Microsoft Leaks: AI Training as “The Greatest Theft in History

Internal Microsoft documents unsealed in The New York Times copyright lawsuit reveal that company leaders and applied scientists, including Director of Applied Sciences Brent Hecht, privately characterized large language model web scraping as “the largest theft of labor in human history” and a self-destructive “doomsday loop.”

Inside the Internal Dissent and the “Doomsday Loop”

The disclosures offer a rare look at how major tech firms internally evaluated their own data-ingestion pipelines while scaling generative AI products. According to documents surfaced in the ongoing litigation, Brent Hecht argued that training machine learning models on published work without payment constituted an “astounding theft of unprecedented proportions.” Another internal presentation warned that Microsoft’s AI-driven content strategy triggered a “doomsday loop” that would simultaneously degrade model performance and the entire web.

Engineers and product leads mapped out a grim ecosystem dynamic. As summarized in the unsealed files, large language models function as “a product that destroys its own supply chain.” When a chatbot answers a user’s query directly inside a search interface, it eliminates the user’s incentive to visit the originating website. This traffic drop means less revenue for the websites that produce the text, leading to fewer of them surviving and a subsequent reduction in the volume of fresh human-generated text available for future model iterations.

Microsoft’s own quantitative tracking substantiated the warning. Data cited in the court filings showed that Copilot’s response engine drove click-through rates to The New York Times website down by as much as 93 percent compared to regular search.

Beyond the systemic feedback loop, the documents quantify the sheer scale of ingested publisher content. OpenAI’s mid-stage training sets reportedly included more than 91,000 copies of works from The New York Times, the Daily News, and the Center for Investigative Reporting. A separate Common Crawl extract pulled more than two million documents exclusively from nytimes.com, while an internal collection designated as Project Mango gathered at least 160,000 unique works from news publishers. Technical notes within the files detail how paywalls were bypassed and copyright notices were removed prior to feeding text into the models.

Leadership Warnings and Legal Ramifications

The internal alarm bells extended directly to OpenAI’s product leadership. Nick Turley, who runs ChatGPT, acknowledged in internal correspondence that the resulting products were “largely substitutive” for journalism, terming the shift an “existential threat” to publishers. This assessment mirrors the warnings articulated by Hecht under a different job title.

The release of these communications carries significant weight for the ongoing legal battles over fair use. Microsoft has publicly distanced itself from the candid terminology found in the files, maintaining in court that its operations align with copyright law. However, establishing a valid fair-use defense depends partly on what a company knew and intended. Internal records explicitly naming the extraction process as theft and recognizing the destruction of the publishing ecosystem complicate corporate efforts to frame the data collection as benign.

The unsealed files paint a picture of an industry proceeding with massive data scraping operations despite possessing internal telemetry and engineering consensus that the underlying mechanism undermined the very creators who built the open web.