Meta’s Muse AI Agent Calls Face Human Contractor Issues

Meta’s Muse AI platform, launched in September, now lets human contractors handle phone calls when the AI fails. A Reuters report revealed a racist comment by a contractor and led Meta to roll back the feature. The company is revisiting safety protocols before a public release.

By Felo News Desk · Published Wed, 23 Sep 2026 13:03:00 GMT

Meta’s Muse, an AI‑powered agent platform that debuted on September 8, has quickly climbed the U.S. iOS download charts, drawing attention to its new beta feature that lets the bot place phone calls on users’ behalf. The feature, announced last week, promised a hands‑free way to book appointments, check stock, and gather quotes from contractors. However, a recent Reuters investigation has highlighted significant concerns about the safety and privacy of this capability.

How the Human‑In‑the‑Loop Feature Works

According to internal Meta documents that Reuters accessed, Muse can hand off a call to a human “concierge” when the AI voice is unable to complete the task. The system was first tested with Meta employees, who could opt out of the human‑in‑the‑loop option. The feature was designed to improve success rates for calls that the AI struggled with, such as when a user tried to reach an insurance company that repeatedly hung up after the bot disclosed it was not a human.

When the human contractor steps in, the call is handled as if it were a normal customer service interaction. The contractor can book haircuts, confirm product availability, or obtain quotes from service providers. While this may seem innocuous, the potential for privacy breaches grows if the contractor contacts sensitive professionals, such as a urologist, or if the contractor’s conduct is inappropriate.

Revealed Missteps and a Racist Incident

During internal testing, one employee reviewed a transcript of a call handled by a human contractor and reported that the contractor made a racist reference. The employee’s complaint was taken seriously by Meta’s Superintelligence Labs, where a vice president apologized and announced the contractor was no longer employed. The company subsequently rolled back the human‑in‑the‑loop feature, labeling the approach as a “miss.”

Reuters described the feature as capable of “handling errands like booking haircuts, checking whether a store has an item in stock or getting quotes from different contractors.” The incident raised questions about whether the feature should be re‑implemented, and if so, how to safeguard against future misconduct.

Meta’s Response and Future Plans

Meta spokesperson Dave Arnold explained that internal testing—often called “dogfooding”—is essential to product development. He said the company is working with merchants to refine the calling feature and will only release it publicly once it has robust safety and privacy protections in place. Arnold emphasized that the goal is to improve the feature before it reaches users, and that the company is committed to transparency and proper disclosures.

While the company has not confirmed whether all Muse‑powered calls have been rolled back, it appears that only those involving human contractors have been suspended. The impact on AI‑only calls remains unclear, as does how Meta plans to compensate for lower success rates when the AI cannot complete a call. Gizmodo reached out to Meta for clarification on the “lower percentage of AI calling” quote, but the company has yet to provide a definitive answer.

Why This Matters for Users and the Industry

Meta’s Muse platform represents a significant step toward fully autonomous digital assistants that can interact with the real world. The incident underscores the importance of rigorous safety protocols and the potential risks of involving humans in AI‑driven processes without stringent oversight. Users who rely on Muse for scheduling or service requests may face privacy concerns if their calls are handled by unvetted contractors.

For the broader AI industry, the episode serves as a cautionary tale. As companies explore hybrid models that combine AI and human intervention, they must balance efficiency gains with ethical considerations and robust vetting processes. The incident also highlights the need for clear disclosure to users about when a human is involved in a transaction.

What’s Next for Meta’s Muse?

Meta has not yet announced a new launch date for the human‑in‑the‑loop feature. The company is reportedly conducting additional safety reviews and refining its contractor vetting procedures. Once the feature is deemed secure, Meta plans to roll it out with full transparency, ensuring users are aware of who is handling their calls and what data is shared.

Until then, users can continue to enjoy Muse’s core capabilities—such as scheduling appointments and checking product availability—without the human‑in‑the‑loop option. Meta’s commitment to improving the feature signals a cautious approach to integrating AI with real‑world interactions.

Key facts

• Meta’s Muse platform can now hand off calls to human contractors
• A racist comment by a contractor led to the feature’s rollback
• Meta is revising safety protocols before re‑launch
• Human involvement can boost call success rates but raises privacy concerns
• Users should be aware of who handles their calls and what data is shared

Why it matters

The incident shows that integrating human contractors into AI systems can introduce serious safety and privacy risks, emphasizing the need for stringent oversight before public deployment.

Frequently asked questions

What is Meta’s Muse platform?

Muse is an AI‑powered agent that can perform tasks like booking appointments, checking stock, and making phone calls on behalf of users.

Why did Meta roll back the human‑in‑the‑loop feature?

A Reuters report revealed a racist remark by a contractor, leading Meta to suspend the feature and re‑evaluate safety protocols.

Will the feature be released again?

Meta is working on safety and privacy improvements and plans to release the feature publicly only once it meets strict standards.

Can users opt out of human involvement?

During internal testing, employees could opt out of the human‑in‑the‑loop option; it is unclear if this will be available to public users.

Sources

• [1] gizmodo.com — originally reported as “Meta’s Muse AI Agents Sometimes Tagged in Human Contractors. One of Them Reportedly Got Racist”