Meta’s Muse AI works and creeps me out
The Verge's Emma Roth tested Meta's new Muse AI productivity assistant, which manages emails and shopping, but felt unsettled by how much data it gathered.
Stock photo for illustration only, not from the actual event
- Meta launches Muse AI as its first dedicated productivity assistant
- The cloud-based AI agent handles tasks like email sorting and shopping
- It can generate images and videos, including mocked-up tech products
- The reporter expressed discomfort over the AI's deep data collection
Meta has officially entered the AI-powered productivity space with the launch of its new Muse assistant, designed to take over routine chores like online shopping, email management, and trip-planning. Emma Roth, a news writer at The Verge, decided to put the tool to the test to evaluate its real-world performance from a company historically focused on entertainment rather than productivity. While the assistant largely delivered on its functional promises, the unnerving depth of personal information the AI managed to autonomously gather cast a shadow over the entire experience.
One of the initial tests involved Tasking Muse—which operates via a cloud-based virtual computer—with sorting through a Gmail inbox and clearing out unnecessary promotional messages. This required linking the assistant to a Google account and granting permissions to read, access, and delete emails. Although mobile setup encountered glitches that redirected the browser back to the website, the desktop version successfully wiped thousands of promotional messages and updates. The process left the user uneasy, given Meta's controversial history regarding user privacy over the years.
Stock photo for illustration only, not from the actual event
Meta maintains that Muse was built to exchange data with third-party applications strictly when necessary for user assistance, and insists user information is never shared with advertisers. Reassured enough to proceed, Roth linked an Amazon account to the assistant and requested workout tops matching specific sizing, styles, and colors. Before executing the transaction, Muse thoughtfully noticed existing items in the cart and checked if they should be removed, successfully purchasing only the requested tank tops upon confirmation.
The rise of autonomous AI agents capable of executing complex multi-step actions across different third-party platforms represents a massive paradigm shift in software usability. However, it also introduces unprecedented privacy and security challenges, as users must surrender profound access tokens to deeply personal digital ecosystems like email and retail accounts.
Shifting to creative features, Muse is equipped to generate AI podcasts, create images and videos, and build interactive documents known as artifacts. When prompted to generate visuals for a product launch event featuring Apple devices, the system successfully rendered a foldable iPhone concept, Apple Watch units, and AirPods on branded stands. However, hallucinations occurred as it mislabeled the App Store as “Photos” and misidentified Maps, Music, and Photos as “Calender.” The model firmly refused requests to generate an image of an “Apple CEO” or specific cartoon characters.
"Muse rattled off very specific interests linked to my Instagram account."
Emma Roth
The experience took a strange turn upon exploring the “Feed” tab, where running Meta's default interest-based prompt generated current event summaries alongside localized news stories from the state where Roth was visiting family. Muse admitted it retrieved this location data from the shipping address on the connected Amazon account. Pushing further by asking what else the AI knew based on linked Instagram and Facebook profiles, Muse rattled off highly specific personal interests with unsettling accuracy.
Source: The Verge
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment