Most client-facing mega-corpo image-creation AIs have a double-layer AI: The first makes an image according to specifications, and the second re-scans the image and checks if it has anything "objectionable" in it, and filters the result if it does. This doesn't necessarily mean you're seeking out objectionable content: Ask it for "a woman giving a speech", and it produces one in underwear because it was fed a million images of porn and lingerie product models, and the second AI will catch the image, image-recognize that it's NSFW, and say it can't make your image for you, even though your request was innocent. Different person asks it, different seed is rolled on the dice, she's clothed, so it publishes the image it makes.

As for the sudoku, that's just luck. LLMs have minimal mathematical ability, they're LANGUAGE models, unless they're cascade-linked into other models (in example, Grok will search Twitter if it doesn't have an answer, most AIs have similar functionality and can pull results from Google, so a published-somewhere solved Sudoku it might grab if the seed roll calls for it to try searching the result instead of hallucinating. The LLM isn't searching itself, though, it is calling a different program to do a search, and interpreting the results. Which is advanced, but not the same thing as it doing it itself. That cascaded program it cannot edit, it cannot evolve, the LLM read-only's it as a tool).