How to — check an AI image before you use it

You have a picture you did not take and you are about to put it in front of other people — a hero shot, a slide, a supplier's "factory photo", an illustration for a story. The job is not to win an argument about whether it is AI (AI 101 — What is a deepfake? covers the version that arrives as a person). The job is to decide whether it is safe to use, and to be able to say why in one sentence.
Two minutes at full size, working in this order, catches more than any detector score.
1. Read the text before you look at anything else.
Text is where image models fail most reliably, because they have learned the shape of letters rather than the letters themselves. So read every word in the frame: shop signs, packaging, a laptop screen, the spine of a book, a badge, a licence plate, the label on a bottle. Then read it again.
What you are looking for is not gibberish alone. Modern generators get short words right often enough to feel safe, which is exactly the trap: a sign that says "COFFE" or "OPEN 24 HOUR" reads fine at thumbnail size and wrong at a glance. Longer text degrades first — a menu, a newspaper, a wall of graffiti. A language you cannot identify, with letterforms that look like Latin but resolve into nothing, is the tell.
2. Count the countable, and zoom on the joints.
Hands get the attention, and they deserve it, but they are not the only place a model loses track of a structure it has to keep consistent. Count what has a known count: fingers and knuckles, teeth in a grin, spokes on a bicycle, keys on a piano, buttons on a coat, candles on a cake, windows on a facade, stripes on a shirt.
Then zoom to 200 percent on the joins — where a finger leaves a palm, where a strap meets a shoulder, where a ring sits on a hand, where a cable enters a socket. Look for a finger that starts somewhere it should not, a fifth limb at the edge of a crowd, jewellery that fuses into skin, a watch without a strap, a hand with no wrist behind it. The defect is usually not the count. It is the connection.
3. Check whether the light came from one place.
A photograph is a record of one lighting situation, and that is the hardest thing for a generator to hold across a whole frame. Pick the strongest highlight and decide where the light is. Then check everything else against it: shadow direction under the table, the shadow the subject casts on the wall, the highlight on a cheek and on a glass on the same table, the specular sheen on a car and the reflection in its window.
Two shadows pointing different ways, a face lit from the front and a shoulder lit from behind, a reflection in a mirror showing a room that is not in the frame, a shadow with no object that could cast it — any one of those is enough to stop. Empty space under an object, where contact shadow should be, is the commonest version and the easiest to miss.
4. Step back and check the scene as a whole.
Zoom out. Now look at the things that only break at the scale of the entire picture. Straight lines that should stay straight — tile grout, railway sleepers, the edge of a desk, the top of a door frame — bending or drifting as they cross the frame. Perspective where the vanishing lines disagree with each other. A crowd in which the same face appears four times, or in which the blurred background figures have no legs. Depth of field that ignores distance: foreground and background both sharp, or a blur applied evenly across a scene with ten metres of depth.
Also check that the picture makes physical sense as a moment. A kitchen with no crumbs, a workshop with no dust, a street with no kerb, a product photographed at an angle no one could hold a camera at. Generated scenes tend to be too clean, and the cleanliness reads as wrong before you can name why. The same instinct applies to a moving frame: if the image you were handed is a still from a clip, AI 101 — How do AI video generators work? explains why the failure moves rather than sits still.
5. Ask where it came from before you ask whether it is fake.
Provenance is stronger evidence than any visual inspection, in both directions. Content Credentials — the C2PA standard, which records capture and edit history as signed metadata — can tell you what camera or model made a file, and what happened to it afterwards. You can read them with the Content Credentials verify tool. For images made with Google's models, an invisible SynthID watermark may be present, and Google's own tools will check for it.
Two rules keep this honest. Absence of credentials proves nothing — screenshots, messaging apps, re-encoding and most social platforms strip metadata, and an honest photograph of an honest event usually arrives with none. And presence of credentials proves only what they say: an image can carry a signed "generated by model X" label and be entirely fine to use as illustration, as long as you label it.
6. Run the claim test last.
Ask why this image exists and who benefits from you believing it. A "leaked" product photo with studio lighting and no fingerprints on it. A single photo of an event that fifty people attended, with no other angle anywhere. A supplier's factory picture that appears on four other companies' sites. An account created three weeks ago posting the only photograph of a fire, a protest or a branded package.
If the image is load-bearing evidence for a claim, the standard is not "does it look real" but "can anyone else confirm this happened". Reverse image search on the full frame and on a cropped detail, plus a search for the same event from another source, settles most of these in under a minute.
The one thing not to do
Do not let a detector score settle it. Detection models are trained on the fingerprints of the generators that existed when they were built, and the published benchmarks treat cross-generator generalisation as the hard problem — GenImage, the million-pair benchmark the field still measures against, makes that its central test, and separately measures how much accuracy falls once an image is low-resolution, blurred or compressed, which is the state most images reach you in. A confidence number is one weak signal among several. The tears in the picture are evidence; the score is a hint.
How you will know it worked
You can point at the specific region that fails and say what is wrong with it — a label that spells nothing, a shadow pointing the wrong way, a hand with no wrist. The image survives a slow pass at 200 percent on text, hands and edges without you finding that. Or you found provenance that says plainly what it is, and you published it with a caption that says so. Either outcome is a decision you can defend. It is the same habit as How to — tell AI slop from writing worth reading: name the defect, not the vibe. "It looked fine on my phone" is not a decision.
What is the tell that gave away the last fake image you were sent? Write to us via our contact page.







