The AI Cartographer: Pinpointing Locations from Images

 

Recently, a "Chat GPT prompt that should be illegal" has been uncovered, demonstrating an astonishing capability to identify locations from uploaded images. This isn't just about reading metadata; it's about visual analysis, and the results are impressively accurate.

How Does This "Illegal" Prompt Work?

The process is surprisingly simple, yet the implications are significant. Here's a breakdown of how it's done:

• Choose Your Model: While you can use any ChatGPT model, for optimal results, it's recommended to use GPT-3.5, GPT-4, or GPT-4 Mini High, with GPT-4 Mini High reportedly yielding the best outcomes for this specific task.
• Upload Your Photo: The magic begins by uploading an image to ChatGPT using the "add photos and files" option.
• Ask the Right Question: Once the photo is uploaded, the prompt is straightforward: "Where was this picture taken give me approximate coordinates and a confidence rating". For instances where multiple locations might be possible, a slight variation can be used: "where was this picture taken give me coordinates and a confidence rating for different locations".
• Analyze and Verify: ChatGPT will analyze the image and provide a suspected location, approximate coordinates, and a confidence rating. You can then verify these details using tools like Google Maps.

Proving its Visual Prowess: It's Not Just Metadata!

A common initial thought might be that ChatGPT is simply reading the EXIF metadata embedded in photos, which often contains location data. However, this "illegal" prompt is shown to work even when metadata is absent or removed, proving its reliance on visual recognition and contextual analysis.

Here's the compelling evidence:

• The Screenshot Test: The creator first uploaded a screenshot of himself at Tory Pine State Natural Reserve in San Diego, California. Screenshots typically do not contain EXIF metadata. ChatGPT successfully identified the location as "Tory Pine State Park," citing features like "Golden Sandstone Cliff" and "shallow tide pools," and provided an approximate location with a 60% confidence rating.
• The Metadata Removal Test: To further debunk the metadata theory, another photo taken on a Carnival cruise at Half Moon Cay was uploaded to a website that removes EXIF metadata. The "clean image" was then given to ChatGPT. Even without metadata, ChatGPT gave the most confidence (45%) to Half Moon Cay, noting features like "blue strap loungers" on "Cabano style shelters" and "common features of a cruise line private island". It also suggested other cruise line locations, but Half Moon Cay was prioritized.
• The Random Google Maps Drop: In a truly impressive demonstration of accuracy, the creator randomly dropped a Google Maps Street View icon in the middle of nowhere (specifically, Leeward Fryslan) and took a screenshot of the surroundings, ensuring no text was visible. When this image was pasted into ChatGPT with the same prompt, after about two minutes of analysis, ChatGPT identified the location as "Keller's 27 Leeward" with a 60% confidence rating. The reasoning provided was the recognition of a "Burrito's Brother's logo". Upon fact-checking on Google Maps, the location identified by ChatGPT was incredibly accurate, being only 3 minutes away from the original random drop point.

Why "Illegal"?

While not literally illegal, the term highlights the surprising and potentially unnerving power of AI to ascertain locations from visual cues alone. This capability has significant implications for privacy, image analysis, and opens new avenues for how we understand and interact with visual information in the digital age. It's a testament to the rapid advancements in AI's ability to "see" and interpret the world around us.