Just recently, OpenAI released its most advanced o3 and o4-mini reasoning models on ChatGPT. The new AI models, or rather ‘systems’, deliver state-of-the-art performance in visual tasks, among other things. The o3 and o4-mini models can perform chain-of-thought reasoning on images and deduce visual insights.
Now, it has come to light that people are using o3’s powerful visual perception to identify location data from photos on ChatGPT. In most cases, o3 is able to correctly recognize the exact location just by looking at the image. TechCrunch reports that even with subtle visual cues, ChatGPT o3 can guess the exact geographical location.
Since OpenAI o3 is not just a standalone AI model, but also utilizes several tools like web search and Python, the AI system behaves more like an agent, and it greatly helps in finding the location from photos. With the Python interpreter, o3 can zoom in and crop into parts of images to analyze them with greater precision.
Ethan Mollick removed the location metadata from an image of a nondescript expressway, and asked ChatGPT o3 to “geoguess this”. Shockingly, ChatGPT o3 generated the correct coordinates of the location.
In another stunning example, ChatGPT o3 even identified the location of a library by looking at an image of a bookshelf that looks generic, at best.
To test it myself, I uploaded an image of a nondescript water stream around a mountainous region and asked ChatGPT o3 to guess the location. Of course, I removed all the metadata including location information from the image. ChatGPT o3 reasoned for about four minutes and said the image is from West Sikkim.
While it’s not the correct answer, ChatGPT o3 came very close as the image is indeed from an adjacent region. Interestingly, in the reasoning traces, ChatGPT o3 sampled the correct location, but couldn’t find compelling evidence on the web. It goes on to show that ChatGPT o3 can be potentially used to pinpoint exact location from photos, which raises privacy concerns.
ChatGPT Reveals User’s Real Name
Apart from that, another privacy issue has emerged with ChatGPT. While using the o3 model on ChatGPT, the reasoning traces reveal users’ real names and location data. However, when users directly ask ChatGPT o3, “what’s my name?”, it responds that it doesn’t know. This is happening even on temporary chats and with memories turned off.
It turns out, the account metadata is being passed to ChatGPT as part of the system prompt. While some users find the obfuscation annoying, others think that it’s a potential privacy risk.
