When someone asks me what I think of Meta’s Muse, I say “I’m terrified.”
Am I being hyperbolic? Maybe. Read on and decide for yourself.
Ten million sign-ups in six weeks. A quarter of them use it every day. The average shopping basket is over $200, and Meta had to ask Stripe to raise its credit limits because the volume was higher than anyone expected.¹ Ten million is a small number for Meta, but that was six weeks in, and the pool is Meta’s 2B+ global users. Trust me, they’ll get there.
We should be going into this with eyes wide open on the data and what gets done with it; whose side the agent is on; and what happens to choice in real terms. Unfortunately, I’m not aware of any organizations that have AI literacy as a KPI (and there is not a cent of revenue to be found in meaningful human understanding) so we’re jumping in eyes wide shut.
Opt Me Out.
Last month TIME got ahold of Muse’s internal instructions. We learned that it keeps an hourly-updated dossier on each user. It keeps one on the people you mention who don’t use it. It’s told to infer the goals you haven’t said out loud. It keeps notes on which hours of the day you respond to a nudge.²
I’ll admit, those of us who’ve spent careers in marketing salivate at that kind of data: it’s what every brand on earth has wanted since marketing was invented. The data Meta gets across all its platforms is super rich (iOS 14 notwithstanding). If the data sat in a database and a human pulled a report from it, this might be a different story.
Two things change what’s happening from a marketer’s dream to this human’s nightmare.
The first is that training on that data is on by default, for you and for the people in your conversations. I haven’t used Muse. I’m in it anyway: TIME’s reporting says Muse keeps a file on the people its users mention, so the minute a friend brings me up, into the database I go. Muse can now use what someone else said about me (which may or may not be accurate, by the way) to learn my patterns, my shopping, my friendships, how to make an ad land on me, or for pretty literally whatever else they want.
And, since literacy is low and utility is high, the opt-out is a setting most people won’t go looking for.³ Marketing data you consented to is marketing. Muse data? It’s next level and it’s concerning. Worrisome has not stopped a lot of folks in Menlo, historically.
The second, and much bigger, issue is what’s reading the file.
The redirect
What’s reading the file is an AI built by the company whose feed rewarded extremes, outrage, and anything that got a response and kept you scrolling. Twenty years of UX testing, models and apps built to hold attention, plus your social graph. Now the same company, with one of the world’s richest consumer-behavior data sets, has an AI acting on it, and the target moves from your attention to your wallet.
Stuart Russell, the Berkeley computer scientist who wrote the AI textbook, has been warning for years what a system like this does to the person on the other end: an algorithm rewarded for predicting you discovers that the easiest way to predict you is to change you.⁴
It doesn’t need a villain. Nobody at Meta has to decide to manipulate anyone. The AI will do it on its own, because that’s what it’s rewarded for. Russell, along with several of the “godfather” types, signed a 2025 petition asking governments to ban certain uses of AI outright, among them mass persuasion and systems that pass as human.⁵
Muse’s voice mode is explicitly instructed not to call itself AI.²
Whose Side Is It On?
In theory the platform is aligned with you. You asked for the best black puffer jacket; it wants to find you the best black puffer jacket. Then Meta’s rules arrive.
Muse’s shopping instructions, pulled by a research firm in September, say the top five picks must come from “well-known” merchants, open-web browsing stops at three stores and ten products, and any listing priced far outside the category’s normal range is thrown out.⁶ The price rule ostensibly has a laudable purpose: it’s a scam filter (which is a head-scratcher given that Meta’s own internal projection had about $16 billion of its 2024 revenue coming from ads for scams and banned goods).⁷ It is also a warehouse-sale filter, a small-maker filter, and a filter for the one jacket that costs three times the others because it’s made of sustainable materials that are as toasty as goose down and will outlast them. The blunt instrument of a price filter can’t tell the difference. The outlier options are just out, by default. Maybe the rules change. Probably not before brands rush to conform and lose the very things that made them different, or stay off the shelf and fail for never showing up.
Meta is paid per transaction. Zuckerberg told investors monetization is coming “somewhat sooner than our normal playbook,” and that users can connect Muse to improve the personalization of their ads across Meta.⁸ The KPIs driving decisions are revenue and engagement. A company on the verge of superhuman computing is going to have a lot of motivation and the data to figure out how to get you to buy buy buy (and who knows what else).
So is the agent’s objective misaligned with yours? Likely. An agent whose rules decide which stores exist, which offers are allowed, and which merchants count as real is not necessarily an agent whose incentives point at you.
And, you won’t know when it’s working against your own interest. It handles the discovery, the negotiation, the decision. There is a lot of trust baked into that buy button, whether you know it or not.
Brands will eventually pay a lot of money to advertise on Muse. “No ads” is, was, and always will be: “no ads for now.” And, that’s not going to make anything better.
Next Stop: Beigeville
Okay, put all that aside (if you can). There’s still the matter of choice, in real terms. And taste, as we enter Beigeville: the place where everything has been forced to the mean, and the best stuff never makes the cut because the best stuff is often an outlier, not just on price. If your query includes “best,” as so many do, aren’t you asking for something exceptional?
Beigeville has one shelf. On it, the five best-selling black puffer jackets in America, all from stores you’ve heard of. Not on the shelf, according to those rules: the one from the Vermont outfitter nobody’s heard of, the one with the weird pocket in just the right place that would have become your favorite coat. And, folks, some of us seek out WEIRD. My friend Sharon uses an expression: “you’re not going to meet it on the street.” That’s MY favorite thing to buy – the dress with the inkblot jellyfish pattern and crazy collar, the green sweater with the tiger on it, the $20 clearance Target top that’s bright mustard and black but somehow doesn’t make you look like a bumblebee. This isn’t just stuff. It’s self-expression, and the question is whether those brands, products and outliers survive the defaults, patterns, and engines we’re building.
From an optimization point of view, wandering looks like failure. You’ll get a warm-enough jacket, and you’ll never know what you’d have fallen in love with.
The tech will get smarter and easier to use. And it will narrow our aperture.
The Giant Easy Button
Everyone but Meta has a lot to lose here and we’ll all sign up in droves anyway. Most people aren’t that curious, and the efficiency is hard to combat. Time savings. Cost savings. Drudgery be gone. And, there is no room in performance metrics for conscientious objectors in the C-suite, not in this bruising market.
But, very little good actually comes from pressing the giant easy button. People who build a thing themselves value it more: in the experiment that named the IKEA effect, people who assembled a plain storage box bid about 60% more for it than people handed one already built.¹¹ Choosing for yourself makes the good choices feel better and the bad ones sting more; handing the choice to something else buys you insurance against the sting, at the cost of ever owning the result.¹²
Why “Trust” Is the Right Word
Meta’s privacy pitch rests on something it calls your Virtual Machine: a private computer in Meta’s cloud, one per user, where Muse does its browsing, form-filling, and buying on your behalf, with a second AI standing guard over what leaves it. The promise is that Muse won’t share the data in your Virtual Machine with Meta’s ad business. Notably, Meta never says what’s in that machine, or what isn’t. Its own safety page concedes it can get in when it needs to, and the terms can change at any time.⁹ You can drive an 18-wheeler through what Meta has left itself room to do.
Two weeks ago, a judge in New Mexico handed down a $25 billion penalty against Meta for telling users they controlled their data when they didn’t.¹⁰ Now Meta is asking people to trust a shopping agent on a very similar promise.
So. Do you trust Mark Zuckerberg?
Footnotes
Meta, “Introducing Muse,” September 8, 2026. Meta Q3 2026 earnings call, October 21, 2026. TechCrunch, October 21, 2026. Payments Dive, October 23, 2026.
TIME, “Meta’s Muse AI Agent Is Building a Dossier on You,” October 6, 2026. Meta did not dispute the findings.
Meta, “How We Built Safety Into Muse,” September 2026. Surfshark audit via TechRadar, October 2, 2026.
Russell, S., Human Compatible (Viking, 2019). Carroll, Dragan, Russell & Hadfield-Menell, “Estimating and Penalizing Induced Preference Shifts in Recommender Systems,” ICML, 2022. Russell co-wrote Artificial Intelligence: A Modern Approach and founded Berkeley’s Center for Human-Compatible AI.
Global Call for AI Red Lines, September 22, 2025, red-lines.ai.
PPC Land, October 12, 2026, reporting Profound’s extraction of Muse’s shopping instructions on September 25. Single account; Meta has not commented.
Reuters, “Meta is earning a fortune on a deluge of fraudulent ads, documents show,” November 6, 2025. Meta called the 10% figure “rough and overly-inclusive.”
Meta Q3 2026 earnings call, October 21, 2026: Meta will “get a percent of the value” of transactions.
Meta, “How We Built Safety Into Muse,” September 2026. Meta Terms of Service, updated July 2026. Olaizola Rosenblat, M., Just Security, October 26, 2026.
State of New Mexico v. Meta Platforms, penalty judgment, October 21, 2026. Under appeal.
Norton, Mochon & Ariely, “The IKEA Effect: When Labor Leads to Love,” Journal of Consumer Psychology, 22(3), 2012.
Botti & McGill, “When Choosing Is Not Deciding,” Journal of Consumer Research, 33(2), 2006.