The takeaway: AI image generators fumble hands and text for the same underlying reason — they learn the statistical look of things rather than their structure. A hand has no "five fingers" rule inside the model, and a word has no spelling; both are just patterns of pixels the model tries to make plausible. That's why the results look almost right, and why they've been getting better without the problem fully going away.

You have probably played the game without meaning to. An image looks convincing for a second — the lighting, the faces, the mood — and then you notice the hand has six fingers, or the shop sign reads RESTAUARNT in confident, letter-shaped nonsense. Hands and text are the two classic tells of a generated image. The reasons they're hard are a neat lesson in what these models are.

The model doesn't know what a hand is

Modern image generators (the diffusion models we explained here) work by learning to turn random noise into a picture that matches a description. Crucially, they learn this from millions of images — not from any understanding of anatomy. There is no rule inside the system that says a hand has five fingers, that fingers bend one way, or that a thumb sits opposite the rest.

What the model has instead is a statistical sense of what hand-ish regions of an image tend to look like. And hands are a nightmare for that approach, and researchers point to a few reasons at once:

• They're wildly variable. A face, in most photos, points roughly at the camera and has the same parts in the same places. A hand can be open, clenched, gripping, waving, seen edge-on, half-hidden behind an object, or interlaced with another hand. The range of "correct" appearances is enormous.
• A right answer is a narrow target. Because so many finger arrangements look locally plausible — each finger, on its own, is fine — the model can assemble a hand that is smooth and well-lit and still has the wrong number of digits. It's optimising for "looks like hand texture," not "passes a finger count."
• They're often small. Hands frequently occupy a tiny share of an image's pixels, so there's less signal to get them right and less penalty during training for getting them wrong.

Faces, by contrast, are consistent, usually prominent, and massively represented in training data — which is exactly why models render faces beautifully and hands badly. It isn't that hands are conceptually harder; it's that they're a high-variation, low-budget, easy-to-fudge part of the picture.

Text is even more revealing

Garbled text is arguably the more telling failure, because it shows the model isn't really "writing" at all. When a generator produces a sign or a book cover, it is not spelling a word and then drawing it. It's producing the visual texture of writing — shapes that have the rhythm, spacing and strokes of letters — because that's what it learned text looks like in images.

The model has, in most cases, no character-level notion that the word on the sign must be CAFE and therefore needs those four letters in that order. Spelling is a discrete, symbolic fact; the image model is working in continuous smears of colour. So it generates something that reads as "text-shaped" and, especially in older models, rarely lands on real words — hence the confident gibberish.

Why they're getting better — and what it tells you

Both problems have eased. Newer systems produce legible short text and mostly-correct hands more often, for a few reasons: far more training data, larger models that capture finer structure, and — for text specifically — better "conditioning," where the system is fed a stronger representation of the exact words you asked for, sometimes via a dedicated text encoder, so it has something closer to symbols to aim at.

But the improvement is gradual and imperfect, and that's the interesting part. Hands and text are hard in the same way: they demand discrete correctness — the right count, the right characters — from a system built to produce continuous plausibility. Image generators are extraordinary pattern-matchers. They're weakest exactly where being convincing isn't the same as being right.

So the mangled hand isn't just a glitch. It's the model showing you what it is: a machine that has learned, in astonishing detail, what the world tends to look like — without ever being told what any of it actually is.