How to spot AI-generated images

AI image generators have improved fast, but they still leave fingerprints. If you know where to look, many fakes fall apart under thirty seconds of attention. This guide walks through the tells that catch most of them, why each one happens, and how to combine your own eyes with a detector for a confident read.

Before the checklist, one idea worth holding onto: image generators don't understand what they're drawing. They predict what a plausible image looks like, region by region, without a model of hands as objects with exactly five fingers, or of text as words with fixed spelling. That's why the mistakes cluster in the same places every time, the places that require counting, spelling, or physical consistency rather than general impression. Once you know that, the tells below stop feeling like trivia and start feeling predictable.

1. Hands and fingers

Hands remain the classic giveaway. Count the fingers, then look at how they bend: AI images often show six fingers, joints that fold the wrong way, or two hands merging where they touch. Rings, watches, and held objects frequently melt into the skin. The reason is that hands appear in a huge variety of poses in training data, fists, waves, grips, overlaps, and the model averages across all of them without a rule that a hand has exactly five fingers. Newer generators have improved here, so a correct hand no longer proves an image is real, but a wrong one is still nearly conclusive. Pay special attention to hands doing something complex: holding a cup, interlacing fingers, gripping a rail.

2. Text and logos

Look at any writing in the image: signs, labels, book covers, shirts, street signage, product packaging. Generators tend to produce letter-like shapes rather than real words, and brand logos come out slightly wrong, a swoosh that isn't quite the swoosh, a wordmark with an extra loop. If the text looks like it was written in a dream, coherent at a glance but gibberish when you actually read it, that is a strong signal. This happens because text is a hard, rule-bound system layered on top of the image, and the model treats letters as visual textures rather than symbols. Zoom in on any text before you decide an image is genuine.

3. Backgrounds and edges

The subject usually gets the most model attention, so the background is where quality slips: railings that do not line up, windows with impossible reflections, crowds of half-formed faces, stair steps that change height, and objects that blend into each other. Follow straight lines across the image, door frames, tiled floors, the horizon, and see if they stay straight and consistent. Architectural regularity is hard for generators because it demands global consistency the model doesn't track. A person in sharp focus in front of a background that quietly falls apart is one of the most reliable composite tells there is.

4. Lighting and shadows

Check where the light comes from. AI images often mix light sources that do not agree: a face lit from the left with shadows falling the wrong way, reflections in glasses and eyes that do not match the scene in front of the person, or a subject that appears lit differently from the environment around them, a sign the person and background were effectively imagined separately. Trace the shadows: they should all point away from a common light source and share the same softness. Catchlights, the bright dots in the eyes, should match in both eyes and reflect the same light.

5. Skin, hair, and texture

AI skin tends toward an airbrushed, waxy smoothness, poreless in a way real skin rarely is, sometimes with an unnatural sheen. Individual hair strands often merge into painterly clumps, and flyaway hairs dissolve into the background instead of crossing cleanly in front of it. Fabric patterns are another good check: stripes, checks, and knit textures that warp, change scale, or lose their weave mid-garment are hard for generators to keep consistent. Teeth are worth a look too, they can blur together into a single ridge or repeat unnaturally.

6. Symmetry that is slightly off

Earrings that do not match, glasses with two different frames, eyes with different catchlights, collars that sit unevenly, sleeves of different lengths. Humans and manufactured objects are roughly symmetric, and generators often lose track of the pair because the two sides are produced somewhat independently. The mismatch is usually subtle, which is exactly why it survives a casual glance and rewards a deliberate one. Compare left to right on anything that should match.

7. The gut check, then a detector

If something feels off but you cannot name it, that instinct is data. The human visual system is very good at sensing when a face or scene is subtly wrong even before conscious reasoning catches up, the "uncanny" feeling. When you get it, slow down and run the checks above. Then run the image through an AI detector for a second, statistical opinion, and crop to the suspicious region, a hand, a face, a patch of background, so the detector scores exactly the part that worried you rather than diluting the signal across the whole frame.

Putting it together

No single tell is proof on its own, generators are improving, and any one of these can appear in a real photo or be absent from a fake. The reliable approach is to stack signals. A waxy face plus mismatched earrings plus a melting background plus a high detector score is about as close to certainty as this gets. One minor oddity in an otherwise consistent image means keep looking, don't rush to a verdict. Treat spotting AI as building a case, not flipping a switch.