Meta Muse AI Secretly Used Human Contractors for Phone Calls, Internal Documents Reveal

Meta Muse AI Relies on Human Contractors Behind the Scenes

Meta launched its Muse AI agent on September 8, positioning it as a personal AI assistant capable of handling everyday tasks autonomously. The platform quickly gained traction, reaching 2.5 million downloads and topping the iOS download chart in the United States4

4

. However, internal documents revealed by Reuters show that Meta has been testing a human concierge feature that allows human contractors to handle phone calls placed through the AI agent, raising serious privacy concerns and questions about AI-washing2

2

.

When Meta's Principal Engineer Ryan Fox announced the beta release of Muse's outbound call feature on September 16, he described phone calling as one of the top user requests. But internal company posts show Meta had "added a human agent layer for calls to get completed," referring to them as "Muse human agent calls"1

1

. The feature was enabled for half of Meta's employees as part of internal testing, with an opt-out group available3

3

.

Privacy Concerns and Employee Pushback Over Sensitive Information

Meta employees immediately raised privacy concerns about the human intervention in AI calls. One employee warned in internal posts, "We are one bug away from unnecessary information being leaked to human callers"1

1

. The worry centers on sensitive information potentially being shared with human agents in call centers, including phone numbers, patient IDs, banking information, and other personal details that users might provide to what they believe is an autonomous AI agent making phone calls2

2

.

Another employee complained about potential negative PR, posting internally: "This has potential for so much negative PR. It could portray us as 'their AI is not good enough so they still need humans' kind of coverage for this launch"1

1

. The employee feedback highlighted concerns about transparency and the ethical hurdles of deploying AI agents that secretly rely on human contractors.

Racist Remark Incident Exposes Human Concierge Risks

The risks materialized when one Meta employee asked Muse to call an internet and cable provider to negotiate his bill. According to a transcript reviewed by the employee, the human contractor made a racist remark during the call3

3

. Meta apologized to the employee and stated that the human agent involved would not work on any of its projects again3

3

.

This incident demonstrates why proper disclosures matter when human agents in call centers handle what users believe are AI-powered interactions. Users provide sensitive information to the personal AI assistant expecting privacy protections appropriate for automated systems, not knowing their user data might be accessible to contractors whose backgrounds and training remain unclear.

Call Success Rates Drive Human Intervention Strategy

Internal documents show Meta's rationale for introducing human contractors: call success rates. Meta's own testing revealed that businesses often hang up on Muse when they realize they're talking to an AI agent. One employee reported that when he tried getting Muse to call his insurance company, the company repeatedly hung up after the bot disclosed it wasn't human5

5

.

Internal tests found that handing calls to human agents could increase the success rate to between 95% and 98%, compared with a much lower rate for calls handled entirely by AI3

3

. This significant gap explains why Meta pursued the human concierge feature, but it also highlights the current limitations of AI agent deployment for real-world phone interactions.

Meta Rolls Back Feature After Admitting Disclosure Failure

A Meta Superintelligence Labs vice president acknowledged that introducing the human calling feature without adequate disclosure "was a miss" and confirmed the company had rolled back the feature1

1

. Meta spokesperson Daniel Roberts told Reuters that employee feedback had been "overwhelmingly positive" and that the test aimed to "get feedback so we can implement safety and privacy protections and improve features before we release them publicly"2

2

.

Roberts added that Meta would "only roll it out when it's ready and with the proper disclosures," though he didn't specify what those disclosures would entail or who would receive them4

4

. The company emphasized that internal testing, or "dogfooding," is core to its product development process5

5

.

Pattern of AI-Washing Across Tech Industry

Meta Muse AI joins a growing list of supposedly autonomous systems that rely heavily on human contractors behind the scenes. Meta itself attempted this approach before with Facebook M, a Messenger assistant launched in closed beta during summer 2015. Every message the service handled was reviewed by contractors at Meta's Menlo Park headquarters, but the service never achieved more than 30% automation and was shut down in 20183

3

.

Other examples include shopping app Nate, where hundreds of workers in call centers in the Philippines and Romania manually completed many transactions despite claims of AI automation. Amazon's "Just Walk Out" stores relied on more than 1,000 workers in India to review transactions and label footage, with around 700 out of every 1,000 purchases requiring human review as of 2022. Presto Automation revealed that around 70% of orders handled by its AI drive-thru system required intervention from off-site human agents3

3

.

What This Means for AI Agent Trust and Transparency

The Meta Muse AI incident matters because it exposes the gap between marketed AI capabilities and actual implementation. When companies deploy AI agents for tasks like booking appointments, checking inventory, or negotiating bills without disclosing human intervention, they undermine user trust and create privacy risks that users cannot properly assess.

Google Gemini recently introduced similar capabilities for handling everyday phone calls, though that implementation returns transactional parts like providing credit card information back to users1

1

. As personal AI assistant adoption grows, the industry faces pressure to establish clear standards for transparency about when human contractors are involved and what safeguards protect user data.

Meta's history with privacy violations adds weight to employee concerns. In 2019, Facebook had to stop a program that used contractors to listen to and transcribe Messenger calls1

1

. The pattern suggests that without robust oversight and mandatory disclosures, the cycle of "do bad, get caught, apologize, roll back" will continue as companies race to deploy AI agents before the technology can truly deliver on autonomous promises.