메타 '뮤즈'는 일 돕기보다 감시에 능한 AI 에이전트

Comment Loader Save Story Save this story Comment Loader Save Story Save this story I saw Meta’s latest app, Muse , cross-promoted on another Meta-owned platform, Instagram , and decided to download the AI assistant. “Hand off everyday tasks like finding deals, booking reservations, and managing your inbox,” the pop-up read. “Muse keeps working while you get on with your day.” It’s free to use, and easily connects to other data sources, like my email and bank account. Muse’s default avatar is a beige-colored cross between an Ewok and a Labubu , with its arms outstretched, presumably to scoop up all my data . Meta’s Muse is a mainstream version of AI agents already popular in Silicon Valley for automating personal tasks, like OpenClaw and Instinct . It’s a hit for the company: the Muse app has been downloaded over 900,000 times in its first week, according to Sensor Tower . The user experience is less like prompting a chatbot , and more like texting a friend with task requests—the AI agent acknowledges messages with a thumbs-up emoji and gets to work in the background. Muse is also available through Meta’s WhatsApp platform. But after a few days with Muse, I’m convinced this AI tool is more obsessed with collecting data about me than actually getting stuff done. “Muse is the first personal AI agent built for everyone, with built-in protections and user controls that put people absolutely in charge of how they use it—any suggestion we didn’t build with that in mind from the beginning is ludicrous,” Meta spokesperson Emil Vazquez tells WIRED. Good Clicks Muse’s ability to surf the web is genuinely impressive. When you ask it to complete a task, Muse uses a “virtual machine” to browse the internet on your behalf, running searches and clicking around on different web pages. When I tested similar tools last year, like the now-defunct ChatGPT Agent , the AI’s clicks were erratic and error-filled. Muse rarely seems to get lost while browsing. For example, I asked Muse to order breakfast from a local bakery called Kahnfections . It’s a popular spot for tourists visiting San Francisco, so I requested one of the most-purchased options for pick-up. Muse visited the bakery’s website and added a biscuit sandwich with garlic aioli, cheddar cheese, egg, and bacon to my cart. I was asked to add a credit card via the Stripe payment processor so my agent could complete the purchase. “One tweak the site forced: cheese is single-select, so I picked cheddar and left a note asking them to add Swiss too if they can,” read Muse’s explanation of the breakfast order when it asked for my final approval. Notably, Muse selected the “no tip” option. A bit rude! (I ended up walking down to the shop, ordering something random and delicious.) Meta relies on its sprawling kingdom of platforms to get the word out about Muse. WhatsApp and Messenger for easy onboarding and integrations; prime placement on Instagram, with a large library of Reels videos for agents to search through. But the Facebook Marketplace tie-in was where Muse impressed the most in my testing. Muse quickly found cheap, local couches for sale within my parameters and offered to message sellers to arrange pick-ups. The agent is designed to follow up with users, so Muse nudged me the next day to consider getting the couch I said was my favorite. Muse stores details about your interactions and preferences in a “Memory” document found by tapping the avatar in the top middle of the screen. “Your curated long-term memory: durable facts, preferences, and commitments,” the description in Muse reads. While you can send a chat request for Muse to wipe memory data and you can manually edit the file, Meta doesn’t currently offer a toggle to turn off the memory feature altogether. Data Collector Under the app’s “ideas” tab, Meta’s Muse constantly reminded me to connect more information to the agent. The day after I said I wanted help saving money for a fancy vacation, Muse offered to “wire your savings tracker to your real balance” if I just tapped a few buttons and linked the data from my checking and savings accounts. “Then the weekly summaries and drift alerts work off real numbers instead of hand-entered estimates,” it said. This experience, where Muse offered further personalization in exchange for more data, was emblematic of my overall impression of the app. “Folks don’t recognize that when you talk to an AI, you are talking to the company hosting the AI,” says Rory Mir, director of open access at the Electronic Frontier Foundation . “These chat windows—that we’re used to being connections between us and another person—are really just us directly putting information into Meta servers about ourselves.” Every suggested interaction with Muse started to feel like a guise for me to upload more data about myself for Meta to see. Its “idea” was to let Muse scan my whole inbox and flag must-read messages. Its “idea” was to snap a pic of important documents for Muse to fill out and file. Its “idea” was to photograph my meals for Muse to estimate the food’s calories. “Tell me when your passport and license expire,” another one read. Great idea, Muse. Muse users are automatically opted in to having their interactions with the agent used for AI model training . Meta says this data is “sanitized” to remove identifying information before AI training, though it’s unclear how the process works. The data collection may include what Muse uses as context when accessing the data sources you have connected, like an email inbox or bank account. You can turn off this data collection in the settings for the Muse app, under Data controls . Disable the toggle that says Help improve our AI models . Consumer advocates see Meta forcing Muse users to opt out of AI data training as a major red flag. “This tells me they have not learned from past mistakes and undermines their argument that you should trust Muse with all your information even further,” says Calli Schroeder, senior counsel at the Electronic Privacy Information Center . Schroeder points to Meta recently opting all adult Instagram users into what was essentially an AI deepfake tool , then pulling the feature a few days later, as akin to Meta’s history of “manipulative tactics” and breaking user trust. Tarek Sheasha, a software engineer and vice president at Meta Superintelligence Labs, defends this data collection as critical for training new models and improving performance, in a blog post about Muse’s release . “We think this is a good default—every Muse user gets a better personal agent as we all collectively use the product and help the model understand the intricacies of human life,” Sheasha wrote. Meta also plans to release “confidential” versions of its virtual machine for users later this year that “cryptographically and verifiably prevent” the company from accessing the data inside. Secret Shopper While your Muse data isn’t directly shared with advertisers, the agent browsing on your behalf can still impact what users see via web-tracking algorithms. Muse’s safety blog suggests that if the agent makes a dinner reservation for you or picks out a product on Facebook Marketplace, then that action could “indirectly influence” the ads you see on Instagram. “I do see this as a continued trajectory of their main business model of controlling what people see, putting ads in front of them, and collecting more data,” Mir says. It’s definitely rational to be anxious about handing over your credit card to an agent . Even if you could trust these AI tools wouldn’t be tricked into spending money in ways you don’t want, outsourcing your shopping decisions can have unintended consequences. Consumer advocates see this purchasing automation as potentially opening the door for companies to prioritize certain brands or other products in an agent’s output, based on sponsorship deals. Meta's Chief AI officer Alexandr Wang hinted in an interview with Axios that it’s looking for more revenue opportu