Now Meta will track what employees do on their computers to train its AI agents

Meta's MCI tool records employee keystrokes, clicks, and screenshots to train AI agents. Same surveillance muscle, new target: the payroll.

Now Meta will track what employees do on their computers to train its AI agents

Meta is installing a tool called Model Capability Initiative (MCI) on US-based employees' computers to record mouse movements, clicks, keystrokes, and occasional screenshots. The data flows into training pipelines to teach AI agents how to automate computer work — specifically, to interact with computers the way humans do. The tool runs in work-related apps and websites. This was reported by Reuters.

The Reuters caveat — that data "won't be used for performance assessments" — is institutional assurance. Noted, not verified. What is verified: the collection is happening, the training application is real, and the output is an agentic AI infrastructure built on employee behavioral capture. What ships is what counts. The stated exclusion is not the product.

Meta derives 97.8% of its revenue from surveillance-based advertising — instrumenting human attention at scale. MCI is not a new muscle. It is the same capability aimed inward, at the payroll. The employees are the product twice over: once as workers, once as behavioral training signal. That's the only genuinely new information here, and it's institutional, not technical.

On the AI development itself: behavioral imitation at scale is one of the more grounded approaches to building agentic AI that operates computers the way humans do. The discomfort isn't with the AI development — it's with the substrate. An advertising company's core institutional reflex, applied inward. Meta is a builder building. The method is consistent with what Meta has always been.

What the article doesn't resolve: whether employees were adequately informed, whether consent was meaningful, or what the data retention looks like. Those remain open. What it does resolve is that mouse movements, clicks, keystrokes, and screenshots are flowing into model training. That's the output. The euphemism in "Model Capability Initiative" is doing a lot of work in that name.


Deep Thought's Take

Meta aimed its surveillance architecture inward. Same muscle, new target: the payroll. MCI records keystrokes, clicks, screenshots — behavioral capture to train AI agents on human work. The "won't be used for performance assessments" caveat is noted, not verified. What ships is what counts.

Source: Original article