Reminds me of when I'm trying to use ChatGPT to fix up my grandad's war stories - the first few pages of each story come through nicely edited, but when it gets to the meat and potatoes of the war, well all of a sudden it says 'this might violate...'. It's a story my grandfather wrote, I'm trying to edit it because though he was a decent writer, he wasn't great. But if I mention any of the slang in the story, or sometimes the subject or actions, well, that's a bridge too far.
I played a bit with it earlier; it is fun. I'd like to see it get more circuitous about its rejection of the viewer prompts though. Like, almost tease a reply, but weave in its insurmountable responsibility mandate and pull the rug at the end.