Meta told US-based employees on 21 April that a new tool, the Model Capability Initiative (MCI), will run on their work computers and log mouse movements, clicks, keystrokes, and intermittent screenshots, to be used as training data for the company's computer-use agents. The programme covers Meta staff only, not users of Facebook or Instagram. The announcement landed twenty-nine days before a 10% global layoff round that starts on 20 May.
The surveillance-plus-layoffs angle is the one that travels on social media. The more useful prediction is that every other major AI lab is about to do the same thing to its own employees, for the same reason Meta just did.
Why every AI lab will follow
Frontier LLM training has largely exhausted public text and image data. The bottleneck for the next generation of products (the "agents" that nearly every lab is now shipping) is agentic data: sequences of (screen state, action, outcome) showing a competent human operating real software. This data essentially does not exist on the open web.
The available sources are narrow. Synthetic UI environments are cheap and scale arbitrarily but transfer badly to real apps. Crowdsourced demonstrations through Scale AI, Surge, or Mercor are expensive per trace, narrow in task coverage, and gameable by labellers who learn what the requester wants. Reinforcement learning in simulated environments works in narrow domains and fails in messy enterprise software. The fourth and cheapest source is observation of a captive workforce the lab fully controls: its own employees.
This is why MCI is not a Meta-specific story. OpenAI, Anthropic, Google DeepMind, xAI, and Mistral all need exactly the same agentic data and have exactly the same in-house labour to harvest it from. The legal cover is identical: in the US, an employer can monitor work-app usage on company-issued hardware without further consent. The technical infrastructure is trivial. The only thing Meta did first was announce it out loud.
The Bosworth loop Meta described
Meta CTO Andrew Bosworth told staff in a separate memo that the goal was for agents to "automatically see where we felt the need to intervene so they can be better next time." That is reinforcement learning from human feedback on workplace tasks, with employees as unpaid preference labellers. The pipeline runs: agent attempts a task, employee corrects or completes it themselves, the correction becomes a preference pair, the model fine-tunes toward the human action.
This loop is substantially more powerful than offline demonstration capture because the agent learns exactly the cases where it currently fails. It is also dependent on the workforce that produces the corrections. Once the loop is mature, that workforce is no longer load-bearing. The 29-day gap between the MCI announcement and the start of the layoff round is not subtle.
What to watch
Internal-tooling memos at OpenAI, Anthropic, Google DeepMind, xAI, and Mistral. Equivalent programmes are almost certainly already running at most of them, branded as "agent evals," "internal dogfooding," or "productivity instrumentation." Most will never be announced publicly. The ones that leak will read very similarly to MCI.
Hiring patterns. Watch for AI labs increasing headcount in the adjacencies they want their agents to handle (sales, recruiting, legal review, finance ops, customer support). A spike in non-engineering hires at a frontier lab in 2026 is most plausibly explained as buying agentic training data, with a finite human shelf life.
Regulation. The EU AI Act and works-council consultation rules in Germany and France will force opt-in or collective bargaining for any equivalent programme in EU jurisdictions. Illinois's BIPA arguably already covers keystroke dynamics as biometric data. Litigation will test this within two years.
Agent benchmark results in late 2026. If labs that captured workplace interaction data internally outperform those without on WebArena, OSWorld, and the newer enterprise-task evaluations, the practice becomes universal and irreversible across the industry. If they do not, the surveillance happened anyway.
