But once a human learns a function their errors are more predictable. And they can predict their own error before an operation and escalate or seek outside review/advice.

For e.g. ask any model "which class of problems and domains do you have a high error rate in?".