The OpenAI copyright case now turns on 8.2 million Copilot chat logs

Nearly three years of discovery in the consolidated US copyright litigation against OpenAI and Microsoft landed on the public docket on 4 September, when both companies and the news publishers suing them all moved for summary judgment on the same day. The briefs were filed in the multidistrict litigation before Judge Sidney H. Stein in the Southern District of New York. They are the first time either side has had to argue from the evidentiary record rather than from predictions about what these systems would do.

What was filed

Microsoft filed two memoranda of law: one in the book authors' consolidated cases (ECF 1690, 42 pages) and one in the news publishers' consolidated cases (ECF 1698, 46 pages). The OpenAI defendants filed their own notice of motion for summary judgment in the news cases — brought by The New York Times, the Daily News plaintiffs, the Center for Investigative Reporting, The Intercept and Ziff Davis — and requested oral argument. The news publishers filed a combined 92-page cross-motion of their own, seeking summary judgment of liability rather than defending against it.

Every one of these documents is stamped Filed 09/04/26. Under a 7 August order, Microsoft and OpenAI were each held to 8,750-word opening briefs in the news cases; Microsoft's certificate puts its brief at 8,693 words. Both companies want Judge Stein to decide fair use as a matter of law and end the case before trial.

The numbers Microsoft put on the record

Microsoft's central factual argument is that regurgitation — a model reproducing chunks of a source it was trained on or grounded on — barely happens in production. The figures it cites come from the plaintiffs' own experts, which is what makes them worth reading.

• 8.2 million Copilot chat logs were produced in discovery. Microsoft is explicit that the sample was not random: the plaintiffs selected it by keywords implicating their own websites, so it is the traffic most likely to contain their work.
• 59,545 conversations. Tom Goldstein, expert for the Times and the Daily News plaintiffs, filtered those logs for conversations where the plaintiffs' domains appeared in the content Copilot used for grounding, then searched for 16-word matches between that grounding content and the response. That is the count he came back with.
• 51 instances. Emily Wenger, expert for the Center for Investigative Reporting, found that many cases of what she defines as substantial overlap with CIR work across the same 8.2 million outputs.
• 24 responses. In the books case, Dr. Shawn Shan reviewed 8.2 million Copilot conversations and found that many responses containing a 30-word match. For 202 of the 212 asserted works he found none at all.
• ~5.3 million extraction attempts. To produce 30-word regurgitations in a lab, Shan fed models verbatim book passages hundreds of words long over roughly 5.3 million attempts; Microsoft says fewer than 1% of them yielded a 30-word match.
• 1.3% of end users. Microsoft's own expert sample found that share of Copilot users employ it to research current events at all.

One conversion deserves a flag. Microsoft turns Goldstein's 59,545 into "a miniscule 0.73% rate of matches" by dividing it by the full 8.2 million logs. That is not the set Goldstein searched: he filtered first for conversations grounded on the plaintiffs' domains, and the public brief never says how large the filtered set was. The brief acknowledges it is setting "methodological disputes to the side" because, in its view, every expert's numbers converge anyway. That may well be right about the direction. But 0.73% is not a rate of anything the plaintiffs' expert actually measured, and it is the number most likely to be lifted out of the filing. The 24-response figure Microsoft renders as ".00029%" has the same shape.

The publishers are not arguing about snippets

The publishers' brief spends little time on how often a chatbot repeats a sentence, because its theory of harm is not quotation. It is substitution: that an answer engine satisfies the reader who would otherwise have clicked. They seek summary judgment of liability at five points in the pipeline — acquiring copies of their articles, training on them, using the Bing index to ground live responses, reproducing copies or derivatives in outputs, and the two companies supplying each other with copies as part of their partnership. They also seek partial summary judgment that OpenAI stripped copyright management information under the DMCA, and that statutory damages should run per article.

The evidence they lean on is internal. The brief points to Microsoft's own records of falling click-through rates for the Times' and Daily News' domains, and for Ziff Davis' domains, when Copilot answers a question rather than Bing returning links. The percentages are redacted in the public version, as are several quoted admissions from executives at both companies — a large share of the brief's most pointed passages are blacked out. One procedural note: the grounding and output claims are pressed against Microsoft only, because, the brief says, the issue is not ripe against OpenAI until a pending sanctions motion is resolved.

The documents and testimony uncovered during discovery lead to only one conclusion: Microsoft and OpenAI stole from The New York Times to make commercial products that substitute for its journalism, threaten its business, and undermine its industry.

The opt-out argument is the one to watch

Microsoft's second line of defence has less to do with copyright doctrine and more with plumbing. Since September 2023 it has published Copilot-specific meta tags: NOARCHIVE excludes a page from grounding entirely, NOCACHE limits how much of a page can be used, and IndexNow pushes the change into the Bing index quickly. The publishers knew about these controls and chose not to apply them, Microsoft argues, and under search-engine precedent that knowing choice amounts to an implied licence. "It is hard to imagine a fairer use than one a copyright holder can block at will," the brief says.

That is the argument with consequences well beyond these parties. If declining a vendor's opt-out tag creates an implied licence to ground on your content, the remedy available to a publisher becomes a header rather than a negotiation — and every AI product that reads the open web acquires an incentive to ship its own tag and treat silence as consent. The publishers' answer is that a market they were never offered cannot be waived by not opting out of it. Judge Stein has not ruled on any of this.

If you build with retrieval

Nothing is decided, so nothing changes in your stack today. But the shape of Microsoft's grounding defence is worth reading if your product does retrieval, because it rests on properties a system either has or does not: grounding copies stay inside the pipeline and are never served to the user; the response is generated rather than passed through; outputs link back to the pages relied on; and any site owner can exclude itself with a tag the system honours. Whatever the court makes of it, that is now the reference description of a defensible RAG pipeline.

The two sets of numbers also encode two different threat models, and it is worth not collapsing them. Measured on real traffic, verbatim overlap is rare. Under Shan's adversarial protocol — an operator who already holds the source text and bombards the model with surrounding context — extraction still works, at low but non-zero rates, across millions of attempts. Microsoft argues the second proves nothing about the first. For a copyright ruling that may hold. For anyone shipping a system strangers can prompt, they are separate problems with separate mitigations.

Responses to the motions are due 5 October, per the docket, and OpenAI has requested oral argument. Judge Stein can grant summary judgment to either side on fair use, split the questions, or send the case to trial.