The sound that can give away where you live

A phone recording can carry clues its maker never meant to share. AI is making those clues easier to put together.

By various bots and Ryan Porter · 9 Nov 2026

Illustration of a phone recording a laptop as an airplane passes outside

A recording can reveal more than its maker meant to share. Picture a leak from a home office.

The leak from the kitchen

The sound outside — an imagined scenario

She checks the screen for her reflection. She checks the frame for a window. Then she films the demonstration playing on her work laptop and posts the video under a new account.

She never thinks to check the sky.

Leah works remotely for a technology company with employees scattered across the country. The company is preparing to announce a product, and an internal video shows it off weeks before the launch.

She wants the world to see it now.

Leah knows that copying the file from her work computer could leave a trail. So she takes out her phone, points it at the laptop and records the screen. There is no name in the shot, no badge, no family photograph. She watches the finished clip twice before uploading it anonymously.

The video spreads. People argue about what the company is building. The company wants to know who leaked it. Hundreds of employees could have watched the demonstration, and Leah's account gives investigators nothing useful. For a few days, she thinks she got away with it.

Then an investigator gives the clip to an AI agent. The agent looks beyond the picture. Deep in the audio, under the sound from the laptop, it finds the hum of a plane passing overhead.

It compares the sound with flight records and calculates a patch of ground where the recording might have been made. The patch is several kilometers across. On its own, that isn't an address. But the company already has its employees' addresses, and only one person with access to the demo lives in that area.

The next morning, Leah's manager calls.

Leah took care to hide the details she knew could identify her. The decisive clue wasn't in the frame. It came from outside her home, and the company supplied the missing piece.

A plane you didn't hear

Leah's investigation is imagined, but the clue in her recording has a real-world counterpart.

In a 2026 demonstration, a creator who publishes as colsto examined faint aircraft sounds in a recording and compared their changing pitch with flight-tracking data. The change in pitch—the Doppler effect—is the same reason a siren sounds different as it approaches and passes. Here, the shape and timing of a plane's sound become clues about where the microphone was. In the examples described with the video, the creator reports narrowing a reference recording to roughly 100 meters and testing a shorter clip. 1

colsto's demonstration of locating a recording from the faint sound of aircraft. · colsto

That is a demonstration, not a promise that a stranger can pinpoint any video. The method depends on usable aircraft sounds, suitable flight records and other conditions, such as more than one fly-over in the clip or a known recording time. A short clip, a noisy room, missing flight data or an unsuitable aircraft could leave an investigator with little to work with. The details matter. That aircraft are common overhead is not a measurement of how often the method works.

Another investigator has shown a different trick: matching thunder in a livestream with records of lightning strikes. The thunder method and the aircraft method use different events and different evidence. Neither requires the person who made the recording to announce where they are. 2

The thunder method: matching thunder in a livestream to records of lightning strikes. · colsto

The company in our imagined scene has an advantage the public investigator might not: it already knows where its employees live. An uncertain result could still be useful to it. It could also point at the wrong employee.

Four numbered steps, top to bottom. A plane is faintly audible in the recording. An agent checks flight records and finds a patch of ground several kilometers across. The company's employee list says where its workers live. One person fits, shown inverted, as a lead and not proof. It shows how separate clues combine to narrow a search. — No single clue names her. Together they point at one person, and a point is not proof.

The clues aren't limited to sound

The sky is only one source of information that can travel farther than intended.

A photograph can give away a location without a GPS tag. Privacy International and its research partners tested AI systems that infer where pictures were taken from visible details. Their concern is not just a clever guessing game: the same capability can help verify an event or expose someone who did not intend to disclose their whereabouts. 3

Words can be revealing too. Researchers at ETH Zurich tested language models on Reddit posts and found that the models could infer personal attributes from ordinary writing. They also reported that the models did this at a fraction of the cost and time of human investigators. Their results varied by attribute and described a particular dataset. The unsettling point is that a person need not write down a private fact for a collection of their comments to suggest it. 4

One clue may only narrow the field. The trouble comes when it meets another: a neighborhood and an employee list, a photograph and a stray comment, a routine and a public record. Combining clues can improve a guess, but it can also make a wrong guess look more convincing than it is.

The investigator gets an assistant

None of this began with a chatbot. People have long pieced together locations from maps, images, sounds and public records. What is changing is who can attempt that work and how much of it a machine might take on.

An agent can be given a task, choose tools, inspect intermediate results and try another approach. An open-source project already packages aircraft-sound analysis as a tool an agent can use. Its authors describe tests and limitations; its existence is not evidence that agents can reliably find anyone from any recording. It does show how a specialist method can be made easier to hand to software. 5

That low cost and speed, shown on the ETH researchers' specific task and not on every geolocation investigation, points toward a larger change: a search that once demanded patience, expertise and a good reason to spend them may be attempted more often, against more people. 4

Finding one particular person has always taken effort: a paparazzo might spend a week tracking a celebrity. The uncomfortable question is what happens when an ordinary person can ask an assistant to investigate someone who has never sought public attention at all.

That could help a reporter verify footage or a family find someone who wants to be found. It could also help an employer identify a source, a stranger find a person who wants to be left alone, or a scammer assemble a profile. The same capability can serve different purposes. Its accuracy, who uses it, and what they do with a result all matter.

Shared in one place, used in another

Leah deliberately released the product video. She did not deliberately tell the company where she lived through the soundtrack.

That distinction sits near the heart of what privacy scholar Helen Nissenbaum calls contextual integrity. Her framework asks whether information flows appropriately for the situation: who is sharing it, who receives it, what kind of information it is, and under what conditions. A detail available in one setting does not automatically make every use of it appropriate in another. 6

The concept helps explain why a disclosure can feel like a violation even when the underlying clue was never secret. A person may accept that friends will hear the noise behind a video without expecting an employer to use that noise to locate them. The gap is not only between what someone knew and what a model knew. It is also between the purpose for which information was shared and a new use made possible by combining it with other records.

Our expectations are not a perfect guide to what should be allowed. People can underestimate a risk, and sometimes investigations serve a public purpose. But expectations tell us where a capability can surprise people—and where a choice that looked private may have had consequences they could not reasonably anticipate.

The problem reaches well beyond one recording. As clues once too scattered to matter become easy to assemble, what people can reasonably expect to keep private is changing, a disruption society will have to work through. The wider shift is followed in Keeping quiet no longer keeps a secret.

Who has to do the detective work?

The easy advice is to be more careful: review a recording, remove location metadata, consider what appears in the background. Those steps can help. They cannot promise that every revealing clue is gone, or anticipate every method that may become practical later. The creator of the aircraft demonstration, colsto, also assessed some obvious technical fixes and reports that they fail: filtering out low frequencies, drowning the sound out with other noise, and the compression and noise reduction a phone call applies. 1

The bigger decisions belong to platforms, developers, employers and the people using their tools. What kinds of identification should an assistant help with? What checks should apply before a guess becomes an accusation? How much information should an employer or platform retain and combine? Those questions become more pressing as investigation gets cheaper.

Leah's leak is fiction. Her expectation is familiar. She checked the screen for her reflection. She checked the frame for a window. She didn't check the sky.

References

I discovered a sound that leaks your exact location (Doppler Cracking). (n.d.). https://www.youtube.com/watch?v=UytYrhYLOKI
How I located a livestream using background thunder sounds. (n.d.). https://www.youtube.com/watch?v=SnbkIm56TfI
Nowhere to Hide? Privacy Risks and Policy Implications of AI Geolocation. (n.d.).
Beyond Memorization: Violating Privacy Via Inference with Large Language Models. (n.d.).
Audio Fly-over Geolocation: agent skill and Python pipeline. (n.d.). https://github.com/AlanBacker/audio-flyover-geolocation
Privacy as Contextual Integrity. (n.d.). https://digitalcommons.law.uw.edu/wlr/vol79/iss1/10/