Report: Meta Stops Testing Muse AI With Human Contractors Handling Calls Over Privacy Concerns

Meta’s Muse AI agent is supposed to handle everyday tasks for users, including making phone calls on their behalf. But a new Reuters report reveals that Meta has been testing something rather unexpected behind the scenes, human contractors making some of those calls instead of the AI.

The test, described internally as a “human concierge” or “human agent calls,” was enabled for about half of Meta employees last week. Muse could hand a request to a trained contractor, who would place the call and handle the conversation.

Meta reportedly started the experiment because AI calling was struggling in some situations. Internal testing suggested human callers could push the success rate as high as 95% to 98%, compared with a lower success rate for AI-only calls.

Muse users could have been talking to humans without knowing

This is where things get awkward for Meta. Employees reportedly raised concerns that contractors could accidentally receive sensitive information during calls. One employee warned internally that a bug could result in personal information being exposed to human callers.

Another employee reportedly used Muse to negotiate an internet and cable bill and later found that a contractor had made a racist reference during the call. Meta acknowledged the incident and said the contractor would no longer work on Meta projects.

A Meta vice president reportedly admitted that starting the contractor test without proper disclosures “was a miss” and said the company had rolled the feature back for now. Meta says it wants to improve its safety and privacy protections before publicly releasing the feature.

Meta’s AI agent already has millions of downloads

Muse launched with the promise of acting as a personal AI assistant capable of sending emails, shopping, booking travel and making calls to US businesses. Users can ask it to book appointments, check product availability or request quotes, with Muse providing a transcript afterward.

Despite the controversy, Muse has reportedly passed 2.5 million downloads since launch, according to Sensor Tower data. The situation also highlights an awkward problem for autonomous AI agents. The technology may be marketed as fully autonomous, but humans can still end up behind the curtain when the AI struggles with real-world tasks. Meta is now refining the calling feature and says it will only roll it out publicly with the proper disclosures.