How to spot an AI-generated photo

A few years ago, AI-generated photos were easy to spot. Faces were waxy, hands were a mess, and any image with text in it gave itself away. Not any more. Today's models make photographs that hold up on a phone, on a billboard and often at full size. They get lighting, skin, film grain and depth of field right often enough to fool professionals, and every new release closes the gap a little more.

Every day we put five AI-generated photos next to one genuine photograph, all at the same size and quality, and ask you to find the fakes. Even players who think they're good at this get caught out. Below are the tells worth checking.

1 Hands, bodies and joints

Hands are the classic tell because they need so much structure to look right: five fingers with the right number of joints, nails on the right side of each tip, tendons, and a thumb that bends the right way. Models are much better at hands than they used to be, but they still slip, especially where a hand touches something. Count the fingers, then count them again in reflections and on people in the background. Watch for limbs that merge into clothing, elbows or shoulders that bend the wrong way, and fingers that melt into the object they're holding.

The newest models rarely draw a six-fingered hand. The mistakes that are left are quieter, such as a hand that looks as if it's holding something but isn't. Check what each hand is actually doing.

2 Text, signs and logos

Writing used to be the most reliable giveaway, because models learned what text looks like without learning the language. That has improved fast, and the best models now get headlines and shop names right. Smaller and longer text is where they still go wrong. Read every sign, menu, number plate, T-shirt and book spine as you would in real life. Look for letters that borrow strokes from their neighbours, words that are almost right, and small print that turns into squiggles when you zoom in.

Read for meaning as well as spelling. A price in the wrong currency, a phone number with too many digits or a date that never happened can all give a fake away, even when every letter is perfect.

3 Reflections, shadows and physics

A photograph records one moment under one set of physical rules. AI images are built to look right in each small area, not to make sense as a whole, so the cracks show in how things relate to each other. Check that a reflection in a window, puddle or mirror matches the scene. Check that each shadow points away from the light and falls on the right surface. Follow lines in the background, such as fences, kerbs, railings and power cables, and see whether they connect to anything. Look for objects that melt into each other, stop halfway or rest at impossible angles.

4 Repetition, texture and crowds

When a model has to draw a hundred similar things, such as a crowd, a field of flowers, a rack of clothes or a tiled floor, it often repeats itself. Faces in the middle distance get cloned. Patterns repeat a bit too evenly. Fences grow far more identical, evenly spaced posts than the scene needs. Zoom out and look for a rhythm: real crowds are uneven, while AI crowds can look like wallpaper. The opposite problem is common too. A texture that should repeat, like brick, knitting or hair, turns to mush up close.

5 Lighting, skin and fine detail

This is the hardest category, because it's where the best models do best. A few things are still worth checking. Catchlights, the small reflections of light in a person's eyes, should appear in both eyes and match the scene. AI catchlights can differ between the eyes or have no light source at all. Skin that looks like plastic (smooth, even and poreless) is a weak clue, because real retouching gives the same look. So are over-saturated colours, a horizon that's too perfect, and background blur that falls on the wrong thing. Jewellery and watches are a good test: clasps, links and watch hands turn to nonsense when you zoom in.

6 Scene logic: does the whole picture make sense?

The tells above are mostly about one spot: a hand, a word, a shadow. But a photo also has to follow the logic of the real scene it shows: which way things point, what follows from what, and which way round an object should be. A model can draw every car, face and brick perfectly and still put them together into a scene that could never happen. These mistakes are the hardest to see at a glance, because nothing looks broken until you look at the whole picture instead of one detail.

Traffic is a good example, because it's full of clues about direction. In a queue, every car faces the same way. One-way signs, lane markings, headlights, the driver's seat and the number plates all agree. In one Real vs Rendered puzzle several AI images had cars sitting in the queue facing the wrong way. Nothing in those images looked obviously AI. One car pointing the wrong way gave it away.

The same test works well beyond roads. Work out which way things face and what they're for, then ask whether it all adds up:

• Travel and flow. Someone facing or walking against a queue, escalator or crowd; planes taxiing straight at each other; a platform where the train doors open onto the track.
• Objects the wrong way round. A kettle with the spout where the handle should be, a key with its teeth on the wrong edge, a plug with pins on the wrong side, a door that opens into a wall.
• Back-to-front details. Badges, flags and lettering that read backwards on vehicles or uniforms. Or a street that mixes left-hand and right-hand traffic, with steering wheels on both sides.
• Impossible counts and collisions. Too many wheels on a vehicle, two objects in the same place, or things balanced where gravity wouldn't allow it.

7 Check the source, not just the pixels

As the pixels get harder to read, the most reliable checks are the ones outside the image:

• Content Credentials. Many cameras, editing apps and AI tools can attach signed information about where an image came from, using a standard called C2PA and shown as "Content Credentials". A verified camera record is strong evidence that a photo is real, and an "AI-generated" label is close to proof that it isn't. But this information is easy to strip out, so a photo without it proves nothing either way.
• Reverse image search and first appearance. Run the image through a reverse image search. A real photo usually leaves a trail: earlier versions, a named photographer, a publication, a Commons or agency listing. A polished scene that appears nowhere before one social media post deserves suspicion.
• Metadata. Check the image's EXIF data. A photo from a real camera often records the device, lens and settings, while AI tools usually leave this blank. EXIF is easy to edit, though, so treat it as a hint, not proof.
• Context and motive. Ask why this image exists, who benefits and why it's appearing now. Dramatic, perfectly framed images that show up with no source at a tense moment are the ones most likely to have been made rather than taken.

The tells are already fading

Everything above is a snapshot of an arms race, not a rulebook. Each new generation of models learns partly from the mistakes of the last, so problems like six fingers, alien signs and impossible shadows are steadily being fixed. Some tools automatically add a "fix the hands" step, and editing models now repair the very flaws that used to give a picture away. Within a few years, expect most of the visual tells in this guide to be unreliable or gone.

That doesn't make looking pointless, but it does make the source matter more. In future, spotting an AI photo will be less about squinting at fingers and more about where the image came from, whether it carries verified capture information, and who stands behind it. Treat the visual tells as a useful first check that's slowly wearing out, and rely more on checking the source.

Be humble about your own judgement, too. Even the best-trained eye can already be fooled, and it's only getting harder. The most useful habit isn't certainty. It's knowing when you can't tell.