Microsoft says virtually nobody was grabbing NYT articles through its chatbot

The copyright fight against Microsoft and OpenAI is moving from rhetoric into spreadsheet warfare. Microsoft says it turned over 8.2 million Copilot chat logs—selected because they hit keywords tied to news publishers' sites—and that the resulting expert work shows vanishingly little regurgitation of the works at issue.

That claim lands as Microsoft asks for summary judgment. If a judge buys the "transformative training, rare output" frame, early dismissal becomes plausible. If not, the consolidated publishers-and-authors case keeps grinding toward trial, with substitution and market harm still on the table.

Microsoft's filing is a classic defense move: flood the record with volume, then argue the plaintiff's nightmare scenario almost never happens. Fewer than 1 percent of those 8.2 million logs, it says, shared at least 16 words with grounding news content—59,545 chats. An Authors Guild–side expert, per Microsoft, found only 24 responses with at least 30 matching words across the same pile, and matches in only 10 of 212 books evaluated. For the Center for Investigative Reporting, Microsoft cites an expert finding 51 instances of "substantial overlap."

Those numbers are designed to sound like proof that Copilot is not a substitute newspaper. They are also carefully scoped. The logs were the ones *most* likely to touch publisher material, Microsoft says—which makes the low hit rate rhetorically powerful, and also means the sample is not a random window into all Copilot use. Sixteen-word and thirty-word thresholds are blunt instruments; publishers care about whether a chat replaces a subscription session, not whether a string-match algorithm lights up.

The New York Times' reply is blunt and unsurprising: discovery supposedly shows theft into commercial products that substitute for journalism. That is the real clash of theories. Microsoft wants the court to treat occasional overlap as noise around a transformative purpose. The Times wants the court to treat training-plus-product as a business that free-rides on reporting and then competes with it. Frequency of verbatim dumps is relevant to one story and almost beside the point for the other.

The likelier read is that summary judgment is a stretch if the case turns on market substitution rather than copy-paste counts. Judges can accept that LLMs rarely spit out full articles and still find a genuine dispute about whether the systems are built on copyrighted works in a way the statute does not bless. Microsoft's numbers help on regurgitation; they do less work on the training-data theory that has always been the publishers' core grievance.

One more pressure point sits outside the expert tables: the Trump administration filed a statement of interest supporting OpenAI in the Times matter this week. Political wind does not decide fair use, but it does change how hard parties push settlement versus principle. Expect Microsoft to keep selling "virtually nobody," and expect publishers to keep answering that the product is the problem even when the paste is rare.

News publishers and book authors sued Microsoft and OpenAI over training and product use of their works; the claims were consolidated before one judge over plaintiff objections. Microsoft's Friday filing is part of its push for an early end via summary judgment rather than a full trial on fair use and substitution.