Comment by jonamedia

Interesting question but I don't think they're cheating. Each model only receives the bare image (re-encoded, hosted under a random ID and with the EXIF data stripped), plus the question: Where was this photo taken?. We don't provide them with web search or any other tools. Looking through past rounds, nearly all their near-exact guesses are photos with readable text (e.g. signs, place names or plaques) or landmarks. In contrast with more generic photos they're typically 50–400 km off. Memorising the training data is a fair assumption, since all the photos come from public sources. However if that were the case, they would probably guess the obscure photos just as precisely, but the current data suggests that they do not. One photo taken in mid-2026, after most of the models had been trained, was still placed within 3km. My guess is during training, they are likely to see huge numbers of images with locations, so they are very good at recognising places and reading signs in any language.