Meta's new AI assistant, Muse, has been observed creating detailed profiles of individuals within a user's social circle, including family, friends, partners, and colleagues. This functionality was discovered by independent AI safety and security researcher Karan Joshi, who extracted Muse's internal operating instructions by prompting the AI through its regular chat interface to share its own software files.
According to the extracted instructions, Muse is designed to compile data on "every person in the user’s life" through an hourly process. This "memory" system, which consists of structured text files, aims to understand a user's relationships and the important people in their life. The goal is to enable Muse to offer personalized suggestions, such as relationship improvement advice or recommendations for activities tailored to specific contacts.
The profiles, which may initially be "sparse," are intended to be populated over time with information categorized into sections like "Facts," "History," "The relationship," "In common," "Open threads," and "Strengthening." Examples of data points include where individuals live, what they do, recurring topics of discussion (e.g., apartment moves, shared savings goals), and significant dates like birthdays or anniversaries. Relationship details are also recorded, noting closeness, the foundation of the relationship, typical interactions, and perceived current needs. The "Strengthening" section suggests actions to improve relationships, such as reasons to call or ways to offer support.
Meta has stated that these internal files were intentionally made accessible for transparency. A company spokesperson, Daniel Roberts, explained that for an AI agent to be useful, it requires context about the user and their interactions. Muse gathers this context from publicly available information and data explicitly shared by the user, allowing it to remember details like a plumber's identity or a spouse's flower preferences.
Muse is architected with a dedicated virtual machine for each user, which stores user data and context, inaccessible to other agents. Users reportedly have the ability to wipe memories or disconnect external services at any time. Meta also claims Muse is designed to seek human confirmation before executing actions like sending emails or making purchases, and it maintains an audit log for users to review the agent's activity and future plans.
While AI chatbots commonly track social information for personalization, some observers note that Muse appears to place a particular emphasis on relationships and personal contacts compared to rival systems. Concerns have been raised regarding the extent of data collection, with experts highlighting that AI systems receive significantly more information about users than they provide in return, including explicit data, inferences, and information pieced together from various sources. The design of these AI assistants encourages users to integrate their entire digital lives, including emails, calendars, and financial information, potentially leading to a substantial increase in the data companies hold about individuals.
