Meta is installing tracking software on US employees' work computers that captures mouse movements, clicks, keystrokes, and occasional screen snapshots. The data feeds into the company's SuperIntelligence Labs effort to build AI agents that can handle computer tasks on their own. Internal memos reviewed by Reuters frame it as a way for staff to "help our models get better simply by doing their daily work." The tracking targets specific work apps and websites, not personal use.
Meta spokesperson Andy Stone says the data won't be used for performance reviews, only model training. He explained that building agents to help people complete everyday computer tasks requires real examples of how humans interact with interfaces: clicking buttons, working with dropdown menus, using keyboard shortcuts. Safeguards are in place to protect sensitive content, according to Stone.
But Meta's approach is more aggressive than what competitors do. Microsoft and Google rely on aggregated, anonymized usage patterns from products like Office 365 and Google Workspace. Anthropic and OpenAI have focused on synthetic data and public datasets for their computer-use agents. Capturing raw keystrokes and mouse trajectories at the individual level is a different beast entirely, and it raises questions about how well this surveillance-based method scales alongside privacy regulations like GDPR and CCPA.
There's also the trust problem. Even with guarantees that the data stays within model training, employees who know their clicks are being recorded will inevitably change how they work. Meta's track record on data privacy and sharing doesn't help. The company is betting that granular behavioral data produces better agents, but they're asking their own workforce to absorb the surveillance cost to get there.