Meta is facing internal debate over its Muse AI assistant after testing revealed that some AI-powered phone calls were completed by human contractors.

The feature was designed to improve task success when businesses resisted automated callers, but employees raised concerns about privacy, disclosure, and user trust.

The situation highlights a wider challenge for the AI industry: balancing automation, human oversight, and data protection as AI agents become more capable.

Meta’s AI assistant Muse enters a new phase

Artificial intelligence companies are increasingly developing AI agents that can perform tasks rather than simply provide information.

These systems are designed to serve as digital assistants that communicate with businesses, manage schedules, complete purchases, and handle everyday activities with minimal user involvement.

Meta’s Muse is part of this broader shift toward personal AI assistants. The company has promoted Muse as a system that can help users manage daily responsibilities by taking action on their behalf.

Unlike traditional chatbots, which mainly respond to questions, AI agents are expected to complete multi-step processes. For example, a user could ask Muse to contact a service provider, compare options, schedule an appointment, or request information from a company.

However, turning these capabilities into reliable real-world services has proven difficult. Businesses may not always cooperate with AI callers, and unexpected situations can require additional support. This challenge led Meta to experiment with a human-assisted system to improve Muse’s performance.

Meta tests a “human concierge” system

To address this, Meta reportedly tested an internal “human concierge” system with employees, in which trained contractors, rather than the AI alone, made phone calls and handled business interactions when Muse hit limitations.

Some businesses had been disconnecting calls once they realized they were talking to an AI, so Meta brought in human operators to smooth out these edge cases before considering a wider release.

Why Meta added human support to AI calls

The move reflects a common gap between technical capability and real-world reliability. AI can generate language and perform digital tasks, but real conversations involve unpredictable questions, refusals, and judgment calls that current models cannot confidently make.

Meta’s internal testing reportedly found human-supported calls succeeded far more often than AI-only ones, with completion rates reaching 95% to 98%.

That gave Meta a way to improve the user experience while exposing where AI agents still fall short, but it also raised new questions about privacy and transparency.

Employees question privacy protections

Some employees worried that routing calls to outside contractors undermined Muse’s stated privacy safeguards, including the isolated virtual environments designed to keep AI sessions separate and protect sensitive data.

Their core concern was that users might assume they were speaking only to an AI, when a human could actually be listening to or accessing their information.

A reported incident increases internal criticism

The discussion around Muse intensified after employees reported concerns about contractor interactions during testing. One employee reportedly used Muse to negotiate a lower internet and cable bill. During the call process, a contractor allegedly made an inappropriate comment.

The incident raised further questions about whether human involvement could introduce risks that AI systems were supposed to reduce.

A Meta executive later acknowledged that introducing human-assisted calls without sufficient disclosures was a mistake. The company temporarily rolled back the feature and apologized to the employee who reported the issue.

The contractor involved was reportedly removed from future Meta projects. Meta’s response suggested that the testing process was intended to identify problems before a broader launch.

Human-assisted AI may become a temporary development stage

Human involvement has always played a role in improving artificial intelligence. Companies use human reviewers to evaluate AI responses, improve accuracy, and identify safety problems. However, AI agents create a different challenge because they are designed to take actions in the real world.

The question is not only whether humans help train AI, but whether humans should secretly participate in tasks users believe are automated.

For companies like Meta, the goal is to create AI systems that eventually require less human support while still maintaining reliability and safety. The Muse experiment demonstrates that achieving truly autonomous AI assistants remains a complex process.

Meta defends the experiment and future plans

Despite criticism, Meta has not abandoned the idea of combining AI systems with human support. The company argues that human assistance can help AI agents perform more effectively during the current stage of development.

Meta spokesperson Daniel Roberts said employee feedback on the experiment had been largely positive and explained that the testing was conducted to identify potential weaknesses before a public expansion.

The company also said it would continue improving safeguards, disclosures, and operational procedures. Meta’s challenge is to balance improving AI performance with maintaining user confidence. If users are unaware that humans are involved, questions about transparency could undermine trust in AI assistants.

Muse gains popularity despite controversy

The debate around human-assisted calling comes as Muse continues attracting attention from users and investors.

According to market intelligence data from Sensor Tower cited in reports, Muse has exceeded 2.5 million downloads and reached the top of U.S. app download rankings. The launch has also strengthened investor interest in Meta’s AI strategy.

The future of AI assistants depends on trust

Meta’s Muse experiment highlights a major challenge for the future of artificial intelligence: building systems that are powerful, reliable, and transparent.

AI assistants are becoming capable of handling increasingly complex tasks, but users will expect clear information about how these systems work. Companies will need to explain when humans are involved, how data is protected, and what safeguards are in place to prevent misuse.

The success of AI agents may depend not only on their ability to complete tasks but also on whether people trust them with personal information and important decisions. As AI technology continues to develop, privacy and transparency will become as important as innovation itself.

TL;DR

• Meta reportedly tested a human concierge system for its Muse AI assistant with employee participants.
• Human contractors handled some phone calls when AI struggled with business interactions.
• Employees raised concerns about privacy risks and possible exposure of user information.
• Meta temporarily rolled back the feature after internal criticism and a reported incident involving a contractor.
• Meta said human assistance was being tested to improve reliability and gather feedback before wider deployment.
• Reports citing Sensor Tower data said Muse surpassed 2.5 million downloads shortly after launch.
• The experiment highlights the challenge of balancing AI automation, human oversight, and privacy.

Don’t forget to follow us for more exclusive content.

If you liked this, you might also like:

• Why Meta’s federal case matters far beyond Facebook and Instagram
• OpenAI pushes Washington to lead international standards for frontier AI safety
• Jonathan Haidt says the Meta lawsuit could force social media giants to fix their platforms

This article was made with AI assistance and human editing.