Experience only teaches when the feedback has three properties

Why years of experience predict so little

It is uncomfortable but reasonably well supported that in a number of professional domains, the number of years someone has been doing the work is a weak predictor of the quality of their judgements. In others it predicts a great deal. The distinction isn’t about the difficulty of the work or the intelligence of the people; it’s about the structure of the feedback the environment supplies.

Learning from experience requires a loop: you act, something happens, you find out what happened, and you can attribute it to what you did. Break any link in that chain and the loop is open. An open loop still produces confidence, because the cases accumulate and the vocabulary sharpens, but the accuracy has nothing pulling it towards the truth.

Delay is the first killer

Feedback arriving months or years after the decision is nearly useless for learning, because by the time the outcome appears you cannot reconstruct the judgement that produced it. Worse, hindsight has already revised your memory of what you expected, so the comparison that would have taught you something is unavailable.

This is why domains with immediate consequences produce reliable expertise and domains with slow ones frequently do not. Somebody who finds out within seconds whether their read was right accumulates thousands of corrections. Somebody who finds out in three years accumulates stories.

Delay also lets the intervening events take the blame. When months separate a judgement from its consequence, any number of other things happened in between, and each of them is available as an explanation for a poor result that the original decision actually caused.

Noise is the second

When outcomes depend heavily on factors outside the decision, individual results carry almost no information about the quality of the choice. Good decisions produce bad outcomes routinely, bad ones produce good outcomes, and any learner updating on single cases is fitting to noise.

The corrective is aggregation, which requires records, which almost nobody keeps. Without a written history you cannot see the pattern across many cases, and the pattern is the only place the signal exists. This is the fundamental reason personal experience in noisy domains is such a poor teacher: the sample size available to memory is small and selected by memorability.

Ambiguity is the third and the worst

Some feedback never arrives in a form that can be argued with. A decision not taken has no observable outcome. A candidate not hired never demonstrates whether they would have worked out. A treatment given to somebody who recovers leaves the question of whether they would have recovered anyway entirely open.

Where the counterfactual is missing, any outcome can be absorbed into the existing account. Success confirms the method, failure is attributed to circumstances, and the practitioner ends up with a theory that has never been exposed to a result capable of contradicting it. This is not dishonesty; it is what happens by default when nothing forces the comparison.

Manufacturing better feedback

Where the environment does not supply the loop, it can sometimes be built. Writing down the expected outcome before acting creates a record hindsight can’t rewrite. Setting a date to check makes the delay a scheduled event rather than an indefinite one. Grouping similar decisions and reviewing them together converts a noisy series of single cases into something with a visible pattern.

The counterfactual problem is harder and only partly solvable. Occasionally you can track the cases you rejected — following what happened to candidates who went elsewhere, or to the option not taken where it remains observable. More often the best available move is to write the reasoning down in enough detail that a later reader, including you, can judge whether it was sound independently of how things turned out.

What good feedback looks like when someone else is giving it

The same three properties apply to feedback between people. It should arrive close in time to the thing it concerns, be specific enough to identify what to change, and describe the action rather than the person. Vague, delayed, characterological feedback is not a weak version of the useful kind; it is a different thing that mostly transmits how the giver feels.

One further condition is often missing: the receiver needs to know what standard is being applied. Feedback compares an action against an expectation, and where the expectation was never stated, the recipient learns only that they fell short of something. Making the standard explicit in advance turns an evaluation into information, which is the whole point of the exercise.