Meta’s Muse AI: Impressive Yet Unsettling
Meta has introduced its new Muse assistant, marking a significant step into the realm of AI-driven productivity tools. This AI agent is designed to alleviate your workload by assisting with tasks like online shopping, managing emails, planning trips, and more. I decided to put Muse to the test, especially considering Meta’s history of focusing more on entertainment than productivity. While the assistant generally performed as expected, the unsettling amount of personal information it gathered about me overshadowed the experience.
My first task for Muse, which operates through a cloud-based virtual computer, was to sift through my Gmail inbox and delete unnecessary messages. This required linking Muse to my Google account and granting it permission to access, read, and delete my emails. The mobile setup was a bit rocky; the Google sign-in screen kept glitching, redirecting me back to the Muse website instead of the app. Eventually, I managed to get it working on my laptop, where it successfully cleared out thousands of promotional emails and updates. However, I felt uneasy about allowing Muse to access all my emails, given Meta’s past privacy concerns.
Meta claims that Muse is designed to exchange data with third-party apps only as necessary for functionality and assures users that it does not share information with advertisers. Despite this reassurance, I felt apprehensive but decided to proceed by linking my Amazon account to Muse. I asked the AI to find workout tops in a specific size, style, and color. Before finalizing the order, Muse noticed other items in my cart and asked if I wanted to remove them, which was a pleasant surprise. After I approved the purchase, it successfully ordered just the tank tops.
I then explored some of Muse’s other features, which include generating AI podcasts, creating images and videos, and developing interactive documents known as “artifacts.” However, when I requested images of a “cartoon mouse with red pants” and a “cartoon guy with a red hat, blue overalls, brown hair, and a mustache,” Muse declined. On the other hand, it quickly generated a video and several images of a product launch event featuring devices with Apple’s logo. One image depicted a foldable device resembling an iPhone, an Apple Watch, and AirPods, all displayed on stands with the Apple logo, while a presenter stood in front of a backdrop featuring the iconic bitten apple. Interestingly, the fake iPhone displayed Apple’s native app icons, but it mistakenly labeled the App Store as “Photos” and misnamed Maps, Music, and the actual Photos app as “Calendar.” Muse also refused to create an image of an “Apple CEO.”
Things took a strange turn when I navigated to the “Feed” tab, where Muse generates a news feed based on a text prompt. I used the default prompt provided by Meta: “Make me a feed about my interests. Keep the tone clear and direct. Ensure it is quick to skim. Try to avoid clickbait.” Muse returned several AI-generated summaries of current events, including details about Apple’s iPhone Duo launch, dangerous surf conditions in California, and tankers destroyed in the Iran war. It also included news stories from the state where I was visiting family, which Muse admitted it sourced from my Amazon shipping address.
Curious about what else Muse might know about me, I asked it to share insights based on the Instagram and Facebook accounts I used to access the AI. It listed my specific interests, which included anime, CrossFit, Labrador retrievers, Florida wildlife, and ’90s/2000s nostalgia—topics that frequently appear in my Instagram Reels feed. When I inquired about the source of this information, Muse explained, “Honestly, Instagram doesn’t show you that exact list anywhere in the app—I read it through the account’s API data, which surfaces more than the UI does.” It noted that the closest users can see is their ad topics, which indicate what advertisers can target. I checked my Instagram settings, and while the ad topics included some broad categories like workouts and music, they lacked the detail Muse provided. My “Your algorithm” settings also didn’t reflect the same interests listed by Muse.
Like other AI tools I’ve tried, Muse performs well with low-stakes tasks. I appreciate that it generates a list of “ideas” offering personalized suggestions, such as comparing products, finding and canceling forgotten subscriptions, and tracking upcoming Pokémon card releases. Muse also maintains a list of “goals” based on interactions with the AI.
While Muse effectively handles mundane tasks in the background, the extent of personal details it gathered from sources I couldn’t access left me feeling uneasy about allowing it to manage my emails and credit card information. This level of concern will be a significant challenge for Meta to address.
