Meta's leaked Model Capability Initiative reveals employee surveillance to train AI, amid 8,000 layoffs
The Platformer-sourced disclosure of Meta's Model Capability Initiative is the most pointed workplace-AI-ethics story of the week. Per the reporting, MCI instruments employees' work across Gmail, Slack, GitHub, Google Workspace and VS Code — including keystroke- and mouse-level activity — and feeds that captured behavior into training data for Meta's internal AI systems. Crucially, employees reportedly have no opt-out. The program landed in the same news cycle as approximately 8,000 layoffs at Meta, magnifying the optics.
The legal context Yale law professor Ifeoma Ajunwa supplied is the part that should worry every knowledge-work company watching: there are essentially no US federal limits on workplace surveillance of this kind. State-level rules (notably in New York and Illinois) require disclosure, but not consent or opt-out. Meta's program, if accurately described, is legal — and that means it is replicable. The developer backlash that lit up X and developer Slacks framed MCI as 'invasive' and a betrayal of the implicit trust assumed in modern software-engineering tooling.
The broader pattern matters. As frontier labs increasingly compete on having proprietary training data — the public web is saturated and synthetic data has known limits — capturing high-skill professional work is the next data frontier. Source D's Human Archive funding story (worker-generated, consented datasets) is the explicitly-consented counterpart to what MCI does without consent. The two stories together describe the labor-data market that will define 2026-27.
For Meta, the immediate consequences are reputational (developer-recruiting damage at exactly the wrong moment), regulatory (expect state AGs and EU DPAs to ask questions), and competitive (Anthropic and Google both have AI-ethics positioning they can now sharpen by contrast). For the rest of the industry, MCI is going to be the case study every legal and HR team uses when writing the next round of acceptable-use policies. Watch for whether any large customer publicly demands assurances that their data-handling vendors do not run similar programs.