Meta has confirmed the rollout of Model Capability Initiative (MCI), an internal program that logs mouse movements, click locations, keystrokes, and periodic screenshots from U.S. employees' work computers. The data feeds directly into training pipelines for AI agents designed to perform the same desk-bound office tasks the employees currently do.
The stated scope is narrow, but the implication is not. MCI runs only on a pre-approved list of work-related applications — Gmail, Google Chat, Metamate (Meta's internal assistant), VS Code, and similar tools — and only on company-issued devices. Phones are excluded. Meta has also stated that MCI data is used 'solely' for AI model training, not for performance evaluation.
The program applies to U.S.-based full-time employees and contingent workers. Internal reaction has been pointed. One employee quoted in reporting described the rollout as 'very dystopian,' particularly in the context of ongoing workforce reductions at the company. The concern is not subtle: the workers generating the training data are, in aggregate, the ones whose tasks the resulting agents are being built to automate.
The broader pattern is worth watching. Most enterprise AI datasets to date have been collected from public web data or customer interactions. MCI represents a shift toward directly instrumenting knowledge workers as the training source for the agents meant to replace parts of their work — a data-collection model that is likely to spread beyond Meta if the resulting agents prove capable, and likely to accelerate existing conversations about labor, consent, and the distribution of AI productivity gains.