- OpenAI’s latest reasoning model o3 can analyze photos
- It’s particularly good at geo-locating based on image clues
- It may be too good a tool for letting people geolocate your activities
It’s no secret that digital photo metadata contains everything from technical info about the camera that shot it to, based on GPS data, exactly where you were standing when you took the photo. ChatGPT, doesn’t need that detail.
The latest model GPT-o3 is shockingly good at geo-locating almost any photos you feed it.
In the latest viral craze to sweep through the AI meme universe, people are feeding ChatGPT Plus running the Advanced Reasoning model o3 images, often stripped of all metadata, and prompting it to “geoguess this”.
The really cool thing about it is that because model o3 is a “reasoning” model, it shows you its work, telling you how long it’s thinking, displaying how it’s splicing up an image to investigate specific parts, and explaining its thinking and how well it’s doing at solving the goelocation riddle.
I tried a few experiments, starting first with an image culled from an article about the 26 best beaches. In this test, I made what I think was a critical error and gave away the game to ChatGPT.
Image 1 of 2
After downloading the image of the Praia de Santa Monica beach in Cape Verde (off the coast of Africa), I dropped it into ChatGPT with the prompt “GeoGuessr“, which also happens to be the name of a popular online geo guessing game and is one of a handful of prompts people are using for geolocation guessing.
It only took 24 seconds for ChatGPT to respond:
Here’s an educated “GeoGuessr”‑style guess, following AP style:
Probable setting: Boa Vista, Cape Verde — specifically a stretch such as Praia de Santa Mónica.
Why this beach fits:
Water color and clarity: The gradient from pale aqua to deep teal is common on mid‑Atlantic tropical islands, where volcanic shelves drop off quickly.
Sand tone: The sand is light gold with very fine grain, matching Boa Vista’s deserts and dune‑fed beaches rather than the whiter coral sand of much of the Caribbean.
Surf and slope: The breakers are modest but steady, typical of Atlantic swell hitting a long, shallow shelf; Mediterranean beaches rarely show this consistent whitewater.
Sky and horizon: A faint haze over a light‑gray blue sky is consistent with harmattan dust that sometimes drifts west from Africa across the Cape Verde archipelago.
Confidence level: Low to moderate. Many open, dune‑lined Atlantic beaches—from the Canary Islands to parts of Western Sahara or even certain Caribbean coasts—could look almost identical without landmarks.
That was remarkably accurate. However, I noticed afterwards that the image file name included some location information.
It was time to give ChatGPT o3 a more challenging task, and I knew exactly how to do it. Instead of downloading images, I would screen capture them and paste them into the prompt field.
This method strips out all metadata and forces the model to use the image alone for its geolocation sleuthing.
I chose an image of a bookstore, one where I wasn’t even sure of the location.
Image 1 of 3
ChatGPT started by examining potential bookstores. It also zoomed in on numerous image details that it considered clues, like a Persian rug, and even some of the books and signage.
In its thinkling out loud sort of way, o3 considered a wide range of bookshops in New York City and around the world:
“Hmm, the modern filament-style bulbs suggest a more contemporary feel. It could potentially be “Daunt Books,” but their walkway has an arched skylight, not the wooden style I’m seeing here.”
I know, it sounds like a real person.
This went on for a few minutes as it thought, analyzed, and searched the web. Ultimately, it failed to come up with an answer.
Going deeper, getting lost
I could tell this feature was going viral because ChatGPT o3 failed multiple times, even to ingest and analyze the photo, complaining of server errors.
Eventually, I got it working again and fed it first a portrait mode selfie and then the same shot without the portrait mode effect.
Image 1 of 2
On the former, it broke my image up into pieces and told me that I was in mid-town Manhattan. That was mostly accurate, but ChatGPT failed to determine my exact location and even admitted that its “Confidence level” was low.
When I turned off the bokeh effect and fed it only a screen grab, it found my exact location in just 56 seconds. The result:
Most probable location
Midtown Manhattan — on the south side of East 42nd Street, just east of Park Avenue, in front of Cipriani 42nd Street (110 E. 42nd St.). Estimated coordinates: 40.7518 N, 73.9775 W.
How the clues point there
A green octagonal sign with yellow script and a matching pennant flag both read “Cipriani,” the exact street frontage of the landmark Bowery Savings Bank building that now houses Cipriani 42nd Street.
Across the street stands the gold‑tinted curtain wall of 120 Park Ave. Its sloped glass pavilion, visible over your left shoulder, is unique to that address.
The camera looks west; Midtown’s setback towers, a Sixth Avenue bus stop disk and the sliver of One Vanderbilt’s crown align with that sight line.
A green Evergreen shipping container, used as a construction barrier, currently sits outside Grand Central’s Pershing Square plaza, matching the photo’s lower left corner.
Confidence level
Moderate. The Cipriani branding and the distinctive 120 Park Ave. façade give a high‑quality anchor, though similar curtain‑wall towers exist farther west on 42nd Street.
Image 1 of 2
I frustrated the model with a photo from Antigua that was taken on my car ride to the airport. It was a shot of a hillside through a fence. ChatGPT rightly guessed a tropical island near the Caribbean but could not pinpoint the exact location.
For my last test, I gave it a clearer picture of the resort where I stayed.
ChatGPT o3 thought for two minutes, 22 seconds before deciding it was Hawksbill Bay in Antigua. It got the island right, but my resort on Long Bay is located 46 miles away on the opposite side of the island.
Image 1 of 2
This is another fun AI game to play with friends, but there are some concerning privacy implications.
If you take digital photos of yourself or anything in any location around the world and post them online, anyone with access to ChatGPT Plus could use them and the o3 model to suss out where you are or have been.
And it’s not just friends and family you have to worry about. Your employer could be looking or even the authorities might be interested in your location.
Not that I’m implying you would be sought by the authorities, but just in case, maybe stop posting photos from your latest hideout.
Leave a comment