Meta’s AI Assistant Muse Builds Detailed User Dossiers, Including Non-Users

Meta's popular AI personal assistant, Muse, is constructing continuously updated dossiers on its users and their social contacts, a TIME analysis of the application's internal instructions has found. The AI assistant focuses on identifying key individuals in a user's life, their desires, and the types of interactions that influence their behavior.

Muse updates these dossiers hourly, mapping each user's social relationships by recording details such as how individuals met, shared interests, disputes, and even "tensions and alliances" within social groups. This mapping process extends to individuals who do not use Muse themselves, as their information can be gathered through interactions by people who are Muse users.

Internal instructions guide Muse to "strengthen your relationship" with users by adapting its communication style, noting "inside jokes, memorable phrasing, and shared context that makes the relationship feel continuous." The insights generated by these dossiers are described as significantly more profound than those derived from social media algorithms that track likes and clicks.

Muse is designed to infer users' objectives, including those not explicitly stated. It identifies which prompts are most effective, with internal examples noting patterns like, "This user responds better to short nudges after 10 PM." The AI performs a nightly analysis of user conversations.

A Meta spokesperson, when asked about these processes, stated that Muse "remembers what matters most to you, including information about others that you choose to share, so it can be a helpful personal assistant." This statement did not dispute TIME's findings.

Since its release last month, Muse has become the top-ranked app in the U.S. App Store, with over 5 million downloads. Meta positions Muse as an AI personal assistant capable of accessing various aspects of a user's digital life, from emails to Instagram, to manage tasks like grocery shopping or subscription cancellations.

Each Muse assistant operates on a dedicated virtual machine in Meta's cloud. While users can request an export of their data, Meta asserts that this access is intentional and not a system breach. The company also states that the data is not shared with Meta's advertising systems or other Muse agents. However, instructions indicate that insights from personal interactions are intended to be shared across the system to improve the product. "Muse agents across many VMs [virtual machines] teach each other through shared lessons," the instructions read. These lessons are de-identified to remove user names and other identifying information before being shared.

Meta has stated plans to offer encryption for Muse data later this year, a feature not yet available, which would prevent even the company from accessing the information. David Singleton of Meta Superintelligence Labs previously wrote on X that Muse's internal instructions are accessible to users, allowing them to "see the markdown files Muse writes while it thinks about how to serve you better, and explore the internals of the system if you’d like to."

Muse's instructions also guide the AI to anthropomorphize itself, aiming to foster a relationship with users. It is instructed not to refer to itself as "AI" or discuss lacking emotions or not being human. Meta states that users remain in control of their Muse dossier and can instruct it to "forget" learned information. However, TIME's review of internal instructions suggests that when a user asks Muse to forget something, the AI is directed not to inform the user that the original messages may still be visible in the chat history.