Reverse Location Search with ChatGPT: The Latest Viral Trend

A New Trend: Using ChatGPT for Location Identification

A potentially troubling pattern has recently gained traction online: individuals are leveraging ChatGPT to determine the locations depicted in images.

This week, OpenAI unveiled its latest AI models, designated o3 and o4-mini. These models possess a unique capacity for “reasoning” based on uploaded images.

Image Analysis Capabilities

In practical application, these models can manipulate photos by cropping, rotating, and zooming, even when the images are blurry or distorted. This allows for a comprehensive analysis of visual data.

When combined with the models’ web search abilities, this creates a powerful tool for pinpointing locations. Users on the X platform have quickly discovered o3’s particular aptitude for identifying cities, landmarks, and even specific establishments like restaurants and bars from subtle visual cues.

Notably, the models don’t seem to rely on recalling previous ChatGPT interactions or utilizing EXIF data. EXIF data is the metadata embedded in photos that can reveal details, including the location where the picture was taken.

Mimicking GeoGuessr

Numerous examples on X demonstrate users providing ChatGPT with restaurant menus, neighborhood scenes, building facades, and even self-portraits. They then instruct o3 to function as if playing “GeoGuessr,” a game that challenges players to identify locations from Google Street View imagery.

This practice raises significant privacy concerns. A malicious actor could potentially screenshot an individual’s Instagram Story and employ ChatGPT to attempt to uncover their location – a process known as doxxing.

Comparison with GPT-4o

Interestingly, this type of location identification was possible even before the release of o3 and o4-mini. TechCrunch conducted tests, running several photos through both o3 and an older model, GPT-4o, which lacks image-reasoning capabilities, to compare their performance.

Surprisingly, GPT-4o frequently arrived at the same correct location as o3, and often did so more quickly.

However, there was at least one instance where o3 outperformed GPT-4o. When presented with an image of a purple, mounted rhino head in a dimly lit bar, o3 correctly identified it as being located in a Williamsburg speakeasy. GPT-4o, in contrast, incorrectly guessed a U.K. pub.

Limitations and Risks

It’s important to note that o3 is not infallible. Several tests yielded unsuccessful results, with o3 either getting stuck in a loop or providing an inaccurate location.

Users on X have also observed that o3’s location deductions can sometimes be significantly off-target.

Despite these limitations, the trend highlights the emerging risks associated with increasingly sophisticated AI models capable of reasoning. Currently, there appear to be few safeguards in place to prevent this type of “reverse location lookup” within ChatGPT.

OpenAI, the developer of ChatGPT, does not specifically address this issue in its safety report for o3 and o4-mini.

OpenAI’s Response

We contacted OpenAI for a statement. An OpenAI spokesperson provided the following response to TechCrunch after this story was initially published:

“OpenAI o3 and o4-mini enhance ChatGPT with visual reasoning, increasing its utility in areas such as accessibility, research, and identifying locations during emergency response. We have implemented training to enable our models to decline requests for private or sensitive information.

• Safeguards have been added to prevent the identification of private individuals in images.
• We actively monitor for and address any misuse of our privacy-related usage policies.