To quote George Santayana, “Those who cannot remember the past are condemned to repeat it.” So why does it so often feel like Big Tech has such a short memory?
The latest hiccup comes from Meta. In mid-September, it rolled out a beta test of a feature for its brand-new Muse AI agent that would place calls to businesses to complete tasks for you, like making reservations and appointments.
According to reports by Reuters and 404 Media on Tuesday, however, this allegedly autonomous agent could route these tasks to a trained human concierge in call centers — without disclosing it was doing so in advance.
On Sept. 16, Principal Engineer Ryan Fox at Meta’s Superintelligence Labs announced the beta release of Muse’s outbound call feature and posted a request on X for feedback, noting that “phone calling” was among the top user requests.
But 404 Media says that, based on internal Facebook records, the company was doing something entirely behind the scenes with its personal assistant. When announcing the features to employees internally, the tech giant said that it “added a human agent layer for calls to get completed,” referring to them as “Muse human agent calls.”
Some Meta employees complained, including one who posted internally: “This has potential for so much negative PR. It could portray us as ‘their AI is not good enough so they still need humans’ kind of coverage for this launch.”
The truth is rarely pure…
The problem with the added human meat layer isn’t just the AI-washing of Muse’s capabilities. Nor is it purely the deception of failing to disclose that there could be a human link in the chain. Some agentic systems introduce human oversight as a “human in the loop” for safety and compliance.
It’s unclear how much emphasis on safety and compliance was on the menu for Meta. But employees raised a bigger concern over privacy, with one warning (per Reuters), “We are one bug away from unnecessary information being leaked to human callers.” And bugs there are.
You generally sacrifice privacy when you use AI agents to perform personal tasks, like making a reservation, a purchase or a medical appointment. But it can feel a lot worse if you discover you’ve just had a complete stranger book your STI test.
It’s also a black box. You don’t know what context that person’s been given, if they’ve seen your chat or know where you live and how to contact you because of it. Even more data might be on offer, given that the company’s glasses are chock-full of Muse.
You’re also providing the so-called autonomous agent (in this case, a human contractor in an undisclosed call center) with sensitive information, such as a phone number, patient ID or banking information.
It’s all part of a growing trend of granting AI models the power to perform tasks as personal assistants. Just this week, Google introduced the capability of its Gemini agent to handle everyday phone calls. That implementation turns over the transactional part of a call — like providing credit card info — back to you. But the risks are still there.
…And never simple
The kicker is that Meta has broken privacy trust many times before. In 2019, Facebook had to stop a program that used contractors to listen to and transcribe Messenger calls.
Because AI agents like these frequently require human intervention to verify that they’re delivering the desired results, a mistaken incident of the “Oh, this employee saw this” variety seems inevitable.
Potemkin AI, where a hidden army of human workers does some of the actual work of models behind the curtain, isn’t confined to calls. In February, it was revealed that Waymo had remote operators supplementing autonomous car operations overseas.
And once again, the rollout of Meta Muse’s outbound call feature has been rolled back. Reuters reported that the company temporarily suspended the feature after a vice president in Meta’s Superintelligence Labs unit admitted “it was a miss” to begin testing the feature without adequate disclosures of the human contractors involved.
But the real fault isn’t a bad memory so much as Big Tech’s history of “do bad, get caught, apologize, roll back, maybe pay a negligible fine. Rinse, repeat.” At some point, after we’ve collectively become exhausted, the cycle is reduced to “do bad, but profit anyway.”
Meta didn’t immediately respond to a request for comment and clarification.
