Also wondering if, as another commenter mentioned, it might be trying to estimate your ___location just by network means.

It absolutely does that - o3 knows your current ___location based on IP address etc. This means for a fair test you need to use a photo taken nowhere near your current vicinity - that's why I added examples for Madagascar and Buenos Aires at the end of my post: https://simonwillison.net/2025/Apr/26/o3-photo-locations/#up...

And of course make sure you turn off geotagging in the exif :)

But really, if Google Street View data (or similar) is entirely part of the training dataset it is more than expected that it has this capability.

Thanks! Looks like I missed that last part somehow :)