Meta's AI Uses Humans for Calls, Sparks Privacy Concerns
Meta has been testing a "human concierge" for its new personal AI assistant, Muse, according to internal company posts seen by Reuters. This involves human contractors quietly handling some phone calls placed via the digital agent. The Facebook and Instagram owner recently informed employees about this test, shortly after publicly launching Muse's phone calling capability.
Muse, which has topped US app download charts since its recent debut, is designed to autonomously perform tasks. These include sending emails, shopping, and booking travel on a person's behalf. Its phone calling feature allows users to instruct the agent to dial US businesses and engage with individuals to manage errands such as booking haircuts, checking stock availability, or acquiring quotes from various contractors.
But this reliance on human intervention has prompted privacy concerns among some Meta employees. They warn that the arrangement could lead to sensitive information being unintentionally shared with contractors operating in call centres. One employee expressed strong reservations, stating the feature carried undue risk and was "one bug away from unnecessary information being leaked to human callers."
An employee who directed Muse to contact an internet and cable provider to negotiate a bill faced an issue. A transcript showed the human contractor made a racist reference during the call. In response, a vice president within Meta's Superintelligence Labs unit admitted the lack of proper disclosures for calls placed by contractors was "a miss." The feature was temporarily withdrawn.
And Meta continues its efforts to improve AI calling performance. The vice president noted that some tests demonstrated human involvement could raise call success rates to between 95 per cent and 98 per cent, surpassing the lower percentages achieved by AI alone. She also extended an apology to the employee affected by the racist comment, confirming the human agent responsible would not work on any future Meta projects.
The Muse test shows Big Tech firms sometimes use human labour to deliver on AI promises, amid public reluctance to engage with robocalls. This practice mirrors a previous Meta experiment some years ago. The company, then known as Facebook, trialled a digital assistant called "M" within its Messenger service. Reports indicated 70 per cent of M's tasks were performed by humans.
So a Meta spokesperson, Daniel Roberts, commented on the "human concierge" plan. He stated the employee response had been "overwhelmingly positive." Roberts explained the test's purpose was to gather feedback to implement safety and privacy safeguards, and to improve features before a public release. He affirmed the potential calling feature would only be rolled out when ready, and with appropriate disclosures.
Muse has recorded over 2.5 million downloads since its release, according to market intelligence firm Sensor Tower. The organisation's stock experienced a surge of more than 20 per cent following the launch, adding over USD 200 billion to its market capitalisation and reaching its highest level in more than seven months. Shares saw a minor decrease of 0.6 per cent recently.
But Meta's initial emphasis on Muse's safety protections raised questions about the human concierge plan. The company assured users each agent would have a secure "virtual machine," isolating sessions and storing sensitive data like passwords separately. Employees internally criticised the human concierge approach, arguing it risked undermining these core selling points.
Meta initiated employee testing of the phone calling function earlier this year, gradually expanding its availability to Muse users shortly after the app's official debut. However, some users encountered difficulties with these calls. One employee reported his insurance company repeatedly disconnected calls when they identified the caller as an AI.
And to address these issues, Meta recently activated the human concierge feature, also known as "human agent calls," for half of its employees. The internal announcement indicated Muse could now transfer requests to a trained agent, who would then place and manage the call. Employees were offered an option to opt out of this programme if they preferred not to participate.
Meta tested human contractors to handle phone calls placed via its AI assistant, Muse.
Employees expressed privacy concerns regarding sensitive information exposure and reported instances of inappropriate human contractor conduct.
Meta acknowledged disclosure shortcomings and has temporarily withdrawn the human calling feature.
The company cited human intervention as a means to achieve higher call success rates for Muse.
Despite internal criticism, Muse has accumulated over 2.5 million downloads, and Meta's stock has shown significant gains since its launch.
Source: Reuters
