Stop counting fingers

For about eighteen months, hands were a reliable tell. Generated images produced six fingers, fused knuckles, thumbs on the wrong side. The advice spread, and it worked.

It stopped working. The advice did not stop spreading.

What that costs

Bad tells are not neutral. They give people a procedure that returns an answer, and the answer feels earned because a check was performed.

So a real photograph with an awkwardly foreshortened hand gets called fake, confidently, by someone who followed the method. And a generated image with clean hands passes, because the check came back clear.

A test that is out of date is worse than no test, because it converts uncertainty into false certainty.

What has actually held up

The tells that survive are the ones tied to things models are structurally poor at, rather than things they were merely bad at in one release.

Text inside the image. Signage, spines, plates, screens, clock faces. This has improved but remains the single most productive place to look, because rendering legible text requires a different kind of consistency from rendering a plausible surface.

Repetition in texture. Crowds, brickwork, foliage, patterned fabric. Look for tiling, mirroring, and faces in the background that are the same face.

Boundaries between things. Hair against a background. A spectacle arm passing behind an ear. A strap crossing a shoulder. Where two objects meet, the model has to commit to an ownership decision for each pixel, and that is where it wobbles.

The physics of light. Two shadows disagreeing about where the sun is. A reflection that does not contain the room. A specular highlight on a curved surface pointing the wrong way.

Small repeated objects. Earrings that should match and do not. Teeth. Buttons. Links in a chain.

Zoom, or do not bother

Almost every reliable tell lives at 200% and vanishes at thumbnail size. Deciding an image is real from a phone-sized preview is not a judgement, it is a guess with a ritual attached.

If it matters, open it properly and look at the edges.

And then stop

Here is the part that will save you the most trouble: forensic inspection is the weakest of the available methods, and it is the one everyone reaches for first.

Ask instead where the image first appeared, who published it, and what they lose if it is wrong. That question is answerable, it does not degrade as models improve, and it is the subject of two other pieces here — on provenance, and on why automated detectors fail.

Looking closely is worth doing. It is not worth relying on.