Meta wants to watch how you click and type. The company plans to install employee tracking software that captures mouse movements, keystrokes, and other workstation telemetry to train its AI agents. The data would feed into imitation learning systems that need real human interaction patterns to replicate how people use software and complete tasks.
The technical setup is more invasive than typical productivity monitoring. Endpoint telemetry agents would capture raw input events (mouse coordinates, click timestamps, keystroke cadence) and sync them with screen captures or DOM snapshots. That creates structured state-action pairs machine learning models use to predict the right inputs for a given software state. It's the kind of data you need if you're building agents that can automate real office workflows instead of just generating text, a capability often discussed alongside telemetry collection concerns in similar tools.
Privacy implications are obvious and uncomfortable. Employees would essentially generate training data all day while doing their actual jobs. The specifics around anonymization and data protection remain unclear. Workers on forums like Hacker News have already expressed concern. There's real tension between Meta's need for human interaction data and the basic expectation that your employer isn't recording every mouse movement. Companies always claim privacy protections exist, but tech's track record on that front isn't reassuring. One recent case involves an agent quietly burning user credits without consent.