A Short History of Photo Geolocation

"Where was this taken?" is one of the oldest questions in photography, and for most of the medium's life the answer lived entirely outside the photograph itself — in a pencilled caption, an album label, or somebody's memory. The story of how that question got answered, then automated, then handed to a model that needs no external clue at all, is really the story of photography quietly absorbing more and more of its own context.

It is a four-act arc, and each act solved the previous one's problem while creating a new one. Notes get lost. Metadata gets stripped. Hobbyist expertise does not scale. And a model that reads pixels is fast, cheap and confidently wrong often enough that it belongs in the entertainment column rather than the evidence column.

For most of the twentieth century a photograph's location existed only where the photographer chose to write it down. Slide carousels came with handwritten labels that fell off. Prints were dated and captioned on the reverse, usually by whoever developed the habit in the family. Travel photographers kept notebooks cross-referencing frame numbers to places, because a roll of film records exposure and nothing else. A small number of specialist camera backs and data modules in the 1980s and 1990s could imprint timestamps, and a handful recorded position for surveying and aerial work, but these were professional instruments. For everyone else, geography was an act of memory.

The consequence is that a huge share of the photographic record from the history of photography is, strictly speaking, geographically orphaned. Family archives are full of beautiful, undated, unlabelled prints. Working out where they were taken now involves the same visual detective work a person — or a model — applies to any unlabelled image: read the cars, read the signage, read the clothes, read the light. The parallel with fictional detection is not accidental, and we followed it properly in visual deduction from Sherlock Holmes to AI.

The shift from film to digital brought Exif, a small block of technical data embedded invisibly in the image file: camera model, shutter speed, aperture, timestamp. GPS receivers in cameras and, far more consequentially, in phones added latitude and longitude to that same block. Two public milestones sit behind this. The Global Positioning System was declared fully operational in 1995, and the deliberate degradation of civilian accuracy known as Selective Availability was switched off in 2000, improving typical civilian precision by roughly an order of magnitude overnight. Camera phones arrived in the same window. Within a decade, the default photograph knew where it was.

This was a real turning point, and it came with an equally real catch. Location metadata is fragile in a way the pencilled caption never was.

• —Platforms strip it. Most messaging apps and social networks remove the metadata block on upload, partly for privacy and partly to shrink the file.
• —People switch it off. Location tagging is a setting, and plenty of photographers turn it off permanently for their own reasons.
• —Screenshots have none. A screenshot of a photo is a new file with a new, empty history.
• —Scans never had it. Every print digitised from the film era arrives with a scanner's timestamp and nothing about the original scene.
• —Editing can drop it. Exporting through some tools rewrites the file and quietly discards the block, which is easy to confirm with a utility such as ExifTool.

The result is a strange inversion. In the era when photographs finally learned to record their own position, a very large share of images in actual circulation carry no reliable location data at all. The gap that the pencil used to fill reopened, and something had to fill it again.

Long before models could do it automatically, people built games around exactly this problem. Browser-based guessing games gave players a street-level view and a map, and the community that formed around them got startlingly good — good enough to name a country from a fence post, a kerb colour and the shape of a road marking. That crowd-sourced expertise proved something important. The visual clues in an unlabelled photograph really are rich enough to reconstruct a location with useful precision. It simply took a trained eye.

What the hobbyists worked out, essentially by brute force, was a hierarchy of evidence. Things a country legislates — which side of the road traffic uses, sign geometry, number-plate proportions, the script on public signage — cut the map cleanly. Things chosen by taste or supplied by global trade — a plastic chair, a car model sold on six continents — are almost worthless. That hierarchy is now the backbone of how any automated system approaches the same task.

The current chapter belongs to vision-capable models such as Google's Gemini, which can look at an ordinary photograph with no embedded metadata and reason through the same categories a skilled human player uses. This is what Raven puts to work on the web: upload a picture and the model returns its best guess at where it was taken, using nothing but what is visible in the frame. No Exif is required, because none is used. The mechanics are set out step by step in how AI guesses where a photo was taken.

What makes this era genuinely different from the hobbyist one is not accuracy but attention. A model does not get bored, does not skip the edges of the frame, and does not decide in advance which part of the picture is the subject. That is why its reasoning sometimes fixes on details a person would never mention — a drain cover, a particular shade of kerb paint, the species of weed growing through a pavement crack. We collected some of the odder examples in the weirdest things AI notices in ordinary photos.

The limits are equally real. A model has no ground truth, no second source, and no way to know whether it has seen a scene before. It produces a probability distribution over places and reports the top of it in fluent prose, which reads as far more certain than it is. Every result should be treated as an entertaining estimate, not an answer.

1. 1.Written down. The location lives outside the image, in a caption, a label or a notebook, and is lost the moment the note is separated from the print.
2. 2.Recorded automatically. The location lives inside the file as Exif GPS data, accurate to metres, and survives only until something strips it.
3. 3.Deduced by people. The location is reconstructed from visible evidence by a trained human eye, reliably but slowly, one photograph at a time.
4. 4.Inferred by a model. The location is estimated from pixels in seconds, at any scale, with confidence that has to be read sceptically.

Each stage is still in use. Archivists still write things down. Phones still tag coordinates. Hobbyists still argue over bollards. And a model now does a fast, imperfect version of all three at once. The same arc runs through the way we take pictures in the first place, traced in from postcards to pixels and in how technology changed the travel photo.

It is a neat arc to sit with. A caption written in pen, then an invisible coordinate pair, then a community trained to read a street corner like a map, then a model doing all of that reasoning in seconds from pixels alone. The question has not changed in nearly two centuries of photography. Only the speed, and the honesty required about the answer, have moved.