Meta's 'Model Capability Initiative': AI Training to Track US Employees' Mouse Clicks, Keystrokes, and Screenshots

Meta is launching a sweeping internal program to harvest its US employees' digital activity—including mouse movements, clicks, keystrokes, and periodic screenshots—to create training data for future AI agents. According to internal memos from the Meta Superintelligence Labs team reported by Reuters, the new 'Model Capability Initiative' software will operate on designated work apps and websites, effectively turning routine office tasks into a live data feed for AI development. The initiative frames employee labor as a direct contribution to model improvement, with one memo stating, 'This is where all Meta employees can help our models get better simply by doing their daily work.'

The program signals a significant escalation in corporate data collection for AI training, moving beyond public web scraping and user data agreements to the intimate, granular capture of workplace behavior. The use of screenshots to provide context for clicks and keystrokes adds a layer of visibility that extends beyond mere metadata, potentially capturing sensitive information displayed on an employee's screen during monitored sessions. While framed as a voluntary contribution to AI advancement, the initiative raises immediate questions about the scope of monitoring, data anonymization, and employee consent within the corporate environment.

The rollout places Meta at the forefront of a contentious frontier in AI development: leveraging internal human-computer interaction as a proprietary training resource. It sets a precedent that other tech giants may follow, intensifying scrutiny around workplace surveillance, data privacy rights for employees, and the ethical boundaries of using behavioral data to build autonomous agents. The move also highlights the intensifying race for high-quality, nuanced training data—a race now extending directly into the office workflows of the companies building the technology.