2025-06-16 17:50:20.620463+02 by Dan Lyke 3 comments
Bellingcat: Have LLMs Finally Mastered Geolocation?
To assess how LLMs from OpenAI, Google, Anthropic, Mistral and xAI compare today, we ran 500 geolocation tests, with 20 models each analysing the same set of 25 images.
Turns out some of the GPT's are pretty good at playing Geoguessr.
comments in descending chronological order (reverse):
#Comment Re: LLMs and Geolocation made: 2025-06-22 04:54:41.061484+02 by: Definitely Not a Bot
Anyone could geolocate as well as that if they trained on the entire corpus of the internet. The AI aspect is just a fancy form of Sorting and Searching.
#Comment Re: LLMs and Geolocation made: 2025-06-17 01:01:46.115885+02 by: Dan Lyke
Last time I looked at Geoguessr they were dropping you into a Street View scenario, and if you looked at the JavaScript state you could figure it out. I don't see a mention of stripping EXIF data out in the Bellingcat article, I suppose they might have missed that, but those folks are generally pretty thorough.
#Comment Re: LLMs and Geolocation made: 2025-06-16 19:28:26.846449+02 by: spc476
I wonder how many images used for Geoguessr have GPS coordinates embedded?