You can understand deeply what you're doing and still make a huge amount of mistakes. This is what we're ironically (in terms of this conversation) seeing AI point out over and over now that companies have mythos in their hands. If I had to bet on it at some point we'll see the shift where if AI hasn't fully reviewed and edited code it's considered unsafe to use. Similar to how we all understand how once self-driving cars are good enough human driving should probably be minimized because we're so bad at it.