The Wikimedia Foundation, the non-profit that runs Wikipedia, asked a federal judge in New York on October 29, 2026 for permission to file a brief arguing that the training of the models behind ChatGPT was likely fair use, but that the chatbot's practice of fetching web pages and reproducing them in its answers was likely not.

The 5,831-word proposed brief, styled "in support of neither party", lands in In re: OpenAI, Inc. Copyright Infringement Litigation, case number 25-md-3143, the consolidated proceedings before US District Judge Sidney H. Stein in the Southern District of New York that include The New York Times Co. v. Microsoft and OpenAI. Wikimedia publicised the filing on October 30, 2026, with statements from its chief executive, Bernadette Meehan, and its general counsel, Stephen LaPorte. Like any amicus filing, it is advocacy from a party outside the case. It argues for an outcome rather than presenting evidence, and, as Wikimedia's own motion acknowledges, "a district court has broad discretion to decide whether to accept an amicus curiae brief."

What makes the filing unusual is where it lands. On training, it lines up with OpenAI and Microsoft. On retrieval-augmented generation, the step in which a chatbot searches the web, copies passages into the prompt alongside a user's question and builds a reply from them, it sides with the publishers. And on scraping itself, it argues that copying a page from a website should not count as a copyright "use" at all - a position at odds with the publishers' bid to hold OpenAI liable at the point of acquisition.

In Short

Wikipedia's parent organisation told the judge in the New York Times v. OpenAI copyright case that letting an AI model learn from copied text should be allowed, but that a chatbot grabbing web pages to build an answer should have to follow the rules of the site it copies. For Wikipedia, those rules mean giving credit and keeping the material under the same open licence, so the argument matters to any publisher whose pages end up inside chatbot answers. The judge has not yet said whether the brief will be accepted, so nothing changes for now, but it adds a third position to a case that will shape how AI search tools treat the sources they draw on.

A brief for neither side

The motion arrived as a three-page letter to Judge Stein from Rhett O. Millsaps II of Lex Lumina LLP in New York, docketed as ECF No. 2361, with the proposed brief attached as Exhibit A. It relies on Local Civil Rule 7.1(e) and on an order the court issued on August 7, 2026 (ECF No. 1663), as modified on September 24, 2026 (ECF No. 2049). That order set a 6,000-word limit for amicus briefs and directed prospective amici to seek leave by motion, which, according to the letter, is why Wikimedia did not ask the parties for their consent. The brief comes in 169 words under the cap.

The standard disclosures follow. "No party or party's counsel authored the proposed brief in whole or in part, and no person other than Wikimedia contributed any money to fund its preparation or submission," the motion states. Alongside Millsaps, the brief is signed by three of Wikimedia's in-house lawyers: LaPorte, Associate General Counsel Jacob Rogers and Legal Counsel Aly Marino.

Wikimedia gives the court three reasons to read it. The first is that it can explain how a fair-use ruling would affect the open-knowledge movement built on Creative Commons licences. The second is that its years of building machine-learning tools give it what the motion calls "a different and more balanced approach to fair use than presented by either party in this case." The third concerns retrieval: the motion says its significance "in disturbing the flow of information between human authors and audiences is not explored deeply by either party," and cites a 2015 decision, U.S. v. Yaroshenko, for the point that a valid amicus "offers a fresh perspective on an unsettled question of law that the actual parties to the litigation have not fully addressed."

That third claim is open to question. When OpenAI, Microsoft and five news organisations moved for summary judgment on September 4, 2026, OpenAI asked Judge Stein to rule that both pre-training and Browse - a ChatGPT function the plaintiffs describe as retrieval-augmented generation - were fair use, while the publishers sought liability at separate stages that included grounding and outputs. Retrieval has been briefed. What Wikimedia adds is the view of an organisation whose content sits on both sides of the line it wants drawn.

There is also a gap between the two documents Wikimedia released. Its announcement is dated October 30 and says the foundation "today filed" the brief; the docket stamps on the motion and on the exhibit both read October 29. The announcement describes the proceedings as 12 combined lawsuits against OpenAI, a count the court papers themselves do not give. And strictly, the court has not yet accepted a brief from Wikimedia: the document is a proposed brief, and it joins the record only if Judge Stein grants leave.

Why Wikipedia has a stake

The brief's statement of interest describes thirteen platforms, known collectively as the Wikimedia Projects: Wikipedia, Wikibooks, Wiktionary, Wikiquote, Wikimedia Commons, Wikisource, Wikiversity, Wikispecies, Wikidata, Wikifunctions, MediaWiki, Meta-Wiki and Wikivoyage. Their content, according to the brief, is "created, edited, and moderated by over 260,000 volunteer contributors per month worldwide," a figure footnoted as of January 2025. The announcement published a day later puts the monthly total at "nearly 250,000 volunteers" and says Wikipedia holds more than 67 million articles across 300 languages. In January, PPC Land reported Wikipedia's 25th anniversary with figures of nearly 250,000 volunteer editors and 65 million articles.

The legal interest is licensing. Most written contributions are released under the Creative Commons CC BY-SA 4.0 licence, which, according to the brief, requires "(i) that credit be given to the original creators of the work and (ii) that remixes, adaptations, or any other derivatives of the original works are made under the same free license terms." That combination is what the brief calls a "copyleft" model. Its force depends on copyright. If a use is fair, the licence never comes into play; for fair-use training, as Wikimedia's announcement puts it, "there would not be a requirement to meet any copyright licensing terms."

The motion spells out the central worry in a single sentence. "When AI-based chatbots use Wikimedia-hosted content without also abiding by the CC licensing terms, humans are directed away and thus disincentivized from engaging with free-knowledge projects," it says.

The second worry is infrastructure. "Sustained increases in non-human traffic from scraping and crawling impact Wikimedia's ability to keep Wikimedia sites online for human readers," the brief states, adding that Wikimedia has put technical safeguards in place and directed high-volume users to a service called Wikimedia Enterprise. Without enforceable copyright, the brief warns, "web publishers will be pushed to put up more walls against bad actors, leading to a more closed internet for people."

That concern has a recent record. On October 5, Wikimedia said software agents it believes OpenAI operates made millions of automated requests and may have played a part in a partial disruption of its Wikidata Query Service in May. The same coverage noted that bots accounted for 65 percent of the most resource-consuming traffic, that bandwidth had risen 50 percent since 2024 because of bot activity, and that human traffic to Wikipedia had declined 8 percent over a year. None of those figures appears in the brief. Its only traffic citation supports the statement that Wikipedia "is viewed billions of times every month."

The commercial relationships are tangled, too. Several AI and search companies, including Microsoft, Mistral AI and Perplexity, became Wikimedia Enterprise partners in the year to January 2026, joining Amazon, Google and Meta, according to PPC Land's anniversary coverage. Microsoft is a defendant in the case Wikimedia is now briefing. OpenAI was not among the partners named.

Four stages, separate tests

The brief's structural argument rests on the Supreme Court's 2023 decision in Andy Warhol Foundation for the Visual Arts v. Goldsmith. Wikimedia reads Warhol as requiring courts to run the four-factor fair-use test separately for each use of a copied work, because "[t]he same copying may be fair when used for one purpose but not another."

Wikimedia "agrees with Plaintiff The New York Times," the brief says, that deploying a large language model can be described in four general stages: data acquisition through scraping, pre-training, fine-tuning or post-training, and output generation. Each must be assessed on its own. "Nor should courts import or export the 'purpose' of a copy's use between upstream or downstream stages of generative-AI product development," the brief argues.

Its model for that approach is Bartz v. Anthropic, the Northern District of California case in which Judge William Alsup found on June 23, 2025 that training on books was fair use but rejected Anthropic's defence of a library built from pirated copies. The brief quotes Alsup's separate holdings that retaining pirated copies "should they prove useful for one thing or another" was not transformative, while "using copyrighted works to train LLMs to generate new text was quintessentially transformative."

Three of the four stages the brief names get their own analysis. Fine-tuning and post-training, the third stage, are listed but not examined.

Scraping as a precondition, not a use

The brief's most novel claim concerns the first stage. "Scraping is simply a means of acquiring copies of works, but acquisition in and of itself is not a use of those works," it argues. What matters, in Wikimedia's account, is what a developer does with the copies afterwards.

The argument cuts in both directions. "Unauthorized acquisition of copies does not mean later uses of those infringing copies are not fair - indeed, such a rule would undermine the purpose of fair use," the brief says. "Likewise, fair uses of copies at one point do not transform other, unfair uses of those copies at another point into fair ones."

Wikimedia is careful not to bless scraping wholesale. "This does not mean that scraping activity overall is benign," a footnote reads, noting that some forms violate site terms of use and state contract or tort law, and may also breach the Digital Millennium Copyright Act and the Computer Fraud and Abuse Act. Another footnote frames the point with an analogy: "a public park can remain free and open while preventing someone from driving a container truck through it."

This is where the brief parts company with the plaintiffs most clearly. According to PPC Land's coverage of the September filings, the publishers' 92-page summary judgment brief sought liability at five stages, beginning with acquisition. Wikimedia's position would take that first stage out of the copyright analysis altogether, leaving any remedy for aggressive crawling to contract law, the DMCA or computer-misuse statutes.

Pre-training: likely fair

On training, the brief concedes that "this issue has been well-briefed by the parties," and spends much of its section on Wikimedia's own experience. Machine-learning tools, it says, help volunteer editors "fight vandalism, detect copyright violations and inappropriate personal information, determine what articles most need their help for improvement, and find new topic areas to work on." It acknowledges a cost as well: LLMs have caused "increased difficulty for volunteer contributors when used poorly," citing Wikipedia's own guideline on the presumptive removal of AI-generated content.

The factor-by-factor analysis then runs largely in OpenAI's favour.

On the first factor, purpose and character, the brief argues that the purpose of works on sites like Wikipedia "is to distill and communicate reliable information about the world from a neutral point of view," while copies used in pre-training serve "to provide linguistic patterns that undergird development of the model." It cites the Second Circuit's HathiTrust decision and Kadrey v. Meta, in which Judge Vince Chhabria granted Meta summary judgment on June 25, 2025 in a ruling limited to the thirteen authors who sued.

The second factor, the nature of the work, is likely neutral. Wikipedia articles are factual, but the brief argues that factual material is still protected and that Wikipedia editors "invest substantial efforts to distill information from other sources."

On the third factor, the amount used, the brief accepts that copying entire works is reasonable for training. It quotes Bartz on the need for "billions of words to train any given LLM" and Kadrey on the observation that a full work "does more to train [an LLM] than would feeding it only half."

The fourth factor, market harm, also favours fair use in Wikimedia's view, because pre-training "does not reproduce expressive content or meaning in a manner substantially similar to and thus substitutive of original works." A footnote anticipates the counter-argument that a general-purpose model helps others build infringing tools, and answers it with the Supreme Court's 1984 decision in Sony v. Universal City Studios: the possibility of furthering infringement "does not mean the technology itself is a violation of copyright law."

On this stage, Wikimedia stands alongside the US government. On September 1, 2026, the United States filed a 20-page statement of interest in the case, urging Judge Stein to reject arguments that training LLMs on copyrighted text violates copyright law.

Retrieval: likely not fair

The brief's case against retrieval begins with a description of how it works. "The RAG system can search the web (or another database that may be built from copies of protected works) to identify, copy, rank, and then paste relevant content into the LLM's context window alongside the user's query," it says. According to the brief, "LLM-based chatbots like ChatGPT depend on RAG to provide timely and accurate responses to user queries that cannot be answered by the base knowledge in the underlying LLM."

Three of the four factors then turn against OpenAI; the second, the nature of the work, is again treated as neutral.

On purpose, copies used in retrieval "do not have a further purpose or different character" from the originals, the brief argues, quoting Warhol. It relies on Thomson Reuters v. Ross Intelligence, the 2025 Delaware decision finding that copying one legal research tool's annotations to build another was not transformative, and on Associated Press v. Meltwater, a 2013 Southern District of New York decision. Retrieval inputs and outputs that reproduce protected expression without complying with licence terms, the brief says, "act[] as a substitute for [the original] sites" without "driving [consumers back] to [those] third-party websites."

There is a carve-out for summaries. Many summaries, according to the brief, use sufficiently original language that "they do not violate the expression in the original work at all," and others can serve a transformative purpose. "Our point is that summarizations should not be treated automatically as unfair either," the brief says.

The commercial dimension carries the brief's least conventional argument. Drawing on Harper & Row v. Nation Enterprises, it argues that profiting "without paying the customary price" weighs against fair use, and that "[t]his 'customary price' can be financial compensation or logistical compliance with limitations of use in licensing terms." Put plainly: for openly licensed content, the price is attribution and ShareAlike, and a chatbot that skips them has not paid.

On amount, the brief describes retrieval as taking "the heart" of an original. A system that "breaks that content into 'chunks,' and then selects the most relevant chunks" for the model is, in the brief's phrase, performing "a 'heart' transplant by repackaging the crux of original sources."

On market harm, the brief contrasts retrieval with the search-engine cases that went the other way. In Perfect 10 v. Amazon.com, a search use was transformed "into a pointer directing a user to the source of information." Retrieval, according to the brief, does the opposite. Users diverted from a Wikipedia entry "are also diverted from the opportunity to further that knowledge via, for example, volunteer editing or donating." It closes the factor with the Second Circuit's 2024 Internet Archive decision: any "short-term public benefits" of substitution "are outweighed not only by harm to...authors but also by the long-term detriments society may suffer."

The footnote that complicates the argument

The brief's conclusion asks Judge Stein to hold that "the copying and reproduction of expressive sources in responses to user queries by ChatGPT's RAG system is not fair use." A footnote earlier in the same document is less categorical.

"There are also services using RAG, including ChatGPT, that do make attempts to consistently cite and link to original sources used," it reads. "When they do, and thus abide by the CC BY-SA licenses of Wikipedia works, such uses are less likely to infringe public copyright licenses at all." The same footnote concludes that "depending on the facts, the use of copies for RAG outputs may or may not be fair."

Read together, the two passages suggest the argument is about retrieval without attribution rather than retrieval as such. The brief does not measure how often ChatGPT fails to credit Wikipedia. Its empirical support for an attribution problem is a single study, "The attribution crisis in LLM search results" by Strauss, Yang, O'Reilly, Rosenblat and Moure, published in Data & Policy in 2026, quoted as finding "a substantial 'attribution gap' between relevant content consumed from websites and attribution practices that threatens the sustainability of the digital content ecosystem."

What compliance would look like

The brief's remedy is narrow. "Compliance with the provisions of Creative Commons or other copyright licenses would make RAG uses of works noninfringing," it says. A ruling to that effect, it argues, would push developers to build licence compliance into their products, for example by "structuring their product design so that outputs using RAG techniques also embed the underlying content's source attribution." In that world, "chatbots using RAG to search and surface information from Wikipedia would more consistently follow attribution terms and/or provide links back to the original source material."

Wikimedia's announcement draws the commercial line more directly. For the Wikimedia projects, it says, the aim is "ensuring that uncredited use of the content we host is limited to model training," and that any retrieval or grounding use "must result in proper attribution" under CC BY-SA 4.0. Scraping, whatever the purpose, "must be done sustainably and in line with our public API rate limits or through our Enterprise API offerings."

Several questions remain unanswered in the filing. The brief asks that retrieval honour both the attribution and the ShareAlike terms of the licence, but it does not explain how ShareAlike would attach to a chatbot's answer - whether a reply drawing on a Wikipedia passage would itself have to be released under CC BY-SA. Nor does it address what "the relevant licenses" would mean for content such as The New York Times's, which is not openly licensed and for which compliance would presumably mean a negotiated deal. The Times's chief executive, Meredith Kopit Levien, said in August that the company has sued three AI companies and licenses its content to Amazon, setting out conditions that include control over how its content is used.

The quotes in the announcement frame the filing as a matter of balance. "We are taking a position in this case to defend the free and open human knowledge that Wikipedia depends on. Expanding access to knowledge through AI is a powerful goal, but it must respect the people, contributors, and communities that make reliable knowledge possible," Meehan said. LaPorte argued that "any conversation around fair use and copyright in the age of AI cannot be limited to for-profit organizations: it must include stakeholders across the entire online information ecosystem."

The case it lands in

The filing arrives with the main motions already submitted. On September 4, OpenAI, Microsoft and five news organisations each moved for summary judgment. According to PPC Land's coverage, the plaintiffs are The New York Times Company, eight newspapers in the Daily News group, among them the New York Daily News, the Chicago Tribune and The Denver Post, Ziff Davis, the Center for Investigative Reporting and The Intercept. Together they assert 10.8 million articles; the Times alone claims 6,030,928, published between 1928 and 2024. The defence motions cited an expert finding of 24 instances of verbatim regurgitation in a sample of 20 million ChatGPT conversation logs, a rate of 0.00012 percent.

Those logs have their own history. In November 2025, Magistrate Judge Ona Wang refused to pause an order requiring OpenAI to hand over 20 million de-identified ChatGPT conversations to the news plaintiffs. Ziff Davis, publisher of CNET, IGN and Mashable, filed its own 62-page complaint in Delaware on April 24, 2025 before its claims were folded into the consolidated proceedings. As of PPC Land's September coverage, both defendants had requested oral argument and no date had been set.

Wikimedia's announcement summarises the two sides in its own terms. The Times, it says, alleges that OpenAI and Microsoft's use of its content "for any purpose - including AI model training - is a direct violation of copyright law." OpenAI and Microsoft, in Wikimedia's account, "assert that training and deploying models based on publicly available data constitutes technological development, and falls within the definition of fair use." The brief's conclusion asks the court to reject both: rather than blanket determinations that "'all uses by OpenAI are fair' or 'no uses by OpenAI are fair,'" Wikimedia wants separate analyses for each copy.

Why it matters for marketers and publishers

For anyone who depends on search or AI referral traffic, the line Wikimedia draws sits in the commercially important place. Training happens once, out of sight, and sends no readers to a publisher. Retrieval happens whenever a chatbot looks something up to answer a question, and it is at that moment that a source is either cited and linked or absorbed. Wikipedia is unusually exposed to that moment: research cited in PPC Land's anniversary coverage found it led all domains cited in Google's AI Mode, with 12.02 percent of appearances in July 2025.

The argument also tracks a pattern in other litigation. When Encyclopaedia Britannica and Merriam-Webster sued Perplexity in September 2025, they alleged copying of their articles as inputs for its retrieval-augmented generation system, and PPC Land noted that the US Copyright Office's May 2025 report had found such systems less likely to be transformative when they generate summaries or abridged versions. In May 2026, CNN sued Perplexity over more than 17,000 works, arguing that copies are made at two legally distinct points, ingestion and output. Wikimedia's brief draws a similar distinction, arguing that retrieval weighs against fair use "whether RAG reproduces an original work solely as an input or in both the input and output phases."

Settlements have so far left the output side open. When Anthropic agreed to pay at least $1.5 billion to authors in September 2025, roughly $3,000 per work for about 500,000 books, the deal released no claims based on the output of AI models.

For publishers that have pursued payment for training data, the brief offers little: it would put training largely beyond copyright's reach. For those whose business rests on readers arriving at their pages, it offers more. If Judge Stein accepts the brief and adopts its framework, the legal exposure for AI search products would concentrate at the point where a user asks a question and a system decides whether to credit, link or substitute. That is also where referral traffic is won or lost. Whether the court will even read the argument is still undecided.

Timeline

Summary

Who: The Wikimedia Foundation, the non-profit that operates Wikipedia, represented by Rhett O. Millsaps II of Lex Lumina LLP and in-house counsel Stephen LaPorte, Jacob Rogers and Aly Marino. The filing is addressed to US District Judge Sidney H. Stein and concerns OpenAI, Microsoft, The New York Times and the other plaintiffs in the consolidated proceedings.

What: A motion for leave to file a 5,831-word amicus brief in support of neither party. The advocacy brief argues that scraping is not a copyright use, that pre-training ChatGPT's models is likely fair use, and that ChatGPT's retrieval-augmented generation is likely not fair use unless it complies with licence terms such as Creative Commons attribution and ShareAlike.

When: The motion was filed on October 29, 2026, under an August 7, 2026 order modified on September 24, 2026. Wikimedia announced it on October 30, 2026. Judge Stein has not yet ruled on whether to accept the brief.

Where: The US District Court for the Southern District of New York, in In re: OpenAI, Inc. Copyright Infringement Litigation, No. 25-md-3143.

Why: Wikimedia says rulings at either extreme would harm open knowledge: a narrow reading of fair use could block AI tools its volunteers rely on, while blanket fair use could let chatbots absorb Wikipedia content without credit, diverting readers, editors and donors from the site. For publishers and marketers, the brief places the legal line at retrieval, the stage at which AI answers either cite and link their sources or replace them.