The AI systems getting heat are not the kind that break down the concepts to their essence. But of course all AI get heat for various reasons.
Some retain copies of copyrighted material in their database. Not only the "learned knowledge of seeing the material". This goes so far that you have to tell the machine that stuff like signatures and watermarks have to be left out in the created images. Because those will turn up. The machine did not really learn how to paint, it learned that paintings ought to have those little markings. Same with the hands and arms. It does not know that humans have 2 arms and 2 legs. So it just puts those around a torso. It was trained with pictures showing multiple humans and cannot differentiate that those third arm was from another person. It does not have internal rigging of how stuff would work.
About banning and filtering, I realized there is a similar situation with photo creation. Especially for magazines. They are not real. They are so heavily edited in photoshop that they could just as well have taken a simple AI recreation of that person and edited that. But do they have fine print, that these photos are not made with a camera but are 90% computer editing, including stuff like editing the body shape of the celebrities? And even with cameras, the persons usually wear heavy make up, faking it in real life as well.
An AI can create real life images, discussion about that does not even apply for those game assets. Putting emphasis on the aritifical aspect. Celebrities are a popular target for putting them in various faked situations. Now this can be done with current non-ai tech just as well. AI just makes it somewhat easier (except the thing with the three arms and such ;-)