The Shocking Truth Behind Next-Gen AI Agents... Outcry as Humans Working as AI Subcontractors Emerge

Meta's recently unveiled personal AI agent, Muse, has been promoted for its ability to make phone calls and handle tasks on behalf of users.

However, it has been revealed that Meta tested a method where some of these calls made by the AI were actually handled by humans, namely external contract workers.

According to internal Meta posts obtained by Reuters, Meta recently tested a so-called "human concierge" function on employees, routing some of Muse's telephone tasks to trained human agents.

Muse is an AI agent designed to autonomously send emails and handle shopping and travel reservations.

Recently, it also introduced a phone calling feature to directly call businesses in the United States to book hair salons, check product inventory, and even obtain price quotes from multiple vendors.

However, a problem arose when businesses would simply hang up upon discovering they were talking to an AI.

One employee reported using Muse to call an insurance company, but the call kept getting disconnected once the other party learned it was an AI.

To address this issue, Meta activated the "human concierge" function for half of all its employees last week.

This involved passing calls that the AI found difficult to handle over to external contract agents, who then made the calls directly to complete the tasks.

Meta stated that when humans made the calls directly, the task success rate rose to 95 – 98%.

However, this sparked privacy concerns.

Critics pointed out that sensitive user information could be unintentionally transmitted to external call center workers.

One employee expressed concern that even a single bug could expose unnecessary personal information to human agents.

Actual problematic cases also surfaced.

An employee requested negotiation over rates with an internet and cable provider, and it was internally reported that the human agent who took over the call made inappropriate remarks regarding race.

Executives at Meta's Superintelligence Lab admitted it was a mistake to launch the test without properly informing users that humans could make calls on their behalf.

Meta ultimately withdrew the human concierge feature, but has not entirely discarded the approach of utilizing human agents.

Meta's stance is that it will launch related features after establishing proper notifications and privacy safeguards.

Reported by Kim Minjeong | Video by Ahn Jun-hyeok | Graphics by Yang Hye-min | Produced by SBS Digital News
※ Please note: This article was translated by AI and may contain errors.