How to Tell if an Image Is AI Generated

Before you zoom into fingertips, ask where the image was published and whether an original exists. A picture posted by an account with no history, uploaded minutes after a news event, or forwarded with a cropped watermark is worth less than the same picture in a photo agency archive. Most viral fakes fall apart at this step, long before the pixels matter.

Reverse image search belongs in the routine, but read the results instead of trusting the first match. If the oldest hit is a stock library, you are looking at a licensed photograph. If the oldest hit is a post captioned 'made with AI', you have your answer. A screenshot of a screenshot hides that trail, which is why a repost is never as informative as the file behind it.

Hands remain the most reliable single giveaway, not because models cannot draw a hand, but because hands interact with objects. Watch the contact points: where a thumb presses into a cup, where fingers wrap a strap, where a ring sits on a finger. Generated hands merge into the object they hold, gain a sixth digit at the edge of the frame, or lose a knuckle.

Teeth and eyes follow the same logic. Individual teeth should be countable in an open smile, and pupils should be round and point in the same direction. Grids of small objects, such as bicycle spokes, zips, laces or the tines of a fork, tend to melt into each other.

Generators handle short words well and long strings badly. A shop sign reading 'Café' is not evidence of anything, but a menu board with eight lines of prices usually is, because the letters drift, duplicate or turn into shapes that merely look like writing from a distance.

Zoom to one hundred percent and read what you see out loud. Real signs carry brand names, opening hours, a phone number with the right digit count, or a language that matches the place. Logos are a second test: a familiar logo should keep its proportions. A roundel that is slightly oval, or a wordmark with a letter borrowed from a different font, is a strong hint.

Every photograph has one dominant light source, even overcast daylight. Pick it, then follow it through the picture. Shadows should fall away from it, highlights should sit on the surfaces that face it, and the catchlight in both eyes should be in the same place. Generated scenes routinely mix two light directions, or place a bright window behind a subject while the face is lit from the front.

Reflections are the second half of the test. A mirror, a window, a lake or a polished table should return the scene you can already see. A reflection that shows the wrong number of people, or the back of a head that should be facing the camera, is one of the clearest tells available. A shadow also has to match the object casting it: same angle, same softness, same length.

Look past the subject for ten seconds. Backgrounds are where models economise, because no one scores them. Count the identical windows on a building, check whether a railing continues behind a person, see whether the third chair at a table has the same legs as the first.

Then check the small physics. A cable should connect to something at both ends. A bicycle should have a chain running to the rear wheel. A mug should have a handle and a bottom. A street should have a kerb that continues. None of these observations is conclusive alone, but two broken ones in one picture are hard to explain away.

Skin from a camera has pores, fine lines, uneven tone and small blemishes. Generated skin tends to be smooth, evenly lit and slightly waxy, especially across cheeks and foreheads. The effect survives compression, which is why faces in generated images often look like they have been gently retouched even where the composition is convincing.

Hair and fabric are texture problems. Individual strands should separate, overlap and catch light at different angles; generated hair often blends into a single mass at the edges. Fabric should fold according to gravity, with seams that follow the body. Logos, buttons and zips on clothing should sit flat rather than wobble.

By the time an image reaches you, it has usually been resized three times and recompressed twice. Fine detail disappears, so the small mistakes you would spot in a full resolution file are no longer visible. That does not make the image real, it simply removes your evidence.

Which is why the practical answer to 'how to tell if an image is AI generated' is not a single check but a habit: source first, then details, then light and background, and always a second observation before you state a conclusion. Practising that habit is what the quiz is for.