From my experience using Meta Muse, it seems to know a lot about me, simply by remembering our conversations. But as it turns out, according to a Wired report, internal instructions pulled from the AI agent show that it can organize information about the people around you too, potentially building a dedicated page for everyone from your spouse and friends to coworkers and people you follow.
I use Muse, so naturally I wanted to know what mine had collected. I asked it what it knew about me; luckily, the answer was surprisingly sparse. But after seeing how Muse is designed to build its understanding over time, that may say more about how much access I've given it than what the agent is capable of learning.
Muse can create a page for every person in your life
Independent AI safety and security researcher Karan Joshi extracted Muse's internal instructions and system prompts by essentially asking the agent through its normal chat interface to copy and share its own software files.
Joshi shared what he found with WIRED, including instructions for an hourly process that can create "a page for every person in the user's life."
And Meta appears to mean just about everyone. The instructions the researchers found mention family, partners, friends, colleagues, collaborators and even people you follow.
Those pages may initially be sparse, but they're designed to fill up over time. According to the report, Muse can organize what it knows into sections covering facts, history, the relationship, things you have in common, open threads and even ways of strengthening the relationship.
That could mean relatively ordinary information such as where someone lives or what they do. But it can get considerably more personal.
"They're trying to know you like a friend, which is honestly pretty creepy," Joshi told WIRED.
There is at least one important guardrail here. Muse's instructions tell it to use the evidence available to it rather than fill in the blanks, with an invented detail considered worse than leaving part of someone's page empty.
You don't necessarily have to use Muse to end up in one
This is the part that caught my attention. Muse isn't only learning about the person who signed up for it. To understand that person's life, it also needs to understand the people in it.
That means information about you could potentially exist inside someone else's Muse simply because you're part of that person's life.
To be clear, this isn't evidence that Meta is secretly creating one giant Muse profile for every person who doesn't use the service. Each Muse user has a dedicated virtual machine where their own data and context are stored, and Meta says other users and agents can't access it.
But it does create a strange privacy situation. Because you might never download Muse, connect an account or tell the agent anything about yourself. Someone you know could still tell their Muse about you, and those details can become part of the context the agent uses to understand that person's relationships.
That's a very different type of AI memory than what ChatGPT and Claude use.
Why Meta says Muse needs this information
According to the report, Meta wants Muse to help you achieve your goals. The company gave an example of Muse remembering that someone who sent you an invoice is actually the plumber who previously worked on your bathroom or remembering which flowers your spouse likes.
But that's exactly the type of memory that makes this extra complicated. A personal AI agent that doesn't remember anyone in your life wouldn't be much of a personal assistant. If I tell Muse that my husband has an appointment Thursday or that one of my kids needs something for school tomorrow, I want it to remember that information the next time it's relevant.
But usefulness and privacy are pulling in opposite directions here. Every additional piece of context can make the agent more helpful while simultaneously expanding what it knows about you and everyone around you.
What Muse knew about me — and my family
After reading about these internal pages, I went directly to my own Muse and asked what it knew about me and my family.
As it turns out, it didn't know much outside of what I've personally told it. It knew I had three kids and that we are a "soccer family." It knew what I do for a living and where we live.
But when I looked at the people in my life (outside of my household), the information became even thinner. No names were attached to anyone. That lines up surprisingly well with Muse's own instructions, which say a person's page may begin sparse and become more detailed over time.
I've also been fairly conservative about what I've connected to Muse. Someone who gives an AI agent access to considerably more of their digital life could potentially give it far more information to work with.
My little experiment shows what Muse can tell me after relatively limited use. The internal instructions show what it's designed to do as that relationship grows.
The bigger issue with personal AI
Personal AI assistants are designed to understand your relationships, remember what's happening in your life and eventually act on that information.
As Oxford Institute for Ethics in AI associate professor Carissa Véliz told WIRED, the concern isn't only the information we explicitly give these systems. It's also what they can infer and what they can piece together from multiple sources.
Meta says Muse users can wipe memories or disconnect external services, and the agent includes an audit log where users can see its activity and future plans. Those controls become even more important as you connect more of your life.
Follow Tom's Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds. Subscribe to Tom's Guide on YouTube and follow us on TikTok.
More from Tom's Guide
- Amazon drivers are getting AI smart glasses — but my warehouse tour reveals their 1-million strong automated robot workforce is the real story
- Meta Muse vs ChatGPT Dots — I compared the new AI agents, and they’re not as similar as they look
- Why Siri never worked like ChatGPT — former Siri AI researcher explains what Apple had to change
