Meta pauses employee keystroke AI training program

Meta paused an internal AI training program after screenshots circulated that showed sensitive employee data was accessible across the company, Business Insider reports. Business Insider says the incident was classified as a SEV 2 on a one-to-five severity scale, and that leaked screenshots allegedly exposed employees' private conversations, performance data, and transcriptions. The program, called the Model Capability Initiative (MCI), was announced in April and used staff keystrokes and mouse movements as training data, Quartz and Business Insider reported. A Meta spokesperson told Business Insider, "We have carefully designed this program with privacy safeguards, and while we have no indication at this time that any data was improperly accessed by Meta employees, we're pausing it while we investigate."

Public reporting describes the MCI as collecting low-level interaction telemetry, namely keystrokes and mouse movements, to create realistic training signals for agents. Quartz reports internal changes circulated by Stephane Kasriel that added a 30-minute local pause option and limited full opt-outs to certain employee groups; those changes respond to battery, bandwidth, and sensitive-data concerns raised by staff. Neither source provides an engineering postmortem or technical root cause for the access gap in the screenshots.

The episode sits at the intersection of workplace surveillance, ML data governance, and internal security controls. Companies that repurpose employee interaction telemetry for model training often increase the attack surface for sensitive data, and oversight failures can rapidly erode employee trust, as the reported petition of over 1,500 signatories illustrates (Quartz). Business Insider frames this pause as one of several recent operational security incidents at Meta, noting prior Instagram account-hijack vulnerabilities and other internal AI-related incidents.

Observers should watch for a public or internal postmortem that documents the scope of exposed artifacts, access controls implicated, and any change to data retention or access policies. Also monitor whether internal governance updates from privacy, security, or legal teams are published or leaked, and whether other large tech employers that collect similar telemetry revise opt-in/opt-out and auditing practices. For practitioners: expect heightened scrutiny of internal telemetry pipelines used for model training, particularly around role-based access controls, dataset minimization, and reproducible audit trails.