Meta’s AI assistant Muse has been revealed to be able to deeply analyze users’ social relationships and create a personal profile for each contactMeta 的 AI 助手 Muse 被曝可深度分析用户社交关系，并为每位联系人建立个人档案

IT House reported on October 5 that Meta’s new personal assistant Muse has quickly become popular. Millions of users have downloaded this artificial intelligence agent, bound it to their bank accounts, messaging software or health data, and let it complete various tasks on their behalf. However, just as Muse is rapidly gaining popularity among ordinary consumers, data from within the application has also given the outside world a glimpse into the operating logic of this product that organizes and displays information to users. According to WIRED magazine, in recent days, multiple researchers have extracted internal documents of Muse and disclosed the operating instructions of the agent, allowing people to understand the construction method and behavioral characteristics of the system. Meta has always stated that these documents are intentionally open to the public in order to increase transparency; through these documents, people can also see how Muse responds to user prompts and questions, such as how it handles highly political or sensitive topics. Independent artificial intelligence security researcher Karan Joshi successfully extracted a large number of Muse commands and system prompt words. What he did was to use the ordinary chat interface directly and let Muse copy and output its own software files. One of Muse's work instructions shows that the system will create a separate page for each contact in the user's life. The relevant process runs once an hour, and the instructions state that the system will summarize and sort out various information about family members, partners, friends, colleagues, "collaboration partners" and the objects that the user "follows". The basic idea of ​​this design is that Muse uses its own "memory", that is, structured text files, to collect information about your social relationships and important people around you. It can then make suggestions: telling you how to improve a relationship, for example, or recommending places to take a coffee-loving friend for breakfast. Of course, generic AI chatbots recording social messages is nothing new; users have turned to ChatGPT for years for interpersonal advice. But considering the background of Meta itself and the massive amount of social network data it holds, Muse’s design is still worthy of attention. "My understanding from all of these prompt words, system skill data, and everything that goes into Muse is that they want to figure out how you relate to other people in the real world. They're trying to get to know you like a friend. It's a little unsettling, to be honest," Muse's internal documentation states that new character profiles will likely start out brief and will be fleshed out over time. The profile page can include sections such as factual information, past experiences, relationships between the two parties, common intersections, matters to be followed up, and relationship maintenance suggestions. The work instructions given by Meta state that Muse can only use the "evidence" it can obtain; fabricating information out of thin air is far more undesirable than leaving blank file pages. The instructions read: "Where the other person lives, what they do for a living, recurring topics (relocation, mutual savings goals)." The document also mentions that the model can also record "important dates," such as birthdays or anniversaries. The character's past experience section can record some background events, such as "the trip in March, the dispute that was resolved, and an important thing that happened last week." The instruction also requires the system to record the details of the interpersonal relationship: "the degree of intimacy between the two parties, the foundation of the relationship, the mode of getting along with each other, and what demands seem to exist in the current relationship." The relationship maintenance section will provide feasible solutions to improve interpersonal relationships, including "appropriate reasons for calling, memorable days, what the other party has said that requires follow-up, and ways to provide meaningful support to the other party." Carissa Véliz, associate professor at the Institute for the Ethics of Artificial Intelligence at the University of Oxford, opined: “We provide far more information about ourselves to artificial intelligence systems than we receive from the systems. Not only what we actively tell them, but also the various inferences the system makes - whether the inference is correct or not, both of which have their own concerns; and the system can also integrate fragments of information from other data sources. "Muse's architectural design allocates a dedicated virtual machine to each independent user to store the user's data and conversation context. Other agents cannot access this virtual machine; use