PolicyLaw, regulation & governance
Wikimedia asks court in NYT v. OpenAI to rule AI training fair use but not retrieval-augmented chatbot output
In an amicus brief in the New York Times-OpenAI copyright case, the Wikimedia Foundation sides with neither party. It asks the court to rule that pre-training language models is fair use but that retrieval-augmented generation must honour content licences.
From the The Verge · AI article of October 30, 2026 and cited sources: Wikimedia wants pre-training and retrieval treated separately. [3] [1] [7]
SOURCES[3] Wikimedia Foundation: Files to protect open knowledge in… wikimediafoundation.org[1] Wikipedia’s operator says a bad AI copyright ruling could… theverge.com[7] The Next Web: Wikimedia asks court to rule AI training… thenextweb.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- The Verge · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 30, 2026 article published
- Covered
- Oct 30 edition
- Evidence
- 8 sources across 8 sites
- 1Wikipedia’s operator says a bad AI copyright ruling could…theverge.com
- 2Unite.AI: Wikimedia Foundation seeks to file amicus briefunite.ai
- 3Wikimedia Foundation: Files to protect open knowledge in…wikimediafoundation.org
- 4GV Wire: Justice Dept. sides with OpenAIgvwire.com
- 5Search Engine Journal: Wikipedia traffic down as AI…searchenginejournal.com
- 6Axios: DOJ's copyright filing took key agencies by surpriseaxios.com
Reach: not measured yet
Wikimedia wants pre-training and retrieval treated separately. Training on copyrighted text would count as transformative fair use, relying on the Supreme Court's 2023 Warhol ruling. Chatbots that fetch source text and serve it to users would not count as fair use when that output replaces reading the original. [3] [1] [7]
Under Wikimedia's approach, the attribution and share-alike terms of CC BY-SA would apply to RAG and grounding in tools such as ChatGPT and Copilot. That would push retrieval systems toward tracking the source of each passage. Wikimedia also wants scrapers to stay within public API rate limits or use its paid Enterprise API. [3]
The brief breaks with the Times, which argues that scraping to build training datasets is itself infringement. On training, it lines up with the Justice Department's September statement of interest. Publishers, including the Association of American Publishers, filed their own briefs, and news outlets have asked the court to set aside the Justice Department's view. [1] [4] [2] [8]
Part of Wikimedia's motivation is pressure on its traffic and servers. It reported that human pageviews from March to August 2025 were about 8% lower than a year earlier, which it blamed on AI answers in search, though changes to its bot detection make the comparison harder. It also reported that bandwidth used by AI bots has grown 50% since January 2024. [5]
Wikimedia's brief and the Justice Department statement are both non-binding filings.
Read the full assessment
If the court adopts the training-versus-retrieval split, licensing and attribution risk would fall mainly on retrieval and grounding pipelines, not on pre-training data.
Executive brief
The Wikimedia Foundation, which runs Wikipedia, filed a brief in the consolidated New York Times-OpenAI copyright litigation that supports neither side. It asks the court to rule that pre-training large language models on copyrighted text is fair use. It also asks the court to rule that retrieval-augmented generation (RAG), where a chatbot fetches source text and serves it back to users, is not fair use and must follow licences such as Wikipedia's CC BY-SA 4.0 (Wikimedia Foundation). Wikimedia argues that a sweeping win for either side would hurt open knowledge (The Verge).
What changed and event timeline
Times sues
The New York Times sued OpenAI and Microsoft. Its case was later combined with other publisher suits in the Southern District of New York (SDNY) before Judge Sidney Stein ().
Wikipedia traffic falls
Wikimedia reported that human pageviews from March to August 2025 were about 8% lower than a year earlier. It blamed AI answers in search, and noted that changes to its bot detection complicate the comparison ().
DOJ backs OpenAI
The Justice Department filed a statement of interest arguing that AI training is transformative fair use and citing national security. The Times criticised the filing (;).
Amici line up
Engine Advocacy, Authors Alliance, Public Knowledge, academics, Digital Content Next and the Association of American Publishers (AAP) asked to file briefs. Wikimedia's motion is dated October 29 ().
Wikimedia announces its brief
Wikimedia proposes splitting the analysis: training is fair use, RAG output is not (;).
Capabilities and access
No model release is involved.
Read the full section
No model release is involved. The disputed uses are:
- Pre-training of the general-purpose LLMs behind ChatGPT, which Wikimedia argues is fair use (TNW).
- RAG and grounding in ChatGPT and Microsoft Copilot, which Wikimedia argues requires a licence (Wikimedia).
- Access routes: Wikimedia wants scraping done sustainably, either within public API rate limits or through its paid Wikimedia Enterprise API (Wikimedia).
Technical analysis for researchers and developers
- Wikimedia separates what goes into model weights (pre-training) from what is pulled in at inference time (retrieved text).
- CC BY-SA requires attribution and that "remixes or adaptations" stay under the same licence.
- Evidence on memorisation is thin. The method behind that analysis is not public (The Verge).
Read the full section
- Legal line follows the architecture. Wikimedia separates what goes into model weights (pre-training) from what is pulled in at inference time (retrieved text). The test it proposes is whether the output replaces the reader's engagement with the original work (The Verge).
- Licence obligations apply to RAG. CC BY-SA requires attribution and that "remixes or adaptations" stay under the same licence. RAG pipelines would need to track where each piece of text came from at the passage level (Wikimedia).
- Evidence on memorisation is thin. Microsoft cites an expert analysis that Copilot rarely reproduces full sentences from grounding content. The method behind that analysis is not public (The Verge).
Claims and evidence
- Microsoft says expert analysis shows full-sentence reproduction is rare (The Verge). Both are party claims, not independent tests.
- Wikimedia says ChatGPT reproduces CC-licensed text without meeting the licence terms (The Verge). This is also a party's assertion. Coverage cites no independent audit.
- Wikimedia cites more than 67 million articles in 300 languages and about 250 million monthly visitors (Wikimedia).
Read the full section
- OpenAI calls regurgitation "a rare bug." Microsoft says expert analysis shows full-sentence reproduction is rare (The Verge). Both are party claims, not independent tests.
- Wikimedia says ChatGPT reproduces CC-licensed text without meeting the licence terms (The Verge). This is also a party's assertion. Coverage cites no independent audit.
- Wikimedia cites more than 67 million articles in 300 languages and about 250 million monthly visitors (Wikimedia). Unite.AI gives 260,000+ monthly volunteers; Wikimedia's post gives nearly 250,000 (Unite.AI).
- Coverage of the brief adds no independent corroboration beyond restating Wikimedia's filing.
Context and prior work
- The brief relies on the Supreme Court's 2023 Warhol ruling on what counts as transformative use (TNW).
- The DOJ's position on training lines up with Wikimedia's. News outlets argue the DOJ view conflicts with the US Copyright Office's analysis (Bloomberg Government).
- Wikimedia had earlier reported a 50% increase in bandwidth used by AI bots since January 2024.
Read the full section
- The brief relies on the Supreme Court's 2023 Warhol ruling on what counts as transformative use (TNW).
- The DOJ's position on training lines up with Wikimedia's. News outlets argue the DOJ view conflicts with the US Copyright Office's analysis (Bloomberg Government).
- Wikimedia had earlier reported a 50% increase in bandwidth used by AI bots since January 2024. A cited study found Wikipedia makes up 47.9% of ChatGPT's top-10 most-cited sources (SEJ).
Limitations, safety and contested findings
- Wikimedia supports exempting scraping and training, while the Times argues that scraping to build training datasets is itself infringement (The Verge).
- Wikimedia warned that the 8% traffic drop is hard to compare year on year because its bot detection changed (SEJ).
- Sources disagree on the filing date: October 29 for the motion, October 30 for the announcement.
Read the full section
- Wikimedia supports exempting scraping and training, while the Times argues that scraping to build training datasets is itself infringement (The Verge). Publishers, including the AAP, filed their own briefs (Unite.AI).
- Wikimedia warned that the 8% traffic drop is hard to compare year on year because its bot detection changed (SEJ).
- Sources disagree on the filing date: October 29 for the motion, October 30 for the announcement.
- An amicus brief does not bind the court. Summary-judgment briefing was still under way (Unite.AI).
Business and practitioner implications
- RAG is the exposure point.
- Passage-level source tracking, attribution in the interface and handling of share-alike terms are worth adding to retrieval stacks before a ruling.
- Wikimedia Enterprise and API rate limits become the standard Wikimedia is pushing for responsible scraping (Wikimedia).
Read the full section
- RAG is the exposure point. If the court adopts Wikimedia's split, licences and attribution would matter mainly for retrieval and grounding, not for pre-training corpora.
- Build provenance in now. Passage-level source tracking, attribution in the interface and handling of share-alike terms are worth adding to retrieval stacks before a ruling.
- Paid access gains weight. Wikimedia Enterprise and API rate limits become the standard Wikimedia is pushing for responsible scraping (Wikimedia).
- Publishers face a split among amici: a public-interest nonprofit that depends on open content is backing fair use for training.
The source trail.
Sources (8)
Wikipedia’s operator says a bad AI copyright ruling could threaten the open web
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.com