Meta Tests Humans to Step In When Its AI Agent Falls Short

by Saviour Simulindi

Meta has quietly pulled back a feature that let human contractors secretly take over phone calls placed by its new AI assistant Muse, after one of those contractors made a racist remark during a call, and after Meta's own employees warned users' sensitive information could be exposed to call-centre workers nobody had told them about.

Here's how it worked

Ask Muse to call a business, book a haircut, check if a store has something in stock, and behind the scenes, some of those calls weren't actually being handled by the AI at all. Real people were quietly placing and working through them instead. This is because some people on the receiving end simply refused to talk to an AI caller, so Meta built in human contractors as a backup to stop those tasks from failing outright.

It didn't stay quiet for long

Employees flagged that sensitive information shared on a call could land with a contractor the user never knew existed. That warning stopped being theoretical fast, one employee testing the feature found the contractor handling their call had made a racist remark. A Meta Superintelligence Labs VP called it "a miss" internally and confirmed the feature has been rolled back for now.

Meta's own spokesperson downplayed it as a normal part of testing, saying feedback overall had been "overwhelmingly positive" and that the company would only bring the feature back "when it's ready and with the proper disclosures."

The timing is awkward

Muse has topped US download charts since launching on 8 September, passing 2.5 million downloads, and helped push Meta's stock up over 20%, adding more than $200 billion in market value. Shares dipped slightly after this story broke.

The bigger picture here says something about AI generally, tools sold to us as autonomous and intelligent are often, quietly, leaning on real people to fill in the gaps the technology can't close yet.