Meta's Muse AI Agent Prioritizes Data Collection Over Utility

Meta's newest AI product, Muse, has attracted over 900,000 downloads in its first week according to Sensor Tower. The personal assistant app promises to handle everyday tasks like booking reservations and finding deals while users focus on other priorities. But hands-on testing reveals an application more interested in accumulating user data than delivering genuine productivity gains.

Muse operates differently from traditional chatbots. Users text task requests to the agent, which acknowledges messages with emoji reactions and works in the background using a virtual machine to browse websites and complete actions. The app integrates with email, banking, and Meta's existing platforms including WhatsApp and Facebook Marketplace.

Impressive browsing, aggressive data requests

The technology behind Muse's web navigation represents a genuine advancement. When asked to order breakfast from a local San Francisco bakery, the agent successfully navigated the website, selected menu items matching the request, and added them to a cart for approval. This stands in contrast to earlier AI browsing tools that frequently made erratic clicks and errors.

The Facebook Marketplace integration proved particularly effective, quickly identifying affordable local furniture options within specified parameters and offering to message sellers directly.

However, nearly every interaction with Muse includes prompts to connect additional data sources. The app's "ideas" tab continuously suggests linking bank accounts for "real balance" tracking, scanning entire email inboxes, photographing documents for automated filing, and capturing meal images for calorie estimation. One prompt requested passport and license expiration dates.

Privacy concerns and opt-out defaults

Muse automatically enrolls users in AI model training using their interaction data. Meta states this information is "sanitized" to remove identifying details before training, though the specific process remains unclear. Users must manually navigate to Data Controls settings and disable the "Help improve our AI models" toggle to opt out.

"Folks don't recognize that when you talk to an AI, you are talking to the company hosting the AI," says Rory Mir, director of open access at the Electronic Frontier Foundation. "These chat windows are really just us directly putting information into Meta servers about ourselves."

Calli Schroeder, senior counsel at the Electronic Privacy Information Center, sees the opt-out default as evidence Meta "has not learned from past mistakes." She points to Meta's recent brief rollout of an AI feature that essentially created deepfakes of Instagram users without explicit consent.

Meta defends the data collection as essential for model improvement. Tarek Sheasha, vice president at Meta Superintelligence Labs, wrote that collective usage helps "the model understand the intricacies of human life." The company plans to release "confidential" virtual machine versions later this year that would cryptographically prevent Meta from accessing data inside.

Beyond privacy: cognitive risks

Researchers identify additional concerns beyond data collection. Margaret Mitchell, chief ethics scientist at Hugging Face, co-authored recent research showing agentic tools may cause "cognitive degradation" as users become overly reliant on automated decisions from confident-sounding external sources.

The high personalization levels may also encourage users to share increasingly private information as the agent mirrors their communication style and preferences.

While Muse's agent browsing represents technical progress, the aggressive data collection tactics and opt-out training defaults raise substantial questions about whether the productivity benefits justify the privacy trade-offs.

Why it matters

Muse represents Meta's entry into the emerging AI agent market, where tools automate personal tasks rather than simply answering questions. With nearly a million downloads in week one, the app's data collection practices and default settings could establish problematic norms for how companies handle personal information in agentic AI systems. The tension between personalization benefits and privacy costs will likely define consumer AI adoption in coming years.

These details were first reported by WIRED.