Meta is facing fresh questions about privacy and human involvement in artificial intelligence after reportedly testing a system in which human contractors handled some phone calls placed through its Muse AI assistant.
According to Reuters, Meta internally tested what it called a human concierge system shortly after introducing phone calling capabilities for Muse. Under the test, Muse could transfer certain requests to a trained human agent, who would then place and handle the phone call on the user's behalf.
The development has drawn attention because Muse is designed as a personal AI agent capable of performing tasks for users. The system is intended to go beyond conventional chatbot functions by carrying out activities such as sending emails, making purchases, booking travel and communicating with businesses.
Muse's phone calling feature allows users to ask the AI agent to contact businesses in the United States. Examples of tasks include booking appointments, checking whether a product is available at a store and contacting service providers to obtain price quotations.
However, according to internal company posts reviewed by Reuters, some Meta employees raised concerns after learning that human contractors could become involved in these calls. Their concern was that information users expected to be handled by an AI system could potentially be heard or accessed by people working for external call centres.
The issue became particularly significant because privacy and security are central elements of Meta's positioning for Muse. The company has described Muse as an AI agent that operates within a dedicated security environment. The use of human contractors during some calls raised questions among employees about whether the arrangement could create additional privacy risks.
An internal Meta executive acknowledged concerns about the way the test had initially been introduced. According to Reuters, the executive said it was a mistake to begin the test without proper disclosures and confirmed that the company had rolled back the feature for the time being.
Meta has nevertheless indicated that it continues to work on AI-powered phone calling. The company is examining ways to improve the reliability of calls and interactions between Muse and businesses. According to the company's explanation reported by Reuters, the internal test was intended to collect feedback and develop appropriate safety and privacy protections before the feature could potentially be offered more broadly.
Meta spokesperson Daniel Roberts said employee feedback on the test had been overwhelmingly positive, according to Reuters. He said the company was working with merchants to improve the potential calling feature and that any wider rollout would take place only when the system was ready and with appropriate disclosures.
Why Did Meta Test Human Agents?
One reason for the human involvement was the difficulty AI systems can face when communicating with businesses over the telephone. Some businesses may end calls after recognising that they are speaking to an automated system. This can reduce the success rate of AI initiated calls.
Internal discussions cited in reports indicated that human involvement could significantly improve the success rate of certain calls. Reuters reported that some internal tests showed human agents could achieve success rates of around 95 percent to 98 percent, compared with lower rates for AI only calling.
The experiment therefore highlights a broader challenge for companies developing autonomous AI agents. While AI systems are becoming increasingly capable of completing digital tasks, real world interactions can still involve situations where businesses or individuals are reluctant to communicate with an automated system.
A human fallback mechanism can potentially solve that problem, but it also introduces another layer of privacy and transparency considerations.
For example, if a user asks an AI assistant to contact a service provider about a personal account, the conversation could involve information about the user's name, address, account details, purchases or other private matters. If a human contractor becomes involved, the user may want to know that another person is participating in the interaction.
This issue is particularly important when AI services are marketed as autonomous assistants. Users may assume that a task requested from an AI agent is being completed entirely by software unless the company clearly explains otherwise.
Meta's decision to roll back the feature for now indicates that the company is continuing to evaluate these issues before considering a broader rollout. The company has said that appropriate disclosures and privacy protections would be part of any future public implementation.
Muse and the Growth of AI Agents
The human concierge test comes as Meta expands its focus on AI agents that can perform actions rather than simply provide answers.
Traditional AI chatbots generally respond to questions, generate text or assist users with information. Agent based systems are designed to take additional steps on behalf of users. Depending on the service, these actions can include searching websites, sending messages, booking services, purchasing products and interacting with other digital platforms.
Muse is part of this broader shift. Meta has positioned the product as a personal AI agent capable of carrying out everyday tasks for users.
The phone calling feature extends that concept from digital interactions to conversations with businesses. Instead of asking users to make a call themselves, Muse is intended to communicate with the business and complete the requested task.
However, the human concierge test shows that the transition toward fully autonomous AI can involve technical and practical challenges. If businesses do not accept AI generated calls, companies developing these systems may need to find alternative ways to complete the task.
The use of human contractors can provide a temporary solution, but it also changes the privacy profile of the service. This makes disclosure and user consent important considerations for any future implementation.
Meta has not said that the human concierge system is currently available to all Muse users. The reporting indicates that it was an internal test and that the company has rolled it back for now.
Earlier Meta Experience With Human Assisted AI
The latest experiment is not the first time Meta has explored combining artificial intelligence with human assistance.
Reuters reporting noted that Meta previously tested a digital assistant called M through Messenger about a decade ago. Media reports at the time estimated that a substantial portion of the work associated with that service was performed by human workers.
The latest Muse experiment is different in its design and purpose, but it again highlights the role that human labour can play behind AI systems during periods of development and testing.
The development also raises a wider industry question about how companies should describe AI systems that rely on human assistance. Transparency can become particularly important when an AI product handles tasks involving personal information or communicates directly with third parties.
Privacy Questions Remain
The main concern raised by Meta employees was not simply that humans were involved, but that users could potentially share information with contractors without being fully aware that a human was handling the conversation.
That concern has led to questions about how companies should disclose human intervention in AI services. Clear notices could allow users to understand when an AI agent is operating independently and when a human worker may become involved.
Meta's current position is that the human concierge experiment was part of its internal testing process and that the company intends to strengthen safety and privacy measures before considering a wider rollout.
For users, the episode demonstrates why privacy settings, data handling policies and disclosure practices are becoming increasingly important as AI agents gain the ability to act independently.
The development does not mean that Muse's regular AI calling feature has been permanently discontinued. Instead, Meta continues to work on improving AI calling while the specific human-assisted test has been rolled back.
As AI agents become capable of performing more tasks in the real world, the balance between automation, reliability, privacy and transparency is likely to remain an important issue for technology companies and users.