Unredacted filings show 91,692 NYT articles in OpenAI training data

Unredacted court filings show OpenAI's mid-training data held more than 91,692 copies of New York Times and Daily News articles, TechCrunch reported September 17. The filings also include internal Microsoft data showing NYT click-through rates fell as much as 93% under Microsoft Copilot, compared to traditional Bing search. The number gives publishers a concrete measure of lost traffic tied to a named product, beyond the raw count of copied articles.

A January 2023 memo from Brent Hecht, Microsoft's director of applied science, calls the scraping "the largest theft of labor in human history." The filings also allege employees bypassed paywalls and stripped copyright notices before the material went into training data.

Neither Microsoft nor OpenAI responded to requests for comment on the filings.

What's still missing

TechCrunch's report doesn't say whether the 91,692 figure covers one training run or is cumulative across model versions. It also doesn't specify which OpenAI models were trained with that data.