That rule is breaking. A machine can now take small, harmless clues from ordinary photographs and posts, such as the hills behind a porch or a shop sign down the street, and combine them into a conclusion the person never said. 12
The clue we held back was never the only clue. We just never had to worry about the rest.
A nosy neighbor who never sleeps
Every street has had one: the neighbor who notices which car is in the drive, who visits on Sundays and when the lights go off. Over years, such a person can piece together a surprising amount. But a nosy neighbor can watch only one street, forgets most of it, and has to want to know.
That last part was the real protection. Finding out about one person took time, money and a reason, so it happened only to people somebody badly wanted to find. If nobody had a reason to look at you, nobody did.
Now imagine a neighbor who has read everything everyone has ever posted, never forgets a word, and can be hired by a stranger for a few dollars.
That is close to what researchers have been testing. At a Swiss university, ETH Zurich, a team gave AI programs the public comments of real users of an online discussion website, Reddit, and asked them to guess things the users never stated: where they lived, how much they earned, whether they were men or women. On that collection of online accounts, the first guess was right 85 percent of the time, at a tiny fraction of the cost and time of people doing the same job. 2
The same group later went further. In a 2026 experiment, a series of automatic AI steps read a person's comments, pulled out identifying details, searched for matches elsewhere online, and judged which was most likely. The aim was to link anonymous online accounts to real identities, at about one to four dollars for each person identified. On one online film discussion board, users could be picked out from the films they had talked about over the years. 4
Those were tests on particular collections of posts, not a machine that can find anyone, and results differed from one kind of fact to another. The researchers kept the computer instructions for the later experiment private so that they would not become a program anyone could run. 24
The price matters more than the cleverness. The End of the Rave, an essay on Pilobolus, Anthus's opinion site, follows that price falling case by case, from newspapers that spent years hunting one anonymous writer to a few dollars for a stranger nobody was looking for. When finding out costs almost nothing, it no longer happens only to people who matter to someone. It can happen to anybody.
The clue we held back was never the only clue.
Ruth's afternoon
The pieces are real
Pictures. Switching off the place a camera saves inside each photo removes one obvious clue. It does not erase the storefronts, road markings or landscape in the picture. In February 2026, privacy researchers, including the group Privacy International, reported that modern AI can tell where a photograph was taken from what it shows, even without that saved place. The report does not claim that every picture can be pinned to an address. The same skill can also help check where news photographs were really taken. Its worry is who else can use it. 1
Words. The ETH Zurich tests show that a collection of ordinary comments can give away where someone lives and what they do, without any one comment saying so. 24
Likes. The idea is older than today's chatbots. In 2013, researchers studied the Facebook likes, the thumbs-up presses people make, of more than 58,000 volunteers. Shown one Democrat and one Republican at random, their computer program picked the right one 85 percent of the time from likes alone. For one gay man and one straight man, it picked right 88 percent of the time. The authors warned then that the same could be done to large numbers of people without their knowledge. Those figures describe the people they studied, not a test that works on any individual. 3
Sounds. Even what a microphone picks up by accident can be a clue. The sound that can give away where you live follows an imagined employee who secretly shares a company video and is found by the faint hum of a plane in the recording. It also describes the real tests behind that idea.
The rules nobody wrote down
Tell your doctor something, and you expect it to stay in the doctor's office. Tell a friend, and you expect a conversation. Tell the whole street that your grandson turned seven, and you expect smiles. In none of those cases do you expect what you said to be gathered up, combined with everything else you have ever said, and used to work out where you live.
Those expectations are rules, even though nobody wrote them down. The privacy scholar Helen Nissenbaum calls the idea contextual integrity: information has a proper place to flow, set by who shared it, who received it and why. Moving it somewhere else, for another purpose, breaks the rule even when nothing was secret. 5
A machine that adds up clues breaks that rule by design. It takes a remark made to friends, a photograph shared with family and a complaint made to a shop, and combines them into something none of those listeners was ever told. No single rule was broken in any one place. All of them were broken at once.
Why this is so hard to feel
The historian Yuval Noah Harari puts his finger on part of the problem in Nexus, his history of how information shaped human societies. Our minds, he writes, evolved to understand death by a tiger. "Our mind finds it much more difficult to understand death by a document." 6
A stranger at the window sets off every alarm we have. A conclusion drawn in a distant room, from a thousand things we did not think were secrets, sets off none. Nobody feels the moment a photograph and a ten-year-old comment are put together. There is no footstep and no broken lock. Often the person never finds out at all.
That is why adapting will be slow. Societies have changed their manners before, and it was never quick. In 1890, two American lawyers, Samuel Warren and Louis Brandeis, complained that quick snapshot photographs and the newspaper business had invaded private life, and argued for a new legal right to be let alone. The argument they started is still going. 7 The camera was one new tool, held by a person you could see. Today's change comes from many tools at once, used from anywhere, on everything we have ever shared, by people we will never see.
A wrong guess hurts too
A guess is not a fact dug up. It is a bet. A machine may place someone in the wrong town, or pin a wrong income or a wrong religion on them. If a landlord, an employer or an angry stranger acts on that bet, the harm is real whether or not the guess was. The 2013 researchers said as much: a false guess can threaten a person's well-being as surely as a true one. 3
Researchers who interviewed 17 people in China about what AI might read from their social-media videos heard the same fears: being identified, being watched and being judged wrongly. Several had taken to editing or holding back what they posted, while doubting their efforts would catch every clue. The study is small; it describes what those 17 people felt and did, not how accurate the machines are. 8
That doubt is the first cost. Sharing a celebration, asking for help or joining a club online becomes harder when a person must also inspect every background and imagine every future use of it.
Nobody can be careful enough
The easy advice is to be more careful: check the photograph, switch off the place it saves, keep the audience small. Those steps help. They cannot make a picture empty of hills, or ten years of conversation empty of clues, and they ask every person to be a detective against machines that are better at it. 8
The bigger choices belong elsewhere. Websites decide how much of what people post can be copied in large amounts by machines; the ETH researchers suggest making that harder. The companies that build these systems decide which questions their machines will answer about a stranger. The people interviewed in China wanted safeguards they did not have to build themselves. None of these fixes is complete, and all of them are choices. 48
A rule nobody wrote down is being broken by a machine that never learned it.
For most of history, discretion was enough. You kept something private by not saying it, and the clues you left behind were too small, too scattered and too dull for anyone to bother collecting. That protection was never a rule. It was the limit of what a nosy neighbor could manage, and the limit is gone.
What takes its place has not been decided, but the first adjustment is already visible, and it is a loss rather than a law. The people interviewed in China already hold back what they post. 8 There is an old saying: dance like nobody is watching. For most of history, a crowd of strangers was exactly that, and the essay on Pilobolus describes a generation that has never known a night out that was not being filmed, and has learned to dance for the phones instead. Nobody voted for that. It arrived one raised phone at a time, which is how a change that cannot be felt arrives, and most of us will not notice until it has already happened to us.
References
- Nowhere to Hide? Privacy Risks and Policy Implications of AI Geolocation. (n.d.). https://www.privacyinternational.org/report/5736/nowhere-hide-privacy-risks-and-policy-implications-ai-geolocation
- Beyond Memorization: Violating Privacy Via Inference with Large Language Models. (n.d.). https://arxiv.org/abs/2310.07298
- Private traits and attributes are predictable from digital records of human behavior. (n.d.). https://pubmed.ncbi.nlm.nih.gov/23479631/
- The more you post, the easier you are to unmask. (n.d.). https://inf.ethz.ch/news-and-events/spotlights/infk-news-channel/2026/05/the-more-you-post-the-easier-you-are-to-unmask.html
- Privacy as Contextual Integrity. (n.d.). https://digitalcommons.law.uw.edu/wlr/vol79/iss1/10/
- Nexus: A Brief History of Information Networks from the Stone Age to AI (2024), chapter 3, “Documents: The Bite of the Paper Tiger.” (n.d.). https://www.ynharari.com/book/nexus/
- The Right to Privacy (Harvard Law Review, vol. 4, no. 5, 1890). (n.d.). https://groups.csail.mit.edu/mac/classes/6.805/articles/privacy/Privacy_brand_warr2.html
- Through Their Eyes: User Perceptions on Sensitive Attribute Inference of Social Media Videos by Visual Language Models. (n.d.). https://doi.org/10.1145/3733816.3760751