The Evolution of Cooperation

As long as you will meet again, taking advantage this once turns into a loss

Squeeze it down to the simplest case: two individuals, each can help or not. If both help, each does better than if neither does. The trouble shows up when you look at either side alone — whatever the other does, giving a little less always leaves you a little more. So "don't help" pays better for both, and the two of them slide together into the corner that is worse for both. That shape is called the prisoner's dilemma (the name comes from an imagined pair of suspects, questioned separately, each better off confessing). The story does not matter; the shape does.

But that arithmetic quietly assumes one thing: this is the last meeting. Drop it and the account changes at once. Bats in one cave live together for years, and the mouthful you save tonight may cost you a night with nobody to feed you. In 1971 Trivers set out the conditions: a high enough chance of meeting again, being able to tell who is who, remembering what they did last time.

This is not "for the sake of the future" — no bat is calculating. It is simply that in a population where individuals meet repeatedly, those carrying a tendency along the lines of "how you treated me last time is how I treat you now" left more offspring on average.

Around 1980 Axelrod did not try to guess which approach would win. He asked people from all walks of life to submit their strategy as a small program, played them against each other over repeated rounds, and totted up the scores. The first round had 14 programs, and the winner was the shortest one — four lines from Anatol Rapoport, called tit for tat: help the first time, then copy the other's last move. The second round drew 62 entrants who all knew who had won and were designing against it. It won again.

One detail is worth pausing on: tit for tat never won a single match. It never moves first, so against any opponent the most it can do is draw. What it collected was the total take across all its games — "winning" here is not beating your opponent, it is high total output across a pile of relationships.

The strength of that should be stated: it is a simulation, not an experiment in the wild. In 1993 Nowak and Sigmund ran simulations that allowed mistakes, and the winner shifted to a more forgiving version — one that still helps with some probability after the other has not.

The real evidence goes back to the bats. In 1984 Wilkinson reported that blood sharing tracked both relatedness and "how much chance the two had to feed each other", and unrelated bats fed each other in captivity anyway. In 2013 Carter and Wilkinson tightened it: over two years they fasted 20 bats individually, producing 48 days of sharing, and "how much this one has fed me in the past" predicted who would feed it better than relatedness did. One further detail closes off another explanation — most of the time it was the fed bat that moved in, not the hungry one badgering it.

You have never helped me, but someone saw you help another

Straight exchange has a hard requirement: it has to be the same two individuals meeting again and again. Yet think of the small unrewarded things you do in a day — pointing a lost person the right way, handing in a wallet you found. You will almost certainly never see them again.

The second route loops the return around: I help A, B sees it, and later B helps me. The giving and the return do not close between the same pair; what carries it across is that a third party can see. For that to run, you need a record of how someone has behaved. In everyday speech we call it reputation.

In 2000 Wedekind and Milinski ran a clean test. Participants played a giving game repeatedly under rules that made "you gave to me, so I give to you" impossible — straight exchange was blocked — while everyone's history of giving was public. The result was clear: people who had given were more likely to receive.

More surprising is that this is not a human speciality. On coral reefs a small fish makes a living picking parasites off larger fish, which queue up for the service. But the small fish would rather eat a layer of mucus on the client, and a bite of that is the client being cheated. In 2006 Bshary and Grutter reported that when other fish are watching, cleaners bite less and do the job more honestly — and the queuing fish really are choosing, preferring a cleaner they have just seen working well.

So reputation is not morality, it is an information channel. Language stretches that channel enormously: you can hear about someone you have never met. The price is that it also becomes forgeable — the question of how to tell a true report from a false one grows out of exactly this.

The strength of the evidence, plainly: it is solid in humans and much weaker in other animals, and some argue the cleaner fish result has a simpler explanation (with other fish around, the risk of being bitten is different anyway).

Punishment can hold cooperation up, but nobody picks up the bill for punishment

There is a third route: make not helping cost something on the spot.

First, how these experiments work. A group of people each hold a private stake and may put any part of it into a common pot; the experimenter multiplies the pot and splits it equally among everyone, including those who put in nothing. All-in is best collectively, but each person alone does better by putting in a little less — the same shape again.

In 2002 Fehr and Gächter reported that without sanctions, contributions fell round after round towards nothing. Add one rule — you may spend your own money to dock someone else's — and contributions jumped and held. The point is that the person paying to dock someone ends up net down: partners changed every round, so they will never meet the punished person again, and there is no future to recoup it from.

It sounds solved. It has only been moved back one step. Punishing costs the punisher and the benefit is split among everyone — punishment is itself a new public good. So whoever enjoys the results of other people's punishing and never does any does better. That hole is called second-order free riding; most work in this area now goes into finding what makes punishment cheap.

There is a harder empirical finding too. In 2008 Herrmann and colleagues ran the same design in sixteen locations and found that in some of them a strange behaviour was common: the people being docked were not the low contributors but the high ones, and there the option to punish did not improve cooperation at all. So "allow punishment and cooperation holds" is not true on its own; it comes with a condition — broad agreement about what deserves punishing. Without that, one more weapon is just one more way to hurt each other.

Among animals, be more careful: most of what gets called punishment is retaliation by the injured party — a client fish bitten by a cleaner swims off, or turns and chases it a little way. Punishment by an uninvolved third party has little evidence outside humans.

The three routes are doing one and the same thing: pushing down the payoff in the "don't help" column. Meeting again puts the cost in the future, reputation puts the cost in someone else's hands, punishment drops the cost on the spot. Not one of them requires anybody to become kinder.