Meta's AI Assistant Muse Secretly Used Human Agents to Handle Calls

Meta is testing a "human concierge" mechanism for its personal AI assistant Muse, according to internal posts. When the calling feature encounters complex scenarios, calls may be transferred to contractor call-center employees without user knowledge, with humans completing the call instead of the AI.

Muse has topped U.S. app download charts two weeks after launch. Its calling feature lets users instruct the AI to dial U.S. businesses for tasks such as booking appointments, checking inventory, and getting quotes. Internal tests told staff that human agents achieved a 95%–98% success rate, higher than pure AI calls.

The test raised privacy concerns among employees. One worker warned that sensitive information could be leaked to outsourced staff; another reported that an outsourced worker used racially discriminatory language during a call made by Muse to a cable company to negotiate a bill.

A Meta vice president of the Superintelligence Lab admitted it was a "mistake" to start the test without adequate disclosure, said the feature has been temporarily rolled back, apologized to the reporting employee, and promised that the contractor involved will no longer work on Meta projects.

A Meta spokesperson said employee feedback has been "generally very positive," and the test aims to collect feedback to implement safety and privacy protections before launch; the feature will only be released when ready and with complete disclosure.