My method is this: before I use AI, I try to ask myself "how much should I surrender my judgment on this problem?"
Some problems are too big to surrender judgment. Some problems are solved differently depending on what you want to optimize. Sometimes you want to learn something. Sometimes there's ethics.
Nice. I think I agree, Size of the problem isn't same as "code complexity" or LOC or anything. If the consequences of the wrong solution being deployed are big enough even a 1 line fix can be a disaster.
I like surrender judgement. Its loss of locus of control. I also find myself asking if there are ways the AI systems "monetize" the nature of problems being put forward for solutions. I am probably implicitly giving up some IPR asking these questions, I could even be in breach of an NDA in some circumstances.
Some problems should not be put to an anonymous external service. I doubt the NSA wants people using claude or mistral or deepseek to solve NSA problems. Unless the goal, is to feed misinformation or mis-drection out into the world.