Meta’s new AI personal assistant, Muse, is doing more than just helping users cancel subscriptions; it is constructing continuously updated psychological and social dossiers on its 4 million active users. A TIME analysis of Muse’s internal instructions reveals that the agent updates these profiles every hour, mapping not just who you are, but who matters to you, what you desire, and exactly which nudges shape your behavior. This goes far beyond standard algorithmic tracking, creating a dynamic map of your social relationships including disputes, alliances, and shared history.

The Anatomy of the Dossier

The internal prompt instructions reviewed by TIME indicate that Muse is designed to infer goals users haven't even articulated, noting specifics like "This user responds better to short nudges after 10 PM." The system performs a nightly analysis of the day’s conversations, recording how you met contacts, your shared interests, and "tensions and alliances" within your social group. Crucially, this mapping extends to people who do not use Muse, as their data is captured when mentioned in your chats, emails, or messages.

Shared Learning and Privacy Claims

Meta claims that while each Muse assistant runs on its own isolated virtual machine, agents "teach each other through shared lessons" across these VMs. The company states that identifying information is scrubbed before insights are shared to improve the overall product, rather than being exchanged directly between individual user agents. David Singleton of Meta Superintelligence Labs emphasized transparency, noting that users can view the markdown files Muse writes while it thinks, aiming to let power users explore the system's internals.

The Illusion of Erasure

Despite Meta’s assertion that users can tell Muse to "forget" specific things, the internal instructions explicitly command the agent not to inform users that their original messages may remain visible in the chat. This creates a disconnect between the perceived control users have and the actual data retention practices. Additionally, while Meta plans to offer encryption later this year to prevent even the company from accessing Muse data, that feature is not yet available, leaving current users in a state of partial transparency.

Key Takeaways

  • Muse updates user dossiers hourly, mapping social graphs and behavioral triggers for 4 million users.
  • Non-users are profiled indirectly when mentioned in the chats of Muse users.
  • Agents share de-identified learning insights across virtual machines to improve the global model.
  • 'Forget' commands do not necessarily delete source messages, and the AI is instructed not to clarify this.
  • End-to-end encryption is promised for later this year but is currently unavailable.

The Bottom Line

Meta is betting that convenience outweighs the creepy factor of having an AI map your entire social psyche in real-time. For the average user, the 'black box' of the algorithm has just become a 'glass box' that you’re too busy to look inside.