Meta's Muse Is Better at Surveilling Than Helping Me
wired ·
WIRED’s Reece Rogers tested Meta’s Muse, an AI agent for tasks such as web searches, shopping, reservations, and inbox management. The app browsed relatively reliably, found local couches on Facebook Marketplace, and followed up, but some actions still required user approval and payment information. Muse stores a long-term memory of interactions, while repeatedly encouraging Rogers to connect email, bank accounts, and personal documents. Users are automatically enrolled in sharing interactions for AI-model training, although Meta says the data is sanitized; users can disable this in Data controls, while memory cannot be fully turned off with a toggle. Privacy advocates warn that the agent’s data practices, tracking implications, and potential commercial incentives could undermine trust and influence what users see. Researchers also caution that highly personalized agents may encourage oversharing, passivity, and dependence that weakens independent decision-making. After the experiments, Rogers deleted Muse, concluding that its appetite for personal data outweighed its practical help.
wired