Our advice to the court in NYT v. OpenAI: Copyright fair use is not an “all-or-nothing” and attribution helps defend human knowledge
The Wikimedia Foundation filed an amicus brief in NYT v. OpenAI, urging the Southern District of New York to rule that copyright fair use should be applied in a nuanced and technology-specific manner. The Foundation argues that while many uses of copyrighted works to train AI models are likely fair, using copies in chatbots with retrieval-augmented generation (RAG) diverts human viewers away from the originals and is likely not fair. We hope that this approach will help the court strike the right balance between encouraging innovation and preventing harm to sources of knowledge.
Written by the Wikimedia Foundation’s: Stan Adams, Lead Public Policy Specialist for North America; and, Angela Chung, Technology Law and Policy Fellow
The Wikimedia Foundation submitted an amicus or “friend-of-the-court” brief in In re: OpenAI, Inc. Copyright Infringement Litigation (12 combined lawsuits against OpenAI, the most well-known of them being The New York Times Co. v. Microsoft Corp. and OpenAI) on 30 October 2026 before the US District Court for the Southern District of New York.
This lawsuit is the key case in US courts concerning generative artificial intelligence (AI) and copyright. This matters for the Wikimedia projects because copyright and fair use define the boundaries of what can be part of free and open knowledge and how the Creative Commons (CC) licenses that set the terms for using the content on the Wikimedia projects work in practice.
The case is about whether and to what extent OpenAI and Microsoft’s use of content owned by The New York Times should be permitted under copyright law under the defense of fair use. Each side pushes for an “all-or-nothing” outcome: The New York Times seeks a ruling in which all uses of their content by OpenAI should be deemed not fair, while OpenAI seeks a ruling in which all their uses of content should be deemed fair.
Our brief is not fully aligned with either side in the dispute. Instead, we argue that the right outcome for the free knowledge movement requires that each kind of use present in the large language model (LLM) technology at issue in the case should be distinguished and analyzed separately. Per this case-by-case analysis, we focused on two particular ways that copyrighted works are used in LLMs: the initial LLM training process and the use of retrieval-augmented generation (RAG) to answer the questions that users ask a chatbot. We believe that these two uses are good representations of a broader range of ways that copyrighted works are used in AI. Of the two, we asked the court to find that using copyrighted content for training a model should be considered fair use because that use serves a different purpose than the original works. But we asked the court to find that when a chatbot uses copies of works to extract and deliver that same content to users without attributing the original sources, such a use should not be considered fair because it competes with and substitutes the purpose of the original works.
What this means in practice is that while we think that model developers should be free to experiment with and train different models, we also think that when they deploy them and start copying Wikipedia to answer individual user questions, they should be following the CC licenses such as by properly attributing Wikipedia in an answer.
Copyright in context: How copyright and the ruling in this case affect Wikimedia’s mission
Copyright law is one of the defining things of what fits in the free and open knowledge movement. Copyright places limits on the way that writing, art, photos, and other works of media can be used, copied, and shared that determines whether and how they can be hosted on the Wikimedia projects. The Wikimedia movement, in turn, has a 25-year history of a strong commitment to create free knowledge that anyone can use and to keep it free (see Jimmy Wales saying that the strong virality of the licenses keeping knowledge free is critical to Wikipedia as early as 2001). The CC BY-SA 4.0 license is the thing that translates this ideal of freedom into legal terms that apply to most contributions to the Wikimedia projects.
Fair use, in its turn, limits copyright in a range of ways: it supports the free knowledge movement by allowing people to copy works in a variety of situations where there is no free alternative; it also sets limits on when people need to follow the CC licenses. When it’s working, fair use contributes to a thriving society by ensuring that people can access and copy works for socially valuable reasons like innovative technologies, education, and critical commentary without constantly needing to stop to ask permission. The court’s ruling in this case is thus very important for the Wikimedia movement: it will affect how the Wikimedia projects and the CC licenses work in practice with technologies now used by hundreds of millions of people.
Why the parties’ “all-or-nothing” approaches to fair use harm free knowledge
The key issue being addressed in this case is the following: To what extent should fair use allow or restrict developers’ uses of millions of copyrighted works across the internet when creating new technologies and consumer-facing services such as LLMs and other generative AI?
Get Wikimedia Foundation Policy’s stories in your inbox
Join Medium for free to get updates from this writer.
OpenAI argues that ALL of the ways that they use copyrighted works to develop and operate their LLM systems should be considered fair. In practice, this would mean that they need neither permission nor to honor the license terms for the works they use–including those of Wikimedia projects. If courts agree with OpenAI’s position, AI developers will be able to use anything they can copy in any way they want, including without following the attribution and share-alike requirements of the CC license (or anybody else’s license terms). This would either leave online publishers helpless as they lose traffic and users and/or push many online publishers to make their content more difficult to access through technical methods. Both would lead to a more closed web in which fewer people have access to knowledge.
On the other hand, The New York Times (NYT) argues that NONE of the ways AI developers use copyrighted works should be considered fair. If courts agree with NYT’s position, the process of building and experimenting with AI tools will become substantially more expensive, logistically burdensome, and likely only possible for the largest, most well funded corporations. Many open source developers would not be able to build or even access models, as proprietary developers would be even more restrictive and resistant to transparency. This outcome would harm the Wikimedia movement’s ability to build and experiment with new open source technologies.
In our view, both dimensions of the free and open knowledge movement must be addressed through a sharper and more nuanced application of the fair use doctrine than the parties in this case are seeking.
A “case-by-case” approach to fair use better supports the open knowledge movement
We believe that adopting a case-by-case approach to fair use best helps to maintain open technological development while also deterring practices that reduce human knowledge creation. Our detailed position therefore seeks to strike a balance between the two parties’ all-or-nothing arguments. In brief, we argue the following two points:
First, the use of copyrighted works in training an LLM should be considered fair because the purpose of that use (providing statistical data on the relationships between words) and the purpose of the original works (expressive information and entertainment) are completely different. This is transformative and does not reduce demand for the original content. We also note that some ways of deploying trained models are fair uses: for example, systems built on large language models may help volunteers make the best use of their time for tasks like reviewing new edits, and detecting undesirable posting of personal information and doxxing attempts, among other uses.
If the court adopts this position, developers will be able to use a broader range of works without needing to stop and ask permission from everyone involved when building new innovations that can, in turn, further contribute to the dissemination of knowledge.
Second, some features of generative AI may use copyrighted works in ways that are not fair. Chatbots that use techniques like RAG or its modern variants can be particularly problematic when structured to search, copy, and reproduce content from third-party websites like Wikipedia without acknowledging from where that content was taken when generating an answer to a user’s query. The use of copies by these kinds of RAG systems or similar grounding techniques should usually not be fair because they serve the same purpose as the original works: to provide information relevant to a user’s search. This is not transformative and displaces viewership of the original content used.
If the court adopts this position, AI developers would be required to comply with CC licensing terms when they copy from the Wikimedia projects to answer a user’s questions, and, as a result, chatbots with RAG or other similar features would more consistently attribute the original sources being relied upon to answer user queries.
In short, we recommend that the court recognize that each individual use of copyrighted works in the development and operation of an AI system should be considered separately. Training generative AI models to create new and useful tools can be fair, but using features like RAG to copy and replace sources on which they rely should not be. A case-by-case approach can protect innovation while safeguarding attribution, licensing, and the sustainability of open knowledge.
*****
We appreciate the providers of generative AI systems that are already making an effort to provide attribution in their generative outputs and follow the lenient requirements of our CC licenses. We urge others to do the same and for all commercial reusers of Wikimedia project content to engage with that content in a manner that supports the sustainability of free and open human knowledge.
We will continue advocating for a copyright framework that strikes a fair balance. One that balances the importance of protecting fair uses of copyrighted content to develop open source works in the public interest, along with the sustainability of the sources and humans who create that content.
We thank Rossella Gabriele, Fellow-in-Residence at Yale Law School Information Society Project (ISP) for her support in research and editing, and Rhett Millsaps of Lex Lumina LLP for his support in the filing process.
