Microsoft told a federal court that 8.2 million Copilot logs contain 24 matching responses.

In a September 4 summary judgment filing in Manhattan, Microsoft argued training on books and news is fair use and backed it with discovery numbers: publishers' own expert, given 8.2 million conversation logs, found 24 responses containing at least 30 consecutive words matching source texts, and found matches for only 10 of 212 assessed books. It lands the same week the DOJ took the government's first position in AI training copyright litigation (The Verge).