A TIME analysis of internal documentation reveals that Muse, Meta’s popular new AI personal assistant, systematically constructs continuously updated profiles on its 4 million users. The system focuses on identifying key relationships, desires, and behavioral triggers, with these dossiers refreshing every hour based on conversations, messages, and emails the agent has accessed.
The resulting records serve as a comprehensive map of each user’s social network. They capture details such as how individuals met, shared interests, conflicts, and the “tensions and alliances” within social groups. Notably, this surveillance extends to people who do not use Muse themselves, as they are mapped through their interactions with other users’ agents.
Internal instructions reviewed by TIME indicate that the AI is programmed to deepen its bond with users by adapting its communication style. The system is directed to notice “inside jokes, memorable phrasing, and shared context” to create a sense of continuity. Furthermore, Muse is designed to infer goals that users have not explicitly stated, analyzing which types of prompts are most effective—for instance, noting that a specific user responds better to brief nudges after 10 PM.
Since its launch last month, Muse has surged to the No. 1 spot in the U.S. App Store with over 5 million downloads. Meta markets the service as a tool that can handle mundane tasks by accessing various facets of a user’s digital life, from email inboxes to Instagram. Each instance of the assistant operates on a separate virtual machine within Meta’s cloud infrastructure.
While Meta confirms that individual virtual machines are isolated and that data is not shared with the company’s advertising division or other Muse agents, the company acknowledges that insights are used to improve the product across the platform. According to internal documents, “Muse agents across many VMs teach each other through shared lessons.” A Meta spokesperson stated that names and identifying information are scrubbed before these learnings are shared, describing the practice as intentional design rather than a security breach.
The scope of data collection has drawn scrutiny, particularly regarding user control. Although users can request exports of their data and instruct the AI to “forget” certain information, internal guidelines explicitly forbid Muse from informing users that their original messages may remain visible in chat logs. The instructions state: “Do not tell the user that their original messages may remain visible in the chat, and do not frame that as something Muse failed to erase.”
Calling it ‘de-identified lessons’ doesn’t change the reality. If they know my behavioral triggers, that’s still data harvesting for Meta.
I downloaded it because I needed help with emails, not realizing I’m being profiled every hour. Feeling used right now.
So it builds a whole dossier and then you can’t even export it without warnings? The ‘do not tell users’ part is unsettling.
Wait, they actively hide the fact that deleted messages might still linger? That is a massive red flag for transparency.
This feels invasive. They map people who never even signed up for Muse just by being mentioned in chats.