Text that almost reads
Shop signs, labels, book spines and jersey numbers with letters that are nearly right, or the same word spelled two ways in one picture.
Why: the model draws the shape of writing, not the spelling.
Guide
To tell if a picture is AI generated, check it in this order: where it came from, whether the file carries a provenance label, and only then the visual tells — warped text, odd hands, objects that melt into each other, and light that disagrees with itself. No single tell proves anything. Three or four together usually do.
Looks come last. Most fakes are caught by where they came from, not by counting fingers.
Run a reverse image search (Google Lens or TinEye). If the earliest copy comes from an account that posts nothing but polished pictures, or the “news photo” appears on no news site, you have most of your answer.
Upload the original file to the Content Credentials verify site. Many AI tools and some cameras attach a signed record of how a picture was made. Google’s SynthID check in the Gemini app looks for the invisible signal Google’s AI tools hide in the pixels.
Open the largest version you can find and look at the places on the list below. Screenshots and re-uploads blur exactly the details you need.
A missing label proves nothing. Most pictures online have none, and a screenshot strips it. A label that is present is strong evidence; an empty result is not evidence that a picture is real.
Image models learn what pictures usually look like, not how the world is built. Wherever a detail follows strict rules, the rules are where they break.
Shop signs, labels, book spines and jersey numbers with letters that are nearly right, or the same word spelled two ways in one picture.
Why: the model draws the shape of writing, not the spelling.
Count the fingers, then check knuckles, nails and rings. Watch for fingers that merge, bend the wrong way, or vanish behind an object.
Why: hands look different in almost every photo, so the pattern is hard to learn.
Glasses arms that sink into hair, straps that end on a shoulder, a cup handle that grows out of a hand, earrings that join the jaw.
Why: where two objects meet, the model has to decide which one is in front.
Shadows that fall in two directions, a face lit from the left in a room lit from the right, or a person with no shadow on a sunny pavement.
Why: each part looks right on its own; the scene as a whole was never lit.
Mirrors, windows, sunglasses and puddles that show a different scene, the wrong pose, or nothing at all.
Why: a reflection has to agree with something outside the frame.
Even, poreless skin with a soft sheen, perfect teeth, and hair that shines the same way everywhere.
Why: models drift toward the most “average” flawless face.
Faces in a crowd that turn into smudges, cars that blend into each other, trees that become texture.
Why: the model spends its detail where you are meant to look.
Tiles, fences, brick, window rows and keyboard keys that skip, double up or bend halfway along.
Why: repeating a pattern exactly needs counting, which models do badly.
Two different earrings, collar points of different shapes, eyes with different reflections in them.
Why: each side is drawn as its own detail.
Chairs with an extra leg, a bicycle with no chain, a guitar with the wrong number of strings, a door with no hinge side.
Why: the model knows how a thing looks, not how it is used.
Glowing cinematic light, dead-centre framing, deep background blur on a phone snapshot, and every colour a little richer than life.
Why: models are tuned toward pictures people like, so everything looks like a poster.
A big moment with one picture and no other photos, no video and no other witnesses. Real events leave a trail.
Why: a generator makes one image; the world makes many.
Tick each tell you can actually see. The meter weighs them; it is a guide, not a verdict.
Nothing ticked yet. Start with the source.
In 2023 a six-fingered hand gave most fakes away. The best 2026 models write clean text on signs and draw hands correctly most of the time, and an edited real photo has real hands to begin with. The visual list above still catches a lot, but it catches less every year.
What does not fade is the trail. A real moment usually leaves several photos from several people. A file made by a mainstream AI tool more and more often carries a label. And an account that appeared last week with a hundred perfect pictures is a story in itself. Check the trail first, and use your eyes to confirm.
Wurt.app is a browser app for making pictures and videos with AI. Make a few pictures in Picwurt, then compare them at full size with the list above: you will quickly learn where to look. It is free to start, and after that you pay as you go, with no subscription.
Open Wurt.appWork in three steps. Find where the picture first appeared, check the file for Content Credentials or another provenance label, then zoom in on the places AI still slips: text, hands, things that touch, shadows, reflections and background people. One tell is a hint; several together are strong evidence.
Yes. The Content Credentials verify site reads the signed history some cameras and AI tools attach to a file, and Google’s SynthID check in the Gemini app looks for the invisible signal Google’s own AI tools add. A reverse image search shows where else the picture appears. All three are free.
No. Most pictures online carry no label at all, and a screenshot or a re-upload usually strips the file’s history. A missing label tells you nothing either way; a label that is present is much stronger evidence.
They give a statistical guess, not a verdict. They can flag a real photo and miss a new AI picture, especially after it has been resized or compressed. Use a detector as one more signal next to the source, the provenance check and your own close look.
An image model learns what pictures usually look like, not how a hand is built or how a word is spelled. Small, rule-bound details are the hardest part to get right, so they are where errors show first. Newer models make far fewer of these mistakes, which is why the source and provenance checks matter more every year.