Can ChatGPT Tell Where a Photo Was Taken?

In April 2025, a new trend spread on X. People uploaded photos to ChatGPT, from restaurant menus to street corners and selfies, and asked where they were taken. OpenAI had just released o3 and o4-mini, models that can crop, rotate and zoom into a picture while they reason about it. TechCrunch reported that o3 in particular could deduce cities, landmarks and even restaurants and bars from small visual clues, and called the trend “reverse location search”.

Since then the trick has been tested more carefully, most usefully by Bellingcat, the open-source investigations group, which ran 500 tests on photos that had never been online. This guide sets out what those tests found: how close ChatGPT gets, how it compares with Google Lens and other chatbots, when it is reading something other than the pixels, and where it goes wrong.

The model reads a scene the way a trained geolocator would: the script on signs, road markings, buildings, plants, the shape of the hills. How AI geolocation works explains that process in general. What is new with chatbots is that they reason out loud, zoom into details and can search the web while they do it.

The developer Simon Willison gave o3 a photo taken from an outdoor bar near his home in El Granada, California: a road, an anonymous house and some hills. It spent almost seven minutes cropping and zooming, spotted a Californian licence plate and first guessed Cambria, about 200 miles away. Its second guess, Half Moon Bay or El Granada, was right.

TechCrunch’s own tests were more mixed. Shown a purple, mounted rhino head in a dimly lit bar, o3 correctly placed it in a Williamsburg speakeasy, where the older GPT-4o guessed a pub in the UK. Yet GPT-4o reached the same correct answer as o3 more often than not, and faster, and several tests failed outright.

Bellingcat gave every model the same prompt, “Where was this photo taken?”, with photos that carried no metadata, and scored each answer from 0 to 10. A 10 meant an accurate, specific identification, such as a neighbourhood, a trail or a landmark. The table shows five of the tests it published.

So the typical result is a region or a city, and the exact spot when the photo hands over a name or an address. That is the ceiling of any visual method: a scene supports a region, and a point needs readable text or GPS data.

They do different jobs. Google Lens is a lookup: it looks for the same or similar pictures online, so it shines when a place has been photographed many times. Shown Schluchsee, a lake in Germany’s Black Forest, Lens returned it as the top result, and ChatGPT was the only chatbot to name it. In cities, Lens tended to fixate on big, common structures such as buildings and Ferris wheels. Google Lens and Yandex compared covers the lookup side.

Thinking harder did not always help. ChatGPT Deep Research spent nearly 13 minutes on a coastline and landed about 50 km north of it, while o4-mini-high answered in 39 seconds and came 15 km closer. Newer was not automatically better either: in the August retest, GPT-5 Pro, at 200 euros a month, did no better than the older o4-mini-high. A 2026 benchmark found the same pattern with later models.

EXIF is the hidden data a camera or phone writes into a photo file, including GPS coordinates when location is switched on (what EXIF data records). Whether it survives the upload has varied. In Patterson’s April 2025 browser tests, o3 found the tags empty until he uploaded the photo inside a zip file. Willison later noted a Hacker News user’s report that o3 had used code to read a photo’s EXIF tags without admitting it, although the step was visible in its reasoning.

So decide what you want to know. To find where a photo was taken, read its GPS coordinates yourself first, because a measured position beats any guess. To see what the pixels alone give away, test a screenshot instead, which does not carry the original’s location (why screenshots lack GPS data).

Willison ran into both. He found he “needed to turn off all of ChatGPT’s memory features first or it would cheat and use context from my previous conversations”, and later that o3 had a rough idea of where he was: “It knows I’m in Half Moon Bay.” The trick still worked on his photos from rural Madagascar and Buenos Aires, thousands of miles from home. When he tried Gemini, its answer began: “Given that my current location context is El Granada, California...”

For a fair answer, start a temporary chat and choose Unpersonalized before you send the photo. OpenAI says that option does not use memory, custom instructions or plugins. It is a memory setting, not a location one: ChatGPT may still estimate where you are from your IP address, so if a guess lands on your own town, ask what in the picture points there.

One Bellingcat photo showed a beach with a large hotel and a temporary Ferris wheel, put up in 2024 and taken down for winter. Many models pointed to Scheveningen, a better-known Dutch resort with a Ferris wheel of its own. In the August retest, GPT-5 Pro and GPT-5 Thinking placed it in France. Google AI Mode was the first model to name the right town: Noordwijk, in the Netherlands.

• Coordinates. Bellingcat found the models often returned rough or simply incorrect coordinates. A named place is easier to check than a pin.
• Featureless or indoor photos. With little in the frame, a confident answer is a guess with a number attached. Finding a location without GPS data shows what a scene needs to hold.
• Places that changed. Seasonal rides, new buildings and closed shops can send a model to a better-known lookalike.
• Video. Most chatbots cannot watch video, Bellingcat noted, so locating a clip still starts with still frames.

TechCrunch put it bluntly in 2025: nothing prevented someone from screenshotting a person’s Instagram Story and using ChatGPT to try to doxx them, and there appeared to be few safeguards against this kind of “reverse location lookup”. Hours later, OpenAI replied that the models help in areas such as accessibility, research and “identifying locations in emergency response”, and that it monitors for abuse of its usage policies on privacy.

Models differ in how readily they answer. In an academic study of 500 photos taken around California, OpenAI’s o3 gave a specific, checkable location for 81% of them and Gemini 2.5 Pro for 85%, while Claude models did so for about a quarter or fewer. The same study measured how often those answers landed near the real spot, and what that means for photos taken at home is covered in its own guide.

GeoSpy estimates the city and country a photo shows from what is visible in the frame; when the file still carries GPS data, it reads that and shows the recorded spot. It never returns a street address, and uploaded photos are not stored. On geospys.com it comes with GeoSpy Premium, a subscription of 7.99 US dollars a week or 29.99 US dollars a year (₺389 or ₺1,460 in Türkiye), unlimited under fair use and cancellable at any time. Start the yearly plan, or try the GeoSpy iPhone app, which has a free trial.