AI Exec: We May Have Pulled Off "The Largest Theft of Labor in Human History"

Lawsuit documents from several news organizations against major AI firms reveal internal admissions that large language models were trained by systematically scraping copyrighted journalism and other high-quality content. Executives acknowledged these models substitute for the labor that creates culture, and company research found AI-driven answers led to about 90 percent fewer clicks to news articles and at least a 30 percent drop in referral traffic after AI “overviews” were introduced. Training sets deliberately concentrated on news and included paywalled material and datasets that prohibited commercial use; ownership metadata and copyright notices were stripped from the corpora. After litigation began, a targeted filter was added that suppressed outputs tied to suing outlets, hampering plaintiffs’ ability to detect copying. An internal test of 20 million ChatGPT responses turned up roughly 400,000 identical 25-word overlaps with one plaintiff’s articles, a level comparable to other major publishers.

The filings argue these practices amount to an unprecedented theft of creators’ labor that fuels massive low-quality content while eroding the economic foundation for journalism, creating a self-reinforcing “doom loop.” Executives flagged the problem even as some celebrated models’ ability to replicate paywalled text or discussed workarounds, and corporate debates acknowledged both legal constraints and the logic of compensating creators - yet implementation favored rapid growth over restraint. The core claim is a classic prisoner’s dilemma: the industry as a whole would benefit from paying for content, but each firm has incentives to take it for free. The documents therefore make the case that relying on firms to police themselves is futile and that external rules are needed to protect creators and stabilize the internet’s information ecosystem.