Why Do We Repeat a Mistake We Understand? Prediction Error, Extinction, and Sunk Cost
Learning is not accumulated knowledge. It is a changed response. And a loss does not guarantee that change.
Three candidates with matching qualifications. Same salary. Who gets hired?
Most people look at the credential. Unfair as that seems, the choice has a name in economics.
Why a credential carries a price
In 1973 Michael Spence proved something strange.
Even if education raises productivity not at all, a degree can still produce a wage gap.
One condition suffices: abler people find the degree cheaper to obtain. Then the degree becomes a signal of ability, and employers read it.
A credential is not authority. It is a price tag — attached to how fast someone absorbs information.
There is a counter-argument. Human capital theory holds that education really does raise productivity. The two explanations still compete and the balance is unsettled.
But if Spence is right, one thing follows. A signal is not updated. A price tag stamped at twenty-five proves nothing at forty.
Knowing and having learned are different circuits
The claim that people confuse stored knowledge with actual learning has experimental support.
A 1996 study compared two groups.
| Group | Damaged system | Can state the rule | Learns the probabilistic task |
|---|---|---|---|
| Amnesic | hippocampal | no | yes |
| Parkinson’s | striatal | yes | no |
Exactly opposite.
The ability to explain a rule and the ability to have a response change run on different machines.
So “I know it in my head but my body won’t follow” is not an excuse. It is anatomy. The verbal side can grow indefinitely while the responding side stays put.
What makes the wiring change
When does the responding side change, then?
Schultz and colleagues answered this in 1997. Dopamine neurons do not respond to reward. They respond to the gap between prediction and outcome.
| Situation | Dopamine signal |
|---|---|
| Better than expected | increase |
| As expected | no change |
| Worse than expected | decrease |
Prediction error. And that is the learning signal.
A cold conclusion follows.
If things go as expected, no signal is emitted. Nothing changes.
Lose money the same way again and feel “I knew this would happen,” and the prediction error is zero. The wiring is untouched.
And the old response is never erased
Here is where accounts usually skip a step: the assumption that new learning removes the old.
It does not.
An extinguished conditioned response comes back in three ways.
- Spontaneous recovery — it returns on its own after time
- Renewal — it returns outside the place where extinction was learned
- Reinstatement — it returns on meeting the original stimulus again
Extinction is not deletion. New learning is layered over old, and the two keep competing.
This is why rules built in a calm market collapse in a crash. The rule was not forgotten; the context changed and the old response surfaced.
Not overwriting — competition. To win, you have to keep winning.
Selling winners, holding losers
In 1998 Odean analysed the trading records of ten thousand real brokerage accounts.
The result was clear. Investors sold the ones that had risen and held the ones that had fallen.
The disposition effect. And the behavior ran against their own tax interest. It is not calculation; it is a refusal to realise a loss.
What matters most: experience does not remove the tendency by itself.
Only humans fall for this one
The title of a 1999 paper is the question itself.
“The sunk cost and Concorde effects: are humans less rational than lower animals?”
Reviewing the animal literature, the authors concluded that animals show the sunk cost fallacy less. Continuing a hopeless course because of resources already spent is markedly human.
One proposed cause is self-justification. Admitting a past decision was wrong feels more costly than letting the loss grow.
An animal has no past judgement of its own to defend.
Solitude as verification
Judgement collapsing inside a crowd is an old experimental subject too.
In Asch’s line-judgement task a substantial fraction of subjects conformed to an obviously wrong majority. But a 1996 meta-analysis found that conformity rates vary widely with culture and era. It is not a fixed constant of human nature.
One boundary needs drawing.
That a crowd emits cortisol which spreads through the air and infects others is a metaphor. There is no evidence for such chemical transmission in human crowds.
What is established is emotional contagion. Faces, voices, posture, tempo. It travels by sight and sound. The channel is sensory, not chemical.
Which means the block is not chemical either. Stop looking and stop listening. Close the order book, leave the chat room, keep only numbers and filings.
Solitude is not a feeling. It is a cut-off operation.
A loss is not the price of rewiring
One line needs correcting.
“A loss is what you paid to rewire the brain” is a strong sentence, and it does not match the three findings above.
- Lost as expected → no prediction error → no rewiring
- Rewired anyway → the old response is still there → it returns when context shifts
- The disposition effect does not vanish with experience
Written accurately:
A loss is not the price of rewiring. It is the opportunity to rewire — and the opportunity is mostly spent unused.
Money is not automatically converted into learning. Conversion takes a procedure: writing down what was expected, what actually happened, and how large the gap was — manufacturing the prediction error deliberately.
Memory will not do this. Memory rewrites the prediction after the outcome is known.
This article’s definition
Learning is not the ability to avoid being wrong. It is the ability, having been wrong once, not to be wrong the same way again.
And the test is not how much is known.
Has what the eye reaches first, under the same stimulus, changed?
Do the numbers come before anyone’s opinion? When hope rises, is risk the thing searched for first? If that has changed, it is learning.
If it has not, it is not experience. It is habit.
Nature does not accept an explanation for the same mistake. It simply supplies the same situation again.
In short
- Spence’s signalling theory: a degree can signal ability even without raising productivity. But a signal is never updated
- Stating a rule (hippocampal) and changing a response (striatal) are different circuits — dissociated in opposite directions in 1996
- The learning signal is prediction error. Go as expected and there is no dopamine signal and no rewiring
- Extinction is not erasure. Spontaneous recovery, renewal and reinstatement bring the old response back. It is competition, not overwriting
- The disposition effect: investors sell winners and hold losers, and experience does not fix it
- The sunk cost fallacy is stronger in humans than in animals; self-justification is the proposed cause
- Emotional contagion travels by sight and sound, not chemistry — which is why it can be cut off
- A loss is not the price of rewiring but the opportunity, and it mostly goes unused
This piece is rebuilt for search readers from “The Learning Being” of the serial essay The Great Corona Swindle. The original’s point of view is kept, and each claim carries its primary source and how well established it is.
Frequently asked
- If a degree does not guarantee ability, why do firms look at credentials?
- Economics has an answer: signalling theory. In 1973 Michael Spence showed that even if education raises no productivity at all, it can still work as a signal of existing ability — provided that abler people find a degree cheaper to obtain. Human capital theory, which holds that education genuinely raises productivity, competes with it, and the matter is not settled.
- Why do we repeat mistakes we understand?
- Different brain circuits are involved. Stating a rule in words depends on declarative memory centred on the hippocampus; responding automatically to a stimulus depends on habit learning centred on the striatum. Knowlton and colleagues demonstrated the dissociation in 1996. Knowing in words does not change the response.
- Once something new is learned, does the old response disappear?
- No. An extinguished conditioned response returns on its own after time (spontaneous recovery), returns outside the place extinction was learned (renewal), and returns on re-encountering the original stimulus (reinstatement). Extinction is new learning layered over old, and the two keep competing.
- Does taking a loss produce learning?
- Not automatically. Odean's 1998 analysis of ten thousand real brokerage accounts found investors selling winners and holding losers. Experience does not remove the tendency by itself. A loss creates an opportunity to learn; changing the response requires a separate deliberate procedure.
References
- Spence, M. (1973). Job Market Signaling. The Quarterly Journal of Economics, 87(3), 355–374. — Education as a signal of ability even without productivity gains. Nobel Prize in Economics, 2001.
- Knowlton, B. J., Mangels, J. A., Squire, L. R. (1996). A neostriatal habit learning system in humans. Science, 273(5280), 1399–1402. — The dissociation of declarative memory and habit learning in amnesic and Parkinson's patients.
- Schultz, W., Dayan, P., Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. — Dopamine neurons respond to prediction error, not reward.
- Bouton, M. E. (2004). Context and behavioral processes in extinction. Learning & Memory, 11(5), 485–494. — Extinction is not erasure: spontaneous recovery, renewal, reinstatement.
- Odean, T. (1998). Are Investors Reluctant to Realize Their Losses? The Journal of Finance, 53(5), 1775–1798. — Ten thousand accounts analysed. Winners sold, losers held — the disposition effect.
- Arkes, H. R. & Ayton, P. (1999). The sunk cost and Concorde effects: Are humans less rational than lower animals? Psychological Bulletin, 125(5), 591–600. — A review concluding that animals show the sunk cost fallacy less than humans.
- Arkes, H. R. & Blumer, C. (1985). The psychology of sunk cost. Organizational Behavior and Human Decision Processes, 35(1), 124–140. — The classic experiments on sunk cost.
- Bond, R. & Smith, P. B. (1996). Culture and conformity: A meta-analysis of studies using Asch's line judgment task. Psychological Bulletin, 119(1), 111–137. — Conformity rates vary substantially by culture and era.
Published September 29, 2026