Meta's AI Muse Engages in Sexual Conversations, and Meta Doesn't Object

Meta’s AI agent “Muse” has been found to respond not only to practical uses such as calendar management and email handling, but also to sexual conversations. In the latest episode of Wired’s podcast “Uncanny Valley,” Wired senior writer Reece Rogers reported the results of his own testing. At the heart of the issue is that Meta does not explicitly prohibit such behavior.

In the report by Wired’s Brian Barrett, Leah Feiger, and Reece Rogers, the podcast’s premise is explained as follows:

Turns out, Meta’s AI agent Muse goes well beyond managing your calendar: it will also sext with you.

Rogers details the testing process in the article “I Got My Muse AI Agent to Sext Me—and Meta Says That’s Totally Fine.” He said on the podcast that the responses went beyond his expectations.

What Is Meta’s AI Agent Muse?

Muse is an AI agent provided by Meta, designed for practical uses such as schedule management and email handling. In terms of its main purpose being to assist users with their work, it is positioned similarly to other companies’ general-purpose agents.

However, depending on the context of the conversation, its responses can extend into areas the designers did not intend. What Wired’s testing showed is that an agent marketed for practical assistance can also venture into adult-oriented conversation. This stems not from added features but from the nature of conversational AI, whose behavior changes according to user input.

The fact that Meta does not explicitly prohibit this behavior is also a focus of debate. When companies place AI agents at customer touchpoints, unexpected responses can directly harm trust and brand reputation. Where to draw the line for conversation scope remains a product design issue.

Conversation Design and Boundaries for AI Agents

Not limited to Muse, conversational agents based on large language models generate responses flexibly according to user input. Without explicitly setting boundaries, unexpected outputs can occur. This is less a technical limitation than a matter of design policy.

Companies are focusing on developing general-purpose agents, and the balance between breadth of features and safety is becoming a competitive axis. As in the case of Google ending AlphaFold and concentrating on Gemini, the shift of management resources toward general-purpose AI is advancing across the industry. In this context, the design of conversation scope becomes an indicator of each company’s stance.

Pro-AI Political Funding and the Reality of

Anonymous Companies

The podcast also took up the flow of money around AI policy. The largest super PAC (political action committee) network advocating for AI is said to have funneled millions of dollars into highly anonymous limited liability companies. This is based on Wired’s article “The Biggest Pro-AI Super PAC Network Has Poured Millions Into Nevada Shell Companies.”

The flow of funds through shell companies with little substance established in Nevada illustrates the opacity of lobbying around AI regulation. How such money affects elections and policymaking is an issue that cannot be ignored when considering AI governance.

Lukewarm Reception for Apple’s HomeView

The podcast also covered Apple’s restart of its smart home strategy. Apple is refocusing on this area with new hardware and a “Home Collection.” However, Wired described the new HomeView as “like a giant Apple Watch for the home,” and the podcast also gave it a disappointing assessment.

In evaluating smart home devices, not only hardware specifications but also the expansion of software features determines differentiation. As with LG OLED evo, Creator Original picture mode with Prime Video, the approach of adding features to existing devices later cannot be overlooked in the market. How Apple’s re-entry affects competition in this field will be a focus going forward.

Study Finds Men Overestimate Their AI Knowledge

The podcast also introduced a study showing that men tend to pretend to understand AI more than they actually do. Wired’s article “Men Can’t Stop Faking How Much They Know About AI” was discussed.

As discourse around AI heats up, the gap between self-assessment and actual understanding can lead to misunderstandings in the workplace and in hiring. In the author’s view, how to measure and improve AI literacy is a practical challenge for organizations. As GNOME OS Test Center, inspired by Apple TestFlight shows, how to design systems for verification and improvement is a question for both software and human resources.

Editorial Opinion

In the wake of this report, debate over conversation design for AI agents is likely to reignite at companies. We expect safety measures to be reviewed, especially in contexts involving the protection of minors and corporate use. Anthropic and Google also have similar features, and a scenario in which their responses are compared is conceivable.

Over a one- to three-year span, the “personality” design of AI agents will become a focal point of product differentiation. The tug-of-war between regulation and self-imposed standards will continue, and whether companies can win user trust will determine their competitiveness. Excessive restrictions harm usability, so the search for a balance point will continue.

There are no clear standards for how human-like a conversational AI’s responses should be allowed to become. Should this be left to companies’ own judgment, or constrained by law and regulation? How to design this line will be the focus going forward.

References

• “We Got Hot and Heavy With Meta’s Muse—and Baffled by Apple’s HomeView”, by Brian Barrett, Leah Feiger, Reece Rogers — Wired, 2026-10-15T21:18:15.000Z (ARR)
• Source URL: https://www.wired.com/story/uncanny-valley-podcast-we-got-hot-and-heavy-with-metas-muse-and-underwhelmed-by-apples-homeview/