Meta Monitors Staff Activity Across Major Sites to Train AI

Meta's push to sharpen its AI tools now includes watching how its own employees use the internet, down to their clicks and keystrokes. Internal messages seen by CNBC show that the social media company is logging staffers' activity on hundreds of sites and apps via a tool called the Model Capability Initiative, or MCI. The list includes Google, LinkedIn, Wikipedia, GitHub, Slack, Atlassian, and Meta's own platforms. At one point, it even featured AI rivals like ChatGPT and Claude.

The project feeds data into AI "agents" designed to handle routine computer tasks, part of Mark Zuckerberg's effort to close the gap on generative AI with OpenAI, Anthropic, and Google. A spokesperson for the Facebook and Instagram owner says Meta needs real-world examples of how people move their mouse, click, and navigate menus, insisting there are safeguards so sensitive material isn't misused. "Data is not used for any other purpose," the rep notes. "This is where all Meta employees can help our models get better simply by doing their daily work," an internal memo notes, per Reuters.

But internal chats seen by CNBC show some workers calling the effort "dystopian," warning it could surface passwords, confidential product plans, or personal details. "They've added model training to their responsibilities without any apparent increase in compensation, and the new work requires them to surrender a sense of privacy that their work isn't being monitored," Gizmodo notes. An internal memo seen by CNBC said the tool records on-screen content but doesn't ingest files or attachments, advising employees worried about privacy not to do personal tasks on their work machines.