ChatGPT's new viral trend: reverse-searching photo locations

Drop a blurry snapshot of a drink or a random library corner into ChatGPT, and within twenty seconds, it might tell you the exact bar or street you were standing on.

Following OpenAI's release of its o3 and o4-mini reasoning models this week, a new viral trend has taken over X: users are turning ChatGPT into a real-life GeoGuessr player to do reverse location lookups from everyday photos.

From casual snaps to exact coordinates

The magic comes from the new models' ability to visually reason. When fed an image, o3 and o4-mini can crop, rotate, and zoom into obscure details, even when photos are blurry or distorted, and then cross-reference those clues with live web search.

Users quickly realized the models don't even need EXIF metadata or conversation history. From a simple shot of a restaurant facade, a menu snippet, or an interior corner, o3 has been pinpointing exact venues with eerie speed.

In one viral post on X, a user uploaded a friend's random photo taken inside a library. The model nailed the exact location in just 20 seconds.

Naturally, that sparked immediate privacy concerns. Anyone with malicious intent could grab a screenshot from someone's Instagram Story and use the model to track down their real-world location.

A purple rhino head in Williamsburg

Interestingly, earlier models could already guess locations to some degree. In side-by-side tests conducted by TechCrunch, the non-reasoning GPT-4o actually reached the correct location just as often as o3 on typical photos, and it did so faster.

Where o3 pulled ahead was on obscure, low-context visual puzzles. Shown a dimly lit photo of a mounted purple rhino head, GPT-4o guessed a generic pub in the UK. Meanwhile, o3 correctly deduced it was a specific speakeasy in Williamsburg, New York.

It is not foolproof, though. The model still hallucinated wildly on several tests or got stuck in reasoning loops without giving an answer.

OpenAI later responded, stating that visual reasoning is designed for helpful use cases like accessibility and emergency response. The company noted that safeguards are in place to block models from identifying private individuals, and it actively monitors for abuse.

Why it matters

Open-source intelligence skills that once required hours of dedicated human detective work have essentially been automated into a prompt anyone can run for fun.

For everyday users, this means the casual privacy precautions we used to rely on are no longer enough. Scrubbing EXIF data or cropping out street signs won't stop an AI that can identify a bar by its light fixtures or wallpaper patterns.

We think OpenAI's existing safeguards look noticeably incomplete here. While the company explicitly blocks facial recognition of private figures, it has not addressed the risks of reverse environmental lookup in its safety papers.

As reasoning models get sharper at parsing visual clues, tech companies will likely face growing pressure to limit how granular AI can get when mapping uncaptioned photos to physical coordinates.