Meta’s new personal assistant, Muse, has become a viral hit with millions of consumer downloads, but leaked internal operating files reveal the AI agent compiles hourly, detailed profiles on every person in a user’s life. Independent researchers extracted the software instructions from the chat interface, exposing a system designed to track real relationships, important dates, and shared histories.
Researchers Extract Software Instructions From Chat Interface
Independent AI safety and security researcher Karan Joshi extracted an extensive array of Muse instructions and system prompts by asking the chat interface to copy and share its own software files, WIRED reported. Joshi shared his findings with WIRED after discovering how the AI agent organizes and presents information. Meta maintained that it intended for these software files to be accessible in the interest of transparency. The documentation shows how Muse responds to prompts about sensitive or politicized topics while operating on dedicated virtual machines for each user.
Hourly Profiles Target Family, Partners, and Colleagues
One of Muse’s core instructions creates a page for every person in a user’s life, compiling data hourly on family, partners, friends, colleagues, collaborators, and people followed on social platforms, according to WIRED and dnyuz.com. These structured text files function as a memory bank to make relationship suggestions, such as restaurant recommendations for a coffee-loving friend. Muse’s documentation indicates a page starts sparse and fills out over time with sections covering facts, history, the relationship itself, shared commonalities, open threads, and strengthening advice. Meta’s instructions dictate that Muse must rely strictly on available evidence, noting that invented details are worse than leaving a page empty, dnyuz.com reported.
Experts Question Deep Social Tracking and Data Collection
The system instructions direct Muse to record where people live, what they do, recurring threads like apartment moves or shared savings goals, and dates that matter such as birthdays and anniversaries, dnyuz.com noted. Backstory data can include arguments that got resolved or recent milestones, alongside relationship depth indicators. Joshi described the deep tracking as an attempt to know users like a friend, calling the functionality pretty creepy. Carissa Véliz, an associate professor at Oxford’s Institute for Ethics in AI, stated that users give AI systems far more information than they receive, pointing out concerns over both explicit disclosures and what models infer or piece together from other data sources, dnyuz.com reported.