Appearance
11.8 — Trust and Reputation
In 1980 Robert Axelrod invited game theorists to submit computer programs to compete in a repeated game. In each round, two programs choose to cooperate or defect. Mutual cooperation pays both moderately well. Defecting against a cooperator pays best of all — for the defector. Mutual defection pays both badly.
Played once, the rational move is to defect. Played repeatedly against the same partner, everything changes, and the tournament was to find out how.
The winner was the shortest program submitted, at four lines. It was called Tit for Tat. Its strategy: cooperate on the first move, then do whatever the other player did last time.
It won again in a second tournament, when everybody knew it had won the first.
Why it won
Axelrod identified four properties, and they generalise well beyond the game.
It was nice. Never defected first. Every high-scoring program in the tournament was nice, and every low-scoring one was not.
It retaliated. Defection was answered immediately. A program that always cooperated was exploited to nothing.
It forgave. One cooperative move from the other side and it went straight back to cooperating. It held no grudges, which meant it did not get locked into mutual destruction.
And it was clear. Its behaviour was simple enough that the other side could work out the rule, which is what allowed cooperation to establish.
Treat this as a source of well-framed intuitions rather than as a proof about human life. It is a simplified model, later tournaments with noise and more complex strategies produced more complicated results, and people are not programs.
What generalises reliably: cooperation is stable when interactions repeat and the future matters; exploitation is stable when they do not. And that single distinction explains most of the difference between how people behave with those they will see again and those they will not.
The shadow of the future
The technical term for the thing that makes cooperation rational.
When you expect to interact repeatedly, defection costs you the future stream of cooperation. When you do not, it costs nothing.
Which predicts, and matches what you observe:
People are more honest with those they will see again. Repeat customers, neighbours, colleagues.
Behaviour deteriorates in one-off interactions. Tourists get worse service and worse prices, everywhere.
And it deteriorates in the last round of anything. Somebody serving out a notice period, a supplier who has lost the contract, a relationship both parties know is ending.
Practical consequences.
Where you want good behaviour, make the interaction repeated and make that visible. "We'll be working together on the next three of these" changes behaviour.
And when you notice you are in somebody's last round, adjust your expectations rather than being surprised.
How reputation does the same work at scale
Reputation extends the shadow of the future to people you have never met. Defect against one person, and information about it reaches others, so the cost is paid across every future interaction.
Three things follow.
Reputation is asymmetric. It is built slowly by consistent behaviour and destroyed quickly by one salient violation. This is not irrational: one instance of deliberate betrayal is much more informative than a hundred instances of ordinary cooperation, because the ordinary cooperation is what everybody does.
Negative information is weighted more. People update more on a single bad report than on several good ones, which is the same asymmetry seen from the receiving end.
And it is domain-specific in principle and general in practice. People should distinguish "unreliable with deadlines" from "dishonest about money", and mostly they do not.
The arithmetic against cunning
Which is the promise made back in 1.7, and this is where it gets paid.
Deception has a declining return. The first instance works well. Each one thereafter carries a rising risk of detection, and detection is not linear in cost — it does not merely lose you that interaction, it re-prices every past interaction in the other person's mind and every future one.
And it has a maintenance cost. Keeping track of what was said to whom is expensive, and it degrades under pressure, tiredness and time (10.6's consistency point).
Meanwhile honesty compounds. Each straightforward interaction lowers the cost of the next one, because the other side spends less effort verifying.
So over a long horizon, in a world where people talk to each other, the honest strategy has the higher expected value — not as a moral claim but as an arithmetic one. The cunning strategy wins in environments where interactions do not repeat and information does not travel, and most of a life is not that environment.
The honest caveat: cunning genuinely does win sometimes, in mobile situations, in large anonymous markets, and where the person can move on before the information catches up. Pretending otherwise makes the argument weaker, not stronger. The claim is about the average over a long life, not about every case.
How trust is actually built
Four components, from work on what makes someone trusted.
Competence. Can they do it?
Reliability. Will they, consistently?
Honesty. Do they tell you the truth, including when it is inconvenient?
And whether their interests are aligned with yours — or at least, whether they will tell you when they are not.
All four are required and they are separately assessable (10.8). Somebody honest and unreliable is not trustworthy for a deadline. Somebody competent and self-interested is trustworthy exactly as far as the alignment holds.
The fastest way to build it, and it is underused: name your own conflict of interest before anybody asks. "I should say that I benefit if you choose this option." It costs almost nothing and it is one of the strongest trust signals available, because almost nobody does it.
How it is destroyed
Not usually by a dramatic betrayal. Mostly by accumulation.
Small unkept commitments. The thing you said you would send. Each is trivial and the pattern is not.
Discovering the thing that was not mentioned. From 10.4, omission. People forgive being told an unwelcome truth far more readily than discovering something was withheld, and the second one is what actually ends relationships and contracts.
Inconsistency between what you say to different people. The most common and the most damaging, and in any connected group it is discovered.
And a broken confidence. One is enough, permanently, and it is correct that it is.
Repairing it
Damaged trust is repairable and the process is specific.
A full account of what happened, without minimising. The instinct to soften is what prevents repair.
Acknowledgement of the effect, not the intention (4.8).
A structural change, not a promise. "I'll be more careful" is worthless. "I've moved this to a shared calendar so it cannot be missed" is a change to the system.
And then time and consistency, with no shortcuts. Trust is rebuilt at the rate it was originally built, which is slowly, and demanding faster forgiveness is itself a further violation.
One thing that helps more than people expect: accepting the reduced trust without complaint. Someone who says "that's fair, check whatever you need to" recovers faster than someone who argues that they should be trusted again.
What to do with this
Assume the interaction is repeated, even when it looks like it is not. It usually is, information travels further than you think, and the strategy costs you almost nothing when you are wrong.
Be nice, retaliate, forgive, and be clear. The four properties translate directly: start cooperative, respond to a real violation once and proportionately, return to normal immediately afterwards, and be predictable enough that people can work out what you will do.
Never break a confidence. The single item on this page with no recovery path.
Declare your interest before being asked. Highest trust return per unit of effort available.
And when you have damaged something, change the system rather than promising to try harder — then accept the reduced trust quietly while it rebuilds.
Next: 11.9 closes Part 11 with what happens when the connection is absent — loneliness, which is not the same as being alone, and which behaves more like hunger than like sadness.