Microsoft exec called AI scraping ‘the largest theft of labor in human history,’ new unredacted filings reveal

Unsealed portions of court filings in The New York Times’ copyright action against OpenAI and Microsoft present internal communications and documents that portray the companies’ content-acquisition practices for model training as deeply problematic. According to the new material, executives and researchers within Microsoft and OpenAI used language underscoring the moral and commercial stakes of large-scale scraping of news publishers’ work, with at least one Microsoft researcher characterizing the scope of copying as unprecedented theft. The filings allege concrete steps by the companies to acquire publisher material, including scraping content via the Bing index, creating datasets such as WebText and WebText2 heavily weighted toward news articles, and harvesting millions of pages from Common Crawl. They also describe purported efforts to bypass paywalls and to remove copyright notices from material before it was fed into training datasets, and they claim the two firms exchanged substantial training data through initiatives called Project Taxi and Project Mango. The documents cited in the filing point to measurable commercial effects: internal Microsoft analysis attributed dramatic drops in click-through rates to nytimes.com when users received Copilot answers instead of traditional search results. Company presenters warned that such substitution could damage both publishers’ revenues and the quality of the web content supply that underpins foundation-model businesses. OpenAI and Microsoft executives are also quoted as acknowledging that models are “substitutive” for publishers’ work and could pose an existential risk to news organizations. The newly disclosed excerpts challenge key elements of the companies’ fair-use defense in the litigation by documenting both the scale of copyrighted content used and internal recognition that models can supplant users’ visits to original sources. The filings state specific counts of allegedly copied works in training sets and describe internal awareness of paywall circumvention methods. OpenAI and Microsoft did not provide comment on the newly unsealed material in the reporting; many underlying exhibits remain sealed and some quoted lines in the filings appear without their original context.