- Openai O3’s latest reasoning model can analyze photos
- This is particularly good in geo-localization based on image indices
- It may be too good a tool to let people geolocate your activities
It is not a secret that digital photo metadata contains everything, technical information on the camera that turned it to, on the basis of GPS data, exactly where you stood when you took the photo. Chatgpt, does not need this detail.
The latest GPT-O3 model is incredibly good in geo-localization of almost all the photos you feed it.
In the last viral craze to sweep the world of the same AI, people feed Chatgpt more performing the advanced images of the O3 reasoning model, often stripped of all metadata, and inciting it to “Geoguess this”.
What is really cool is that because the O3 model is a model of “reasoning”, he shows you his work, telling you how long he thinks, displaying how he spares an image to investigate specific parts and explain his thought and how much he is to resolve the Goelocation enigma.
I tried some experiences, starting first with an image drawn from an article on the 26 best beaches. In this test, I did what I think is a critical error and I gave the game to Chatgpt.
Picture 1 of 2
After downloading the image of Praia de Santa Monica beach in Cape Verde (off the coast of Africa), I left it to Chatgpt with the “Geogessr” prompt, which is also the name of a popular online geo gaming game and is one of the guest handles that people use for the geolocation guess.
It only took 24 seconds for Chatgpt to answer:
Here is an educated supposition of “geoguessr” style, according to the AP style:
Probable adjustment: Boa Vista, Cape Verde – In particular a section like Praia de Santa Mónica.
Why this beach adapts:
Water color and clarity: The gradient of the pale aqua to a deep sarcelle is common on the medium tropical islands, where volcanic shelves fall quickly.
Sand tone: The sand is in light gold with a very fine grain, accompanied by the deserts of Boa Vista and the beaches derived from dune rather than the white coral sand of most of the Caribbean.
Surf and slope: The circuit breakers are modest but stable, typical of the swell of the Atlantic hitting a long shallow shelf; Mediterranean beaches rarely show this coherent living water.
Heaven and horizon: A slight mist on a light gray blue sky is compatible with Harmattan dust which sometimes derives west of Africa through the Cape Verde archipelago.
Level of confidence: Low to moderate. Many beaches of the Atlantic open and doubled from the dunes – from the Canary Islands to certain parts of the Western Sahara or even certain Côtes du Caribbean – could seem almost identical without a benchmark.
It was remarkably precise. However, I noticed that the name of the image file included location information.
It was time to give Chatgpt O3 a more difficult task, and I knew exactly how to do it. Instead of downloading images, I would capture and colmis in the fast field.
This method eliminates all metadata and obliges the model to use the image alone for its geolocation detective.
I chose an image of a bookstore, the one where I was not even sure of the location.
Picture 1 of 3
Chatgpt began by examining potential bookstores. He also zoomed out on many image details which he considered as clues, as a Persian carpet, and even some of the books and signaling.
In his way of thinking, O3 considered a wide range of bookstores in New York and in the world:
“HMM, the modern filament style bulbs suggest a more contemporary feeling. It could be potentially” Dunt Books “, but their bridge has a well -fitted light, not the wooden style I see here.”
I know, it looks like a real person.
It lasted a few minutes as he thought, analyzed and searched the web. In the end, he failed to answer.
Go further, get lost
I could say that this feature became viral because the Chatppt O3 failed several times, even to ingest and analyze the photo, complaining of server errors.
Finally, I made it work again and first nourished it a selfie in portrait mode, then the same blow without the portrait fashion effect.
Picture 1 of 2
On the first, he broke my image in pieces and told me that I was in the middle of the city of Manhattan. It was above all correct, but Chatgpt failed to determine my exact location and even admitted that his “level of confidence” was low.
When I turned off the Bokeh effect and I only fed him, he found my exact location in just 56 seconds. The result:
Most likely location
Midtown Manhattan – On the south side of the 42nd East rue, just east of Park Avenue, ahead of Cipriani 42nd Street (110 E. 42nd St.). Estimated contact details: 40.7518 N, 73,9775 W.
How the clues point there
A green octagonal panel with a yellow script and a matching flag flag both read “cipriani”, the exact facade of the rue du Bâtiment de la Banque d’épargne Bowery which now houses Cipriani 42nd Street.
On the other side of the street stands, the Rideau de Tired Gold of 120 Park Ave. Its sloping glass pavilion, visible on your left shoulder, is unique at this address.
The camera looks west; Midtown’s backpins, a bus stop disc of the sixth avenue and the crown of the crown of a Vanderbilt align on this line of view.
A green persistent leaf expedition container, used as a construction barrier, is currently outside the Pershing Square Plaza Grand Central, corresponding to the lower left corner of the photo.
Level of confidence
Moderate. The Cipriani brand and the distinctive facade 120 Park Ave. Give a high quality anchor, although similar curtain walls exist further west on the 42nd street.
Picture 1 of 2
I frustrated the model with an antigua photo that was taken during my car trip to the airport. It was a photo of a hill through a fence. Chatgpt rightly guessed a tropical island near the Caribbean, but could not determine the exact location.
For my last test, I gave him a clearer image of the station where I am stayed.
Chatgpt O3 thought for two minutes, 22 seconds before deciding that it was Hawksbill Bay in Antigua. It has done the island well, but my seaside resort on Long Bay is located 46 miles from there across the island.
Picture 1 of 2
It’s another fun game to play with friends, but there are implications for confidentiality.
If you take digital photos of yourself or anything in a place from around the world and you publish them online, anyone who has access to Chatgpt Plus could use them as well as the O3 model to determine where you are or have been.
And these are not just friends and the family you need to worry about. Your employer could seek or even the authorities could be interested in your location.
Not that I imply that you would be wanted by the authorities, but just in case, may stop publishing photos of your last hiding place.
