--- title: "Wikimedia Tells Court AI Training Is Fair Use but Grounding Isn't" description: "In the New York Times v. OpenAI case, Wikipedia's operator says training AI on your content can be fair use, but AI answers that copy it back need a license." url: https://publishmyblog.com/ai/wikimedia-ai-fair-use-brief/ section: "AI" author: "Saeed Shahriar" published: 2026-11-02T08:35:00.000Z --- # Wikimedia Tells Court AI Training Is Fair Use but Grounding Isn't The nonprofit behind Wikipedia takes neither side, and draws the same line publishers care about: the answer box, not the training run. Wikimedia wants a New York court to split AI copyright in two. **Training on your work can be fair use, it says, but chatbots that copy it back to readers should follow your license.** The nonprofit behind Wikipedia [filed its amicus brief on October 30](https://wikimediafoundation.org/news/2026/10/30/wikimedia-foundation-amicus-brief-new-york-times-v-open-ai-case/), and **it takes neither side**, as [The Verge's report on the brief](https://www.theverge.com/policy/1022880/wikipedia-openai-new-york-times-copyright) notes. ## What the case is about The brief went to the Southern District of New York, in 12 combined lawsuits against OpenAI. **The best known is The New York Times against Microsoft and OpenAI**, filed in 2023. Then there's the question the court has to answer: is using copyrighted content to build large language models fair use? If you publish anything online, your content sits inside that question too. Wikimedia expects the ruling to carry weight. Few courts have dealt with it so far, and it calls the court and the parties significant, so others will look here. The two sides aren't close: - The Times says any use of its content, training included, breaks copyright law. - OpenAI and Microsoft say training and deploying models on public data is fair use. ## Wikimedia's middle line So where does Wikimedia land? It won't treat every use the same. Training on copyrighted content is fair, it argues, when the model is something new and different from the sources. **No license terms would apply.** Grounding is the other case. That covers retrieval-augmented generation (RAG) and similar methods, where a chatbot finds your content on the web and hands it back to the person asking. Here, Wikimedia says, **that use isn't fair**, and the AI company should respect your license. On Wikipedia that's CC BY-SA 4.0. It asks for credit to the creators and keeps adaptations under the same license, which means **an AI answer built on a Wikipedia article would have to say so**. Why should you care about Wikipedia's license? **That second case is where your visitors go**, and The Verge says Wikimedia sees AI siphoning visitors from its own sites. ## Why Wikipedia sits in the middle It uses AI too. **The brief says LLMs help its editors fight vandalism** and catch copyright violations, The Verge reports. That's why it wants neither extreme: - A sweeping win for publishers could cost Wikipedia tools its editors use. - A sweeping win for AI companies could mean a more closed web, with less source material to go round. It'd also like scraping done within its API rate limits, or through Wikimedia Enterprise. - Wikipedia has more than 67 million articles in 300 languages. - Nearly 250,000 volunteers work on it every month. ## Why it matters Of course, an amicus brief is only an outside voice, and **the judge is free to ignore it**. Still, I think it draws the right line for you. **Bing made the same split** when it [separated AI training from AI search grounding](https://publishmyblog.com/seo/bing-splits-ai-training-and-grounding/). My read: training costs you little on the day it happens. Grounding is where you lose the click, because the answer box hands your reader what your page would have. If courts land anywhere near this line, I'd expect the money you can make from AI as a publisher to sit in licensing the answers that reuse your pages, rather than in training runs. So what can you do now? Treat training and grounding as **two separate decisions in your crawler settings**, since I think the court may end up treating training and grounding differently too. **Bing already lets you split them.** Meanwhile, Google's planned [shared AI training opt-out in robots.txt](https://publishmyblog.com/seo/google-will-back-one-robots-txt-rule-for-ai/) covers your training side. Check your settings today. Then watch for the ruling. It'll matter. ## Common Questions **What did the Wikimedia Foundation file in the New York Times v. OpenAI case?** An amicus brief, filed on October 30, 2026 in the Southern District of New York, in the 12 combined lawsuits against OpenAI. It takes neither side and argues that training on your work can be fair use, while chatbots that copy it back to readers should follow your license. **Does Wikimedia think AI companies can use Wikipedia without credit?** Only for training. For retrieval-augmented generation (RAG) and similar methods, Wikimedia says AI companies should respect its CC BY-SA 4.0 license, which asks for credit, and should scrape within its API rate limits or through Wikimedia Enterprise. **When will the court rule on fair use in the OpenAI copyright cases?** The sources don't give a date. The court will decide whether using copyrighted content to build large language models is fair use, and Wikimedia expects the ruling to carry weight with other courts. ## Sources - [The Wikimedia Foundation files to protect open knowledge in New York Times v OpenAI case](https://wikimediafoundation.org/news/2026/10/30/wikimedia-foundation-amicus-brief-new-york-times-v-open-ai-case/), Wikimedia Foundation - [Wikipedia’s operator says a bad AI copyright ruling could threaten the open web](https://www.theverge.com/policy/1022880/wikipedia-openai-new-york-times-copyright), The Verge