Free quiz: spot the AI image vs the real photo. Daily challenges sharpen your eye — share your rank card.
Free with no sign-up, and the analysis runs inside your browser — files never leave your device. On our public gate (30 real + 83 AI images) it measured 94.7% at original quality and 87.6% on recompressed copies, with zero real photos confidently mislabelled as AI. Treat it as an indicator, not proof: the newest commercial generators evade every detector.
No detector catches 100% — your own eye is the last line of defense. Ten questions a day train you to spot AI-image artifacts, deepfake tells and scam formulas. Questions are seeded by the date, so the whole world plays the same set on the same day.
Every question comes from what this site actually checks: generation artifacts (hands, text, lighting), deepfake video-call verification, ticket/giveaway/romance scam formulas, account impersonation tricks, and provenance knowledge like C2PA and EXIF. The explanations after each answer are the real guide.
Scores stay on your device. Share your streak and grade (S–D) with the share button: the share text contains no answers, so no spoilers.
The ten daily questions are fixed by a value derived from the date. Everyone opening it on the same day gets the same questions in the same order, which is what makes sharing a score meaningful — and what makes refreshing to fish for easier questions impossible.
Practice mode is the opposite: personal. It remembers the axes you have been getting wrong lately and draws from those more often. Miss repeatedly on hand shapes and scenes with hands appear more frequently. Only about half the set is drawn that way, though, because drilling one axis alone dulls the eye.
Records stay on this device. There is no account and no server storage, so opening it elsewhere starts from zero. The weakness record also decays over time, so an axis you struggled with once does not follow you forever.
Many of the clues in these questions are disappearing generation by generation. Finger counts, mangled text, mismatched earrings were characteristic of models from a few years ago and rarely appear in current ones. Memorise only those and you will soon be out of date.
What lasts is physics. Whether shadows agree with the light source, whether reflections match the actual arrangement of the scene, whether perspective converges on a single vanishing point. These stay fallible for as long as models draw plausibility rather than understanding a scene.
So the point of the game is a viewing order rather than a score. Do not start at the face: start at the straight lines in the background, the shadows on the floor, the places where hands meet objects. That habit is what separates people who know where to look when a real check comes back ambiguous.
Moving past a question shows the answer along with what you should have noticed. That screen, rather than the ones you got right, is what stays with you. A lucky guess teaches nothing for the next similar scene; a wrong answer whose reasoning you read changes the next judgement.
Questions divide into a handful of axes: hands and joints, shadows and light, surface texture, text inside the image, file provenance, frame continuity in video, and account posting patterns. As misses accumulate on an axis, practice mode serves it more often.
One thing to hold on to: the score here is a map of weaknesses, not proof of skill. A low score is information about what to practise, not a bad result. The goal is knowing where to start when a real check comes back uncertain.
Two people played the daily ten separately and compared. The questions and their order were identical, and they missed different ones. The daily set is fixed by a value derived from the date, which is what makes the comparison meaningful. The useful part is not the score gap but where the misses diverged — one missing shadows and the other missing hands means each has a different axis to practise.
Images with an extra knuckle, or background signage melting into gibberish, get high answer rates — once you have seen the tell, you cannot unsee it. The explanation goes further and says why the tell exists: generators fit the overall impression before local structure, so the places with the strictest rules — hands, text — break first.
Answer rates fall noticeably on portraits from newer models, where finger counts are right and text stays intact. That drop is the point of the quiz. Knowing the easy tells are disappearing is what stops you from concluding "nothing looks off, so it is real".
A quiz gives you a labelled question and a promise that an answer exists; nine out of ten feels like competence. A real timeline offers neither the label nor the hint that one of these is generated. In practice almost everything you scroll past is real and the task is to catch the rare one. Treat the score as a way to keep your eye in, not as a substitute for checking.
Images with verified labels cannot be bundled for licensing reasons, and external links rot into dead questions. Instead the quiz teaches the judgment you use in the wild — and you can practice on real images any time with the home checker.
Only in your browser (localStorage). There is no server account or leaderboard.
No. These scenes are chosen precisely because people find them hard, so low scores are normal. Faced with a genuinely suspicious image you can take your time, look at several clues, and combine the checks and provenance tools on this site. The quiz exists to train that way of looking.