Meta tracks US employees' clicks and keystrokes to train AI agents

The buried angle here is not the surveillance itself but what it reveals about the data bottleneck: Meta apparently cannot get sufficient behavioral signal for agent training from synthetic data or external sources alone, so it turned its own employees into a labeled dataset. That is a meaningful admission about where agentic AI training actually stands.

MIT Technology Review's piece on 'Treating enterprise AI as an operating layer' argued that competitive advantage in AI is shifting toward whoever controls the infrastructure where AI is deployed and refined. Meta's move fits that thesis precisely: the training pipeline is now part of the operational stack, and employee behavior is raw material for it. InsightFinder's April funding round (story 1) pointed to a parallel problem, that diagnosing how AI agents fail requires deep observability into real-world usage. Meta is apparently solving the training-data side of that same problem through internal surveillance rather than external instrumentation. What neither story addressed is the labor and consent dimension, which is where regulatory exposure will likely concentrate.

Watch whether any EU-based Meta employees raise formal complaints under GDPR in the next 60 days. A regulatory inquiry in Europe would force disclosure of what data was collected and under what consent framework, turning an internal policy into a public governance test case.