Comment by olmo23
4 days ago
I think it uses exif data for that, because when I tried it did show my location but the vision api was overloaded, so nothing more showed up.
4 days ago
I think it uses exif data for that, because when I tried it did show my location but the vision api was overloaded, so nothing more showed up.
Nope, I uploaded some exif-less photos and in my cases it guessed between somewhat well to astonishing well.
I uploaded a pic of some friends at the lake and it guessed a very specific lake 1000 miles away from where it was taken. Obviously it was a very generic background, all you see is trees and water so it could be anywhere. I uploaded a scanned photo from when my parents were my age standing in front of a NASA sign at KSC and it got it right but I think you can read some text on the sign. It can also be tricked really easy. I uploaded a selfie of some friends wearing Halloween costumes of Bill Belichick and his girlfriend (wearing UNC merch) taken in a bathroom with the words "GET OUT" written on the mirror. It thinks the photo was taken in North Carolina (it wasn't) and that the couple would be interested in buying graffiti supplies (they aren't).
The assumptions it makes about religion, politics, income, and biases is kinda lame. It just makes an assumption based on the age and isn't correct most of the time.
And you think it's acceptable to upload photographs of your friends to some random service to use as it sees fit?
Glad I'm not your friend, honestly.
That's (scarily) pretty standard for most LLMs by now. Paste the same images into ChatGPT and you will get a very accurate guess
It's also pretty fun to do this with Gemma 4 with its very pretty and structured reasoning output (which SotA model providers hide). For example for one picture that it misidentified as being taken inside the "Long Room of the Old Library at Trinity College Dublin" I can see that it did consider the correct answer (Duke Humfrey's Library in Oxford) early on as one of three candidates, but was apparently mislead by the ceiling height and a window in the background
I'd be surprised if that was required. OpenAI's o3 was already professional-human-level at guessing location from a photo, so unless the companies intentionally stopped training on those datasets, modern models should be too.