Leaked Audio: Meta Trained AI on Employees Gmail, Code, and Chats — Then Laid Off 8,000

A leaked audio recording obtained by More Perfect Union has captured Meta CEO Mark Zuckerberg telling staff at an April 30 all-hands meeting that the company has been quietly monitoring employee activity across Gmail, GChat, internal AI assistant Metamate, and the VSCode coding environment to harvest training data for its frontier AI models. The recording surfaced on May 19 — the day before roughly 8,000 employees began receiving layoff notices.

Internally, the program is known as the Model Capability Initiative. According to the leaked audio and accompanying reporting, it logs mouse clicks, keystrokes, and occasional screenshots from an approved list of applications used by full-time staff. "The AI models learn from watching really smart people do things," Zuckerberg said on the tape, arguing that the average intelligence of Meta employees is "significantly higher than the average set of people that you can get to do tasks" through traditional data-labeling contractors.

The layoffs themselves rolled out by time zone starting at 4:00 AM Singapore time on May 20, working through European and US offices over the following hours. Meta says about 7,000 of the displaced workers are being reassigned into AI roles, with around 2,000 already integrated into the company's Applied AI and Engineering team. When employees asked Meta CTO Andrew Bosworth whether they could opt out of the tracking, his one-sentence answer, according to people in the room, was that there is no opt-out on a corporate device.

Backlash was immediate. More than 1,000 employees have signed an internal petition demanding the company halt the Model Capability Initiative, and flyers urging colleagues to refuse to participate appeared on office walls in Menlo Park the same morning the layoff emails went out. Labor lawyers quoted by Common Dreams argue that the combination of pervasive workplace surveillance and same-week layoffs may invite scrutiny from the National Labor Relations Board and from European data-protection regulators, given the scope of monitoring across Gmail-based corporate accounts.

The episode arrives at an awkward moment for Meta, which has spent the past year positioning its Superintelligence Labs and the new Muse Spark model as the centerpiece of a roughly 45 billion AI capex push. Rivals at OpenAI and Anthropic rely heavily on outside contractors and licensed corpora to teach agentic behavior; Zuckerberg's pitch on the tape — that employees themselves are the best source — turns Meta's own workforce into a training corpus, and now into a public-relations problem.