Meta’s Muse AI Quietly Used Human Contractors to Make Calls, Raising Privacy Questions

Meta's new personal AI assistant Muse has been testing a "human concierge" system in which contractors quietly place some phone calls on behalf of users, exposing a significant gap between the company's vision of autonomous AI and the human labor still needed to make some of its most ambitious features work reliably.

The test was disclosed to Meta employees last week, shortly after the company publicly launched Muse's phone-calling capability, according to internal company posts reviewed by Reuters.

Under the system, Muse could hand a user's request to a trained human contractor, who would place the call and handle the interaction with the business. The employee posts said users could ask Muse to perform tasks such as booking appointments, checking product availability or obtaining quotes from contractors.

The arrangement immediately raised privacy concerns inside Meta. Employees warned that contractors working in call centers could potentially hear sensitive information during conversations that users believed were being handled by an AI system.

A vice president in Meta's SuperIntelligence Labs acknowledged in an internal post that it "was a miss" to begin testing contractor-placed calls without proper disclosure. The company subsequently "rolled back this feature" for the time being, she said.

Meta said the experiment was designed to gather feedback and improve the system before any broader release.

A Meta spokesperson said employee feedback had been "overwhelmingly positive" and that the purpose of the test was to develop safety and privacy protections.

"We're working with merchants to continue improving this potential calling feature, and will only roll it out when it's ready and with the proper disclosures," Meta spokesperson Daniel Roberts said.

But the development reveals a fundamental challenge facing the current generation of AI agents. Companies are increasingly presenting these systems as digital workers capable of carrying out tasks independently, but the reliability needed for real-world interactions can still require human intervention.

In Meta's case, the internal testing suggested that human involvement could dramatically improve performance. The vice president said some experiments showed human-placed calls achieving success rates of between 95% and 98%, compared with a lower success rate for calls handled entirely by AI.

That difference helps explain why human contractors remain useful even as AI companies market agents as autonomous.

Phone calls are particularly difficult for AI systems because they involve unpredictable conversations, accents, interruptions, ambiguous requests, and situations in which the person on the other end may not know they are speaking to an AI system. A human can adapt immediately when a conversation moves outside the agent's expected workflow.

The problem is that the use of humans changes the nature of the product being offered.

A consumer who asks an AI agent to negotiate a bill or make an appointment may reasonably assume that the interaction is being conducted by software. If a human contractor takes over without clear disclosure, the privacy expectations surrounding the service are different.

That issue is especially significant for Muse because Meta has positioned privacy and security as central to the product.

At launch, Meta said each Muse agent would operate inside its own secure virtual machine, essentially a dedicated computer in the cloud designed to isolate a user's session. The company also said sensitive information such as passwords would be kept in separate secure storage.

The calling feature was designed to extend that model into the physical world. Users can instruct Muse to call U.S. businesses and handle routine tasks, after which the system provides a summary and transcript of the conversation through the app.

The human-concierge experiment shows that securing the computing environment is only part of the privacy challenge. Once an AI agent needs to interact with people outside its digital environment, protecting information also depends on how much human intervention sits behind the system.

The development also recalls Meta's earlier attempt to build an AI assistant.

About a decade ago, Facebook tested a digital assistant called M through its Messenger service. Media reports later estimated that humans performed around 70% of the tasks assigned to the system. The difference today is the scale and sophistication of the technology. Modern AI agents can perform far more tasks autonomously, while Meta has access to billions of potential users through Facebook, Instagram and WhatsApp.

Muse has already gained significant traction. The app has accumulated more than 2.5 million downloads in the U.S. since its launch and has topped American app download charts for two weeks, according to data from Sensor Tower.

The product is central to CEO Mark Zuckerberg's broader push to deliver what he has called "personal superintelligence" to Meta's enormous user base. Muse can perform tasks including sending emails, shopping and booking travel, potentially shifting AI from a tool that generates information to an agent that takes actions on behalf of its user.

That transition makes reliability more important.

An inaccurate answer from a chatbot can usually be ignored or corrected. An AI agent that makes a phone call, cancels a service, negotiates a contract, or makes a purchase can create consequences outside the chatbot itself.

Meta began testing the phone-calling feature internally in August before gradually expanding access after Muse's public launch. The human concierge system was then enabled for about half of Meta's employees last week, according to an internal announcement.

Employees who did not want to participate could join an opt-out group.

However, the experiment highlights a difficult trade-off in building autonomous agents for Meta. Human intervention can make an AI system considerably more effective, but it can also undermine the perception that the service is genuinely autonomous and introduce new privacy and disclosure obligations.

The company has now pulled back the human-concierge feature, at least temporarily. Meta says it intends to continue improving AI calling and will only release the capability more broadly with appropriate disclosures.

The episode may prove less important for Muse's immediate rollout than for the wider economics of agentic AI. As companies promise software that can act independently in the real world, the hidden question is how much human work is required to make those promises reliable. If agents still need people to step in whenever interactions become difficult, the technology may be autonomous in appearance while remaining partly human-operated underneath.