When News Became Training Data: What the Unsealed Filings Say

The unsealed filings in the New York civil case have revealed internal messages and documents. The filings say OpenAI's mid-training datasets held 91,692 copies of works from the Times, the Daily News and the Center for Investigative Reporting. They also describe a Common Crawl-derived set that included more than two million nytimes.com documents and note that Microsoft and OpenAI moved training data through projects like Project Mango and Project Taxi.

News plaintiffs said the materials make clear that the products can be substitutes for original reporting, while OpenAI head Nick Turley warned the tools posed an existential threat and wrote that the products are largely substitutive, period. The filings say OpenAI researchers discussed plans to bypass paywalls and that some training data had copyright notices removed. Microsoft presentations described a doom loop that would hurt publishers and the web, and Microsoft data has been cited showing that Copilot answers cut click-through rates to nytimes.com by as much as 93%. Project Mango is said to contain at least 160,903 unique works, and company leaders including Satya Nadella have been deposed about licensing paywalled content.