Exploring Muse, Meta’s Latest AI Assistant
Recently, I noticed that Muse, Meta’s newest application, was being cross-promoted on another Meta platform, Instagram. Intrigued, I decided to download the AI assistant. A pop-up claimed, “Take control of everyday tasks like finding deals, making reservations, and managing your inbox. Muse keeps working while you carry on with your day.” The application is free to use and connects seamlessly with other data sources such as my email and bank account. The default avatar for Muse resembles a beige hybrid of an Ewok and a Labubu, with arms outstretched, presumably to gather all my data.
Muse serves as a public-facing version of the AI agents already popular in Silicon Valley, designed to automate personal tasks, akin to OpenClaw and Instinct. The app has seen tremendous success, amassing over 900,000 downloads in its first week, as reported by Sensor Tower. The user experience feels less like instructing a chatbot and more akin to texting a friend with task requests: the AI agent acknowledges messages with a thumbs-up emoji and quietly gets to work in the background. Muse is also accessible via Meta’s WhatsApp platform.
Impressive Navigation Capabilities
After using Muse for several days, I became convinced that this AI tool is more focused on collecting data about me than on completing tasks. Emil Vázquez, a spokesperson for Meta, asserts, “Muse is the first personal AI agent designed for everyone, with built-in protections and user controls that allow individuals to have absolute control over how they use it; any suggestion that we have not designed it with that in mind from the outset is ridiculous.”
The ability of Muse to navigate the web is genuinely impressive. When tasked with completing an action, Muse employs a “virtual machine” to surf the internet on my behalf, performing searches and clicking through various websites. In contrast to similar tools I tried last year, like the now-defunct ChatGPT Agent, Muse navigates with far greater accuracy. For instance, I asked Muse to order breakfast from a local bakery called Kahnfections, a popular spot for tourists in San Francisco. I requested one of their best-selling items for pickup. Muse accessed the bakery’s website and added a sandwich with garlic aioli, cheddar cheese, egg, and bacon to my cart.
I was prompted to add a credit card via the Stripe payment processor to allow my agent to complete the purchase. “A small adjustment the website required: only one type of cheese can be selected, so I chose cheddar and left a note requesting Swiss cheese as well, if possible,” Muse explained as it sought my final approval for the breakfast order. Notably, Muse opted for “no tip,” which felt somewhat impolite.
Integration and User Experience
Meta leverages its extensive platform empire to promote Muse. WhatsApp and Messenger facilitate onboarding and integrations, while a prime spot on Instagram, complete with a vast library of Reels videos for agents to explore, adds to its visibility. However, I was particularly impressed by Muse’s integration with Facebook Marketplace during my trials. Muse quickly located affordable sofas for sale in my area that matched my criteria and offered to message sellers to arrange collection. The agent is designed to keep track of users, reminding me the next day to consider purchasing the sofa I mentioned as my favourite.
Muse retains details about your interactions and preferences in a document labelled “Memory,” which can be accessed by tapping the avatar located at the top centre of the screen. “Your selected long-term memory: lasting data, preferences, and commitments,” reads Muse’s description. While you can request chat-based deletion of memory data and edit the file manually, Meta currently does not provide an option to completely disable the memory feature.
Data Collection Concerns
In the “Ideas” tab of the application, Muse persistently reminded me to connect more information to the agent. The day after I expressed a desire to save for a luxury holiday, Muse suggested linking my savings record to my actual balance by simply pressing a few buttons to relate my current and savings accounts. “This way, weekly summaries and deviation alerts will be based on actual figures rather than manually entered estimates,” it indicated.
This experience, where Muse offered greater personalisation in exchange for more data, characterised my overall impression of the application. Rory Mir, Director of Open Access at the Electronic Frontier Foundation, clarifies, “People don’t realise that when you talk to an AI, you’re talking to the company that hosts it. These chat windows, which we’re used to seeing as connections between us and another person, are merely a way to directly input information about ourselves into Meta’s servers.”
Perceptions of User Interaction
Each suggested interaction with Muse began to feel like an excuse for me to upload more personal data for Meta’s scrutiny. One “idea” was to allow Muse to scan my entire inbox and highlight messages I shouldn’t miss. Another involved photographing important documents for Muse to fill out and archive, while another suggested taking pictures of my meals for Muse to calculate calorie intake. “Let me know when your passport and driving licence expire,” suggested another prompt.
Muse users automatically consent to having their interactions used for training AI models. Meta claims that this data is “anonymised” to remove identifying information prior to AI training, although the specifics of this process remain unclear. Data collection may include what Muse uses as context when accessing the data sources you’ve connected, such as an email inbox or bank account. You can disable this data collection in Muse’s application settings under the Data Controls section by switching off the option labelled Help us improve our AI models.
Consumer Advocacy and Privacy Issues
Consumer advocates view the requirement for Muse users to opt out of data training as a significant red flag. Calli Schroeder, Senior Legal Advisor at the Electronic Privacy Information Center, remarks, “This indicates that they haven’t learned from past mistakes and further undermines their argument that you should trust all your information to Muse.” Schroeder highlights that Meta’s recent inclusion of all adult Instagram users in what was essentially a deepfake tool based on AI reflects the company’s history of “manipulative tactics” and its tendency to betray user trust.
Tarek Sheasha, a software engineer and Vice President of Meta Superintelligence Labs, defends this data collection as essential for training new models and enhancing performance. “We believe this is a good default setting: every Muse user benefits from a better personal agent as we collectively use the product and help the model understand the complexities of human life,” Sheasha stated. Meta also plans to release “confidential” versions of its virtual machine later this year for users that would “cryptographically and verifiably” prevent the company from accessing the contained data.
The Impact of Automated Purchasing
While your data from Muse is not directly shared with advertisers, the agent navigating on your behalf may still influence what users see through web tracking algorithms. Muse’s security blog suggests that if the agent makes a dinner reservation or selects a product on Facebook Marketplace, that action could “indirectly influence” the ads you see on Instagram. Mir warns, “I see this as a continuation of their core business model: controlling what people see, showing them ads, and collecting more data.”
Understandably, sharing your credit card information with an agent can be unsettling. While you might trust that these AI tools won’t trick you into spending money unwisely, delegating purchasing decisions can lead to unintended consequences. Consumer advocates argue that this automation of shopping could open the door for companies to prioritise certain brands or products in the agent’s recommendations, based on sponsorship agreements. Meta’s head of AI, Alexandr Wang, hinted in an interview with Axios that the company is seeking additional revenue opportunities with Muse beyond just advertisements, without disclosing specific details.
The Joy of Personal Decision-Making
Personally, I relish long online shopping trips where I can spend hours browsing a website and exit with nothing in my cart. There’s a whimsical inefficiency in searching for leather jackets I may never buy. By exploring the internet on my own, I gradually build up my sense of taste and understanding of what I truly like. If Muse does everything for me, I miss out on the process that leads to my decision-making.
Schroeder envisions a future where AI makes all your important decisions about tastes, preferences, where you go, where you eat, what you do, and what vacations you take. “In a sense, that’s the joy of being human: to try things and discover what you like. That’s why the idea of leaving all that to a machine feels quite terrifying.”
Potential Risks of AI Dependency
Relying on agents like Muse from Meta could diminish our own functionality as these tools collect data about us. Margaret Mitchell, a researcher and Chief Scientist of Ethics at Hugging Face, explains, “The design of AI agents is intended to make users disengaged, which makes users more passive.” Mitchell recently co-authored a research paper examining how such tools are not effectively designed for human oversight and may actually impair users’ ability to evaluate what the agent is doing on their behalf.
Users of agents may come to overly depend on these tools, even more than when interacting with a standard chatbot, as agents have access to more information and can perform more tasks. One potential