ChatGPT can now localize image locations much more precisely

Determining the location where a photo was taken just by looking at it – in an analog way is a popular pastime for some people, but for ChatGPT users it is now often a matter of a few clicks and a few seconds. As OpenAI has now confirmed, its Large Language Models (LLMs) ChatGPT o3 and o4-mini are now able to perform this task.

There are numerous testimonials online from users who give ChatGPT a randomly selected image and ask: "Where in the world was this image taken?" They give the LLM menus, self-portraits and more. ChatGPT apparently provides correct answers very often.

ChatGPT apparently uses images from map services

The tests conducted by heise online with o4-mini were only partially successful. In the case of a photo with a striking mountain and lake landscape, ChatGPT immediately recognized the location correctly. However, the exact coordinates and the Google Maps location were not entirely precise. The situation was different with a random photo taken in a German residential area, without any clues such as signs with street names or bus stops. ChatGPT spent around one and a half minutes zooming in on various places in the image that it thought would provide clues. In the end, it apparently provided – with an incorrect location based on street images from map services –, referring to architectural details that it thought it had recognized in the image.

ChatGPT apparently does not use EXIF data – i.e. metadata of a digital image, which often also contains location data. The LLM also does not appear to use its knowledge from past conversations. Although it was also possible to find out the location of images with earlier models, ChatGPT now analyzes the images much more precisely. For example, individual image sections are examined more closely and it applies these more detailed findings to a web search. OpenAI also describes this capability in a press release on the new image reasoning functions of o3- and 04-mini. However, there is no explicit mention of localization, which many users are currently trying out.

The new possibilities also raise the question of the potential for misuse. It would probably be very easy to find out the home addresses of some social media users. A spokesperson for OpenAI told the US news portal TechCrunch that the models are trained to reject requests for private or sensitive information. Protective mechanisms are designed to prevent the model from identifying private individuals in images. Abuse of the data protection guidelines would be actively monitored and action taken if necessary.

(nen)