Meta wants to build AI agents that can use computers like humans do. So they're watching their own employees use computers. Every keystroke, every mouse click, every screen on Google, LinkedIn, Wikipedia, GitHub, Slack, and hundreds of other sites. The program is called the Model Capability Initiative, run out of Meta's Superintelligence Labs under Alexandr Wang, whom Meta hired from Scale AI last year. The data feeds Muse Spark, the new model Meta released this month.
A Meta spokesperson told CNBC that building agents to "help people complete everyday tasks using computers" requires "real examples of how people actually use them." They say safeguards protect sensitive content. Employees aren't buying it. Multiple workers called the surveillance "dystopian" in internal messages viewed by CNBC, raising concerns about passwords, personal health information, and immigration details getting swept up. One internal memo suggested employees "can control what shows up on your screen by not doing personal work on your work computer." Problem solved, apparently.
The legal picture is messy. Capturing employee inputs on third-party sites like Google and LinkedIn likely violates those platforms' terms of service, which prohibit automated data collection. The practice could expose Meta to liability under the Computer Fraud and Abuse Act, which covers exceeding authorized access, and the Electronic Communications Privacy Act, which governs interception of electronic communications. California's CCPA and Europe's GDPR both demand explicit consent and data minimization for employee monitoring. Broad surveillance of external sites may not clear those bars.
One Hacker News commenter called it the "Soylent Green of dogfooding." That about captures it. Meta is consuming its own workforce's digital behavior in a bid for training data. The company has lagged behind OpenAI, Anthropic, and Google in generative AI. Zuckerberg went on a spending spree last summer to close the gap. Whether logging every click and keystroke of the people building your products is justified is a question Meta seems to have already answered.