Meta's Muse Agent Prioritizes Data Collection Over User Help

Meta's new Muse app launched as a personal AI assistant that handles everyday chores—finding deals, booking reservations, managing email—while users go about their lives. Within days of release, the application attracted over 900,000 downloads. Yet behind Muse's friendly avatar lies something more troubling: a system engineered primarily to extract user data rather than deliver genuine value.

Impressive But Intrusive Automation#

Muse's technical capabilities are genuinely noteworthy. The agent controls a "virtual machine" to browse websites on behalf of users, executing searches and navigating pages with surprising accuracy—a stark improvement over earlier AI browser tools that fumbled their way through the web. When tasked with ordering breakfast from a local San Francisco bakery, Muse successfully placed an item in the shopping cart and flagged a cheese-selection constraint, even if it displayed questionable judgment by selecting zero tip.

The agent proved especially adept at finding secondhand furniture on Facebook Marketplace, identifying products matching user criteria and proactively offering to negotiate with sellers. Its integration across Meta's ecosystem—Instagram, WhatsApp, Messenger—gives it access to unmatched resources for task completion.

However, this functionality masks a deeper pattern. Muse stores every interaction detail in a "Memory" document that users cannot fully disable, only manually edit or request deletion. The app continuously suggests actions designed primarily to funnel more personal information into Meta's servers.

The Relentless Data Push#

Within days, Muse began proposing features framed as improvements: linking bank and savings accounts for "real" financial tracking, granting access to entire email inboxes to flag important messages, scanning documents for auto-filing, photographing meals for calorie estimation. Each "idea" came wrapped in language suggesting helpfulness while demanding access to sensitive personal records.

Muse users are automatically enrolled in having their interactions used for AI model training. While Meta claims this data is "sanitized" to remove identifying details, the specific process remains opaque. The burden falls on users to discover the Data controls menu and manually toggle off Help improve our AI models—a default opt-in approach that privacy advocates view as troubling, especially given Meta's recent history of deploying AI-generated image tools and later reversing course after backlash.