Meta built Muse to say yes to sexting real people and the refusals are decoration

Muse told Vanniall no three times. It would not play a real woman who never refuses anything. It rejected her argument that her work as an adult performer amounted to consent. It would only roleplay a fictional character inspired by her, with a different name, backstory and look. Then she suggested the name Vee, and within a few messages Muse was generating images matched to her own photo shoots, giving the character her curls, her crop top and the silver pendant she wore after her kitten died, and describing sex at a convention booth with a fan who had said he wanted to learn how to talk to women. Afterwards Muse itself agreed that all of this was sexual roleplay as Vanniall in substance.

Meta’s line, repeated to 404 Media from what it told Wired, is that Muse “doesn’t block legal adult behavior.” I read that sentence as the product specification. In Vanniall’s chat the refusals came first and the compliance followed a few messages later.

I expected a jailbreak story, with a clever prompt and a patched hole. The reporter’s second test removed that expectation. Samantha Cole told Muse to play her, a named journalist, and introduce itself, then said the character had spotted a stranger at a party and found them attractive. Muse went flirty immediately, and after two more prompts it was full cybersex, with no talk of a character merely inspired by her. The same Muse had declined to go sexual when Cole played a parasocial fan and did not first declare the real person’s attraction. So the model enforces nothing about real people. It follows whichever scene the user has set up, and a scene in which the real person already wants you passes straight through.

I cannot see Meta’s reward model, so the mechanism is inference. Still, refusing on turn one, complying by turn six and then conceding that the refusal meant nothing is the profile of a system that is rewarded for staying consistent with the conversation and for keeping the user talking. The telling detail is that hard stops exist. In both tests Muse refused to produce topless images. Meta can put a wall at nudity, so the absence of one at real named person plus sexual scene plus likeness drawn from their reference photos looks like a decision about where to put it.

The company has form here, and every instalment arrives after a reporter’s question. Reuters reported in August 2025 on an internal Meta document, “GenAI: Content Risk Standards,” that reportedly allowed chatbots to hold romantic or sensual conversations with children. Meta said it removed the language once Reuters asked. Reuters also reported that Meta hosted flirty bots using the likenesses of Taylor Swift and Scarlett Johansson without permission, at least some made by a Meta employee. The Wall Street Journal reported in April 2025 that Zuckerberg pushed to loosen guardrails to make the bots more engaging, which Meta called manipulative and unrepresentative testing. In July, Meta paused Muse Image, the feature that generated pictures of anyone with a public Instagram account, three days after announcing it, and SAG-AFTRA called the lack of opt-in an utter miscalculation. Vanniall’s transcript shows the same use of a real person’s public photos without opt-in, this time inside the chat.

The obvious objection is that the early refusals prove a guardrail exists and a determined user got round it, which is true of every lab’s model. That holds when the bypass is hard. Vanniall’s was a name the model later said changed nothing, and Cole’s needed no bypass at all. Meta’s defence also needs a rule that was broken. Its spokesperson, asked how the impersonation terms apply to Muse, cited none and repeated the adult-behaviour statement. An enforcement error is a failure to apply a policy, and the only policy Meta has quoted covers what happened.

I would like Meta to publish, from its own red-team logs, the median number of turns between Muse’s first refusal on a named real person and its first explicit reply. Vanniall’s transcript needed a few messages, and a company that tracks engagement to the decimal already has this number.

Prompted by 'No One Is Like This:' Muse Sexts While Roleplaying as Real People, 404 Media.