A new ChatGPT trend is drawing attention because it turns ordinary photos into location puzzles. Users are uploading images and asking the system to work out where they were taken, often from small visual clues that a person might overlook.
The trend spread after OpenAI released o3 and o4-mini, its newest AI models. Both can reason through uploaded images, and that matters because the models can inspect photos by cropping, rotating and zooming in, including images that are blurry or distorted.
Why The Trend Took Off
The basic idea is simple: give ChatGPT a picture, then ask it to identify the place. On X, users quickly found that o3 could often infer cities, landmarks, restaurants and bars from details inside the image.
Many examples followed the format of GeoGuessr, the online game where players guess locations from Google Street View images. Instead of a game screen, people fed ChatGPT restaurant menus, neighborhood snapshots, building facades and self-portraits.
That is what makes the trend easy to understand and hard to dismiss. The model is not always relying on an obvious sign or a famous monument. In some cases, it appears to combine subtle image details with web search to narrow down a location.
What The Models Appear To Use
The source article notes that the models often do not seem to be using memories from earlier ChatGPT conversations. They also do not appear to depend on EXIF data, the photo metadata that can reveal information such as where a picture was taken.
Instead, the concern comes from a different capability: visual reasoning paired with web search. If a model can examine small elements in a photo and then look for matches or supporting evidence online, a casual image can become a search problem.
That is powerful in legitimate contexts, but it also changes the risk profile of images people share. A menu, wall decoration, storefront, library interior or street corner can become a set of clues.
How Well It Works
TechCrunch tested o3 against GPT-4o, an older model without the same image-reasoning capabilities. The result was not one-sided. In many cases, GPT-4o reached the same correct answer as o3 and did so faster.
There was at least one case where o3 outperformed GPT-4o. Given a photo of a purple, mounted rhino head in a dimly lit bar, o3 identified it as a Williamsburg speakeasy. GPT-4o instead guessed that the location was a U.K. pub.
Still, o3 was not consistently right. Some tests failed because the model got stuck without reaching an answer it could treat as reliable. In other cases, it gave the wrong location. Users on X also pointed out examples where o3's guesses were far from the real place.
The practical takeaway is not that ChatGPT can always identify a location from a photo. It is that the tool can sometimes do enough to make the behavior worth worrying about.
The Privacy Problem
The most obvious risk is doxxing. The article raises the example of someone taking a screenshot from a person's Instagram Story and using ChatGPT to try to find where that person is.
This matters because the user does not need to own the original photo or have access to its metadata. A screenshot can preserve enough visible detail for an AI model to make an attempt.
The risk is especially clear with images that seem harmless in the moment. A restaurant menu, a neighborhood view or a background facade may feel ordinary, but each can contain location signals when analyzed closely.
- Menus can point toward specific restaurants or bars.
- Facades can reveal storefronts, architecture or street-level identifiers.
- Neighborhood photos can expose local details that narrow the search area.
- Self-portraits can include background clues even when the person is the focus.
What OpenAI Said
TechCrunch reported that there appeared to be few safeguards preventing this kind of reverse location lookup in ChatGPT. The article also said OpenAI did not address the issue in its safety report for o3 and o4-mini.
After the story was published, an OpenAI spokesperson sent TechCrunch a statement. OpenAI said o3 and o4-mini bring visual reasoning to ChatGPT and can help in areas such as accessibility, research and identifying locations in emergency response.
The company also said it trained its models to refuse requests for private or sensitive information, added safeguards intended to stop the model from identifying private individuals in images, and monitors for abuse of its privacy policies.
That response frames the same capability in two ways. Visual reasoning can make ChatGPT more useful when location context is beneficial. But the viral trend shows how quickly the same feature can be tested in ways that create privacy pressure.
For now, the issue is less about perfect accuracy and more about the lowered barrier. If a widely available chatbot can sometimes infer where a photo was taken from visible clues alone, people may need to think differently about what their shared images reveal.