Meta has begun deploying software on its employees’ computers in the United States that records their keystrokes, mouse movements and, occasionally, on-screen activity, harvesting the data to train artificial intelligence models designed to automate routine office tasks. The initiative, internally dubbed the Model Capability Initiative, marks a significant escalation in the use of workplace surveillance as raw material for AI development. A spokesman for the company told the BBC that the project is essential for building “agents to help people complete everyday tasks using computers”, adding that the tool includes safeguards to protect sensitive content and is not used for any other purpose. Yet the announcement has unsettled many within the company. One employee, speaking on condition of anonymity, described the prospect of having their smallest interactions monitored and fed into a machine-learning pipeline as deeply intrusive, even if the data is ostensibly anonymised and compartmentalised.

Viewed from Washington, the practice occupies a legally ambiguous space. At the federal level, employers in the United States enjoy broad latitude to monitor workplace activity, provided they disclose the practice and do not violate specific privacy statutes or collective bargaining agreements. This permissive environment stands in sharp contrast to the European regulatory framework, where the General Data Protection Regulation imposes strict proportionality and consent requirements on employee monitoring. Analysts in London note that while Meta’s initiative is currently limited to its American workforce, the logic of training AI on real human behaviour will inevitably raise questions about whether similar programmes could be extended to operations in Europe or elsewhere, triggering scrutiny from data protection authorities.

The tool’s ultimate ambition is to create AI agents capable of independently performing complex workflows — booking meetings, sorting emails, filling forms — by learning from the very patterns of human interaction with software. This approach, sometimes called “imitation learning”, requires vast quantities of granular behavioural data. Meta’s internal memorandum, seen by Reuters and reported by Il Sole 24 Ore, makes clear that the MCI software runs on company-issued computers and internal applications, capturing not only clicks and keystrokes but also screenshots of the user’s display, all to build a corpus of natural computer-use examples. The company insists that the data remains within a controlled environment and is segregated from other business functions, but trust among employees is fragile.

Looking ahead, the controversy touches on a deeper tension in the AI industry: the hunger for high-quality training data versus the expectation of workplace privacy. If Meta’s experiment succeeds, it could accelerate a trend in which employees’ daily digital labour becomes the invisible feedstock for the very automation that may eventually replace them. European regulators, already wary of Big Tech’s data practices, are likely to view the scheme as a test case for the boundaries of consent under the GDPR. Meanwhile, in the United States, the absence of a comprehensive federal privacy law leaves workers with little recourse beyond their employer’s goodwill. As Meta pushes forward, the question is no longer whether AI can learn from human behaviour, but at what cost to the human beings providing the lessons.