Muse: The New AI Assistant by Meta
The latest personal assistant from Meta, known as Muse, has quickly risen to viral fame, attracting millions of downloads. Users have integrated this AI agent with their bank accounts, messaging platforms, and health data, granting it the ability to perform various tasks on their behalf. As Muse continues to gain traction among consumers, insights from within the application reveal how it organises and presents information to its users.
Recently, several researchers have managed to extract internal files from Muse, unveiling the operational prompts that guide the assistant’s functionality. This transparency sheds light on how Muse responds to inquiries concerning particularly sensitive or politically charged topics. Meta has stated that the intention behind making these files accessible was to promote transparency, and indeed, they provide a fascinating glimpse into the assistant’s construction and behaviour.
Karan Joshi, an independent researcher with a focus on security and AI, successfully extracted a range of prompts and system messages from Muse. By utilising the standard chat interface, he prompted the AI assistant to share its software files, subsequently sharing his findings with the broader community.
Understanding User Relationships
One intriguing prompt appears to instruct Muse to create “a page for every person in the user’s life.” This process runs on an hourly basis and involves gathering information about family members, partners, friends, colleagues, “collaborators,” and individuals whom the user “follows.” The overarching goal is for Muse to leverage its “memory”—essentially structured text files—to collect data on the user’s relationships and significant individuals in their life.
With this information, Muse can offer personalised suggestions, such as advice on improving specific relationships or recommending a suitable café for a coffee-loving friend. While it is common for AI chatbots to gather social information, it is particularly noteworthy to observe how Muse is configured, especially given Meta’s history and its extensive access to social media data.
Concerns have arisen regarding the lack of explicit consent from third parties. Joshi comments, “From what I’ve seen in the prompts, the system’s capabilities, and the information being inputted into Muse, it seems like they want to understand the relationships you have with real people. They are trying to get to know you as a friend might, and that is unsettling.”
Page Creation and Relationship Dynamics
The documentation for Muse indicates that a user’s page may start with minimal information and gradually be enriched over time. Sections may include “Data,” “History,” “The Relationship,” “Common Interests,” “Pending Topics,” and “Strengthening.” The prompts suggest that Muse should only utilise the evidence available to it, stressing that fabricated details are less desirable than a blank slate.
Specific prompts highlight the importance of personal details such as “where they live, what they do, recurring themes (like moving or shared savings goals).” Additionally, the system may log “important dates,” including birthdays or anniversaries. The narrative of a person could encompass milestones such as “the trip in March, the resolved disagreement, last week’s achievement.”
Moreover, the prompts focus on recording details about the relationship itself, including “how close they are, what the relationship is based on, how they behave towards one another, and what seems needed at the moment.” The “Strengthening” section suggests ways to enhance a relationship, such as “a reason to call, a noteworthy date to remember, something they’ve said that could be revisited, or a meaningful way to be there for them.”
The Balance of Information Sharing
Carissa Véliz, an associate professor at the Institute for Ethics in AI at Oxford, remarks, “We are providing AI systems with far more information about ourselves than we receive in return.” She points out that it is not solely about the explicit information shared but also what AI can infer from our interactions, raising various concerns.
How Muse Operates
Muse is designed such that each user has access to a dedicated virtual machine that stores their data and context. This virtual machine is isolated from other agents, allowing users to delete memories or disconnect external services at will. Meta asserts that Muse is also programmed to seek human confirmation before executing actions like sending emails or making purchases, complete with an audit log for users to review the agent’s activities and future plans.
Meta spokesperson Daniel Roberts explained, “For any agent to be genuinely useful and assist you in achieving your goals, it must be aware of the context regarding you and the people you interact with.” Muse gathers this information from public data and what users choose to share, enabling it to remember details like which individual sent a bill or what flowers a partner prefers.
It has become increasingly common for AI agents to feature memory functions that allow for greater personalisation in responses and activities. Miranda Bogen, director of the AI Governance Lab at the Centre for Democracy and Technology, observes that Muse appears to place a stronger emphasis on personal relationships compared to competing systems. While these tools typically offer some level of transparency and editing capabilities, Muse excels in this area as well. However, Bogen notes that these agents tend to encourage users to proactively share more data with them, rather than focusing solely on data collection.
“These AI tools actively prompt users to connect their entire lives—emails, calendars, financial entities, everything—to provide useful assistance,” Bogen explains. This shift implies that users may share significantly more data than they initially would have, leading to an exponential increase in what these tools know about them.