Maybe that's because we never really thought about it. In hindsight, it's only logical. For an artistic rendering, correctness doesn't matter much nor does understanding the model. For flying a plane or driving a car or just transforming code from one language to another, it very much does.
It's like some of the stuff that GPT-3 generates, it sounds extrememly plausible and realistic but then you read it properly and realise that parts just don't make any sense.
However I don't think these things are show-stoppers, I think we'll see a lot of artists getting good at providing the "feedstock" and guidance to these systems to deal with those quirks and generating some really interesing work.
Symbolic AI dominated the field for 50 years: thanks partially to an accident of history (the highly influential work of Minsky, Newell, etc) and partially due to lack of the data needed to try anything else.
Now we’ve seen essentially 50 years of research on data-driven methods compressed into 5. In retrospect it makes sense that applications to which symbolic reasoning is especially ill-suited (including artistic rendering) would be the lowest-hanging fruit.
Wouldn't this create more jobs for artists? It's just a different tool
Getting higher level languages and various libraries/frameworks didn't make coders obsolete, because there was still tons of work to do at a "higher level" that required algorithmic/engineering-style thinking. Yes, previously difficult tasks got a lot easier, but that just meant that even more ambitious projects could be tackled, and that was useful.
In contrast, it may be the case with art that the things a relative newbie can spit out with the help of image generation models are "good enough" for most use cases. Yes, a real artist would be able to squeeze out even more, but many companies may be willing to go with the option that's enormously cheaper and still seems to produce okay results.
What we're going to find out is whether novice users of image generator AIs frequently get stuck with inadequate results. IMHO the answer is likely to be "no", especially after we've iterated on the tools for a few more years.
It's not clear that you'll see something similar here, that easier-to-produce high complexity art will greatly increase demand for said art.
(The “cow tools” Far Side comic is a dumb example.)
Since these AIs can’t count, that extra weird thing is probably going to be people with too many fingers. I actually have a personal collection of weird Midjourney images because if you ask for wide aspect ratios it starts generating multiples of the same thing and fusing them together…
Francis Bacon has a quote that is apt about how he paints the things that cannot be described with words.