Companies Are Selling Workers’ Private Messages and Emails as AI Training Data

A growing number of defunct and struggling companies are monetizing their internal archives — including employee emails, Slack messages, and private workplace communications — by selling the data to artificial intelligence firms hungry for high-quality training material. The practice has raised urgent questions about consent, privacy law, and the fate of personal information workers assumed would remain confidential.

The emerging marketplace for corporate communications data reflects a fundamental shift in how artificial intelligence developers are sourcing training material. As large language models such as those produced by OpenAI, Anthropic, Google, and Meta have scaled up, they have largely consumed the open internet — books, Wikipedia, news archives, Reddit threads, and publicly available code repositories. Industry researchers, including a widely cited study from the research group Epoch AI, have projected that the stock of high-quality public text data could be effectively exhausted between 2026 and 2032. That scarcity has driven a race for previously untapped reservoirs of human writing, and few sources are considered more valuable than authentic workplace dialogue, which captures natural reasoning, problem-solving, negotiation, and domain-specific expertise that synthetic data cannot easily replicate.

Bankruptcy proceedings have become a particularly fertile ground for these transactions. Under U.S. bankruptcy law, a debtor’s data holdings are generally considered part of the estate and can be liquidated to satisfy creditors, subject to certain privacy restrictions established after the 2000 Toysmart case, in which the Federal Trade Commission intervened to block the sale of customer data that had been collected under explicit privacy promises. However, employee communications — as opposed to consumer data — occupy a far murkier legal terrain. Most U.S. workers sign agreements acknowledging that anything sent through company servers belongs to the employer, and courts have repeatedly upheld that workplace privacy expectations are minimal. That legal framework, designed decades ago for surveillance and liability purposes, was never contemplated as a license to feed personal exchanges into machine learning pipelines.

Regulators in both the United States and Europe are beginning to take notice. The European Union’s General Data Protection Regulation imposes strict limits on processing personal data for purposes not disclosed at the time of collection, and California’s Consumer Privacy Act and related state laws grant individuals rights to know how their data is used. Several state attorneys general have signaled interest in examining whether the sale of employee communications to AI developers constitutes a material change in purpose that requires renewed consent. Legislation introduced in Congress, including proposals tied to broader AI accountability frameworks, could eventually require disclosure when personal communications are included in training corpora, though no federal statute currently mandates it. Labor advocates have also raised concerns that workers displaced by the very AI systems trained on their words have no compensation mechanism — a dynamic some legal scholars have compared to unpaid data labor.

What to watch next is whether courts, regulators, or Congress move to close the gap between 20th-century workplace privacy doctrine and 21st-century AI economics. Pending litigation over training data — including high-profile copyright cases brought by authors, news organizations, and artists — may indirectly shape how employee communications are treated, as judges grapple with questions of consent, derivative use, and fair compensation. In the meantime, employment attorneys are advising workers to assume that anything typed on a company device could one day surface inside a model’s parameters, while corporate boards are beginning to weigh whether the short-term revenue from data sales is worth the reputational and legal exposure such transactions may invite.