AI art has the same problem as ChatGPT has - it doesn't have any fundamental understanding of what it is drawing, it's merely "predicting" a likely outcome given some starting point. For both algorithms, the starting point is a "prompt" that defines what sort of special solutions the algorithm should use, and then it just throws random dice and mixes and matches the solution patterns using the random numbers.
Because of this, ChatGPT has a fundamental by-design tendency to hallucinate and ramble nonsense or simply make stuff up, and Stable Diffusion will always make stupid errors in anatomy or produce nonsensical distortions and errors that have to be corrected and fixed by "prompt engineering" to guide the algorithm to give you the answer or result that you want. Even then, you're merely improving your odds of getting the answer you want, not actually eliminating the chance of error. These faults cannot be fixed or improved upon because they're caused by the very principles by which these algorithms operate (Markov chains). It is not given that anyone will "eventually" come up with anything better either: insisting that AI will inevitably get better is just begging the question.
As you try to get the algorithm towards more and more detailed and fine output, you will reach a point where it's just faster and simpler to fix the mistakes yourself rather than attempting to trick the AI to behave. It's a neat trick, but very limited in the end, because the only way to "improve" is to set up templates and automated tests for the outcome and simply run the random number generator longer, discarding all the attempts that went wrong.
Creating those tests then becomes a chore in itself, because you have to figure out a way to define things like, when are the fingers of a character appropriate to the intent of the picture… YOU have to start giving the algorithm exact details to match and spending enormous amounts of processing power to meet all your requirements at the same time by random chance.