Meta employees are confronting the surveillance technology their company perfected for billions of users after management installed keystroke-tracking software on their work computers this week.
The new "Model Capability Initiative" tool captures mouse movements, clicks, typing activity, and occasional screenshots from US-based staff as they use designated work applications including Gmail, GChat, VSCode, and internal tools like Metamate (Meta's AI assistant for employees). According to internal communications viewed by Business Insider and Reuters, the data feeds directly into Meta's AI training pipeline to teach models how humans handle software interfaces.
"This makes me feel like a lab rat," one employee wrote on the company's internal forum.
The post quickly drew hundreds of replies, with one user asking, "How do we opt out?" The question became the top-rated comment as others flooded the thread with angry reactions.
Company representatives confirmed there is no option to opt out of the tracking on work-provided laptops.
Meta frames the initiative as essential for developing AI agents that can complete routine digital tasks like dealing with dropdown menus or using keyboard shortcuts, areas where current models still struggle despite excelling at research and coding.
"For agents to understand how people actually complete everyday tasks using computers, we need to train our models on real examples," an internal memo stated.
The tracking software operates only on a designated list of work-related applications and websites rather than monitoring all computer activity. Company officials emphasized that safeguards protect sensitive content and that collected data will not be used for performance monitoring or any purpose beyond model training.
Internal backlash arrives as Meta prepares significant workforce reductions reportedly targeting about 8,000 global employees starting May 20. These cuts represent part of a broader pivot toward AI-driven roles where executives aim to automate tasks previously handled by human workers.
Meta's approach mirrors similar efforts across the industry where companies seek real-world interaction data for AI development. OpenAI introduced its "Operator" tool in early 2025 capable of using web browsers on behalf of humans while Anthropic debuted comparable computer-use technology in October 2024.
