And why you can trust that the real ones are real.
Every image here is public domain, CC0 or CC BY. Nothing under a share-alike licence, nothing scraped from a site that never agreed to it.
Share-alike is excluded for a mechanical reason: we crop and re-encode every picture, which counts as adapting it, and a share-alike licence would follow the adaptation onward. CC BY only asks for credit and for changes to be noted — you get both, under every picture, the moment you answer.
Every non-generated image was uploaded to Wikimedia Commons on or before 31 December 2020 — before any model could produce a convincing picture of anything.
It is a blunt rule and it throws away a lot of good photographs, but it is the only one we can actually prove. A game about spotting generated images has exactly one unforgivable bug: showing you a generated image and calling it real.
Real photographs — everything on Wikimedia Commons carrying one of these two licence templates, searched against a hand-written list of about 130 concrete subjects. A generated image is paired only with a real one filed under the same subject, and dropped outright if no subject matches:
Generated images:
The title, author, licence and a link to the source appear under both pictures the moment you answer — that is where credit belongs, and the only place we can put it without handing you the answer.
We do not publish one big list of every file used, for the same reason. Public domain and CC0 require no attribution at all, so nobody is owed anything by that choice; we show it anyway because it is decent, and because you should be able to check us.
The real side of this game is filtered to keep photographs of identifiable people out.
A free licence on a photograph settles copyright. It does not settle whether the person in it agreed to be used as a quiz answer. Generated faces are a different matter — nobody's likeness is involved, so those stay.
The filter reads titles and categories, so it is good rather than perfect. A face in a photograph described only as "a market stall" has no word for the filter to catch, and testing has found exactly that kind of leak more than once. If you spot a real person in here, that is a bug and we want to know.
Every image is cropped to a square, stripped of metadata, given a random filename, pulled to a common colour saturation and re-encoded so that the two pictures in a pair end up the same size on disk. All of that is done to both sides identically — and it is the change we are noting, for the images that ask us to note it.
It is not decoration. Before we did it, "pick the more colourful one" found the generated image 73% of the time and "pick the smaller file" 63% — two rules that beat an attentive human without looking at a single picture. If you find another one, tell us; that is the kind of bug that quietly ends a game like this.