The concept behind GeoSpy is simple: create a system capable of identifying the location of an image based on its pixels. At first, this sounds like a pipe dream. But if you try it, it’s eerily efficient.

The concept of geolocation is not new. Whether it’s social media posts, satellite imagery, or metadata embedded in photos and videos, experts in Open Source Intelligence (OSINT) have been doing important work geolocating photos for law enforcement, humanitarian efforts, and monitoring crisis situations. Everyone who’s been playing or watching GeoGuessr will also know just how much info you can draw from things like bollard design, architecture, or even vegetation color.

But training an AI to do it is different.

GeoSpy was made by a firm called Graylark Technologies from Boston. It’s been open to the public for months, up until 404Media tried contacting them about people who wanted to stalk women.

We only got to try it on a few photos, and while it didn’t get everything right, it was sometimes surprisingly on point. More in-depth reviews also found good results. So here is a tool that the police (or criminals) could use in real-time and with minimal training. Like ChatGPT, GeoSpy is available to anyone willing to pay. Is that a good thing?

More tools like this one will pop up

The key to training such an AI comes from data. Big, big, sets of data, millions of photos. It’s a common misconception that the stunning progress in the field of machine learning comes from better algorithms. Granted, the algorithms have gotten better, but that’s only a small part of the improvement. Instead, it’s bigger datasets that power such improvements. The GeoSpy website confirms this is the approach they used:

“There are many approaches to tackling this problem, and the GeoSpy team has explored them all. Each method comes with trade-offs, but the core idea is simple: if you train an AI system with enough photos tied to ground-truth locations, the AI can learn to identify subtle clues that would take a human years to master.

For instance, consider the unique paint colors used on fire hydrants in some cities. While not particularly useful on its own, combining this detail with a visible street name or a coffee shop sign in a photo could lead to a confident guess about the location. Now imagine repeating this process across thousands of elements in an image. Over time, AI learns to recognize these patterns far more efficiently than a person ever could. GeoSpy leverages this capability, making it a powerful tool for geolocation.”

This makes it pretty likely that other such tools can pop up soon. GeoSpy may have a competitive advantage (it’s already attracting significant investments), but there’s no reason other companies and AIs can’t develop similar tools.

So how concerned should we be?

Perhaps the most impactful feature is searching in a city. If you know someone or some image is in a particular city, you can search where on the map that image is likely to be. For now, only a few cities are included, but the database will likely increase soon.

This is, on one hand, very useful.

Faktisk Verifiserbar, a Norwegian cooperative fact-checking organization that focuses mostly on fact-checking conflict areas using OSINT techniques, has been using GeoSpy. They say it enables them to extract “unique features from photos and matches those features against geographic regions, countries and cities.”

No doubt, police can also put this to good use when investigating serious crimes. You can also use it for fraud detection (like detecting listed properties that may not be in the claimed city). But what’s stopping the police (or any organization, for that matter) from building a database on people not involved in any criminal activity? The tool also isn’t perfect — if law enforcement ends up relying on it, it can make faulty or questionable decisions.

But, tools like GeoSpy are also a stalker’s paradise. You can see that on GeoSpy’s discord server as well, as 404Media points out.

Already, people are using the AI to determine the address of their “friends”.

“Why do you want to find some girls’ house,” one Discord member said to another after such a request.

“I have interview for private investigator job. I might get paid to stalk people soon.”

“Bro. Wtf. Not in my server.” That last line came from Daniel Heinen, the founder of Graylark and GeoSpy.

Granted, Heinen has taken visible steps to address the problem of stalking with GeoSpy. He strikes back against dubious requests and has now made the demo version not publicly available. The version is “available exclusively to qualified law enforcement agencies, enterprise users and government entities,” the website now reads.

But are we comfortable with this kind of technology even in restricted hands? For all the help it can provide, this technology carries the potential for significant misuse, from stalking to privacy invasions on a scale that’s hard to imagine.

The problem lies not just with the technology but with how humans use it. History shows that powerful tools often find their way into the wrong hands, no matter how many restrictions are in place. And once a tool like GeoSpy is out in the world, it becomes nearly impossible to control its spread or ensure ethical use. Even with the best intentions, the risk of abuse looms large.

Of course, there’s nothing restricted to GeoSpy. It’s the same type of question we have to deal with for multiple types of technology. The answer may decide what the future of our society is shaping up to be.