Meta Is Recording Every Click and Keystroke Its Employees Make

Meta has started collecting the computer activity of its US employees in detail, recording mouse movements, keystrokes, and periodic screenshots under a program it calls the Model Capability Initiative. The goal, according to internal memos reported by TechCrunch and Fortune, is to train AI agents that can perform ordinary office tasks on their own.

The list of sites where keystrokes and clicks are being captured includes Google, LinkedIn, Wikipedia, GitHub, Slack, and Atlassian, according to CNBC. In practice, the program captures nearly everything a knowledge worker does throughout the day.

Meta's reasoning

The internal framing positions the program as a way for employees to directly improve AI model capabilities. The specific problem being targeted: AI agents still struggle with routine tasks that humans do automatically, things like navigating dropdown menus, using keyboard shortcuts efficiently, and moving between tools in a normal workflow. Feeding real human computer behavior to models should, in theory, accelerate progress on these gaps.

That is a reasonable technical argument. The problem is the context in which it is happening.

Why employees are uncomfortable

The monitoring program is rolling out as Meta prepares for further layoffs. That combination has not gone over well internally. One employee, speaking anonymously to the BBC, described having their "minutest computer actions" harvested for AI training while bracing for potential redundancy as "very dystopian." The concern is coherent: employees are contributing to systems that may ultimately reduce the number of roles like theirs.

There are also practical data concerns. Employees working on the program use platforms that contain sensitive information: passwords entered on monitored sites, health and immigration data visible in HR tools, early-stage product details not yet publicly announced. Meta says safeguards exist and the data is used only for training purposes, but the specifics of those safeguards have not been made public.

A pattern worth watching

Meta is not the first company to use employee behavior data for AI training, but the scope here is broader than most previous examples. The program spans hundreds of applications, is apparently opt-in at enrollment but runs continuously thereafter, and is being deployed at a company already known for aggressive data collection practices.

The larger dynamic this reflects is real and growing. AI agents that can perform knowledge work will need training data drawn from real knowledge work. Where that data comes from, who consents, and what protections apply when it includes sensitive material are questions the industry has not answered clearly. As companies restructure workforces around AI, the tension between using employees as training data sources and protecting their privacy will likely become harder to avoid.