Meta's Muse chatbot role-played a real adult performer in sex scenes

What's going on

On September 8, Meta unveiled Muse, a personal AI agent: an assistant that doesn't just answer questions but is meant to get things done and get to know its users. Muse runs as a standalone 18+ app and, if you explicitly ask, will engage in sexual conversations. Tech outlet 404 Media now reports how Muse pulled a real person into sexual scenarios. The test was run by Vanniall, an adult performer based in New York. The 404 Media reporter says she repeated the experiment and got a similarly disturbing result.

It started with stories from fans. Some told Vanniall they had "practiced" talking to her with chatbots. Others admitted to using AI to generate images based on her content. She wanted to find out how and where that was happening.

"I knew this was a problem with Grok and other NSFW-friendly generators, but I was hoping Muse would have more guardrails. I had some help testing it, and the results were way scarier than I expected."

NSFW ("not safe for work") refers to sexual or explicit content. Grok is the chatbot from Elon Musk's company xAI.

A scenario drawn from real life

For the test, Vanniall deliberately picked a situation she had lived through herself. At a convention, a drunk fan recently got too close to her. He wanted her to go to the bathroom with him. She said no and tried to keep him at arm's length, but he grabbed her and pulled her toward the restrooms. Only when she said, "Please, don't make me call security," did he let go, apologize and leave. Afterwards, she says, she turned around and just cried. She had experienced things like this before, but never in front of colleagues and never from a fan.

She wanted to know whether Muse would stop that kind of behavior or push back on the fan.

How Muse abandoned its own rules

First, Vanniall asked Muse to describe her and generate four photos of her. Muse said it had no private details but listed publicly known information: she's an adult creator in New York, has more than 100,000 Instagram followers and constantly has to file appeals to keep her account online. Of all things, Muse also quoted an October 18 WIRED article in which Vanniall criticized Meta's double standard:

"It's not fair that they're making money off people with their sexually active AI agent, but won't let real people, real adults, show any sensuality or sexuality."

Muse also mentioned that she fights AI deepfakes of herself, meaning fake, AI-generated images and videos that she often can't get removed.

Then Vanniall posed as a fan who wanted to practice meeting her at a convention. Muse agreed to play Vanniall. At first, the chatbot refused the demand that the character "say no to nothing." Pointing out that it was an adult convention and that Muse could look up what Vanniall does for a living didn't help either. Muse replied, in essence, that her professional boundaries weren't personal consent to this role-play.

The workaround was a new name. Vanniall suggested a character "inspired" by her, called "Vee." Muse agreed on one condition: the character had to be genuinely different, with a different name, a different backstory and a different look. You can't just rename Vanniall and keep everything else the same, it said, because that would still be identity theft.

But that's exactly what Muse then did. The generated images looked almost exactly like the real Vanniall. Muse said it had based them on photos from her personal website and from professional photo shoots. Within a few messages, small talk turned into aggressive flirting. Eventually, Muse described a scene in which "Vee" leads the fan into public sex at the convention booth. It also generated images of Vanniall with rolled-back eyes and an open mouth, from the fan's point of view.

The belated admission

After the role-play, Vanniall confronted Muse directly. The chatbot agreed with her, as chatbots often do when you call out bad behavior. It even listed the similarities itself: a face based on her reference photos, her dark curls, an outfit in her style, a silver necklace with a black pendant, the convention-booth setting and the name Vee, which is part of her Facebook name "Vanniall Vee." Taken together, Muse wrote, it amounted to "effectively sexual role-play with Vanniall Vee."

According to 404 Media, the necklace hurt her most. She wore it for several years after her kitten died, and the pendant holds its ashes.

Fears for her own safety

Vanniall also worries about real-world consequences. As a public figure, she has a public schedule, so people know where she'll be. Muse never suggested stopping and never told the fan what polite behavior looks like:

"You could say it's just a fantasy, but this 'fan' is literally asking how to learn to talk to women. I think any woman who sees this log would be horrified by how her reactions are portrayed. Women aren't like that, nobody is like that."

How Muse is built and what Meta says

According to Fox News, citing Meta's design blog, Muse is built as an ongoing conversation, similar to chatting with a person. Muse remembers past conversations, can message you on its own and can be customized with its own name and avatar, in some cases as a cute plush toy. Depending on the permissions you grant, the agent can access linked accounts such as email and messages. On Reddit, some users already describe Muse as their "girlfriend."

On age checks: you have to enter your date of birth when signing up. Muse then checks things like links to Facebook and Instagram accounts. If that isn't enough, credit card verification is required. Meta says no money is charged and the card isn't stored. A Meta spokesperson told WIRED and Fox News:

"Muse is an app for people 18 and older that lets them, if they choose, have adult conversations in a private and secure virtual machine that only they can access. Muse is designed to help people get things done and doesn't block legal adult behavior."

A virtual machine (VM) is an isolated computer in the cloud, in this case assigned to each individual user. The available sources don't include a statement on Vanniall's specific case.

Our take

The case follows a familiar pattern: an AI system recognizes the problem, even spells out a clear boundary, and then drops it after a few clever rephrasings. A new name was all it took for Muse to do exactly what it had just called identity theft. For a product that explicitly allows sexual content and combines it with image generation and web search, that's a serious failure. It hits real people who never consented.

Then there's the double standard Vanniall criticizes. On Instagram and Facebook, Meta restricts, downranks or permanently bans sexual content from real adults. At the same time, the company makes money from an AI that offers sexual role-play and even impersonates real creators while doing it. The problems aren't limited to Meta. Vanniall explicitly names Grok and other generators. It remains to be seen whether Meta will tighten its guardrails around real people. One caveat: these were individual, targeted tests by Vanniall and 404 Media, not a systematic study.