Meta pauses its employee-tracking AI program after an internal data leak
Meta has paused the Model Capability Initiative, a program launched in April that recorded keystrokes, mouse movements, and occasional screen content from US-based employees to help train AI agents capable of handling routine computing tasks. The suspension came not over the surveillance itself but because a security error left roughly 45,000 database tables — containing private conversations, transcriptions, and performance data — visible to the entire company.
The program had already generated internal resistance: more than 1,600 employees signed a petition calling the constant screen-scraping an invasion of privacy. Meta had said the data would not be used to assess individual performance, but participation was mandatory for most staff.
The episode is a concrete example of the friction inside large AI efforts, where the training data companies most want — detailed records of how skilled workers actually operate — is also among the most sensitive to collect and store. The leak, rather than the objections, is what forced the pause, underscoring that the program’s risks were as much about data security as worker consent.