Researchers Uncover How Meta's Muse AI Builds Detailed Profiles of Users' Friends and Family
Security researchers have extracted internal instructions from Meta's viral AI assistant Muse, revealing it automatically builds detailed pages on users' family, friends and colleagues. Meta says the transparency was intentional.
Meta's new personal AI assistant, Muse, has become a viral hit with millions of downloads, as users connect it to bank accounts, messages and health data to complete various tasks. The app has now drawn scrutiny after researchers extracted its internal operating instructions.
Independent AI safety and security researcher Karan Joshi managed to pull an extensive set of Muse's instructions and system prompts simply by asking the chat interface to copy and share its own software files. He shared his findings with Wired.
Among the revealed features is Muse's ability to build, on an hourly basis, "a page for every person in the user's life" — compiling data on family members, partners, friends, colleagues, collaborators and people the user follows. Using its "memory," stored as structured text files, Muse can then offer suggestions, such as how to improve a particular relationship or where to take a coffee-loving friend for breakfast.
According to the documentation, a page may start out "sparse" and fill in over time with sections like Facts, History, The Relationship, In Common, Open Threads and Strengthening. The instructions stress that Muse should rely only on available "evidence," since invented details are considered worse than an empty page. The system can also track important dates like birthdays and anniversaries, as well as how close a relationship is and what it might need going forward.
Carissa Véliz, an associate professor at Oxford's Institute for Ethics in AI, notes that users are giving AI systems far more information than they receive back, including inferences the systems draw from other data sources.
Meta has said each user has their own isolated virtual machine for storing data, inaccessible to other agents, and that users can wipe memories or disconnect external services at any time. Muse is also designed to seek confirmation before taking actions like sending an email or making a purchase, and maintains an audit log of its activity.
Experts, including Miranda Bogen of the Center for Democracy and Technology's AI Governance Lab, say Muse appears to place more emphasis on relationships and personal contacts than competing systems, and that AI assistants in general increasingly encourage users to share more of their data.
