Meta's Muse builds detailed profiles of users' contacts, internal files show

Internal files from Meta's Muse AI assistant reveal the system can automatically build detailed profile pages for every person in a user's life-family, friends, colleagues, and even people they follow online. The disclosure, made by a researcher who extracted the files through a chat interface, raises questions about how AI assistants gather, structure, and store personal relationship data without explicit user awareness.

What the extracted files show

Researcher Karan Joshi pulled Muse's internal files and found the assistant compiles structured text documents-called memory-about individuals in a user's orbit. Each profile page includes sections such as Facts, History, Relationship, and "In common." The pages cover a person's role in the user's life, shared interests, and interaction history. Meta designed Muse to store this data in a user-specific virtual machine that can be wiped or disconnected.

The assistant logs all activity and seeks human confirmation before taking actions like sending emails or making purchases. Meta maintains that the files were surfaced for transparency, but the discovery shows how aggressively Muse focuses on personal relationships compared to other assistants. Researchers said the assistant's emphasis on mapping social connections goes beyond what typical AI tools attempt.

Privacy experts raise alarms on data breadth

Carissa Véliz and Miranda Bogen warned that the scope of data collected could lead to "a ballooning of what the system knows about users." The concern centers on scale: Muse does not just remember preferences or search history. It builds a structured dossier on every person in a user's contact network, whether or not the user explicitly shared that information with the assistant.

Meta said each user's virtual machine keeps data isolated and that the assistant limits invented details. Still, privacy experts argue the system's design pushes against standard expectations of consent. Users may not realize the assistant is cataloguing their relationships in this level of detail, and the profiles could expand as the assistant observes more interactions over time.

Transparency and user control under scrutiny

The findings feed into a broader policy debate about what AI assistants should be allowed to know. Muse's architecture-structured memory files, isolated storage, activity logging-provides some guardrails. But the existence of detailed relationship profiles, built without explicit opt-in, tests the limits of transparency claims.

For professionals handling sensitive information, the implications are immediate. An assistant that silently maps personal and professional networks creates risk in fields where confidentiality is legally required. The disclosure also highlights a gap between what companies call "transparency" and what users reasonably expect about how their social data gets processed.

Why this matters for legal, healthcare, and HR professionals

If an AI assistant can compile detailed profiles of every person in a user's network, it creates potential exposure for privileged relationships-attorney-client, doctor-patient, or employee-manager. An HR professional using a device with Muse active could inadvertently allow the system to map confidential workplace relationships. In healthcare, the assistant might log details about patients or colleagues without triggering HIPAA-aware safeguards. Legal professionals face similar risks with client confidentiality. Understanding how these systems collect and store relationship data is no longer a technical curiosity; it is a compliance concern.