How to tell if a photo is AI-generated: 5 things to check
Most advice about spotting AI images is two years out of date. Here is what still works, what no longer does, and the skill worth teaching instead.
You have probably heard that you can spot an AI image by counting the fingers. That advice was good in 2023. It is close to useless now, and repeating it teaches children a test that will fail them.
The checks below are more durable, because they come from something models are structurally bad at rather than from one artefact that got patched.
1. Look for things that repeat
Image generators match patterns; they do not understand variety. So they clone. Three identical trees in a park, three identical butterflies in a garden, a crowd where four people in the background are the same person.
Real scenes are never that tidy. Real trees differ in height and lean; real crowds contain no duplicates. Scan the background, not the subject — the subject gets the model’s attention, and the filler is where it economises.
2. Read any text in the image
Signs, book pages, labels, shop fronts. Generated text still tends to slide into shapes that look like letters until you actually read them. This remains one of the fastest checks available, and it takes a second.
3. Check the light and the shadows
Ask one question: where is the light coming from, and does everything agree? Shadows falling in two directions, a shadow that points the wrong way for the movement, a candle flame visible through the person standing in front of it. Models assemble a picture that is locally convincing and rarely check it globally.
4. Check the physics
Liquid running upward out of a beaker. A thrown ball drifting up mid-flight instead of arcing down. A reflection that does not match what is in front of the mirror. Each object looks right on its own; the relationships between them are where it falls apart.
5. Check whether things belong together
A palm tree in an Arctic scene. A fish larger than the tank it is swimming in. Generators blend training data from wildly different sources and have no sense of scale and no sense of place. This is the check children are often best at, because it needs world knowledge rather than visual forensics.
Extra checks for video
Video is harder to fake and therefore easier to catch. Four things worth knowing:
- Blinking. Real people blink every few seconds. Generated faces often forget entirely, or blink at strange intervals.
- Faces during head movement. Turning is where models struggle most — watch for a face briefly smearing or warping as it moves.
- Lips against audio. Mouth movement that does not match the words, or barely moves at all.
- Background people. Generated crowds reuse the same walk cycle. If four people in the background stride in perfect unison, the crowd is filler.
The honest caveat
Every tell on this page is temporary. Hands got fixed. Text is getting fixed. Blinking will be fixed. A checklist of visual artefacts is a depreciating asset, and teaching it as the skill sets a child up to be confidently wrong in about eighteen months.
So teach the checks — they are useful today, and they build the habit of looking closely. But teach the durable question alongside them:
Provenance does not depreciate. An image with no traceable source, arriving with a story attached, that is designed to make you feel something strongly, deserves suspicion regardless of how clean the fingers look. That question will still work when every artefact on this page is gone.
A four-minute exercise
- Find two images of a similar scene, one you know is generated and one you know is not.
- Give your child sixty seconds with each, and ask them to say which and why.
- The “why” is the entire exercise. A right answer with no reason is a coin flip.
- Then ask the harder question: if you could not tell, how else could you find out?
That last step is the one that moves a child from spotting to checking — and checking is the part that lasts.
Read next
Your child can practise this in the app
FutureMinds has two levels on exactly this — one for still images, one for video, both built around real tells. Try a question from it now, no signup.
Try a real question →