Wikimedia Amicus Brief Splits AI Training From RAG

The Wikimedia amicus brief filed in the biggest AI copyright case in New York sides with neither OpenAI nor the authors and news organisations suing it. The Wikimedia Foundation, which runs Wikipedia, argues that training a general-purpose language model on copied works is likely fair use, but that chatbots which fetch and reproduce web content at answer time should follow the licences that content carries.

The foundation announced the filing on 30 October. The court record on CourtListener shows it went in late on 29 October, US time, as docket entry 2361 in In re: OpenAI, Inc. Copyright Infringement Litigation before Judge Sidney Stein in the Southern District of New York. That case combines a dozen suits, the best known being The New York Times Company’s claim against Microsoft and OpenAI.

In brief

Wikimedia’s brief backs “neither party” and asks the judge to rule on each use of copied works separately.
It says pre-training ChatGPT’s underlying model was fair use, but copying in ChatGPT’s retrieval feature is not unless it follows licences such as Creative Commons.
The foundation has asked for leave to file; the court has not yet ruled on that request or on the case.

What the Wikimedia amicus brief asks the court to do

Strictly, Wikimedia filed a letter motion asking permission to submit an amicus brief, with the proposed brief attached. It was signed by Rhett Millsaps of Lex Lumina, with the foundation’s general counsel Stephen LaPorte and two other in-house lawyers, and it certifies 5,831 words against the court’s 6,000-word limit.

Its complaint is with both camps. “Wikimedia is concerned that the parties’ positions in this case push for overbroad and simplified applications of the fair-use doctrine,” the brief says. Fair use, it argues, should not be read so as to lock every use of a work behind proprietary terms. “Neither should fair-use doctrine serve as a green light to snub licensing terms and fragment the flow of knowledge between human authors and audiences.”

Its answer is to judge each use on its own. It relies on the Supreme Court’s 2023 ruling in Andy Warhol Foundation v. Goldsmith, which held that the same copying can be fair for one purpose and not for another. “Each distinctive use should be assessed with independent fair-use analyses,” it says.

Scraping, training and retrieval augmented generation

The brief splits the life of a chatbot into stages. First, it says, “the act of scraping should not be treated as a ‘use’ of a copyrighted work”: collecting copies is a means of acquiring them, and what matters is what is done with them afterwards. A footnote adds that scrapers should still take care not to overload websites, because “a public park can remain free and open while preventing someone from driving a container truck through it.”

Training comes next. “The use of copyrighted works to train a general purpose LLM is likely to be fair,” the brief says, because it is “likely transformative and does not inherently create market harm by substituting for the underlying works.” Wikimedia leans on its own volunteers’ experience of building machine-learning tools that fight vandalism and spot copyright problems. “It would be contrary to the purposes of copyright law to block these capabilities by labeling LLM development as categorically not fair use,” it argues.

Retrieval augmented generation, or RAG, is where the brief draws its line. A chatbot using RAG copies text from sites such as Wikipedia or the Times, “selects and pastes the most relevant passages or ‘chunks'” into the model’s context and “often substantially reproduces the original content in its generated output,” the brief says. Wikipedia’s text is published under the CC BY-SA licence, which requires attribution and that reusers share alike.

So the brief’s specific request is that the court “rule that while the pre-training of ChatGPT’s general LLM was fair use, OpenAI’s use of copies in ChatGPT’s RAG feature is not fair and must abide by the terms of the relevant licenses, such as the Creative Commons license for Wikimedia Project content, to avoid infringement.” A footnote softens that: services, ChatGPT included, that consistently cite and link to their sources are “less likely to infringe” at all.

Why the RAG line could matter more than the training fight

This is TSN’s analysis. Much of the case, and the publishers’ brief filed the same day, turns on whether training is fair use. Wikimedia concedes that point to OpenAI and moves the argument to what happens when a chatbot answers a question. That is a narrower target, and probably an easier one for a court to grasp, because retrieval copies specific passages at a specific moment rather than absorbing billions of documents into model weights.

It also offers a remedy that does not need anyone to pay. “Compliance with the provisions of Creative Commons or other copyright licenses would make RAG uses of works noninfringing,” the brief says, and a ruling requiring it “would encourage developers to build license compliance into RAG features.” For an open licence, compliance mostly means credit, a link and the same licence downstream. That is a technical fix rather than a cheque, which may make it more palatable to a judge wary of choking innovation.

The weakness is the brief’s own footnote. If a chatbot that links its sources is less likely to infringe, much of the harm Wikimedia describes could be fixed by better citations alone, and OpenAI can point to the links ChatGPT already shows. The fight over what counts as proper attribution in an AI answer has barely begun. Wikimedia’s own tensions with OpenAI predate this filing: earlier this month the foundation said it had found what it called “rogue” OpenAI agents on its sites.

What happens next

Judge Stein must first decide whether to accept the brief. He is weighing summary-judgment motions from both sides, as TSN reported when the publishers filed, and newer suits such as USA Today Co.’s case against OpenAI have asked to be treated as related to the same proceedings. Elsewhere, the Third Circuit rejected fair use for a non-generative legal research tool. TSN tracks the wider litigation on its AI copyright lawsuits hub.

The brief ends with a warning to both sides: “fair use principles should not be used to greenlight the displacement of human-produced knowledge in the name of innovation alone.” Its own test for that is simple enough to check in any chatbot answer: whether a reader is sent back to the page the words came from.