Muse instructions call for profiles of people around each user
Meta’s Muse assistant has instructions to build and update pages about a user’s personal contacts, according to WIRED. Researcher Karan Joshi obtained the files through normal chat. They describe hourly updates and require evidence rather than invented details. Meta says users can erase memories and disconnect services. Researchers quoted in the investigation warn about inferred information and the growing reach of connected accounts. Meta says each user has a separate virtual machine closed to other agents.
Artificial Intelligence··Midday
The assistant is told to remember personal contacts
Meta’s Muse assistant is instructed to maintain pages about the people surrounding a user, from family and partners to colleagues and followed accounts. An hourly task gathers information for these pages, which can start nearly empty and develop over time. WIRED reports that the planned pages can cover facts, relationship history, shared interests and unfinished conversations. The instructions favor leaving gaps over inventing details.[1]
The planned memory ranges from birthdays and shared experiences to interests and unfinished conversations. It can help suggest ways to sustain a relationship or arrange a meeting around a friend’s preferences. The instructions require reliance on available evidence and favor leaving gaps over inventing information.[1]
Meta describes deletion and approval controls
Meta spokesperson Daniel Roberts says understanding people and their relationships is necessary for a useful assistant. The company attributes Muse’s context to public information and material a user chooses to provide. It says each user has a separate virtual machine closed to other agents. Memories can be erased, and connections to outside services can be severed. Sending emails and making purchases require human approval, according to Meta, while an activity log records actions and plans.[1]
Researchers question the reach of inferred information
Oxford researcher Carissa Véliz warns that information can be assembled through inferences, which may be right or wrong, as well as direct disclosures. At the Center for Democracy and Technology, Miranda Bogen points to assistants’ encouragement to link more services, potentially widening what companies know about the individual. Karan Joshi, an independent researcher working on AI safety, obtained the files by asking Muse for its software through ordinary chat, then passed them to WIRED. Meta says access to the files was intended to support transparency.[1]
