Microsoft is bringing receipts to its copyright fight with The New York Times and book authors. In new legal filings, the company says an analysis of 8.2 million Copilot chat logs — handpicked because they were most likely to contain publishers’ content — found that just 59,545 responses shared at least 16 words with news articles used to ground the model. That’s under 1 percent. For books, the numbers are even skimpier: only 24 responses out of those same 8.2 million conversations contained at least 30 matching words, and just 10 of 212 evaluated books showed any matches at all. An expert for the Center for Investigative Reporting found 51 instances of ‘substantial overlap’ with CIR work, Microsoft says.
The context here matters. These logs weren’t random — they were filtered specifically for keywords tied to the plaintiffs’ websites. Microsoft’s argument is essentially: even when you stack the deck against us, Copilot barely reproduces anything. That’s the fair use case in a nutshell. The company contends that while LLMs train on copyrighted material, the resulting systems serve a fundamentally different purpose than the original works, and rare verbatim overlap ‘hardly undermines the transformative purpose of LLM training.’
The Times isn’t buying it. Lead counsel Ian Crosby fired back in a statement: ‘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.’ That’s a direct shot at the substitution argument — if Copilot can answer questions with NYT-quality journalism without sending readers to nytimes.com, the damage is real even if verbatim copying is rare.
Microsoft filed for summary judgment on Friday, asking the judge to end the case before trial. The publishers and authors had their claims consolidated under one judge despite their objections, and the Trump administration this week filed a statement of interest supporting OpenAI in the NYT case. If the judge denies summary judgment, discovery continues and the case grinds toward trial. For publishers watching this closely, the stakes are existential: if 1 percent regurgitation is the benchmark, what does that mean for the next generation of models that may be far more prone to reproducing training data? Microsoft’s own numbers show the problem is small today — but the legal precedent set here will outlast any single model.
