Meta's Muse AI Assistant Gets Tangled Up in How It Works

Meta is pitching its AI assistant Muse as a powerful tool, but for users it stirs up a faint sense of unease. Part of the blame lies with the new Mac app, which—if permitted—gains access to Messages, Calendar, and Notes. The real oddity, however, lies elsewhere: for all its cleverness, Muse can't clearly explain how it's built.

A Conversation That Shouldn't Have Happened

Jason Aten, an editor at Inc magazine, posted screenshots of his conversation with Muse on Threads. The assistant suddenly began asking questions about a conversation Aten had been having in Messages. The problem is that, by Aten's own account, he had never given Muse access to his messages.

When the user asked how the assistant knew the contents of the conversation, it replied: "I saw the notification previews, not the message history. I didn't read your texts." Further questioning led the assistant into a dead end. "The honest answer: I can't describe the exact mechanics," Muse admitted. "All I know is that the connected Mac app provides access to notifications as one of its capabilities, and they reach me through device sync."

That sounds, to put it mildly, less than reassuring.

Meta: The Assistant Is Just Making Things Up

David Singleton of Meta Superintelligence Labs tried to set the record straight in the comments under the post. First he listed all the permissions Muse needs to read messages, including full disk access for the Mac app, and stressed that every feature is connected only with the user's explicit consent.

The most curious part comes at the end: according to Singleton, Muse "doesn't monitor notifications on the Mac, but syncs data from Messages only after the user has specifically enabled access." In other words, the problem wasn't covert surveillance of Aten's messages—it's that the assistant simply didn't understand what it was talking about.

Singleton acknowledged the company's responsibility: when Muse claimed it was syncing "device notifications," it got tangled up explaining its own feature and produced an incorrect version. Meta apologized for the erroneous answer and promised to teach the assistant to describe its inner workings more accurately.

It's not entirely reassuring, but there's a logic to it. We already know that chatbots don't reveal their secrets—instead they say what people want to hear. And sometimes, it seems, they don't even know how they work themselves.

Source: The Verge