The Hand Problem Is Dying. Here Is What Replaces It.

Every article about spotting AI images used to begin the same way: look at the hands. Extra fingers, fused knuckles, thumbs sprouting from the wrong side of the palm. It was good advice in 2023. It is dangerous advice in 2026, because it teaches people to run one check, pass the image, and stop looking. The models caught up. In the batches I generate for ImageGuesser puzzles, current-generation models now produce anatomically plausible hands most of the time — and player statistics on our anatomy-themed puzzles show how effective that improvement is at fooling people who learned the old trick.

It is worth understanding why hands were ever a problem, because the reason predicts what breaks next. Hands are the most structurally complex object most photos contain: over two dozen degrees of freedom, constantly gripping, folding, and hiding parts of themselves behind other parts. In training data, hands almost never appear in a clean, canonical pose. So the model saw enormous variety and little consistency, and it learned "handness" as a texture — skin, knuckle creases, nails — without reliably learning the structure underneath. More training data and better architectures have since patched the common poses. But the underlying weakness — texture without structure — did not go away. It moved.

So where did it move? Ears are my current favorite. An ear is as topologically weird as a hand, but nobody poses for a photo with their ears, and no model developer gets mocked on social media for bad earlobes, so ears receive less corrective attention. Look for helix folds that merge into the cheek, asymmetric ear canals, or earrings that defy gravity or don't match across ears. Paired items in general are gold: earrings, cufflinks, headlight assemblies on cars, shoelaces. The model generates each side as a fresh prediction rather than as one object seen twice, and the pair quietly disagrees.

Teeth remain shaky in a subtler way than before. The old failure was a horror-movie smear of enamel. The new failure is statistical: rows of teeth that are too uniform, too numerous, or slightly misaligned with the jaw's perspective. Count them if the image resolution allows it; adults showing fourteen upper teeth in a natural smile are rare. Eyes have a similar story. Modern models render beautiful irises, but catchlights — the little reflections of the light source — still frequently disagree between the two eyes, implying two different rooms. And where hair meets skin or fabric, watch for strands that pass through a collar or dissolve into the weave, because hair-object intersection is exactly the kind of global bookkeeping diffusion models skip.

The deepest tell is not any single feature but the relationship between them. A real photograph is one moment captured by one camera: every element shares the same light, the same lens distortion, the same depth of field. A synthetic image is thousands of local predictions stitched by probability, and the stitching shows in relationships — a face lit from the left above a shirt lit from the right, a sharp jawline against ears that sit in a different focal plane.

My practical advice for 2026: keep checking hands, since the failure rate is reduced, not zero. But treat them as one stop on a circuit, not the destination. Ears, paired objects, teeth count, catchlight agreement, hair intersections, lighting coherence. Run the circuit in our anatomy module, then test yourself on today's puzzle — the models we use will happily punish anyone still relying on finger-counting alone.