Microsoft Internally Called AI Scraping 'the Largest Theft of Labor in Human History,' Unsealed Filings Show

A set of newly unsealed court filings has exposed, in unusually direct language, what Microsoft said about AI training data when it thought the public was not listening. According to TechCrunch, the unredacted documents show the company internally used the word "theft" to describe OpenAI's data practices.

The filings lay out three linked facts: both companies scraped paywalled Times content, used that material to build training datasets, and warned internally that the practice would gut publishers. The gap between the private assessment and the public posture is the sharpest thing in the record.

The most quoted line goes further than a legal complaint. AI scraping was described internally as "the largest theft of labor in human history," a framing that moves the argument from copyright compliance toward labor and value distribution, and shows the risk was understood inside the company rather than discovered later.

For the litigation, documents like these matter because they speak to knowledge and intent. Internal warnings are close to ideal evidence on that point, because a publisher does not have to prove scraping was technically possible, only that the defendant understood what it was doing and could foresee the harm.

The Times detail matters for a second reason. Paywalls are the core of many news organizations' business model, and if content behind them was folded into training data, the bargaining position of publishers in licensing talks narrows considerably.

The broader pattern is familiar: AI developers and publishers have spent years operating in a gray zone of license later, comply later. What the filings suggest is that the trade-offs were not invisible inside these companies. They were priced in and accepted.

Watch two things next. First, how widely other pending copyright cases cite these documents. Second, whether Microsoft and OpenAI adjust their public language in response. Either way, the training-data problem is getting harder to describe as a question of technical neutrality.