New court filings in the three-year-old lawsuit between The New York Times, OpenAI, and Microsoft reveal internal admissions regarding AI training practices. Documents suggest that companies bypassed paywalls and removed copyright notices to build training datasets through mass scraping.
Internal Microsoft communications from January 2024 describe the impact of the Copilot search engine on The New York Times as a doom loop. Data indicates that click-through rates for the publisher dropped by up to 93% compared to traditional Bing search results.
OpenAI and Microsoft allegedly utilized specific initiatives, Project Taxi and Project Mango, to exchange training data. One dataset alone contained at least 160,903 unique works from news publishers, while another Common Crawl-derived set included over 2 million documents from nytimes.com.
Microsoft CEO Satya Nadella testified that paywalled content should be licensed for training purposes. Meanwhile, internal memos from Microsoft’s Director of Applied Science, Brent Hecht, characterized the scraping of labor as the largest theft in human history.
