Scary AI Can Look at Photos and Figure Out Exactly Where They Were Taken

There's no hiding from this AI.

Pinpoint PrePIGEON

A trio of Stanford graduate students have made a powerful AI that can guess the location of a wide variety of photos with remarkable accuracy.

Known as Predicting Image Geolocations (PIGEON), the AI is trained on Google Street View and can effortlessly pinpoint where photos were taken, even outwitting some of the best human "geoguessers."

The developers claim their AI can correctly guess the country where a photo was taken 95 percent of the time, and usually within a startling 25 miles of the real location.

They also note some of its potentially game-changing applications, such as assisting in biological surveys or quickly identifying roads with downed power lines.

For all its very useful potential, though, it sounds like a privacy nightmare waiting to happen, with some experts fearing the abuse of such AI tools in the hands of the wrong people.

"From a privacy point of view, your location can be a very sensitive set of information," Jay Stanley at the American Civil Liberties Union told NPR.

Guess to Success

The students were inspired by the online game GeoGuessr, which drops players into a random location on Google Street View and has them try to guess where they are by pinning it on a map.

To create PIGEON, they took a neural network called CLIP, made by ChatGPT creator OpenAI, that learns about images through text and trained it using Street View.

"We created our own dataset of around 500,000 street view images," Silas Alberti, one of the Stanford students who developed the tool, told NPR. "That's actually not that much data, [and] we were able to get quite spectacular performance."

To put it to the test, the developers pitted their AI against Trevor Rainbolt, who's perhaps the best known geoguesser and who regularly goes viral for pulling off feats like tracking down the location of old family photos.

In a video on his YouTube channel documenting their faceoff, PIGEON regularly — though not always — beats Rainbolt, and watching it will give you a sense of the ease at which it operates. The developers note that the AI hadn't seen any of the specific locations prompted by the game before in its dataset, too.

FAInd Me

There's no doubt that PIGEON's potential is astounding, even more so when you consider the tiny budget with which it was made. It's a testament to how even small teams can make powerful AI tools, which by extension highlights both the technology's seemingly limitless horizons and the challenge of safely developing it.

"The fact that this was done as a student project makes you wonder what could be done, by, for example, Google," Stanley told NPR.

Stanley fears the government and corporate surveillance that this technology could make even more powerful. Of course, such entities no doubt have little trouble spying on us already, but stalkers could also abuse these tools to track down unwitting people using photos shared online. And that, unfortunately, is as much of a consequence of living in our digital age as it is of our impending AI one.

More on AI: Microsoft’s Stuffing Talking Generative AI Into Your Car