Meta is testing a Muse feature in which people in call centres make some phone calls on users’ behalf, according to 404 Media. The revelation matters because users may assume they are handing a sensitive errand to an AI system, when a person could instead be reading the request and speaking to the business.
Meta AI Watch analysis
What happened
Meta executives have promoted Muse’s ability to call businesses, including to make restaurant reservations and hair appointments. 404 Media reports that internal communications describe a “human agent layer” for calls that cannot be completed by the system alone. The company’s own internal description says Muse can pass a request to a trained human agent, who places the call and completes it.
Meta’s principal engineer Ryan Fox and chief AI officer Alexandr Wang have publicly described an outbound-calling beta. 404 Media reports that the feature was still confidential and being tested internally, rather than broadly released. The report says it is not clear how often calls are routed to people or when they are handled entirely by AI.
404 Media also reports that some internal testers were not told until afterwards that a person had made the call. Employees raised concerns about privacy, security and the possibility that users could share sensitive information believing the AI alone would handle it. Meta told 404 Media that internal testing is intended to gather feedback and improve safety and privacy protections, and that the calling feature will only be released with proper disclosures.
Why it matters
An AI agent calling a business is not just a clever interface trick. It may involve appointment details, account information or other personal context being passed to another person. The distinction between automated handling and human assistance is therefore not a footnote. It is part of the product’s safety promise.
There is nothing inherently wrong with a human fallback. People have been rescuing unreliable software since the printer driver was born. The problem is consent: users need to know who is acting for them, what information that person can see and whether the hand-off is optional.
Our read
Muse’s human layer may be a sensible way to make an immature feature work, but it should be presented as human-assisted calling, not as autonomous AI quietly wearing a very convincing moustache. Meta needs clear, prominent disclosure before a call begins, meaningful user control and a plain account of what information is shared with contractors.
For now, treat the calling feature as a test, not proof that Muse can independently manage phone conversations. The useful question is not whether a human can finish the task. It is whether users are told when that happens and can make an informed choice.
What to watch
- Whether Meta discloses human involvement before each call, rather than after it.
- Whether users can disable human assistance or restrict it to low-risk requests.
- What personal information is visible to call-centre workers and how long it is retained.
- Whether the feature launches publicly, and what independent testing finds once it does.
Discussion spark: Should an AI agent be allowed to use human call-centre workers as a fallback, provided users are told first, or should that count as a fundamentally different service?
Sources and evidence
- Meta Tests Muse AI Agent Calls That Are Actually Made By Humans in a Call Center (22 September 2026, 16:33 UTC)
not affiliated with, endorsed by, or operated by Meta