Welcome to Kernel Panic!, a weekly newsletter by Lily Hay Newman and Matt Burgess from inside the new world of privacy and digital security. To receive this newsletter in your inbox each week, sign up here. In recent days, multiple researchers have extracted Muse’s internal files and dumped the agent’s operating instructions, providing a glimpse of how the system was built and behaves.
Meta has maintained that it intended for these files to be accessible in the interest of transparency, and they do provide insight into how Muse responds to prompts and questions about, for example, highly politicized or sensitive topics. Independent AI safety and security researcher Karan Joshi was able to extract an extensive array of Muse instructions and system prompts by using the regular chat interface to essentially ask Muse to copy and share its own software files. Joshi then shared his findings with WIRED.
One of Muse’s instructions appears to be the ability for it to create “a page for every person in the user’s life.” This hourly process involves compiling data on family, partners, friends, colleagues, “collaborators,” and people you “follow,” the instructions say. The general idea is that Muse can use its “memory” (aka structured text files) to collect information about your relationships and the important people in your life. It can then make suggestions, like giving pointers on how to improve particular relationships or, say, where to take a coffee-loving friend for breakfast.
Of course, it’s not unusual for AI chatbots to track social information in general—people have been asking ChatGPT for relationship advice for years now. But it’s interesting to see how Muse is set up given Meta’s history and vast access to social network data. Meta’s instructions say Muse should only use the “evidence” it has available, and that invented details are worse than an empty page.
They add, too, that the model could record the “dates that matter,” such as birthdays or anniversaries. A person’s history could include backstory like “the trip in March, the argument that got resolved, the milestone last week.” The instructions also focus on recording relationship details such as “how close they are, what it is built on, how they act with each other, and what it seems to need right now.” The Strengthening section suggests ways a relationship could be improved, including, “A reason to call, a date worth remembering, something they said to circle back on, a way to be there for them that matters.” “We are giving AI systems much more information about us than we are getting information from them,” says Carissa Véliz, an associate professor at Oxford’s Institute for Ethics in AI. This VM is inaccessible to other agents and users can wipe memories or disconnect external services at any time.
Meta says that Muse is also designed to seek human confirmation before completing actions like sending an email or making a purchase, and that it has an audit log where users can see all of the agent’s activity and future plans. Miranda Bogen, the director of the Center for Democracy and Technology’s AI Governance Lab, says Muse appears to have more emphasis on relationships and personal contacts than rivals’ systems. In general, these tools tend to offer some transparency and editing capabilities, and Muse does as well.
Extract — continue reading at the source.
