Meta is deploying a new internal software suite, the Model Capability Initiative, to capture granular user interaction data from employee workstations. According to an internal memo reported by Reuters, the system will log keystrokes, mouse movements, clicks, and periodic screenshots from US-based staff and contractors. The data collection targets specific work applications like Gmail, GChat, and Meta's internal tools, explicitly excluding personal phones.

The stated engineering objective is to gather high-fidelity behavioral datasets to train AI agents capable of autonomous task execution. The focus is on capturing nuanced interactions—dropdown selections, keyboard shortcuts, workflow patterns—that current models cannot reliably infer. In a company channel, CTO Andrew Bosworth framed this as essential for building agents that employees will later direct, though the specific model training pipelines were not detailed.

Reaction from technical staff was immediate and negative, with internal posts expressing discomfort and requests for opt-outs, which Bosworth stated were unavailable for work laptops. This initiative unfolds alongside significant workforce reductions, with plans to cut 8,000 roles.

From a system design perspective, the move highlights the intensive data requirements for next-generation AI. It also surfaces stark trade-offs between data utility and employee privacy. Meta asserts the data is siloed from performance reviews and manager access, with safeguards for sensitive content. However, privacy experts like the Center for Democracy & Technology's Eric Null warn such pervasive monitoring risks encoding structural biases and sets a concerning precedent for workplace surveillance under the guise of model development.

Source: CNET