Trust asymmetry

Why does one broken promise cost more than many kept ones earn?

Someone keeps their word to you thirty times, then breaks it once. On a ledger the balance is overwhelmingly positive — and yet something has changed that thirty more kept promises will not quickly undo. People usually explain this by calling themselves oversensitive, or saying trust is fragile by nature. But suppose it is neither. Suppose the asymmetry is the correct response to what the two kinds of evidence actually tell you. What would have to be true for that to be the case?

Start by being precise about what trust is. It is not warmth and it is not a reward for good behaviour. It is a prediction: your estimate of how likely this person is to keep their word next time, when keeping it will cost them something and you will not be able to check. Once you see it as a prediction, the question stops being about fairness and becomes one about evidence. How much does each observation tell you?

So run the comparison. Imagine two kinds of person — one who will keep their word under real temptation, one who will not — and ask what each does on an ordinary day. The reliable one keeps almost every promise. But so does the unreliable one. Breaking promises is expensive: it costs reputation, relationships, future dealings. An unreliable person is not someone who defects constantly; such a person would be easy to spot and would run out of counterparties. They keep their word whenever keeping it is cheap and defect when the payoff is large enough. Most of their observable record looks exactly like the reliable person's.

Now the asymmetry is visible. A kept promise is something both types produce nearly all the time, so observing one barely distinguishes them. A broken promise is something one type does routinely under temptation and the other almost never does. The evidence is lopsided because the behaviours are lopsided, not because your feelings are.

Put rough numbers on it to see the shape. Say a reliable person keeps ninety-nine promises in a hundred and an unreliable one keeps ninety. Then a kept promise is about 1.1 times more likely from a reliable person — a nudge. A broken promise is ten times more likely from an unreliable one. Nine kept promises in a row raise a fifty-fifty prior only to about seventy per cent; a single break drops it below twenty, lower than where you started.

Notice what makes evidence strong here. Not the size or drama of the act, but how differently the two types produce it. Which gives the sharpest version of the point: a promise kept when keeping it was easy is worth almost nothing as evidence, because everyone keeps those. A promise kept when breaking it would clearly have paid is worth a great deal. The information lives entirely in the temptation.

That is why people who need to be believed do not simply assert their reliability — assertions are free, and a liar can make them as cheaply. They do something expensive instead: put up money they lose on default, invest visibly enough that walking away would hurt, turn down a rival offer where you can see them do it, tell you something that would damage them if you repeated it. A signal is credible only when it is cheap for the honest type and costly for the dishonest one. Anything a defector could imitate at low cost tells you nothing, however sincerely delivered — which is why "trust me" carries no information at all.

The reasoning above connects three things usually discussed separately. The asymmetry of evidence explains the asymmetry of feeling; the asymmetry of feeling explains why people bother with costly signals; and the cost of the signal is what makes it evidence in the first place. One mechanism seen from three sides.

Two honest corrections. First, this is not a claim that people weigh evidence optimally. Psychologists have documented a broad negativity bias — bad events weigh more than comparable good ones across many domains — and some of what happens after a betrayal is that, not calculation. The argument here is narrower: even a perfectly calibrated observer with no bias whatever would show this asymmetry, so it should not be treated as simply a flaw. It is mostly the shape of the evidence and partly the bias, and telling those apart in a particular case is difficult.

Second, the model treats "reliable" and "unreliable" as fixed types, and people are not. Trustworthiness is contextual and it changes. So the correct inference from one defection is not that the person is a defector, but that the probability has moved sharply — and how sharply should depend on how tempting the situation was, whether they concealed or disclosed it, and what they did afterwards. This is why repair is possible but slow, and why it works through actions with a price attached. An apology costs nothing; restitution costs something; accepting a constraint that will bite you later costs the most, which is why it moves belief furthest.

Worth marking a boundary: none of this is about enforcement. Contracts, collateral and credit records solve the same problem by making defection punishable rather than informative. Between people who know each other there is usually no enforcement mechanism at all — only inference — which is why the inference has to do so much work.

• Why hostages, bonds and deliberately burnt bridges make promises believable.
• How reputation among third parties changes the arithmetic when defection becomes visible to everyone.

Every term the collection defines is gathered in the glossary.