Meta's AI assistant Muse was recently promoted as capable of making phone calls on behalf of users, handling everything from restaurant reservations to salon appointments. But behind the scenes, a significant portion of those calls is being placed not by the AI itself but by human agents working in call centers.

What Meta Announced and What 404 Media Found

Earlier this month, Meta executives publicly celebrated Muse's expansion into outbound calling. Ryan Fox, the principal engineer leading the project, posted on X that the beta had been broadened to cover US businesses. Alexandr Wang, Meta's chief AI officer, echoed the announcement with his own post about the phone beta.

Beneath those public messages, however, the internal picture told a different story. According to internal communications reviewed by 404 Media, Meta told its own employees that the company had "added a human agent layer for calls to get completed" and that the human-assisted version was "ready for company dogfooding," the industry term for internal product testing before public release.

How the System Actually Works

An internal message board post seen by the outlet describes the feature as follows: Muse does not simply dial a number on its own. Instead, it hands user requests to a trained human agent, who then places the actual phone call, works through the conversation with the business, and completes the requested task. The system is designed to report back to the user with a transcript and a summary.

Meta has not disclosed how frequently calls are routed to humans versus handled entirely by the AI, making it unclear whether the human layer is a rare fallback or the default mode of operation.

Employee Concerns About Privacy and Perception

Meta employees raised pointed objections when they learned about the human agent layer. The most immediate concern was privacy. When a user shares personal information to arrange a doctor's appointment or another sensitive matter, that data is being transmitted to a human being, not just to an AI system. One employee noted that Meta's justification, which cited extensive contractor training for data security, amounted to nothing more than a policy claim rather than an actual security mechanism.

There was also the question of transparency. Some testers reported that they were not informed a human being was making the call until after it had already taken place. The assumption that one was interacting with a secure AI system was not communicated before the conversation began.

An employee summed up the broader risk in blunt terms on the internal board: if this feature launched publicly with humans operating behind the scenes by default, it would provoke a media backlash portraying Meta as an AI company whose technology was not advanced enough to do the job on its own.

The Broader Pattern of AI Products Using Hidden Human Labor

This is not an isolated case. Several companies have faced scrutiny after it emerged that their AI products relied on human workers to perform tasks they claimed were automated. The practice, sometimes called "human-in-the-loop" marketing, erodes user trust when it is not disclosed upfront.

What makes Meta's situation notable is the scale and the visibility of the announcement. The company positioned Muse as a flagship AI product at its developer conference, and the calling feature was framed as evidence of AI capability in the real world. For the human layer to exist underneath that pitch, without being part of the public narrative, raises questions about how Meta evaluated the risk of such a disclosure gap.

One Researcher Who Actually Used the Feature

Not all feedback was negative. Ravid Shwartz Ziv, an AI researcher at Meta, posted on X that Muse had successfully called a business's customer service on his behalf, navigated the phone tree, waited on hold, and spoke with a human representative who resolved the issue. He noted that the representative on the other end did not appear to know they were talking to a system that involved AI at any point.

Other users reported more mixed results, with some claiming the call function failed entirely or that the person on the receiving end hung up when the call came through.

Meta's Response

A Meta spokesperson characterized the internal testing phase as a normal part of the development process, stating that employee feedback is essential for implementing safety and privacy protections before any public rollout. The company said it is working with merchants to refine the feature and that it would only release it with appropriate disclosures.

Whether that commitment to transparency holds once Muse leaves the testing phase remains to be seen. For now, the gap between what the public-facing announcements describe and what the internal records reveal is the most striking aspect of this story.