Wikipedia's owner wants AI training protected, but not copy-paste answers
Wikimedia's amicus brief in the NYT-OpenAI fight splits AI training from chatbot retrieval. Here is why ANI's Delhi case and Indian AI startups should pay attention.
Key points
- The Wikimedia Foundation filed an amicus brief in the New York Times copyright case against OpenAI and Microsoft, backing neither side.
- The brief, filed on 30 October, says AI training should be fair use but chatbots supplying copies of articles should need a licence.
- A ruling along those lines could be cited in ANI's Delhi High Court case and would put retrieval-based AI products under more licensing pressure.
Photo: Christopher Bowns / Wikimedia Commons, CC BY-SA 2.0 (cropped). Representational image.
The News
The Wikimedia Foundation, the nonprofit that runs Wikipedia, has filed an amicus brief in the copyright litigation brought by The New York Times and other publishers against OpenAI and Microsoft. The filing, lodged on Friday 30 October, backs neither party. Instead it asks the court to separate two uses of copyrighted text.
On one side sits training: Wikimedia argues that scraping should not count as infringement, and that pre-training large language models on copyrighted works without a licence should qualify as fair use where the model produces something genuinely new. On the other side sits retrieval: when a chatbot fetches copies of material in answer to a user's question and hands it over in a way that replaces a visit to the original, Wikimedia says a licence should be required.
For Wikipedia, honouring that licence would mean crediting the source under its Creative Commons terms, which also oblige any adaptation to carry the same licence. Chief executive Bernadette Meehan framed the filing as a defence of open human knowledge, while general counsel Stephen LaPorte warned a careless ruling could stifle open-source technology.
Why It Matters
The Times alleges that OpenAI infringed its copyright by scraping its site to build training datasets. OpenAI and Microsoft maintain that their use is fair, with OpenAI describing verbatim regurgitation as "a rare bug". Wikimedia warns that an outright publisher win could let rights holders block transformative uses, while an outright AI-industry win could amount to "a green light to snub licensing terms".
Wikimedia sits on both sides of the trade. Its sites are among the most heavily used sources of training data, yet the chatbots built on that data divert readers away from the volunteer communities that maintain the articles. Yet its editors use language models to fight vandalism and spot copyright violations. The split it proposes mirrors the reasoning US courts applied to Google Books in 2015, where wholesale copying for search was allowed because the output did not substitute for the books themselves.
Indian Angle
The closest Indian parallel is already before the Delhi High Court. News agency ANI sued OpenAI there in November 2024, alleging that its reporting was used to train ChatGPT without permission, and several Indian publishers and music labels later sought to join the proceedings. A US ruling on training versus retrieval would not bind an Indian judge, but it is likely to be cited in Delhi.
That caveat matters because Indian law is structured differently. Section 52 of the Copyright Act, 1957 sets out a closed list of "fair dealing" exceptions rather than the open-ended, four-factor fair use test used in the US. There is no express exception for text and data mining, so an Indian court may find it harder to adopt Wikimedia's generous view of training. The government has also been weighing a statutory licensing model with royalties for AI training.
For Indian builders, from enterprise chatbot vendors to Indic-language model makers such as Sarvam, retrieval features are the more exposed part of the stack. Products that reproduce publisher content on demand face a stronger licensing argument than models trained once on scraped text.
FAQ
What exactly has Wikimedia asked the court to do?
It wants the court to avoid a blanket ruling either way. In its view, pre-training on copyrighted works without a licence should be fair use when the model creates something new, but serving copies of articles that replace reading the original should require a licence and attribution.
Does Wikimedia support OpenAI or The New York Times?
Neither. The brief formally takes no side and criticises both parties for pushing overly broad readings of fair use. It warns that a sweeping win for publishers or for AI companies could harm volunteer-run open knowledge projects such as Wikipedia.
Does a US ruling apply in India?
No. A US decision does not bind Indian courts, and India uses a narrower fair dealing regime under Section 52 of the Copyright Act, 1957. Judges hearing similar disputes, including ANI's case against OpenAI in Delhi, may still treat a reasoned US judgment as persuasive.
What should Indian AI startups watch for?
The line between training and retrieval. If courts accept Wikimedia's approach, features that reproduce publisher content in real time are likely to need licences, so startups offering news summaries or search-style answers should review their content sourcing, licensing and attribution now.
This story was reported by The Verge. Read the full original coverage at The Verge.