Imagine asking an AI agent to sort out a bill, then learning that a human caller handled the request. The help may still be useful. You might also have shared different details if you had known. In an internal test of Meta’s Muse , 404 Media reported that at least one tester learned only afterward that a person had made the call. That missing choice is where the privacy argument begins.
404 Media’s September 22 reporting on Muse’s human callers is the starting point for this examination of Meta’s record of hidden human review and privacy claims. At the time, Meta said the feature remained in internal testing and would receive proper disclosures before public release. The report did not establish how often people handled calls or that this human-caller system had shipped to consumers. It did show a tester making a decision without knowing who would carry it out.
Who gets the details of the errand?
Meta’s Muse launch announcement describes a dedicated virtual machine, user controls and an audit trail. It says data in that machine is excluded from Meta’s ad systems. A human handoff raises another question alongside those protections: who gets the task details? A booking can include a name and schedule. A billing problem can include an account issue. Someone may welcome help with either and still want to approve the person receiving that information.
Internal posts reviewed by 404 Media described trained human agents completing some calls. The outlet could not determine their share of calls or everything contractors could see. Meta said testing was intended to catch safety and privacy problems before release. A test is a good place to discover that gap. The public product should then make the handoff visible while the user can still decline it.
Transcription consent left another question unanswered
An earlier audio feature shows how easily consent to a function can be mistaken for consent to everyone behind it. In 2019, Facebook confirmed that contractors listened to and transcribed Messenger voice clips to check automated transcription. Users had opted into transcription, but outside human listening had not been made clear. Facebook said it paused the practice after reporting. A person can want words converted into text and still care who hears the recording.
The issue extends into Meta's AI hardware. An investigation by Svenska Dagbladet and Göteborgs-Posten interviewed more than 30 workers and others connected to a Kenyan subcontractor reviewing material from Meta's AI glasses. Workers described seeing intimate footage from homes, including bathroom scenes. Meta later ended its work with that contractor . Meta said human review improves its AI and that it obtains user consent. That response leaves the trust problem intact: people wearing the glasses, and people around them, have little practical visibility into who may later see what was captured.
The destination of the data is another separate choice. In 2025, Meta announced that conversations with Meta AI, including voice chats, would inform recommendations and ads , with stated limits for sensitive topics and some WhatsApp accounts. That policy concerns other Meta AI interactions. Muse’s dedicated-machine promise should be assessed on its own terms, without treating every Meta product as one data flow. Readers need to know both who can handle a request and which later uses apply to that particular service.
Follow the stated purpose of the data
In 2019, the Federal Trade Commission imposed a $5 billion Facebook settlement after alleging violations of an earlier privacy order and deception about users' control over their data. The agency said Facebook misled people about access through friends' apps, the route at the center of the Cambridge Analytica scandal . It also alleged that Facebook used phone numbers supplied for account security in advertising without disclosing that use in the security flow. The lesson is specific: Meta's predecessor has used data supplied for one stated purpose to serve another.
The European Commission’s €110 million fine in 2017 concerned misleading information during the WhatsApp acquisition review. Facebook had said reliable automated account matching was impossible. The Commission found that the technical possibility existed and staff knew it. In a separate decision in July 2026, Meta withdrew a Muse Image feature that let people reference public Instagram accounts in generated images. These episodes concern different technologies. Together they give users a reason to seek a specific explanation of access and control before connecting another account.
Safety promises now come with court-enforced conditions
On August 26, California's attorney general announced a settlement of up to $17 billion with Meta and a coalition of attorneys general over alleged harms to children. A court approved the consent judgment later that day . The allegations concerned compulsive-use design, data from children under 13 and misleading safety claims. The agreement includes youth time limits, overnight blocks, age checks, an independent auditor and restrictions on false or misleading safety claims. It does not establish that every allegation is true. It does put concrete, enforceable conditions around assurances that families had been asked to trust.
Separately, a New Mexico court ordered $942 million in penalties and remedies after a jury found widespread violations involving minors. Meta said it would appeal. Neither case is about Muse, but both challenge the idea that Meta's public-facing safety language can be accepted on faith.
Did Facebook and Instagram listen to conversations?
The ad for something you only discussed aloud is a familiar source of suspicion. In a peer-reviewed survey across three countries , 77.7 percent of 300 US respondents reported an ad seemingly connected to an offline conversation. Their observation is worth investigating, but the survey could not identify the data path behind each match. It therefore cannot establish secret ambient microphone recording by Facebook or Instagram. An explanation should follow the information that produced the ad rather than fill the gap with a guess.
Known audio and conversation-data practices already give users concrete questions to ask. Messenger clips reached contractors. An opt-in Android feature uploaded call and text history , meaning call and message records rather than the spoken content of calls. Meta also says voice conversations with its AI can affect advertising. Keeping those mechanisms distinct makes it easier to identify which permission or policy needs to change.
One supposed smoking gun turned out to be a warning about marketing deception itself. A 2024 Cox Media Group pitch deck reported by 404 Media invoked Facebook among ad partners while claiming an “Active Listening” service could target ads from conversations near devices. In 2026, the FTC found the service used no voice data . It bought and resold email lists. The deck does not establish that Meta listened through phone microphones. It shows how readily the ad industry could sell the fear of such listening while hiding the data trade it actually used.
Put the choice before the call
For Muse, start with the point where the request leaves the agent. Before a human receives it, the user should be able to see who handles it, what information goes with it and what record remains. They should be able to decline that route. A transcript delivered afterward may help them understand the call, but it arrives too late to make the original disclosure choice.
For the next agent connected to a personal account, inspect three decisions before starting the task:
- Recipient: whether a model, employee or contractor receives the request.
- Information: which account details, recording or transcript that recipient can see.
- Choice: whether the user can decline or limit that handoff before disclosure happens.
The Google and Meta calling comparison applies a related test to interruption and takeover during a call. Meta can make its own handoff understandable at the same practical moment, while the user still has a choice.
