Why do AI image generators mess up hands and fingers so badly?

They don't really understand what a hand is. An AI image generator like Midjourney or DALL-E learns by studying millions of images paired with text descriptions. It learns patterns, not anatomy. A hand looks wildly different from photo to photo — open, closed, pointing, holding something, partially hidden. From the AI's pattern-matching view, a hand is just a cluster of skin-toned shapes with a highly variable number of visible appendages. It's trying to predict a plausible arrangement of pixels, not build a 3D model with five fingers. That's why you'll often see six or seven fingers, or fingers that melt into each other. The AI is essentially guessing at a very complex, flexible shape without any hard rules to guide it. Think of it like drawing a bicycle from memory. You know it has two wheels, handlebars, and a frame, but if you've never studied the chain mechanism, your drawing might have the chain connecting to the wrong spot. The AI has the same problem, but with knuckles and joints. Newer models are getting much better at this, not because they suddenly understand hands, but because they've been trained on a flood of images specifically captioned to highlight correct hand anatomy. It's a brute-force fix, but it's working. The real insight here isn't about hands — it's a perfect window into how these models actually think, which is to say, they don't think at all. They remix visual patterns with no underlying concept of a physical world.