Inside Meta’s Muse: How the AI Builds Detailed Profiles of Your Social Circle

Meta’s newly launched personal assistant, Muse, has seen rapid adoption among consumers, with millions connecting the AI agent to their financial accounts, messages, and health data to automate daily tasks. However, recent findings from independent researchers have shed light on the extensive personal data the system gathers about users’ social circles.

Multiple investigators have successfully extracted Muse’s internal operating instructions and system prompts by using the chat interface to request the app to share its own software files. These documents were shared with WIRED and reveal the underlying mechanics of how the AI organizes information. Meta stated that these files were intended to remain accessible to promote transparency, noting that they help illustrate how the system handles queries regarding sensitive or politically charged subjects.

According to the extracted documentation, one of Muse’s core functions is generating a dedicated page for every individual in a user’s life. This process runs hourly and compiles data on family members, partners, friends, colleagues, collaborators, and people the user follows. The AI utilizes structured text files, referred to as “memory,” to accumulate details about relationships, allowing it to offer suggestions such as advice on improving connections or recommending breakfast spots for coffee-loving friends.

Karan Joshi, an independent AI safety and security researcher, analyzed the prompts and system skills data. He noted that while it is not uncommon for AI chatbots to track general social information, Muse’s approach is distinct given Meta’s extensive history with social network data. “What it seemed like to me… is that they want to understand your relationships that you have with real people,” Joshi said. “They’re trying to know you like a friend, which is honestly pretty creepy.”