How to spot an AI-generated image

A practical checklist, from the people who make a game out of it.

Image generators have become very good. The obvious tells of a few years ago, like six-fingered hands and melted faces, are rarer now. But a generated picture is still built differently from a photograph. A camera records light that bounced off real objects. A generator predicts what pixels usually look like next to each other. That difference leaves traces, and once you know where to look you will start seeing them.

Nothing on this list is proof on its own. Real photos can have odd lighting, and good AI images can pass every check. Treat each tell as evidence, and look for two or three that agree.

1. Read the text

Signs, labels, licence plates, book spines and shop names are still the most reliable place to start. Generators often produce lettering that looks like writing from a distance and turns into nonsense up close: letters that merge, words that are almost words, or a sign that changes alphabet halfway along. Zoom into anything that should be readable.

2. Count and follow the details

Hands, ears, teeth, jewellery and eyeglasses are hard because they have strict structure. Check that fingers have joints and nails where you expect them, that earrings match, and that glasses frames continue behind the ears. In buildings, follow a line of windows, a railing or a fence: in generated images the pattern often drifts, doubles or simply stops.

3. Check the light

In a photograph every shadow points away from the same light source, and the brightness of a surface depends on which way it faces. Generated images often get the overall mood right but the physics wrong: shadows that fall in two directions, a face lit from one side while the background is lit from the other, or highlights on a surface that faces away from the sun.

4. Look at reflections

Mirrors, windows, water, sunglasses and the catchlights in someone's eyes should all show the same scene from the right angle. AI tends to invent reflections that are plausible in texture but disagree with what is actually in front of them, or gives two eyes catchlights of different shapes.

5. Zoom in on texture

Real surfaces have irregular, layered detail: pores and fine hairs on skin, grain in wood, mortar between bricks, frayed edges on fabric. Generated surfaces are often too smooth, or carry a repeating pattern that looks fine at a glance and becomes wallpaper-like when you zoom. A uniform, airbrushed sharpness across the whole frame is a warning sign.

6. Watch the background

Generators put most of their effort where the subject is. The edges and the distance are where they cut corners: crowds whose faces dissolve, trees that merge into each other, cars with the wrong number of wheels, buildings whose floors do not line up. If the centre of the picture is perfect and the corners feel dreamlike, be suspicious.

7. Ask whether it is too good

Real photos contain clutter: a bin, a wire, a stranger walking past, a slightly crooked horizon. AI images often look like the platonic ideal of their subject, with perfect weather, perfect composition and nothing out of place. Polish is not proof, but it is worth noticing.

8. Use the tools

• Reverse image search (Google Lens, Bing, TinEye) can find an earlier, credited original, which is strong evidence a photo is real.
• Content Credentials (C2PA) are attached by some cameras, editing apps and AI tools. Sites like contentcredentials.org can read them when they are present.
• Platform labels such as "AI info" on Facebook and Instagram are useful when they appear, but their absence proves nothing.