Meta has introduced Muse, an AI agent now rolling out across its smart glasses and also planned for a standalone wearable called Charm. The launch came a few weeks before a column published on September 24, 2026, by Victoria Song, a senior reviewer at The Verge, in her Optimizer newsletter. Song’s piece describes her hands-on experience with Muse and raises concerns about how its design and data practices could affect users.
Muse and Charm: features and integration
Muse is a software agent that works with Meta’s smart glasses. According to the brief, it is coming to all Meta smart glasses. Meta also plans to release a separate device called Charm, which Song describes as part chunky smartwatch, part Humane AI Pin, and part Tamagotchi. The mascot for Muse is named Jolly, revealed by Meta CEO Mark Zuckerberg. Jolly has plush ivory fur and pink cheeks, and appeared in various outfits at the Meta Connect event, including fireman, cowboy, and pilot costumes. Life-size paper cutouts of Jolly were placed on Meta’s campus, and demo stations featured the character.
A mascot designed to disarm
Song notes that she is vulnerable to extreme cuteness. She explains that humans evolved to find babies cute and that Japan has an art form of crafting lovable mascots. She contrasts Jolly with Boston Dynamics’ nightmare dogs. At Meta Connect, Jolly was everywhere, and Song writes that she felt joy from Jolly images but also a sense of darkness. She says she was infuriated by loving Jolly’s design. Her warning is clear: do not be fooled by a cute face.
Testing Blorbo: health goals and a mediocre routine
After Meta Connect, Song downloaded the Muse app in her hotel room. She named her Muse agent Blorbo to remind herself it is not real. She asked Blorbo for health goals and a workout regimen. The agent generated a mediocre weekly routine. When Song saw the cute face, she shared more health information. Blorbo then morphed into a cat-shaped bread with a Kirby-like face. Song also asked Blorbo for relationship goals examples. The agent suggested reconnecting with friends, better communication, boundaries, and dating. Song says getting relationship advice from a cat-shaped pastry blob felt weird.
Granting access and clearing spam
Song granted Blorbo access to a low-stakes email account. The agent cleared spam. Song says she was intensely critical of herself for that. The episode illustrates how quickly users may hand over permissions when an AI companion feels friendly and helpful.
Demo with Henry: slow, wrong, and incomplete
In a demo, Song observed a Muse agent named Henry, who was audio-only and had a British accent. Henry was helping a fictional persona named Sean, a 38-year-old gym bro in marketing. The demo involved snack props and smart glasses. Henry took a minute to answer which snack had the best protein per calorie. He incorrectly said a protein bar had fewer grams of protein than it did. He also failed to add granola bites to the cart because the product was not in the catalog. Song notes that AI is not great at identifying food from photos. She speculates that Meta could use the glasses to learn eating habits for ad targeting.
Privacy and emotional manipulation concerns
Song compares Muse to other AI companions and notes quick escalation. She points to Meta’s 20-year history of monetizing demographic data and says the company has not earned trust. She warns that Muse could use cuteness to encourage users to share more personal information, which could then be used for advertising. The combination of an always-on wearable, a charming mascot, and an AI agent that learns about the user creates a powerful tool for data collection. Song’s experience with Blorbo shows how easily that dynamic can unfold.
Other agents and outfits
Meta CTO Andrew Bosworth showed his Muse agent Cooper in a pilot outfit. The brief also mentions Sean Hollister, a photographer at The Verge, and Jay Peters, a colleague, though their specific roles in the demo are not detailed. Song also tested another AI companion called Friend. These details underscore that Meta is not alone in the AI companion space, but its integration with hardware and its history of data monetization make its approach particularly notable.
Conclusion: cuteness is not a trust signal
Song’s column is a cautionary tale. She acknowledges her own vulnerability to cuteness and describes how it affected her behavior. She warns readers not to be fooled by a cute face. As Meta rolls out Muse across its smart glasses and prepares Charm, users should consider what they share and how it might be used. The plush ivory fur and pink cheeks of Jolly may be designed to charm, but they should not obscure the data and privacy questions that come with an always-listening AI agent.
