TL;DR:

  • Meta is installing tracking software called the Model Capability Initiative on US-based employees’ computers to capture mouse movements, clicks, keystrokes and occasional screen content, using the data to train its AI agents.
  • The company says MCI data will not be used for performance assessments and that safeguards protect sensitive content, though it has not elaborated on what is excluded.
  • Meta plans 10% global layoffs starting 20 May, following similar AI-driven cuts at Amazon (30,000 corporate roles) and Block (nearly half of staff in February).

Meta’s internal memos, seen by Reuters, reveal a shift in how frontier AI labs generate training data: rather than scraping external content, they are instrumenting their own workforces. The Model Capability Initiative (MCI) is designed to close AI model weaknesses in areas like dropdown-menu navigation and keyboard shortcuts — the mechanical tasks that still trip up “computer-use” agents when they try to complete knowledge-work tasks autonomously.

The Agent Transformation Accelerator pipeline

MCI is one component of a rebranded “AI for Work” programme now called the Agent Transformation Accelerator (ATA). CTO Andrew Bosworth’s internal memo frames the destination clearly: “Our agents primarily do the work and our role is to direct, review and help them improve.” That explicitly positions human employees as labellers and reviewers of AI output in the near term, rather than as primary task-doers. It is the most candid public statement yet from a frontier AI company that its workforce restructuring is aimed at training its own successor systems.

Why UK and EU employers cannot copy this pattern

Reuters cites labour-law professors who are clear that the practice would likely be unlawful in Europe. In Italy, using electronic monitoring to track employee productivity is explicitly prohibited. German courts permit keystroke logging only in exceptional circumstances — typically a suspicion of a serious criminal offence. More broadly, the practice would likely breach the General Data Protection Regulation’s legitimate-interest and data-minimisation tests, and in the UK specifically, the Information Commissioner’s Office’s employment-practices code would create additional enforcement risk. Meta’s MCI is structurally not available to US-headquartered firms’ UK or EU subsidiaries without substantial redesign.

Looking forward

For UK businesses watching Meta’s strategic arc, two signals matter more than the privacy angle. First, the direction of travel — AI labs are now willing to treat their entire workforce as training-data infrastructure, accepting the reputational cost. Second, the productivity claim is now load-bearing: Meta’s 10% layoff decision is explicitly conditioned on AI agents being able to absorb displaced workload. If that does not materialise on schedule, either the cuts reverse or product quality visibly degrades. Either outcome will inform whether UK firms — facing tighter labour protections and lower talent-turnover tolerance — pursue more modest AI-workforce restructuring patterns through 2026.

Share this article