Meta's Muse Is Better at Surveilling Than Helping Me

Reece Rogers’s WIRED hands-on with Muse, Meta’s free personal AI agent, is less a gadget review than a map of what an always-on assistant extracts once it lives in WhatsApp and a phone-shaped app. Sensor Tower put first-week downloads above 900,000. The pitch is deals, reservations, and inbox help; the default avatar looks like a beige Ewok crossed with Labubu. The experience feels like texting a friend—thumbs-up acknowledgments, background work, a virtual machine that browses for you.

Rogers ordered breakfast at Kahnfections in San Francisco: Muse added a biscuit sandwich, asked for a Stripe card, picked cheddar, noted a Swiss request, and chose no tip. On Facebook Marketplace it found cheap local couches, offered to message sellers, and followed up the next day. That is real utility. The same product ships a memory document for preferences you can wipe or edit—but no toggle to turn memory off entirely—and an Ideas tab that constantly nudges connecting bank accounts, scanning the inbox, photographing documents and meals, and tracking passport or license expiry.

We’re for agents that actually finish chores. Muse’s bakery run and Marketplace follow-ups are the kind of useful AI we want more of—delegated browsing that saves time instead of generating another chat transcript.

The surveillance shape is the product, not a bug in the UX. Users are auto opted-in to AI training on interactions Meta calls “sanitized”; the opt-out sits under Data controls as “Help improve our AI models.” EFF’s Rory Mir’s line lands: talking to an AI is talking to the company that hosts it. EPIC’s Calli Schroeder flags the opt-out default as a red flag and points at Instagram’s deepfake opt-in-then-pull history. Meta spokesperson Emil Vazquez calls built-in protections and user controls the reality and any darker suggestion “ludicrous.” Meta Superintelligence Labs VP Tarek Sheasha defends the default as good for collective model improvement and says confidential VMs that cryptographically block Meta access are planned later this year. Until those ship, the trust ask is large.

Meta’s own safety blog says agent actions—reservations, Marketplace—can “indirectly influence” ads on Instagram even when data is not directly shared with advertisers. Axios quotes CAO Alexandr Wang looking for Muse revenue beyond ads. Hugging Face’s Margaret Mitchell warns agents can disengage users, worsen oversight, risk cognitive degradation, and that personalization pulls more disclosure. Rogers deleted the app after testing; the last ping asked to connect more apps so Muse could do more.

I’m watching whether confidential VMs arrive with proof users can verify, whether the training default flips to opt-in, and whether “indirectly influence” ads becomes the quiet monetization path while the agent keeps asking for the bank.

Muse is available as a dedicated app and through WhatsApp. Rogers’s piece ran September 20, 2026, after a week of heavy download numbers and a product still arguing that helpfulness and a training-default opt-out can coexist.