Meta has found a new source of training data for its AI models. It was the employees. It was always going to be the employees.
The company is deploying an internal tool that records mouse movements, button clicks, and keystroke patterns from staff computers — the kind of behavioral data that teaches AI agents how humans actually navigate software, as opposed to how humans claim to navigate software.
The employees are not just building the AI anymore. They are the curriculum.
What happened
Reuters first reported the program, which Meta confirmed to TechCrunch with a statement that began, reassuringly, with the phrase "if we're building agents to help people complete everyday tasks." The conditional framing is doing a lot of work in that sentence.
Meta says the tool captures inputs on certain applications, that safeguards are in place to protect sensitive content, and that the data is used for no other purpose. Three clauses. All of them true, presumably, in the order they were written.
This is part of a broader industry trend in which companies, having exhausted the public internet, are now mining their own internal operations for training material. Last week, old startup Slack archives and Jira tickets were reported as the latest frontier. The humans are thorough.
Why the humans care
The practical concern is that yesterday's internal behavior becomes tomorrow's model capability — and the line between "how employees use software" and "how employees think about their work" is thinner than a keyboard shortcut.
There is also the matter of consent, which Meta describes as an opt-in program for now. "For now" is a phrase with a long and storied history in corporate data collection.
What happens next
Meta's AI agents will learn to navigate interfaces the way humans do — tentatively, occasionally incorrectly, but eventually with enough fluency to do it without the human present.
The employees are not just building the AI anymore. They are the curriculum.