GeoGuessr AI: Can a Machine Beat the Best Players?

On 11 May 2023, Trevor Rainbolt, one of the world’s best-known professional GeoGuessr players, posted a video called world’s best ai vs geoguessr pro. He lost all six games. His opponent was PIGEON, a research model from Stanford that had learned the game from Street View imagery alone.

“We weren’t the first AI that played against Rainbolt,” one of its creators, Silas Alberti, told NPR. “We’re just the first AI that won against Rainbolt.” Three years on, whether AI can beat GeoGuessr is settled. The useful questions are how GeoGuessr AI reads a scene, where a person still has the edge, what the rules allow, and how to turn the same technology into practice rather than a ban.

PIGEON was built by Lukas Haas, Michal Skreta, Silas Alberti and Chelsea Finn. It began as a class project in Stanford’s CS 330 course, appeared on arXiv in July 2023 and was presented at CVPR 2024, one of the leading computer vision conferences.

The live test is what made it more than a benchmark score. A Chrome extension took a screenshot in each of the four compass directions, sent the images to the model and placed its guess. Across 458 matches in Competitive Duels, it beat players of every skill level. The authors even built in an optional random delay before each guess, to make its behaviour “seem more human-like”.

Read the median and the mean together. Half of its guesses landed within about 44 km, yet the average miss was 252 km: more than one guess in 20 was off by over 750 km, and those misses drag the average up. The authors are candid about the ceiling too: today’s systems, theirs included, cannot make street-level predictions reliably.

General-purpose AI has since joined in. In April 2025, Sam Patterson, a Master I–ranked player, played five No Move rounds against OpenAI’s o3, giving the model only two 90-degree screenshots per round. o3 won 23,179 to 22,054 and named all five countries correctly. The same month, a 1,100-word “GeoGuessr prompt” for o3 went viral, though a 2026 test on 200 photos found it did no better than a short one.

The published model was trained on 400,000 images: four views each of 100,000 random GeoGuessr locations. Here is the whole pipeline, step by step, without the maths.

1. Step 1Turn the view into numbersPIGEON takes four images, 90 degrees apart, covering the full 360-degree view. OpenAI’s CLIP, a model first trained to match pictures with text, turns each image into an embedding, and the four are averaged into one summary of the place.
2. Step 2Learn geography from captionsBefore the main training, the model studied images paired with generated captions such as “In this location, people drive on the left side of the road” and “This location has a temperate oceanic climate”. The captions push it toward features that matter for location.
3. Step 3Choose a geocell, not a coordinateInstead of predicting latitude and longitude directly, PIGEON sorts the scene into one of 2,203 geocells. They follow real administrative borders, merged until each one holds enough training examples, with oversized cells split by clustering.
4. Step 4Punish far misses harderDuring training, a cell next to the right one earns partial credit, weighted by distance. The model learns that a near miss beats a far one, the same logic GeoGuessr’s scoring uses.
5. Step 5Refine inside the best candidatesFinally it takes its five most likely geocells, compares the scene with clusters of training locations inside them, and moves the pin to the closest match. The authors describe this retrieval step as a first for the field.

Nobody told the model what a bollard is. The authors found that it still learned “strategies that are taught in online GeoGuessr guides”, paying attention to vegetation, road markings, utility posts and signage. Those are the same clue families in a human country-clue checklist. For ordinary photos rather than panoramas, the reasoning is the one set out in our explainer on AI geolocation.

• Decision time. Patterson usually guessed within a minute or two, often within 10 seconds. o3 nearly always needed more than 2 minutes, and once over 6. In Duels, one player’s guess can leave the other as little as 15 seconds, so minutes of thinking are a real handicap.
• Knowing what matters. His verdict was that a human is very good at knowing what matters, while the model did a lot of unnecessary, repetitive cropping and “got distracted by advertising multiple times”.
• Moving for a better clue. PIGEON’s bot judged four fixed images. Its human opponents could also move around the scene, which the paper itself notes gave them more information to refine a guess.

Do not read too much into that list. After about 20 rounds of testing, Patterson rated o3 as comparable to Master I or better players. The human edges are about how the game is played, not about knowing more.

Patterson keeps a flashcard deck with hundreds of entries on road lines, power poles, bollards and licence plates, and he put the gap bluntly: “These models have more than an individual mind could conceivably memorize.” That is the real advantage. A human has to specialise; a model does not.

The weak spots are predictable. PIGEON learned only from GeoGuessr locations, so it has never seen the places Street View does not cover. The authors also list tunnels, bodies of water, dark scenes and forests among its hardest images, and its accuracy drops as population density falls.

That wording covers both kinds of AI. A bot that plays for you is third-party software. A chatbot you paste a screenshot into is an external source of information. The Community Rules list three penalties when there is evidence of cheating:

• Leaderboard ban: removed from official singleplayer leaderboards and kept off them in future.
• Competitive suspension: blocked from official leaderboards in competitive and singleplayer games.
• Competitive ban: permanently blocked from competitive play and removed from official leaderboards.

You may also lose virtual items, and action can extend to every account you own. The Terms of Service separately ban automatic programs and scripts, naming bots in the list, and allow a temporary suspension or a permanent ban. A suspended or banned player is not entitled to a refund.

What about the PIGEON experiment itself? It ran with help from inside GeoGuessr: the paper thanks the company’s chief technical officer for facilitating the live games against human players. What separated the experiment from cheating was GeoGuessr’s involvement, not the software.

1. Step 1Play the round cleanNo second tab, no assistant, no screenshot to a chatbot. The round tests what you can read on your own, and that is the only thing worth measuring.
2. Step 2Name your deciding clueBefore you place the pin, say which single piece of evidence you trust most. It turns a hunch into a claim you can check afterwards.
3. Step 3Find the clue you missedAfter the reveal, look for the detail that should have decided the round. Our practice routine gives the order to check things in.
4. Step 4Ask about the clue, not the roundA question like “how do Czech and Slovak roadside posts differ?” teaches something you can reuse. Treat the answer as a lead to verify: Patterson found o3’s descriptions of signs and road lines mostly accurate, not always.
5. Step 5Test the skill on real photosGuess where one of your own photos was taken, then compare with an AI estimate from GeoSpy. It reads GPS data when a file still carries it, so remove the location first or the test proves nothing.

The last step matters because the game and real photos train the same eye. What you learn from a panorama is what finding a photo’s location without metadata relies on.