Meta Halts Muse AI Calling Test After Human Contractors Handled Calls Without User Disclosure

Meta Platforms halted an internal test of its Muse AI assistant's calling feature after Reuters reported on Tuesday that complex calls were routed to outsourced call-center workers without users' knowledge. A Superintelligence Labs vice president acknowledged launching the test without adequate disclosure was a "mistake," according to internal posts cited by Reuters. The "human concierge" mechanism, open to roughly half of employees with an opt-out, transferred difficult calls to trained contractors who placed them and returned transcripts and summaries. Internal data showed human-placed calls succeeded at rates of 95%-98%, versus a "lower percentage" for pure AI calls. Privacy exposure and a racially charged incident prompted the shutdown. An employee said an outsourced worker made racial remarks during a call, and the vice president apologized and pledged that worker would be excluded from Meta projects. Employees warned users might share sensitive data unknowingly with humans, and that a single bug could leak information. Meta spokesperson Daniel Roberts said feedback was "overwhelmingly positive." A decade ago, Meta's Messenger assistant M similarly relied on humans for roughly 70% of tasks.