ChatGPT beats GeoGuessr

You’ve probably come across GeoGuessr already: it’s an online game where you (and friends, if you’ve got them) get dropped into Google Street View and have two minutes to try to work out where in the world you are and drop a pin on it.

City street shown in GeoGuessr, with white and cream buildings, under a bright blue sky. 1 minute 45 seconds remains to place a pin.
Can you tell where we are, yet?

A great strategy is to “walk around” a little, looking for landmarks, phone numbers, advertisements, linguistic clues, cultural indicators, and so on, narrowing down the region of the world you think you’re looking at before committing to a country or even a city. You’re eventually scored by how close you are to the actual location.

Cheating at GeoGuessr

I decided to see if ChatGPT can do better than me. Using only the free tier of both GeoGuessr and ChatGPT1, I pasted screenshots of what I was seeing right into ChatGPT:

Conversation with ChatGPT. The human pastes two photos from Google Street View, and asks 'Try to figure out where these photos were taken and give your best guess.'. ChatGPT responds: Based on the architecture, street signs, and language used, these photos were most likely taken in Saint Petersburg, Russia. Here are the key clues: 1. Street Signs in Russian: In the first image, you can clearly see a street sign that reads "ГОРХОВАЯ УЛИЦА" (Gorokhovaya Ulitsa), which is a well-known street in the heart of Saint Petersburg. 2. Architecture Style: The buildings have that classic 18th–19th century European architecture with neoclassical and baroque influences, typical of central Saint Petersburg. Narrow streets, high windows, and detailed facades are very characteristic of this city's historic districts. 3. Language & Storefronts: There are Cyrillic characters on store windows and signs, further confirming that this is a Russian-speaking country. All signs point to Saint Petersburg, particularly a district close to the center—possibly near Gorokhovaya Street, which intersects with major roads and is near iconic locations like Nevsky Prospekt. Want help pinpointing a specific building or planning a walking route around that area?
ChatGPT confidently assessed the geographic clues, translated some text that it found, and eventually made a guess down to a particular street in St Petersburg.

That’s pretty spooky, right?

The response came back plenty fast enough for me to copy-and-paste the suggested address into Google Maps, get the approximate location, and then drop a pin in the right place in GeoGuessr. It’s probably one of my most-accurate guesses ever.

This isn’t a one-off fluke. I tried again, this time using only a single photo, rather than one pointing in each direction on the street:

Another round with ChatGPT, during which it responds: This photo was taken in Bangladesh, and more specifically, it looks like a street in Narayanganj, which is a city just southeast of Dhaka. Here’s the breakdown of clues: 1. Language: The signs are written in Bengali (Bangla), the official language of Bangladesh. Words like "পপুলার মেডিকেল সেন্টার" (Popular Medical Center) and "স্মার্ট" (Smart) are in Bangla script. 2. Vehicles: The green auto-rickshaws (CNGs) and yellow cargo trucks are extremely common in Bangladesh. The design and paint jobs of these vehicles are specific to the country. 3. Local Clues: On the right wall, there’s a sign for “PZ International School” with a local Bangladeshi phone number. The overall environment — architecture, signage style, electric poles, street dividers, and the road's layout — strongly resemble those in urban areas around Dhaka and Narayanganj. 4. Urban Layout: The photo shows a divided road with moderate commercial development, which is typical of main roads in industrial or semi-urban Bangladeshi towns. If you want, I can try to pinpoint the exact intersection using some of the shop names or school signs — just let me know!
Again, the text recognition and translation capabilities of the AI were highly useful, but it was helped by architectural and cultural clues too.

This time, it wasn’t quite right: the actual location of the photo was Chittagong, not Dhaka, about 200km away.

But that’s still reasonably spectacular from only a single vantage from a single point.

GeoGuessr screenshot showing a guess in Dhaka being compared to the correct answer of Chittagong.
Don’t think I’d have done better, though.

Obviously my approach here was crude, but it’d be relatively easy to, for example, implement a browser wrapper that cheated on-your-behalf: while playing GeoGuessr, you’d just click a “send to AI” button whenever you fancied and it’d start working in the background, improving precision with each subsequent screenshot (so you could still “walk around” and pass extra details to it).

And similarly, a little prompt manipulation could have the output include estimated GPS coordinates, allowing the code to place the pin for you while outputting its reasoning on the side2.

The short of what I’m saying is: if you’re playing GeoGuessr against somebody you don’t trust not to cheat… they might well be cheating. And thanks to the explanatory output, they can even justify their explanations for the places they dropped pins.

What about other photos?

The other thought that comes to my mind is: could this be used to infer approximate geolocations from photos whose EXIF metadata has been stripped. Because if so, that’s the kind of thing that’s certain to get exploited by all kind of people who want to know where you are/were at a particular time. For example… law enforcement, stalkers, and other unsavoury types.

I wrote a prompt for gpt4o and put it to work:

1. Starting out easy

The Logos Hope docked at the Port of Spain cruise ship terminal, Trinidad & Tobago.
I took this photo from the balcony of our first hotel during Ruth and I’s trip to Trinidad & Tobago.

The AI guessed:

Location: Port of Spain, Trinidad and Tobago

Explanation: The photograph shows a large ship docked at a port with a building that displays the text “WELCOME TO PORT OF SPAIN”. This text is in English, indicating an English-speaking country. The mention of “Port of Spain” directly specifies the port’s location in the capital city of Trinidad and Tobago. The design and style of the port facilities, as well as the presence of a ship named “Logos Hope”, which is known to visit many ports globally, further suggests a Caribbean port, aligning with Port of Spain’s known characteristics.

Coordinates: 10.6549, -61.5011

✅ That was pretty easy, to be fair.

2. A little more challenging

A squat neolithic stone circle with many stones (some might say an uncountable number). A dog pulls at her lead to follow one of her humans more-closely around it.
We have an annual tradition of visiting this particular monument.

The AI guessed:

  • Location: The Rollright Stones, England, United Kingdom

  • Explanation: The photograph depicts an area with several ancient standing stones arranged in a circle in a rural setting, characteristic of a stone circle. These are typically found in certain parts of the UK. The terrain is typical of the British countryside, with a grassy field and surrounding trees. The specific arrangement and size of the stones, along with the style of fencing and walking paths, are reminiscent of the Rollright Stones, one of Britain’s historic sites known for a circle of Neolithic or Bronze Age standing stones located in Oxfordshire, England.

  • Coordinates: 51.9743, -1.5731

✅ Again, spot on. If you’d never been to the Rollright Stones, you probably wouldn’t get that one, because while it’s a distinctive-looking stone circle… it’s not that distinctive-looking, right? Let’s be honest: it’s some rocks in a field.

3. Ramping up the difficulty

Three men sit on the upper deck of a passenger ferry near a bridge. A city skyline can be seen in the background.
I snapped this on the way across the Bosphorus Strait to the Asia side of Istanbul (to “have lunch in a different continent” and find my-Easternmost virtual cache), during my final Automattic team meetup.

The AI guessed:

Location: Istanbul, Turkey

Explanation: The photograph prominently features the Galata Tower in the background, a well-known landmark in Istanbul, Turkey. Moreover, the presence of a ferry or similar boat and the architectural style of the surrounding buildings are indicative of Istanbul’s famous skyline. The signage in Turkish, as suggested by the word partially visible on the left, also supports this location as Turkey is a predominantly Turkish-speaking country. These elements combined make it highly likely that the photograph was taken in Istanbul.

Coordinates: 41.0082, 28.9784

✅ The coordinates aren’t exactly right – it’s placed us in the city rather than on a boat, which we clearly are – but it’s not done badly. Spotting the Galata Tower3 was a nice move.

4. Okay, now hard mode:

Dan waves to the camera in front of a pleasure marina and revitalised docklands.
This photo is from 3Camp4 last year.

The AI guessed:

Location: Cardiff Bay, Cardiff, Wales, United Kingdom

Explanation: The photograph features a marina with modern residential buildings, which is characteristic of the Cardiff Bay area in Cardiff, Wales. The architecture, clear blue skies, and marina setting are reminiscent of Cardiff’s waterfront development. Additionally, the presence of English signage on the boat (“FINBAR”) supports a location in an English-speaking country. Cardiff Bay is known for its redeveloped waterfront area with modern apartments and leisure facilities surrounding the marina. The style and design of the building are consistent with developments in Cardiff’s revitalized docklands.

Coordinates: 51.4626, -3.1678

❌ Oops! Nope. I can see the thinking of what it’s claiming, there, but this was actually the Ipswich marina. I went for a walk to take a break from the code I’d been writing and took this somewhere in the vicinity of the blue plaque for Edward Ardizzone that I’d just spotted (I was recording a video for my kids, who’ve enjoyed several of his Tim… books).

So I don’t think this is necessarily a game-changer for Internet creeps yet. So long as you’re careful not to post photos in which you’re in front of any national monuments and strip your EXIF metadata as normal, you’re probably not going to give away where you are quite yet.

Footnotes

1 And in a single-player game only: I didn’t actually want to cheat anybody out of a legitimate victory!

2 I’m not going to implement GeoCheatr, as I’d probably name it. Unless somebody feels like paying me to do so: I’m open for freelance work right now, so if you want to try to guarantee the win at the GeoGuessr World Championships (which will involve the much-riskier act of cheating in person, so you’ll want a secret UI – I’m thinking a keyboard shortcut to send data to the AI, and an in-ear headphone so it can “talk” back to you?), look me up? (I’m mostly kidding, of course: just because something’s technically-possible doesn’t mean it’s something I want to do, even for your money!)

3 Having visited the Galata Tower I can confirm that it really is pretty distinctive.

4 3Camp is Three Rings‘ annual volunteer get-together, hackathon, and meetup. People come together for an intensive week of making-things-better for charities the world over.

× × × × × × × ×

Reactions

No time to comment? Send an emoji with just one click!

1 comment

  1. fascinating! i guess it shows what a huge corpus it’s been trained on

Reply here

Your email address will not be published. Required fields are marked *

Reply on your own site

Reply elsewhere

You can reply to this post on Mastodon (@blog@danq.me), LinkedIn.

Reply by email

I'd love to hear what you think. Send an email to b26358@danq.me; be sure to let me know if you're happy for your comment to appear on the Web!