Yeah, I always felt the same way — that's exactly why I find it appealing when the AI throws unexpected results at me. It still has to stay safe, of course.
I'm currently waiting to hear back from @signal-response, before I take the tool public anywhere. But I'd be glad if people were interested in it.
In the meantime, though — if that sort of thing appeals to you, feel free to stop by my page. I've already released a nice little AI-driven roleplay game that goes in exactly that direction.