Meta Trained AI on Employee Activity Before Laying Them Off

An audio recording from a Meta all-hands meeting, surfaced by eWeek, captures Mark Zuckerberg explaining that the company has been monitoring employee activity across Gmail, internal chat tool Metamate, GChat and VSCode, and using the captured behaviour to train AI systems. The reasoning given in the recording is blunt. The model "learns from watching really smart people do things," and Meta engineers are better training subjects than third-party contractors.

The recording is dated 30 April. The first details of the surveillance programme reached the press in the same window that 8,000 Meta employees received layoff notices, with around 7,000 others reassigned into AI-focused teams. CNBC reported that the cuts touch roughly a tenth of the workforce.

Surveillance, then severance

The juxtaposition is hard to ignore. Engineers whose code and correspondence had been quietly feeding model training were then told the model would be doing more of their work going forward. In an internal memo summarised by CNBC, Zuckerberg framed the cuts as a response to an environment where "success isn't a given" and where AI capability decides which companies stay competitive.

Meta has long used internal data for product work. What is unusual here is the framing of human staff as raw material for the systems that will replace some of them. Eastern Herald reports that further reductions are expected later in the year.

Open questions on consent

Meta has not publicly addressed how staff were notified that their email, code commits and chat traffic were being used as training data, or what consent process applied. Employees joining a Big Tech company generally sign expansive monitoring acknowledgements as part of onboarding, so the legal ground is probably solid. The cultural ground is shakier.

For workers in adjacent companies, the story is a useful prompt. If a competitor's internal monitoring extends to keystroke-level capture of engineering work, "we don't sell your data" no longer fully describes the surveillance picture. The relevant question is what gets fed into the training pipeline, and whether the people generating that data know about it.