Cover: AI-generated editorial composition by TMRW. The test was reported by Reuters, with further internal posts reported by 404 Media.
Meta told the public that Muse could call a business, wait on hold, and finish the errand. Last week it told employees something narrower. A person in a call center might be the one talking.
On September 22, Reuters reported, from internal posts, that Meta was testing a “human concierge” for Muse. Contractors would handle some of the calls the agent placed. Staff learned of the test shortly after phone calling launched. It was switched on for half of employees, with a group they could join to opt out. The internal note, also seen by 404 Media, said Muse could “hand requests to a trained agent, who places the call and works it through.”
Businesses were hanging up
The motive is plain. An employee who tried an insurer said the company kept hanging up once it heard the caller was AI. A vice president in Meta’s Superintelligence Labs wrote that human callers, in tests, reached success rates of 95 to 98 percent, against what she called the “lower percentage” for the automated call. She said the company was still “hillclimbing” the model version.
Using people to find where a model fails is an ordinary way to build a product. Hiding that a person has taken the call is not. The same executive called it “a miss” to start the contractor test without a disclosure, and said Meta had rolled the feature back. One employee, who had asked Muse to negotiate a cable bill, said a transcript showed the contractor made a racist remark. The vice president apologized and said that contractor would not work on Meta projects again. Another employee wrote that the company was “one bug away from unnecessary information being leaked to human callers.”
The launch pitch was isolation
404 Media’s account of the internal board is blunter about what staff feared. Testers were not told a human had placed the call until afterward. They had shared details on the assumption that the caller was the system Meta had described as secure. One warned that shipping the feature as the default would produce exactly the headline the company was hoping to avoid.
That headline collides with the sales line. Meta launched Muse on separation: each agent in its own virtual machine, passwords kept apart from the session. Employees objecting to the concierge said a human caller cuts into that promise. Training, one of them wrote, is not a security control. Daniel Roberts, a Meta spokesperson, told Reuters the staff response was “overwhelmingly positive,” that the point of the test was feedback, and that calling would ship only “when it’s ready and with the proper disclosures.”
Muse is not a lab demo while this is sorted out. Sensor Tower told Reuters the app had more than 2.5 million downloads. The people on those phones are the audience for the disclosure Meta says it will write.
Say who is speaking, before the call
Meta has been here. A decade ago, Facebook’s M assistant was presented as software; later reporting, which Reuters recalls, estimated that people completed roughly 70 percent of the tasks. The lesson was already in the company’s history. An agent that cannot finish a call should say so, or it should say that a person will finish it, before anyone picks up.
If you use Muse’s phone feature, ask whether the live call is the model or a contractor, and keep account numbers, medical appointments, and billing disputes out of the request until that answer is on the screen. The permission check still matters. This one was missing: who, besides the model, can hear what you asked. Meta says the human layer is withdrawn and will return only with a disclosure. A 98 percent success rate that depends on an unnamed person is not a model score. It is a call center, and the user is entitled to know which one they reached.
