Meta paused its controversial AI training initiative after sensitive employee data was inadvertently relocated to an unauthorized internal location, raising security concerns within the company. The incident prompted Meta to halt the project while it conducted a thorough internal review.
The program, known as the Model Capability Initiative, was designed to collect keystrokes and mouse movements from most of Meta’s US employees to improve its AI models. Despite built-in privacy safeguards, the initiative faced significant backlash from employees due to mandatory participation and privacy worries.
Meta’s Chief Technology Officer, Andrew Bosworth, revealed in an interview that the data involved was never breached externally but had been mistakenly placed where it should not have been by one of the researchers handling downstream data. This relocation made some transformed employee data accessible internally beyond the intended circles, though there was no suspicion of malicious intent.
Originally launched in April, the Model Capability Initiative aimed to gather varied data to enhance AI training. However, it largely generated repetitive data, limiting its usefulness. To address employee concerns and improve data diversity, Meta had introduced expanded opt-out options for staff a few weeks after launch. Bosworth noted that employees could pause data collection individually, highlighting an effort to balance AI development with employee privacy.
The internal fallout was significant, with Bosworth acknowledging historically low employee morale linked to the program. Screenshots reviewed by media outlets indicated that the leaked data was widely accessible internally, accelerating the decision to suspend the initiative in June.
A Meta spokesperson confirmed the pause, emphasizing that no evidence existed of improper external access but underscoring the company’s commitment to investigating the issue thoroughly. Requests for further comments on the investigation or the future of the program were declined.