How to tell whether an image was made by AI
The tells that still hold up, the ones that no longer do, and why AI image detectors should not be trusted on their own.
SamImageGenerator ·
Detection is now a practical skill rather than a curiosity, and the honest starting point is that no single test is reliable. What works is a stack of weak signals plus the context the picture arrives in.
Tells that still hold up
| Look at | What gives it away |
|---|---|
| Text in the scene | Signage, labels and book spines drift into near-words |
| Hands and counts | Fingers, teeth, chair legs, railing posts come out wrong in number |
| Where things meet | Hair against a background, glasses against a cheek, a strap crossing a shoulder |
| Physics of light | Two shadows in different directions, a reflection that does not match |
| Background people | Sharp in front, melting behind |
| Surface texture | Skin and fabric that are uniformly smooth, everywhere, at every distance |
Tells that have stopped working
"Too perfect" is no longer a signal — grain, blur and camera artefacts are easy to ask for. Neither is a watermark's absence, since cropping removes it. And a picture that looks like a phone snapshot can be entirely generated; naive realism proves nothing at all.
Why detectors are not the answer
Automated AI image detectors return a probability, not a verdict, and they are wrong in both directions: real photographs get flagged after heavy editing or compression, and generated images pass after a screenshot and a resize. A number with a percentage sign feels like evidence and is not one. Use one as a hint alongside everything else, never as a conclusion.
The stronger evidence is not in the pixels. Where did the image first appear, and when? Does anyone else have a photo of the same event from another angle? Does the file carry capture metadata, or content credentials? A picture with no history before yesterday is a bigger red flag than any artefact.
The other side of the same skill
Knowing the tells is also how you make generated images that hold up. Avoid text inside the picture, keep hands out of the frame or out of focus, ask for one clear light direction, and label AI imagery where the platform asks for it. Honest labelling costs nothing and is the difference between a useful tool and a liability. Try it on the home page, see what real prompts produce on Explore, and read the working limits in the prompt guide.