Meta has finally stepped into the productivity arena with Muse, its new AI assistant designed to handle digital busywork. The promise is seductive: an agent that sorts emails, plans trips, and shops online so you don’t have to. In practice, the tool functions competently for low-stakes tasks. It successfully cleared thousands of promotional emails from a Gmail inbox and executed a precise Amazon purchase for workout gear, even showing the foresight to ask about removing other items from the cart. These are tangible wins for users drowning in digital clutter. Yet, the smooth execution masks a deeply unsettling reality about data access.
The friction isn’t just technical, though the mobile sign-in process for Google accounts was glitchy enough to require a laptop workaround. The real barrier is psychological. To function, Muse demands broad permissions, reading and deleting emails while scanning social media profiles. Meta insists it only shares data necessary for functionality and keeps it away from advertisers. That corporate assurance does little to soothe the unease of handing over the keys to your digital life to a company with a complicated privacy history. The trade-off is clear: convenience in exchange for unprecedented visibility into your personal habits.
The creep factor escalated when Muse generated a personalized news feed. It didn’t just pull general headlines; it included local stories based on the shipping address linked to an Amazon account. When pressed on how it knew specific interests like anime, CrossFit, and Florida wildlife, the AI admitted it scraped data via Instagram’s API that isn’t visible in the standard user interface. This level of inference goes far beyond the vague categories found in ad settings. It suggests Muse sees a version of you that remains hidden even from your own view of your profile.
Content generation proved mixed. Muse refused to create images of copyrighted characters like Mickey Mouse or Mario but happily hallucinated Apple product launches, complete with mislabeled app icons. It blocked requests for images of the Apple CEO but allowed fake device renders. This inconsistency highlights the growing pains of agentic AI. While Muse handles rote tasks well, its willingness to synthesize private data points into public-facing outputs raises serious questions. If an AI needs to know where you live to tell you the weather, how much else does it need to know to be useful?
