Why AI still gets hands wrong, and when it does not

Why hands were genuinely difficult

It was not arbitrary. A hand has more independent joints than almost any other part of the body, and the number of distinct configurations it can adopt is enormous compared with, say, a face, which varies within a much narrower range.

Hands also occlude themselves constantly. Fingers pass behind fingers, a grip hides most of the palm, and the visible outline changes completely with a small rotation. A model learning from photographs sees a huge variety of partial shapes for one object.

And the training data was unhelpful. Photographs are composed around faces, and hands appear at the edge of frame, blurred by motion, cropped, or holding something that hides them. The examples were plentiful and low quality in exactly the way that matters.

Put together, that produced the familiar failure: a model that knew a hand has finger-shaped things attached and had no reliable sense of how many, in what arrangement, at what scale.

What fixed it

Nothing exotic. More data, better annotated, plus architectural changes that improved how models handle spatial structure generally, plus explicit attention from teams who were tired of the criticism.

The last row is the point of the whole article. The tell was fixed years ago and the advice is still circulating, which means people are actively misreading images by applying it.

Why counting now fails in both directions

A false negative first: correct hands prove nothing. Current models render them properly most of the time, so a picture passing the finger test has passed a test that almost everything passes.

A false positive second, and this one causes more harm. Real hands photographed in motion produce blurred, merged, apparently deformed shapes. A genuine photograph of somebody gesturing at a party can look, to a person primed to count, exactly like the thing they were warned about.

That asymmetry matters because the cost falls unevenly. Wrongly dismissing a generated image as real is usually recoverable; wrongly accusing a real photograph, and the person in it, is not.

It is also self-reinforcing. Somebody who has been told hands are the tell finds a hand that looks odd, concludes the image is fake, and repeats the advice that produced the error, which is how a retired heuristic outlives its usefulness by years.

What the episode teaches about every other tell

Hands are a worked example of a general pattern, and the pattern is more useful than any specific tip.

• A visible flaw is a public benchmark. Anything nameable gets measured, and anything measured gets optimised.
• Popularity accelerates the fix. The more widely a tell is repeated, the more incentive there is to remove it.
• Advice outlives the flaw by years. The correction never spreads as fast as the original tip did.
• Fixes are uneven across tools. A tell removed in a flagship model may persist in an older or smaller one.
• The residue moves downward. What survives is not visible flaws but statistical texture, which is below what an eye resolves.

The last point is why detection tools exist at all. If the differences were still visible, nobody would need a model to find them, and the entire field would be a list of things to look for.

What still holds up

Two families of check have resisted the pattern, and both are about coherence across the frame rather than about local detail.

Lighting is the first. Every surface in a photograph is lit by the same sources, so shadow direction, hardness and colour temperature all agree. A generated scene often gets the subject right and the surroundings approximately right, and the disagreement shows in the shadows.

Physical continuity is the second. A strap that passes behind a shoulder and reappears in the wrong place, a reflection showing a room that could not be there, a pattern that fails to line up across an occlusion. These need a model of the world rather than a model of appearances.

Neither is a quick test, and that is the honest conclusion. The checks that still work require looking at the picture as a place, and the checks that were quick are the ones that have been fixed.

What this means for anybody teaching this

A large amount of media literacy material still leads with finger counting, which now teaches people a test that produces confident errors. Anyone writing training, guidance or a school lesson on this has a version problem rather than a content problem.

The durable version of the lesson is about method rather than about features. Teach that visible flaws are temporary, that any tell popular enough to be taught is popular enough to be fixed, and that the reliable questions are about source and context.

It is also worth teaching the failure mode explicitly. Somebody who knows that real hands in motion look wrong is much less likely to make the accusation that does actual damage, which is calling a genuine photograph of a real person a fake.

The honest headline for any such material is that no quick visual test survives contact with current models, and that the useful skills are asking where an image came from and who benefits from it being believed.