Meta spied on its employees to train the AI replacing them

Three weeks between the audio and the layoff emails

April 30, 2026, Meta all-hands. Mark Zuckerberg calmly walks employees through the logic of the Model Capability Initiative: capture what they do at their workstation to train AI models. The quote from the audio released on May 19 by the More Perfect Union collective is unambiguous: "We are using this to feed a very large amount of content into the AI model, so that way it can learn how smart people use computers to accomplish tasks."

On May 19, the day the audio dropped, 8,000 Meta employees got their termination email. Roughly 10% of headcount. CTO Andrew Bosworth justified it internally the same week: the cuts fund the $125 to $145 billion in 2026 AI capex, a figure Meta raised in late April and which knocked its stock down 6% in after-hours trading.

Three weeks between the audio and the emails. Between the capture of work and its automation, the same sequence.

What the Model Capability Initiative collects

According to sources who reviewed the internal documentation (The Register, Common Dreams, eWeek), the tool ships preinstalled on company-issued laptops. It records keystrokes, clicks and mouse movements, and takes periodic screenshots. The monitored application scope is limited to an "approved" list including Gmail, GChat, the VSCode code editor, and Metamate, Meta's internal AI assistant.

On opt-out, Bosworth's answer fits in one sentence: "There is no option to opt out on your corporate laptop." No arbitration, no gray area, no role-based exemption. If you work at Meta on Meta hardware, your workflows feed the dataset.

In the audio, Zuckerberg stresses that "no human is looking at or watching what people are doing" and that content is "stripped out as much as possible." The program is allegedly not designed for surveillance or performance tracking. That nuance carries real legal weight in the US. On the technical side, the line is thinner than it sounds: this is still a system capturing keystrokes and screenshots without individual consent from the people on screen.

Why this data, and not something else

The industrial timing is precise. The big labs are racing to ship agents: AI systems that can execute end-to-end the kind of task a skilled employee handles across five tools, ten tabs and three conversations. Training those agents requires granular examples of real workflows. Not manuals, not job descriptions, not synthetic prompts: click-by-click data, context, app switches at the moment a decision turns.

That is exactly what MCI produces. Zuckerberg's justification in the audio, on picking Meta employees rather than external contractors, completes the picture: "the average intelligence of the people who are at this company is significantly higher than the average set of people that you can get to do tasks if you're working through these contractors." Dataset quality scales with operator quality. So you record the best operators available, i.e., the in-house engineers.

The dataset → model → budget cut sequence becomes legible. The $145 billion in capex funds the data centers where the models trained on the work of the 8,000 employees pushed out on May 19 will run. The accounting on the operation is cold: employees deliver training data, the company converts that data into automation, and automation pays back part of the capex by deleting their roles.

The petition, the flyers, and Meta's silence

The New York Times had already run a piece on May 8 about an internal "revolt" around the program. The release of the audio on May 19 turned a contested HR memo into recorded evidence.

In the days that followed, employees taped flyers in meeting rooms, vending machines and restrooms across Meta offices. A formal petition went up, citing the protections of the US National Labor Relations Act. It has now passed 1,000 signatures, and a unionization push started in the UK.

Meta's public response has been tight: no statement, no confirmation, no formal denial. The audio's authenticity has not been officially validated. The Week flags the unresolved chain of custody; The Register writes outright that "Meta has not confirmed the authenticity." Even setting the audio entirely aside, what remains is the May 8 NYT coverage, the Bosworth memo confirmed by AI research teams, the 1,000-plus signatures and the flyer photos. The factual substance does not depend on the .mp3 file.

The European regulatory gap

That leaves the question of direct interest to tech companies operating in France. The MCI as described would be illegal under three overlapping regimes.

Article L1222-4 of the French Labor Code requires employers to inform workers, before deployment, of any system capable of monitoring their activity. The MCI, without prior notice and without opt-out, breaches that obligation head-on. The CNIL has already fined comparable systems: keyloggers, periodic screen captures, inactivity detection through mouse movement. A 2024 precedent set a 40,000 euro fine on a system that flagged the absence of keystrokes over 3-to-15-minute windows, deemed "particularly intrusive."

Then on August 2, 2026, the AI Act enters its full application phase. Regulation (EU) 2024/1689 classifies AI systems used in HR and performance tracking as "high-risk." That triggers a mandatory information and consultation obligation with the works council before any deployment, with penalties up to 35 million euros or 7% of global turnover. Case law is already ahead: in July 2025, the Créteil judicial court issued a summary order suspending AI tools in a trade press company, finding that "artificial intelligence is a new technology" under article L.2312-8 of the Labor Code.

Multiple sources suggest Meta is already excluding its European employees from MCI on GDPR grounds. The company has not confirmed this, and the claim should be treated with caution. If it holds up, it tells you something useful: Meta knows the program is defensible in the US and indefensible in Europe.

The window closing on August 2 is the one in which European Big Tech subsidiaries can still attempt this kind of silent capture, betting on the enforcement lag. After that, the expected legal cost shifts the calculus.

The question has changed

Public debate on AI long organized itself around one line: AI will replace jobs. The Meta sequence documents a more precise mechanic. Jobs are not replaced wholesale by a system arriving from outside. They are first converted into training data, quietly, on employer-issued machines. Then they are deleted to fund the infrastructure that will run the model trained on their own work.

That loop, not just the job losses, is what the 1,000 signatures on the Meta petition are trying to make legible. And that loop is what the AI Act, from August 2 onward, forces companies to bring before the works council before any deployment in Europe.