May 11, 2026

The New York Times reported that Meta is logging what its employees type, how they move their mouse, what they click, and what appears on their screen. The program is called the Model Capability Initiative, and the stated purpose is training AI on how people complete computer tasks.

Hundreds of employees objected. An engineering manager asked how to opt out. Meta’s CTO Andrew Bosworth answered: “There is no option to opt out on your corporate laptop.”

Zuckerberg clarified that this is not surveillance. It is just teaching AI “how smart people use computers.” The data is not used for any other purpose and safeguards are in place.

I find it hard to add much to this. Meta has spent twenty years building a company on collecting user data at scale, with no meaningful opt-out, under the consistent framing that it is for a better experience, not surveillance. The playbook is unchanged. The only novel element is the audience.

Employees are now on the receiving end of the same architecture they helped build.. and they are, by the Times’ account, miserable.

There is a related story that belongs in the same picture. Meta introduced internal dashboards to track employees’ token consumption. The Times reports that this led some employees to build so many AI agents that others had to introduce agents to find agents, and agents to rate agents. Separately, an employee built an unofficial leaderboar, called “Claudeonomics”, ranking the company’s 85,000 employees by tokens consumed.

Token consumption as a productivity metric is the AI equivalent of measuring lines of code. Both are easy to observe, easy to report, and easy to game. They measure activity, not output. An agent left running overnight generates tokens. A long prompt generates tokens. Asking the same question five ways generates tokens. The predictable result is exactly what the Times describes: people optimizing the metric instead of the work, up to and including building agents whose only job is to manage other agents.

As usual, the metric selects for the behavior of optimizing the metric