Meta Deploys Keystroke Tracking for AI Training

Meta has begun deploying an internal tracking tool, the Model Capability Initiative (MCI), on U.S. employees to collect mouse movements, clicks, keystrokes and occasional screenshots to train AI agents. The effort sits inside a larger program rebranded as Agent Transformation Accelerator (ATA). Meta says the collection is limited to work apps and will not be used for performance reviews, but the announcement provoked significant employee pushback and privacy concerns.

The program captures low-level UI interaction data to teach models how humans actually perform tasks, for example using keyboard shortcuts and choosing from dropdown menus. The internal memo states this is intended to improve agent capabilities for routine desktop workflows. Key technical attributes disclosed so far include:

• •Captures: mouse movements, clicks, keystrokes, and occasional screen snapshots on work-related apps and websites
• •Program name and scope: Model Capability Initiative (MCI) within the Agent Transformation Accelerator (ATA)
• •Promised limits: data collection restricted to work apps and not for performance reviews, per Meta internal memos

This is a concrete example of a major tech firm instrumenting its workforce to generate proprietary training data at scale. The data type here is behavioral UI telemetry, which can yield high-signal supervision for agents that simulate user interactions. That is valuable because models trained on logged, real-world human interactions can close gaps in automation for GUI-level tasks. At the same time, employee surveillance raises privacy, consent, and compliance questions. The move intersects with ongoing debates over workplace monitoring law, data minimization, and trust in AI-driven productivity tooling. The CTO and internal memos frame the work as efficiency and model improvement; employees view it as intrusive. As one memo put it, "This is where all Meta employees can help our models get better simply by doing their daily work."

How Meta operationalizes safeguards, opt-out mechanisms, retention limits, and de-identification will determine legal exposure and employee trust. Regulators and privacy-conscious enterprises will closely monitor whether behavioral telemetry collected for model training meets consent and data-protection standards.