The feeling that starts it
You talk about a thing. You have not searched for it, not typed it, not messaged anyone about it. Within a day an advertisement for it arrives. Almost everyone has had this happen, and the inference is immediate and reasonable: the phone was listening.
It is one of the most widely held beliefs about modern technology, and unlike most claims of covert surveillance it is testable. The microphone either sends audio or it does not. Researchers have gone and looked.
What the testing found
The most thorough attempt is a 2018 study from Northeastern University, published at the Privacy Enhancing Technologies Symposium under the title Panoptispy. The team — Elleen Pan, Jingjing Ren, Martina Lindorfer, Christo Wilson and David Choffnes — ran more than 17,000 Android applications, including the major social and shopping apps the theory usually names, on real handsets and captured everything each app sent over the network.
They found no audio being exfiltrated. Not encrypted audio, not compressed audio, not fragments. The specific mechanism the theory requires — the microphone opening unbidden and speech leaving the device — did not appear in any app they tested.
What they found instead is the part that tends to get dropped in the retelling, and it is not reassuring. Some applications were recording the screen and sending video of it to third parties, without disclosure. One food-delivery app transmitted screen recordings to an analytics firm; the recordings included a screen where users entered personal details. The researchers also found apps holding media permissions far broader than anything their function required.
So the finding was not that nothing is happening. It was that something else is happening, and the microphone was the wrong thing to watch.
Why the ads are accurate anyway
The stronger objection to the listening theory is that the listening would be redundant.
An advertising system already knows a great deal without any audio: every page you load that carries a tracking pixel, your location history and therefore which shops you stand inside, your purchase records bought from data brokers, the contacts you uploaded, and — decisively — the same information about everyone you know. If a friend searches for a product on the sofa next to you, and the system knows you are two people who are repeatedly in the same place, the advertisement that follows does not require anyone to have heard the conversation. It requires only that the conversation and the search were about the same thing, which is usually why the conversation happened.
There is also a counting problem underneath the experience. People notice the advertisement that follows a conversation and do not notice the thousands that do not, or the conversations that produce nothing. The impression is real. It is not evidence of the mechanism.
The 2019 admissions, and what they were actually about
In 2019 the belief received a substantial and genuine shock, and it is worth being precise about what was revealed, because it is often stated more broadly than the record supports.
That summer The Guardian reported that Apple contractors grading Siri recordings regularly heard material users never intended to send — medical details, business dealings, couples having sex — captured when the assistant activated by mistake. The source, Thomas le Bonniec, had worked on the programme in Cork and later went public in a letter to European regulators. Within weeks similar reporting landed on Amazon's Alexa transcription teams and on Google Assistant, where Belgian broadcaster VRT NWS obtained more than a thousand recordings, including some made without any wake word at all.
Apple suspended the grading programme worldwide and apologised, saying it had not been "fully living up to our high ideals" and making human review opt-in.
What that episode established is that voice assistants activate accidentally, that the resulting audio was retained, and that human beings listened to it. What it did not establish is the advertising claim. None of the revelations involved speech being converted into advertising categories, which is the specific thing the theory asserts. Apple had told the US House Energy and Commerce Committee the previous year that iPhones do not listen for advertising purposes and that the company does not sell Siri data — a denial that survived 2019 intact, because the grading story was about quality review, not ad targeting.
In January 2025 Apple agreed to pay 95 million dollars to settle Lopez v. Apple, a class action brought in 2019 over exactly these unintended activations. The complaint alleged both that human contractors had reviewed the recordings and that the material had been used to sharpen ad targeting. Apple settled while denying all of it — denying wrongdoing, and denying that Siri data was ever sold or used to build marketing profiles. A judge granted final approval, and payouts of up to twenty dollars per device began reaching claimants.
The part that complicates everything
Then, in September 2024, the picture got harder to summarise.
The technology outlet 404 Media obtained and published a pitch deck from Cox Media Group, a large American broadcaster, marketing a product it called Active Listening. The deck's own language is the striking part: it advertised the ability to "capture real-time intent data by listening to our conversations," and to "pair this voice-data with behavioral data to target in-market consumers." It listed prices — figures around one hundred dollars a day for weekly data pulls within a ten-mile radius, and roughly double that for twenty.
CMG had named Google, Amazon and Meta as partners in earlier marketing of the product. After 404 Media approached Google for comment, Google removed CMG from its Partners Programme. Amazon said it had not worked with CMG on the product, and Meta opened a review.
No evidence has been produced that Active Listening ever worked as advertised, or that any device microphone was ever used this way. It may well have been a sales document describing a capability the company did not have — advertising technology has a long history of overstating what it can do. But something real is established by it regardless: a substantial media company believed this was a service it could sell, wrote the pitch, and found the claim plausible enough to put in front of buyers.
Where this leaves the claim
The honest summary has three parts, and they do not resolve into a single verdict.
The specific mechanism — apps silently opening the microphone and shipping your speech to advertisers — has been looked for directly, at scale, and has not been found. That is the strongest empirical result available on the question.
The surveillance the theory is reacting to is nonetheless real, and in some respects worse than the version people fear, because it is continuous, legal, disclosed in terms nobody reads, and does not require a microphone. Screen recording without disclosure, location histories, cross-device graphs and brokered purchase data produce the uncanny accuracy without anyone listening to anything.
And the companies have earned the suspicion. Accidental activation was denied in spirit and then conceded in fact. Human review was undisclosed until journalists found it. A broadcaster wrote the pitch deck. Each of those is a real event, and together they explain why a claim that keeps failing its direct test keeps refusing to die.
The belief persists because it is the closest available metaphor for something true. People can feel the profiling and cannot see the machinery, so they reach for the one sensor they know is in the room.