You ask your AI agent to negotiate a cable bill. A contractor picks up the phone, speaks to the company on your behalf, and makes a racist remark during the conversation. You find out none of this happened the way you assumed, not unlike cases of secretly tracking users without their knowledge.
That is not a hypothetical. Reuters reported it on September 22, 2026, roughly two weeks after Meta launched its personal AI agent, Muse, in the United States.
The Reliability Gap That Pushed Meta Toward Human Labor
The business logic behind human-assisted calling is real, but it does not explain what Meta failed to disclose.
Businesses were reportedly reluctant to engage with AI callers, which created a real reliability problem for Muse’s phone-calling feature. The feature was designed to handle routine tasks like booking appointments or checking product availability. According to Reuters, human-handled calls reached a success rate of approximately 95–98%, compared with substantially lower rates when AI completed calls on its own.
That gap explains the business logic. It does not excuse what came next.
Meta tested a human concierge feature with roughly half its employees during the week before Reuters’ report. An internal announcement said Muse could “hand requests to a trained agent, who places the call and works it through.” Participants in the test were not clearly told this handoff could happen.
One contractor, handling a bill-negotiation call for a Meta employee, made a racist reference during the conversation. Meta apologized to the employee and said the contractor would not work on any of the company’s projects again.
The incident matters beyond the offense itself. It demonstrates that human involvement inside a product users experience as software can introduce unpredictable conduct. No safety monitor, no Sentinel agent, no virtual machine architecture can catch that reliably in real time.
A responsible rollout of any feature like this requires, at minimum:
- Explicit disclosure before a human takes over a call
- User consent specifically for human handling of personal conversations
- Clear limits on what information contractors can access or hear
- Auditable records of every instance of human involvement
- A straightforward way to cancel and revert to AI-only handling
A vice president at Meta Superintelligence Labs acknowledged the lack of disclosure “was a miss,” and Meta rolled the feature back. Spokesperson Daniel Roberts said employee feedback had been “overwhelmingly positive” and described the test as groundwork for building proper safety and privacy protections before any public release.
This Script Has Been Performed Before
Meta is the latest in a long line of companies that marketed automation while quietly relying on human labor.
This is not a new story with a new company. It is a recurring feature of consumer AI products.
Meta itself ran Facebook M starting in 2015, a digital assistant that contractors reviewed at launch while the company described it as AI supervised by people. Shopping app Nate presented itself as an automated checkout tool. The U.S. Department of Justice later alleged that workers in the Philippines and Romania manually completed between 60% and 100% of purchases during parts of 2021. Presto Automation disclosed in 2023 that roughly 70% of orders through its AI drive-through system required intervention from off-site human agents, a dynamic that echoes debates around AI age laws and accountability in automated systems.
The pattern repeats often enough to notice: the marketing promises automation, and the operations deliver labor.
Meta has not said it is abandoning the human concierge feature permanently. Its stated position, per Reuters, is that the feature may return once proper protections are in place. Hybrid AI-human systems can achieve meaningfully better results than fully automated ones, and that is worth acknowledging.
What is not acceptable is launching a system where contractors speak on your behalf, hear details of your personal finances, and operate without your knowledge. Before Muse picks up that phone again, you should know exactly who is on the other end of the line.
