After the phase when people used ChatGPT to create action figures from their photos, the new online entertainment seems to be all about location tracking. At first glance, it looks like a game, users upload photos of restaurant menus, selfies or street views and ChatGPT tries to find where the photo was taken. Some even “turn” ChatGPT into a player of GeoGuessr, a popular online game where you have to guess the location based on images from Google Street View.
This may sound harmless, but it also has a serious “dark side.” OpenAI recently released its new o3 and o4-mini models, which have advanced “visual reasoning” capabilities. This means the AI can zoom, rotate, crop, and analyze even blurry images to extract useful information. When these capabilities are combined with the model’s other capabilities, the result is an incredible ability to identify locations.
All it takes is a simple Instagram photo, and the AI can tell which city you’re in, which neighborhood, and even which exact address. In one test, the o3 model was able to tell that a photo of a purple rhino head in a bar was taken at a place in Williamsburg, New York, while GPT-4o mistakenly thought it was a British pub. However, while o3 is more accurate in some cases, it’s not infallible, sometimes getting stuck trying to determine the location.
OpenAI says these capabilities could help in a variety of areas, from accessibility for people with disabilities and scientific research to emergency response. They also claim that the models reject requests for private or sensitive information and that safeguards are in place against identifying private individuals in photos. However, the AI’s ability to guess a location from a photo based on small details (signs, interiors, scenery) still remains potentially dangerous, especially in the hands of those who want to harm someone.