retrieved answers differently
In his explanation of the argument, Foundation General Counsel Stephen LaPorte describes training as learning patterns from large collections of material. The Foundation argues that this use may qualify as fair use when the resulting model is something new and different from the works used. How the training material was acquired remains a separate legal question.
A chatbot can also consult outside material when answering a question. Retrieval-augmented generation, usually shortened to RAG, connects generation to retrieved sources beyond the model’s training. Grounding is the broader idea of connecting an answer to supporting information; RAG is one way to do it.
Several actions can happen within that process:
- Retrieval finds material relevant to a request.
- Reproduction copies protected expression from a source.
- Substitution describes an answer serving a role that the original source would otherwise fill.
Retrieving a page can support an answer without copying its protected expression. Assessing substitution, infringement and applicable licence conditions requires examining what the answer supplies and how it is used. The Foundation argues that products which retrieve and reproduce copyrighted content in their responses should respect source licences.
Fair use involves more than whether a model creates something new
Fair use is a US copyright doctrine that can allow use without permission. The statutory test considers four factors: the purpose and character of the use, the nature of the copyrighted work, the amount and significance of what was used, and the effect on the work’s potential market or value.
The Foundation’s emphasis on transformation and source substitution speaks to that analysis. Neither concept is the entire test, and its public position should be read as legal advocacy. It does not establish a court’s determination of these uses or the rules in other countries.
What Wikipedia’s open licence requires
Open licensing gives permission under stated conditions. Wikipedia’s copyright policy explains that contributors hold copyright in their contributions; the Foundation does not own all the content or have blanket authority to waive its terms.
Wikipedia article text is generally available under Creative Commons Attribution-ShareAlike 4.0, or CC BY-SA 4.0, with additional licensing details and exceptions described in the policy. Images and other media have their own terms. A photograph appearing in a Wikipedia article may use a different licence, and some material is included under a copyright exception rather than an open licence.
Under the CC BY-SA 4.0 legal code, sharing licensed material requires appropriate attribution, relevant supplied notices, a licence text or link, and an indication of modifications where applicable. The licence allows reasonable ways to provide that information depending on the medium and context. Wikipedia’s text-reuse policy describes attribution methods including a link to the article.
Simply adding “Source: Wikipedia” is therefore an incomplete general instruction for compliance.
ShareAlike adds conditions when someone shares an adaptation they produce. That adaptation must be licensed under CC BY-SA 4.0 or later, or a BY-SA-compatible licence. The legal definition of an adaptation depends on a modification that requires permission under copyright or similar rights. It does not automatically encompass every AI answer, model or application.
There is also an important starting question: is licensed permission needed for the use? Section 2 of CC BY-SA expressly preserves copyright exceptions and limitations. Where an exception such as fair use applies, the licence and its conditions do not apply. Nor does attaching an open licence create copyright protection over public-domain material.
Why Wikipedia’s operator is defending room for AI training
The Foundation’s explanation connects both parts of its argument to the sources and people that sustain Wikipedia.
Its volunteers rely on journalism and research to build articles. The Foundation worries that overly restrictive treatment of training would concentrate AI development among organisations able to buy access, making open-source development harder.
It also warns that overly broad fair-use protection could encourage publishers to restrict access and weaken the licence conditions that keep open knowledge reusable. These are the Foundation’s concerns about possible consequences, rather than measured effects established by this announcement.
For Wikipedia, attribution also connects readers with the underlying articles and their contributors. The Foundation fears that answers replacing those visits could reduce visibility and opportunities to recruit volunteers who maintain the knowledge that AI services draw on.
