META’S Muse AI assistant reportedly handed some phone-call tasks to outsourced human contractors during an internal test, without clearly telling users that people might take over. Muse is designed to contact businesses to make bookings or request quotes. When the AI struggled, contractors could manually dial the business and handle the conversation. The feature was available to roughly half of Meta employees in the test, with an opt-out option, but had not been released publicly.
The hand-off exposed contractors to information users had provided to Muse, potentially including names, phone numbers, the purpose of the call, billing details and account information. The article says a Meta employee found that a contractor made a racially discriminatory remark during a call to an internet and cable provider about negotiating a bill; the remark was not made by the AI.
Meta vice-president Connor Hayes said human intervention raised the call success rate to 95%–98%, compared with fully automated calls, because some businesses reportedly ended calls when Muse identified itself.
Meta acknowledged that testing the feature without advance disclosure was a mistake and has temporarily suspended it. The contractor involved in the reported incident has been removed from the project and will not work on any Meta project, according to the company.