Meta installs employee monitoring software to train AI agents on workplace tasks

Meta is deploying monitoring software called the Model Capability Initiative on computers used by U.S.-based employees, according to reporting by Reuters cited in this column. The tool captures mouse movements, clicks, keystrokes, and occasional screenshots to provide training data for AI agents designed to perform computer tasks. Meta CTO Andrew Bosworth outlined the intended future state in internal communications: agents would handle primary work tasks while employees would focus on directing, reviewing, and improving those systems.

The company states that collected data will not inform performance evaluations and that safeguards exist for sensitive content. However, internal employee discussions obtained by Platformer reveal substantive privacy concerns. Workers questioned whether the tool would inadvertently capture personal identifying information, health data, or financial details — concerns compounded by the fact that Gmail and other personal-use applications are approved monitoring contexts. When pressed on privacy review procedures and data safeguard specifics, Bosworth provided minimal detail.

A key distinction separates U.S. and European implementations. Meta cannot deploy MCI to European employees due to GDPR protections and worker safeguard requirements. This creates a geographic divide in workplace surveillance standards, effectively making GDPR-protected regions unavailable for the data collection effort.

The monitoring initiative arrives concurrently with a major workforce reduction. Meta confirmed a 10 percent reduction affecting approximately 8,000 employees, alongside a decision to leave 6,000 open positions unfilled. Zuckerberg has signaled that future organizational structures will require smaller teams, with single highly talented employees accomplishing work that previously required larger groups. Among employees, speculation about larger cuts remains common.

The program reflects a broader shift in how AI training data is sourced. As publicly available text has been extensively mined for model training, companies face pressure to generate novel datasets. Meta's approach — systematized collection of employee workflow data — may establish a template for how technology companies extract value from their own workforces to build automation systems that could eventually displace those same workers.