An AI can copy what a photograph looks like, but not how one is made. It leaves marks: grain that is too clean, detail spread too evenly, colours lined up in a way no lens can produce. CowZX runs ten checks after those marks and shows you, one by one, everything it found.
Image was downscaled for analysis; sensor-level tests carry less weight.
Ten checks, each after a different mark, each answering on its own. Every one of them shows you a picture of what it saw, so you can look at the evidence yourself instead of taking a number on faith.
Most AI tools write their own name inside the file, sometimes along with the text that produced the image. This check looks for those marks — and for the data a real camera records, like the model, the aperture and the moment the photo was taken.
A generator assembles the picture by enlarging a small draft several times over, and each enlargement echoes at regular spacings. Stripping the content away and looking only at what repeats makes those echoes stand out. JPEG files cannot be checked: their own compression grid drowns the signal.
A camera measures one colour per point and calculates the other two, leaving a fine checker pattern where red and blue sit on opposite corners. Enlarging an image leaves a checker pattern too, but the same one on all three colours — so this check reads the pattern's shape, not just its presence.
Every real photo has a fine random grain that gets stronger in the bright areas and never fully disappears. AI images usually come out too clean, or carry a fake grain that looks identical across the whole picture.
When a picture is enlarged twofold, each pixel ends up closer to its partner inside the pair than to the neighbour just outside it, and that imbalance survives everything except compression. A measured picture has no reason to prefer one side. JPEG files cannot be checked: compressing a smooth area produces the same imbalance on its own.
An AI builds the picture by enlarging a much smaller version of it, and that enlargement leaves marks repeating at regular intervals. Breaking the image down into its patterns makes them jump out as bright dots that no real scene has.
No lens is perfect: it bends red and blue slightly differently from green, so edges pick up thin coloured fringes that grow as you move towards the corners. An AI draws the three colours perfectly on top of each other, which is optically impossible.
Every device squeezes an image in its own way and signs it in the process. A picture that came out of an AI was saved once, by ordinary software, and the damage from squeezing it again spreads evenly instead of concentrating on edges and textures.
An AI fills the entire canvas with invented tiny detail, in about the same dose everywhere, because that is how it learned to look sharp. A real scene has genuinely empty places — sky, a wall — sitting right next to crowded ones.
Inside the file an image is a long table of numbers, and in natural pictures those numbers follow a proportion known for almost a century. Generating, enlarging or heavily retouching an image all push that proportion off.
Generators tend to stay inside a tidy range: highlights that never burn white, blacks that never crush, fewer distinct shades overall. A real picture burns out its bright spots and spreads across a far wider set of colours.
An AI assembles a scene without ever placing a lamp in it, so parts of the picture can end up lit from directions that do not go together. On its own this check is easily fooled — a real room also has several light sources.
Read this part before you use the number to accuse anyone.
The percentage says how much the evidence leans one way, not how likely you are to be right. Read the ten checks underneath it before you conclude anything.
A screenshot, a heavy edit, or a trip through WhatsApp or Instagram wipes out exactly the marks these checks look for. A real photo that went through them scores higher than it should. This is where most false alarms come from.
Print an AI image, photograph the print, and every mark disappears. What you get is a genuine photo of a fake scene, and no amount of pixel analysis can tell you that.