1 It does look vaguely like Mellissa Joan Hart (the actress that played Sabrina in the TV series) but you did specify Archie style so I'm not sure what's going on. Perhaps Archie Comics is not in the data base. 

2 Yeah, the assumption that an AI is trained enough to pick up all the cultural references is usually wrong.
But as the long-running gag in The Monkees went, "If you hum a few bars I can fake it."

3 With images, it's all about how the training data was tagged. A culturally-aware human might write a detailed set of tags covering every aspect of an image, but they won't do that for 1,000,000 images, so instead the model-makers feed pictures to an image-recognition model and tell it to create tags. If the tagging model wasn't trained on certain concepts (characters, actions, art styles, "events which Communist China denies", etc), then it can't create tags for them.

This is why a lot of popular anime-aware models are trained on danbooru, where the tagging is crowd-sourced by obsessive fans...

The major online LLMs seem to be sufficiently culturally hip that you can ask them for a detailed description of a character or costume, which you can then feed to an image generator. For instance, Qwen Image has a vague idea of what Sean Connery looks like, but didn't know how his character was dressed in Highlander, so I asked ChatGPT to describe that costume and pasted it into the prompt.

-j

4 Hmmm, I wonder if I can ask Grok to help me construct a prompt to feed into Grok....