Microsoft says Copilot rarely repeats New York Times text in chats

In short: Microsoft says its Copilot chatbot almost never repeats long chunks of text from publishers and book authors, based on an analysis of millions of chat logs shared in a lawsuit.

What happened

Microsoft filed new court papers in its copyright dispute with publishers, including The New York Times, and with book authors. The plaintiffs say Microsoft and OpenAI used their writing to build AI products and that the chatbots can sometimes spit back copyrighted text.

As part of the evidence-gathering process in the case, Microsoft says it provided 8.2 million Copilot chat logs to an expert hired by news publishers. Microsoft says these chats were picked because they included keywords that pointed to the publishers’ websites, meaning they were more likely than average to include publisher text.

Microsoft says the review found 59,545 chats that had at least 16 words in common with news content used to “ground” Copilot. In simple terms, “ground” means the chatbot checks or leans on specific source material (like looking things up in a reference binder). Microsoft also says an expert for the Center for Investigative Reporting found 51 cases of “substantial overlap” with its work. In the authors’ case, Microsoft says an expert found 24 chatbot responses with at least 30 matching words, and only 10 of 212 books checked had any matches.

Microsoft argues these numbers support its claim that training AI on copyrighted works should count as “fair use,” because the AI is used for a different purpose than the original articles or books. Microsoft is asking the judge for summary judgment, which would end the case early without a full trial.

Why it matters

This case could help decide what rules AI companies must follow when they use articles and books to build chatbots. For regular people, it could affect what chat tools can do, how reliable they are, and whether news and book creators get paid when their work is used.

Source: The Verge AI