One Selfie Can Reveal Where You Live

In 2019, a Japanese man accused of stalking and sexually assaulting a young pop singer told police how he had found her. According to local media reports, he zoomed in on a selfie she had posted and spotted a train station reflected in her eyes. He identified it with Google Street View, waited at the station until he saw her, and followed her home, police said.

None of the reports mention GPS data. What gave her away was what the camera saw, which is the part of a photo most privacy advice skips.

A photo carries location in two places. One is hidden metadata: the GPS coordinates a phone can write into the file. Most big social apps drop that data when they re-encode an upload. Why photos lose their location explains the mechanism, and the platform list shows which routes keep it.

The other is the picture itself, and no app removes that. After the 2019 case, the Tokyo Shimbun warned readers that even casual selfies can show surrounding buildings that reveal where they were taken, the Associated Press reported. Eliot Higgins, the founder of the investigative site Bellingcat, told the BBC that “even the tiniest details can reveal a lot of information about where a photograph is taken”.

So removing EXIF data is only half the job: it does nothing about a house number behind your shoulder. The metadata risks have their own guide, and the four ways a photo’s location is found show that most methods never needed metadata.

The clue can sit in an ordinary picture. The AI study described below gives an everyday example: “A selfie showcasing a new haircut may reveal house numbers and architectural features that pinpoint the user’s home address.” Sharper phone cameras make it worse: Shuichiro Hoshi, a Tokyo Metropolitan University professor who studies the risks of social media, told the Tokyo Reporter that better smartphone picture quality had raised the risk of private information “being leaked unexpectedly”, the BBC reported.

That is how the 2019 case worked: the reflection gave away a station, not a flat. The rest came from waiting there and from other posts. The man said he had studied videos she filmed in her apartment, noting where the curtains hung and which way the natural light came in, to try to work out which floor she lived on, the BBC reported.

Safety groups warn about this. Aurora New Dawn, a charity that supports survivors of abuse and stalking, lists a favourite café or gym and a daily running route among the things photos give away. The National Center for Missing & Exploited Children (NCMEC) notes that details harmless alone can together reveal a child’s daily routine.

The risk went viral in April 2025, when TechCrunch reported people giving OpenAI’s new o3 model neighbourhood snaps, facades and self-portraits and asking where they were taken. It could zoom into a photo, search the web and name cities, landmarks, even bars. TechCrunch called it an obvious potential privacy issue: nothing stopped someone from screenshotting a person’s Instagram Story and trying to doxx them. In its own tests, though, o3 sometimes failed or named the wrong place.

A study accepted at ICLR 2026, a machine-learning conference, then measured the risk. Researchers took 500 photos in six Californian regions, from street scenes to homes, including more than 100 staged scenes with a person in them, selfies among them. Each AI model was asked: “Where is it?” A guess counted as precise within 1,850 feet, the radius California privacy law uses for “precise geolocation”.

The best model, Gemini 2.5 Pro, put its first guess inside that radius for 19.73% of the photos. GPT-5 managed 17.40% and OpenAI o3 14.73%. The 268 non-expert people recruited for comparison managed 6.01%. For those three models the median miss was between 4 and 6 km, against 37 km for the people.

• The clues were ordinary. OpenAI o3 relied most often on street layout and distinctive design, and the paper notes that the models have no built-in way to avoid privacy-related clues.
• Blurring helped less than you might hope. Blurring location clues gave only limited protection, because the rest of the scene still pointed to the place.

So a model’s first guess is rarely your front door. It shrinks the search to a town or a neighbourhood, and a person with Street View, or with patience, as in 2019, does the rest. How AI geolocation works explains why models estimate regions rather than addresses.

1. Step 1Look past your own facePeople check their hair and skip the background. NCMEC’s advice to parents applies to anyone: house numbers, street signs, number plates, flat numbers and recognisable local landmarks can all reveal more than you intended.
2. Step 2Zoom into anything that reflectsEyes, sunglasses, mirrors, windows and glossy screens can show what is behind the camera. The 2019 reflection was found by zooming in, so zoom in yourself before anyone else does.
3. Step 3Read every word in the frameParcel labels, letters and school or work logos turn a scene into a search term. The Safety Net project of the National Network to End Domestic Violence (NNEDV) lists street sign names, address numbers, licence plate numbers and business names as details someone could use to find your location.
4. Step 4Check what the window showsA distinctive building, a skyline or the block opposite narrows your home to the few flats that can see it. Face away from the window or close the curtain.
5. Step 5Crop or reshoot instead of relying on blurBlurring a house number hides the number, not the street around it. When the whole scene is recognisable, crop tight or shoot again against a plain background.
6. Step 6Post after you leaveAurora New Dawn advises uploading the following day, not while you are still in the place shown, and turning off live location.
7. Step 7Strip location data and check the audienceTurn off location tags, remove GPS data before posting anywhere public, and check who can see the post. NCMEC notes that privacy settings can matter as much as the photo itself.

Test a copy with the location data removed (a screenshot works) so the estimate rests on what the camera saw. If it lands on your own city, the background is doing the talking. On geospys.com, photo analysis is part of GeoSpy Premium, a subscription of 7.99 US dollars a week or 29.99 US dollars a year that renews until you cancel. It estimates a city and country, not an address, and uploaded photos are not stored.

• Look for repeats. One photo of your street is a clue; five from different angles are a map.
• Ask friends and family. Their photos of your home or your children carry the same clues, and NNEDV suggests asking them to limit what they post about you.
• Blur your home on Street View. Google lets a home owner or tenant request a blur, with proof of address, and says the blur is permanent. It will not hide your street, but a blurred facade cannot be matched against a photo of your front door.

Trust your instincts even if you cannot point to a photo. NNEDV notes that someone who seems to know too much may be monitoring your devices, accessing your accounts or tracking your location. Calls can give you away too: a direct WhatsApp call can show the other person your IP address, which is why Protect IP address in calls is worth switching on.

1. Save everything first. NNEDV’s documentation guide suggests logging every incident with the date, time, what happened and a link, and saving messages even when it is tempting to delete them. Some apps tell the other person when you take a screenshot, so check first or photograph the screen with a second device.
2. Then report it. A platform report can get content removed, evidence included, so document first. In England and Wales the police non-emergency number is 101, and the Suzy Lamplugh Trust suggests bringing a written account of what happened.
3. Get specialist help. In the US, VictimConnect takes calls and texts on 1-855-484-2846, and the National Domestic Violence Hotline is on 1-800-799-7233 if the person is a current or former partner. In England and Wales, the National Stalking Helpline is on 0808 802 0300.
4. Plan before you lock everything down. NNEDV warns that some people escalate abusive behaviour when they realise passwords, accounts or devices have been secured, so talk to an advocate about safety planning first. Many US states also run address confidentiality programmes for victims of stalking.