Copyright and artificial intelligence: the legal fight moves to the outputs

On 4 September 2026 the deadline passed for filing motions for summary judgment — the request by which a party asks the judge to decide without a trial because the relevant facts are not in dispute — in the consolidated proceeding «In Re: OpenAI, Inc. Copyright Infringement Litigation» before federal judge Sidney H. Stein, at the U.S. District Court for the Southern District of New York. The public docket on CourtListener lists the motions filed by The New York Times Company and the expert analyses supporting the defence, signed by Chris Callison-Burch. Microsoft moved for summary judgment on the argument that training on journalistic texts and books amounts to transformative use protected by fair use, stating that Copilot "does not replace their protected expression". On the other side, OpenAI insists in its own brief that "Nobody owns facts, just as nobody owns language". The publisher plaintiffs ask instead that fair use be ruled out both for the way the paywalled articles were obtained and for the training carried out on the copies, while the publishers' lawyer Steven Lieberman describes the evidence filed under seal in these terms: "The stuff we were forced to file under seal is scorchingly hot. Not just a smoking gun; they're a smoking bazooka." What the sealed material actually contains cannot be publicly verified: Lieberman's remains a party's assessment of documents that nobody outside the case has been able to read.

The defence argument tries to shift the centre of gravity of the analysis away from the material absorbed during training and towards how often the models actually generate protected content. According to Microsoft's brief, an expert appointed by the publishers examined roughly 8.2 million Copilot conversation logs selected through keywords tied to the websites of the news organisations involved — that is, by construction, the subset most likely to contain their works: the percentages that follow do not describe Copilot's traffic as a whole. Within that sample, 59,545 records (about 0.7%) shared at least 16 words with the articles, while on the book side there are 24 responses with at least 30 matching words (0.00029% of the conversations analysed), relating to 10 of the 212 books at issue and leaving 202 volumes with no matches at all. Still according to Microsoft's brief, an expert for the Center for Investigative Reporting identified 51 instances of substantial overlap with CIR works in the same dataset. In the background sit the Sensor Tower estimates reported by the New York Daily News, which point to more than a billion users for ChatGPT in June and a cost of $6,800 to generate a million 500-word articles in journalistic style.

We have not read the full PDF of Microsoft's brief: at the time of verification the public docket on CourtListener did not surface that specific motion among the searchable entries, and part of the material is filed under seal. The figures given here come from the filed text as reported by the New York Daily News, The Next Web and The Verge, outlets that had access to it. The measurement thresholds (16 words for news, 30 for books) are likewise the ones applied in the defence's own analysis, and no technical challenge from the plaintiffs on method and parameters has become public yet. The judge has made no decision, oppositions are due by 9 October 2026, and no hearing date has been set.

Whether the right measure is what comes out in the output or what went in to train the model is precisely what judge Stein will have to rule on: the two briefs ask for fair use to be applied to two different questions.

— Olya