I think it's similar thing with hands, we immediately see the problem with the hands because we know how they work and recognise when the generated hands wouldn't work.

This expands to all kind of machinery actually. The moment you start reasoning about the output it falls apart. I only assume at some point there would be another AI model which is trained on how things work and will guide the current AI stuff.

Like the way ControlNet guides StableDiffusion to produce output the way the user desires.

It also happens with human artists.

Artists are trained in things like perspective, shadows, etc... They often devote a good amount of time drawing people right maybe, study basic principles like mass distribution, but they don't know everything about the underlying principles. For example, I have a painting of a harbor with a few sailboats, it looks great until you realize that sails take the wind from different directions, as if each boat had their own wind. Unless the artist really has a strong focus on realism, sometimes involving collaboration with experts in the field of what they are drawing, you are going to find mistakes like these everywhere.

It happens with human artists that don't bother understanding the things they draw. What is just a way of saying that yes, human can emulate those NLP¹ AIs, but we can choose not to. (And no, it doesn't take a strong focus on realism. Our world has all kinds of mechanics that most people understand.)

1 - We do need a name for this characteristic, because NPL AI is anything that talks with humans