I worked on a research topic regarding this for about a year. The task was to produce a location estimate from a photo. In a relatively simple first try, I took photos taken from airplanes and used a globe model to render views. All it took was to have an approximate location (for instance: a city or a type of building) to generate hypotheses for the location. In only few iterations the location could be matched.

I then improved the matching algorithm and implemented one that could match perspective views to top down views. I could take a friend's Facebook picture and just tell the software the city the photo was taken in and it would give me the GPS location in under an hour. Sometimes a simple clue like the geometric arrangement of windows or solar panels would be a unique enough clue to generate the exact position. All it took was time and some brute force computation.

And never forget that the location at the time of taking the photo can directly be encoded in the DCT coefficients by your phone. Either by your phone app or your hacked phone app. This information is very hard to detect by Facebook or Instagram, but if you expect it is there, reconstruction is trivial. Imagine finding a GPS location from the coefficients that points to Ukraine... Then you can have a second look at the photo and boom. Literally.