← Back to feed

Meta is installing new surveillance software on US-based employee computers as part of an initiative to build AI agents capable of performing work autonomously, according to internal memos obtained by Reuters and shared within Meta’s SuperIntelligence Labs team on Tuesday.

The tool, designated Model Capability Initiative (MCI), captures mouse movements, clicks, and keystrokes across work-related apps and websites. It also takes periodic screen snapshots. The stated purpose is to help AI models learn to replicate human-computer interactions that current systems struggle with — selecting from dropdown menus, executing keyboard shortcuts, navigating complex UI flows.

“This is where all Meta employees can help our models get better simply by doing their daily work,” one memo states.

The Agent Transformation Accelerator

MCI sits within a broader internal programme CTO Andrew Bosworth rebranded this week as the Agent Transformation Accelerator (ATA), replacing what was previously called the “AI for Work” initiative. In a separate memo circulated Monday, Bosworth wrote that Meta would increase internal data collection as part of the push.

“The vision we are building towards is one where our agents primarily do the work and our role is to direct, review and help them improve,” Bosworth wrote, adding that agents should “automatically see where we felt the need to intervene so they can be better next time.”

Meta’s move is direct: it is using its own workforce as a source of high-quality, labelled behavioural data to close the gap between AI models that can talk about tasks and agents that can actually execute them in a real computer environment.

Stakes and Context

The rollout comes as Meta simultaneously:

• Announced 10% global layoffs starting May 20 — the largest headcount-to-compute trade in its history
• Deployed MTIA 400 custom inference silicon targeting Nvidia for GenAI workloads
• Committed $21B to CoreWeave for cloud capacity through 2032
• Locked a 1GW+ MTIA deal with Broadcom through 2029

That constellation of moves signals that Meta’s actual workforce reduction goal is an increase in AI-executed work volume, not a net reduction in output. MCI is the data acquisition layer that makes that possible.

Safeguards and Legal Exposure

Meta told employees the captured data would not be used for performance reviews and that “safeguards” are in place. The company has not published the technical specifications of those safeguards or clarified data retention policies.

Privacy experts interviewed by Reuters said the programme raises substantial concerns. The combination of keystroke logging with periodic screen captures goes beyond typical enterprise productivity monitoring into behavioural surveillance of the kind usually associated with call centre or fraud-prevention contexts.

The deployment is currently limited to US-based employees, which likely reflects the EU’s stricter employment data protections rather than a policy preference.

The Data Problem This Solves

OpenAI (Codex Chronicle on Mac), Google (Project Jarvis/Mariner), and Anthropic (Claude Managed Agents) are each building computer-use agents, and each faces the same bottleneck: training data that reflects real human workflows on real computers is scarce and expensive to generate through synthetic means.

Meta’s MCI programme is a structural shortcut. With tens of thousands of engineers and knowledge workers on staff, the company can generate millions of labelled interaction sequences per day at near-zero marginal cost. The tradeoff is that the people generating those sequences did not sign up to be training-data subjects as a condition of employment.

Whether that tradeoff survives scrutiny — internal or regulatory — is the open question. With EU AI Act full enforcement 16 weeks away, the geofencing to US-only operations suggests Meta’s legal team already has a view.