Meta recently implemented a human-driven concierge service to supplement its Muse AI assistant's phone-calling capabilities. This experiment, involving human workers managing calls on behalf of users, aimed to improve reliability during a recent testing phase.
The 95 to 98 percent success rate gap
The decision to introduce human intervention was driven by the technical limitations of autonomous agents. While the Muse AI agent is designed to autonomously book travel, shop, and send emails,users frequently encountered friction when attempting to use the phone-calling feature. For instance, some employees reported that insurance companies would hang up the moment they recognized the caller was an AI.
To combat these failures, Meta enabled a "human agent calls" feature for half of its employees last week. as the report notes, this move was intended to bridge the gap between automated attempts and successful task completion. Testing indicated that while AI calling struggled with consistency, human-made calls achieved success rates between 95 and 98 percent. Some media reports even suggested that humans were performing as many as 70 percent of the agent's tasks during the trial.
A racist remark and the risk of leaked data
The reliance on human labor introduced significant ethical and security vulnerabilities that have alarmed Meta's internal workforce. Employees raised concerns that sensitive personal information could be unintentionally shared with contractors working in call centers. One staff member warned that the company was only "one bug away from leaked data," questioning if the feature's benefits outweighed the inherent privacy risks.
The human element also introduced behavioral risks that AI does not possess. One employee, who had asked Muse to negotiate a utility bill, discovered through a transcript that a human contractor had made a racist reference during the call. In response to the backlash, a Meta vice president in Superintelligence Labs apologized to the employee and promised that the specific agent involved would be removed from all Meta projects.
A $200 billion market cap boost and the "M" assistant echo
Despite these internal controversies, the Muse AI app has demonstrated massive commercial momentum. According to Sensor Tower, the app has surpassed 2.5 million downloads since its release. This popularity has coincided with a significant financial windfall for the parent company; Meta's stock has climbed more than 20 percent since the launch, adding over US$200 billion in market capitalization.
This strategy of using humans to mask AI shortcomings is a pattern Meta has encountered before. The current situation harks back to an experiment from a decade ago when Facebook trialed a digital assistant called M within the Messenger app. In both instances, the company has navigated the tension between the promise of total autonomy and the reality of human-dependent service.
The missing disclosures in Meta's "hillclimbing" phase
As Meta moves forward, several critical questions remain regarding the transparency of its AI operations.. While spokesperson Daniel Roberts stated that the company is working with merchants to improve the feature and will only roll it out with "proper disclosures," the specific nature of these disclosures remains unverified. It is unclear how Meta will balance the need for human assistance with the user's right to know they are not speaking to an algorithm .
Furthermore, while Meta emphasized that each agent would operate within a secure virtual machine to protect passwords and sensitive data, the introduction of human contractors complicates this security architecture. Whether Meta can successfully "hillclimb" its AI calling technology without compromising user privacy or professional standards remains the central challenge for the Muse project.
Comments 0