Internal documents newly made public in a three-year‐old lawsuit show that Microsoft and OpenAI executives privately condemned their own practices around content scraping as deeply harmful to news publishers and journalists. The unredacted details, derived primarily from New York Times legal briefs, reveal the companies grappling with accusations ranging from “theft” to “existential threat” to the livelihoods of creators whose work powers AI models.
Genes of the Case: What the Documents Expose
The lawsuit, initiated by a major news publisher, alleges that Microsoft and OpenAI built their foundational AI models using mass scraping of copyrighted content—including material hidden behind paywalls—without proper licensing. Documents show admissions from Microsoft leadership that their AI systems, particularly the “Copilot answer engine,” have driven click‐through rates for the publisher so low—by as much as 93%—that it threatens both its web traffic and financial foundation. Internal memos warn of a looming “doom loop” in which the very content that feeds AI models could be undermined by the models themselves.
One particularly stirring internal characterization came from a senior Microsoft scientist in early 2024, who described the scraping efforts as “the largest theft of labor in human history.” Meanwhile, OpenAI leaders reportedly acknowledged that the models produce outputs that are increasingly substitutive of news content, especially when accessed through chatbots instead of visiting publisher websites. There were also explicit internal discussions about efforts to evade paywalls, stripping copyright notices from source material, and building datasets disproportionately weighted toward news‐site articles.
Scale, Strategy, and a Legal Reckoning
The filings quantify the scale: one dataset dubbed “mid‐training” contains over 91,600 works from the plaintiff publishers, while another, assembled under a project codenamed Mango, includes more than 160,900 unique pieces. The companies also allegedly pulled millions of articles from public sources like Common Crawl, and integrated content from news sites in ways that appear to sidestep licensing norms. Senior leadership repeatedly conceded under oath that models had been trained on paywalled material, and that the models’ outputs often undercut traffic—and revenue—for original publishers.
Legally, the battle centers on whether this sort of content scraping qualifies as “fair use.” While many courts have trended toward more generous interpretations of fair use for training large language models, the new admissions in these documents clash with key fair-use criteria—especially the idea that the use should not compete with or harm the market for the original work.
Executives from both Microsoft and OpenAI declined to comment in response to the newly revealed filings. The case continues, and it may well set landmark precedents for how AI models can use copyrighted text going forward.
Bridging technology, law and ethics, this case underscores the growing tension between innovation and intellectual property rights. As AI models rise in capability and ubiquity—and increasingly rely on third-party content to learn—the need for clear agreements with content creators becomes more urgent. What this means in practice: regulators and publishers will likely push for licensing standards, transparency, and possibly statutory reforms. The outcome of this case could reshape how AI is built, who gets compensated, and what kind of content is available freely in the digital ecosystem.
