The Evolution of Trust

Imagine a machine standing between you and a stranger. Each of you holds a coin. If you put your coin in, the stranger receives three coins. If the stranger puts a coin in, you receive three. You both choose at the same time, and you cannot talk.

If you both cooperate, you each end up better off. But look at it from your side alone. If the stranger cooperates, you do best by keeping your coin: you get their gift and lose nothing. If the stranger keeps their coin, you still do best by keeping yours, because otherwise you give something for nothing. Whatever the stranger does, keeping your coin pays more.

The stranger reasons the same way. So both keep their coins, and both miss the gain they could have shared.

This small puzzle has been studied for more than seventy years. It explains why trust is fragile and, under specific conditions, why it can grow and last. Nicky Case's 2017 interactive The Evolution of Trust made the research widely known; the results behind it reach back through decades of game theory. ^[1]

When Cooperation Loses

The situation above is the prisoner's dilemma. Its defining feature is a conflict between two kinds of rationality:

• Individual rationality: for each player, betraying (called defecting) pays more, whatever the other does.
• Collective rationality: if both cooperate, both do better than if both defect.

The central question of this whole field is how cooperation can arise among self-interested individuals when betrayal is always individually tempting. The short answer, developed below, is: when the same individuals meet again and again.

Where the Dilemma Came From

In 1950, the mathematicians Merrill Flood and Melvin Dresher at the RAND Corporation were testing the new theory of games, which RAND studied partly for its possible uses in nuclear strategy. They wanted to know whether real people would play the "equilibrium" strategies predicted by theory. They had two colleagues play a game with this structure 100 times in a row. The players cooperated far more often than the theory predicted. ^[2,3]

The same year, the Princeton mathematician Albert Tucker was asked to present the game to an audience of psychologists at Stanford. To make it vivid, he told a story: two prisoners, questioned separately, each offered a lighter sentence for testifying against the other. The story gave the game its name. ^[3]

The Formal Description

Each player chooses to Cooperate (C) or Defect (D). The payoffs are given four letters:

• $R$, the Reward for mutual cooperation;
• $T$, the Temptation to defect against a cooperator;
• $S$, the Sucker's payoff for cooperating with a defector;
• $P$, the Punishment for mutual defection.

The payoff matrix, listing (row player, column player), is:

The game is a prisoner's dilemma when two conditions hold:

The first condition creates the trap. Because $T > R$, defecting against a cooperator pays more than cooperating. Because $P > S$, defecting against a defector pays more than being exploited. So defection is better in every case: it is the dominant strategy, and mutual defection is the game's only Nash equilibrium.

The second condition makes cooperation worth having. It says that steady mutual cooperation pays more, on average, than taking turns exploiting each other. Without it, two players could do better by alternating betrayals.

In the coin machine above, keeping your coin while the other gives yields $T = 3$; both giving yields $R = 2$ (you give one, get three); both keeping yields $P = 0$; giving to someone who keeps yields $S = -1$. Check: $3 > 2 > 0 > -1$, and $2 \cdot 2 = 4 > 3 + (-1) = 2$. It is a true prisoner's dilemma.

Established result. In a single, one-shot prisoner's dilemma between purely self-interested players, defection is the rational choice. Nothing later in this article changes that.

Playing Again: The Shadow of the Future

Everything changes when the game repeats. If you will meet the same person tomorrow, today's betrayal has a cost: they can respond.

This requires one more quantity. Let $w$ be the probability that, after any round, there will be another round with the same partner. The political scientist Robert Axelrod called this the shadow of the future. When $w$ is close to 1, the future matters a great deal. When it is close to 0, the game is effectively one-shot.

There is an important subtlety. If both players know exactly when the game ends, cooperation can unravel backward. On the last round there is no future, so both defect. Knowing that, there is no reason to cooperate on the second-to-last round, and so on back to the first. The formal theory therefore predicts defection throughout any finite game of known length. ^[3] Real people often cooperate anyway, and the theory of cooperation mostly concerns games with an uncertain end.

Axelrod's Tournaments

Around 1979, Axelrod ran an unusual experiment. He invited game theorists, economists, psychologists, and mathematicians to submit computer programs that would play the repeated prisoner's dilemma against each other in a round-robin tournament.

The first tournament had 14 entries plus a program that moved at random. The winner was the simplest entry, submitted by the psychologist Anatol Rapoport. It was called Tit for Tat: cooperate on the first move, then copy whatever the other player did on the previous move. ^[4]

Axelrod published the results and ran a second tournament. This time 62 entrants knew that Tit for Tat had won and could try to beat it. Tit for Tat won again. ^[13]

Axelrod identified four properties shared by successful strategies:

1. Nice: never be the first to defect.
2. Retaliatory: respond to a defection, so exploitation does not pay.
3. Forgiving: return to cooperation once the other side does.
4. Clear: be easy to recognize, so others can learn that cooperating with you pays.

Tit for Tat never "beats" anyone in a single match; at best it ties. It won because it did well on average across many partners, getting high mutual rewards with cooperators and avoiding heavy losses to exploiters.

Why Tit for Tat can hold its ground

In 1981, Axelrod and the evolutionary biologist W. D. Hamilton published a paper in Science that joined these results to evolutionary theory. ^[5] They asked when a population using Tit for Tat could resist invasion by any other strategy. Their answer was a condition on $w$:

In words: the probability of meeting again must be large enough that the one-time gain from betrayal, $T - R$, is outweighed by what the betrayer loses afterward. The first term protects Tit for Tat against constant defectors; the second protects it against players who alternate defection and cooperation. For the coin machine, this gives $w \geq \max(1/3,\,1/3) = 1/3$.

Axelrod and Hamilton also argued that cooperation based on reciprocity can start even in a mostly selfish world, provided the cooperators arrive in small clusters that interact mostly with each other. ^[5] This built on Robert Trivers's earlier theory of reciprocal altruism, which proposed that helping unrelated individuals can evolve when help is likely to be returned. ^[6]

The Problem of Noise

Tit for Tat has a weakness that the tournaments hid. In the real world, mistakes happen. A message is misread, an action fails, a cooperative gesture is taken as an insult.

Suppose two Tit for Tat players are cooperating and one accidentally defects. The other retaliates on the next move. The first then retaliates for the retaliation. The two fall into an echo of alternating punishment, and one more error can lock them into mutual defection. Tit for Tat, it turns out, is too quick to punish and too slow to repair.

Researchers found two main solutions.

Generosity

In 1985, Per Molander showed that adding a small amount of unconditional generosity, sometimes cooperating even after the other defects, improves performance when errors occur. ^[7] In 1992, Martin Nowak and Karl Sigmund simulated evolution with many strategies and random errors. Tit for Tat did not win in the end. It served as a stepping stone: it cleared out exploiters, after which a more forgiving strategy, Generous Tit for Tat, took over. ^[8]

Win-Stay, Lose-Shift

In 1993, Nowak and Sigmund studied an even simpler rule. Win-Stay, Lose-Shift (also called Pavlov) repeats its last move if it got a good payoff ($T$ or $R$) and switches if it got a bad one ($P$ or $S$). Two such players who fall into mutual defection both get a bad payoff, so both switch back to cooperation together. The strategy corrects its own mistakes. In their simulations it outperformed Tit for Tat. ^[9] It has its own weakness, though: it keeps trying to cooperate with unconditional defectors every other round, so it needs some Tit-for-Tat-like players around to keep exploiters rare.

Five Routes to Cooperation

In 2006, Nowak summarized the field in Science as five mechanisms that allow natural selection to favor cooperation. ^[10] Each mechanism has a simple rule comparing the cost to the helper, $c$, with the benefit to the receiver, $b$:

The symbols mean:

• $r$: genetic relatedness between helper and recipient;
• $w$: probability of another encounter with the same individual;
• $q$: probability of knowing someone's reputation;
• $k$: average number of neighbors in a social network;
• $n$: group size; $m$: number of groups.

Direct reciprocity is the mechanism behind Tit for Tat and The Evolution of Trust. The rule says the same thing as Axelrod and Hamilton's condition in a simpler form: cooperation can evolve when the chance of meeting again is larger than the cost-to-benefit ratio of helping. The other four rules show that repetition is not the only route. Reputation, family ties, and stable local networks can also make cooperation pay.

Evidence

The theory has been tested in several ways.

• Laboratory experiments. Dal Bó and Fréchette had human participants play repeated prisoner's dilemmas with a random stopping rule. Cooperation did emerge and grow with experience when the probability of continuing was high and the reward for cooperating was large. But it emerged under stricter conditions than the basic theory requires: people did not reliably cooperate just because cooperation was theoretically possible. ^[11]
• Biology. Reciprocal altruism has been used to explain behavior such as cleaning symbioses and some alarm calls, although how common strict reciprocity is in animals remains debated. ^[6]
• Computer models. The results of Axelrod, Nowak, and Sigmund have been reproduced and extended many times; they are established facts about the models. How far the models describe real societies is a separate, harder question.

The Popular Explainer

In July 2017, the designer Nicky Case released The Evolution of Trust, a free interactive guide based on Axelrod's work and the later research on noise. ^[1,12] It uses the coin machine described at the top of this article, with payoffs that satisfy the prisoner's dilemma conditions.

The player meets a set of characters whose behavior corresponds to strategies from the research literature: a copycat (Tit for Tat), an unconditional cheater and an unconditional cooperator, a grudge-holder that never forgives a single betrayal, a tester that probes its partner before deciding, a more forgiving copycat that waits for two betrayals before retaliating, a character that follows Win-Stay, Lose-Shift, and a random player.

The simulations let the player change three things and watch the population evolve:

• The number of rounds per match. With too few repeated interactions, cheaters take over.
• The payoffs. If mutual cooperation is not rewarding enough, cheaters take over.
• The rate of mistakes. With no mistakes, the copycat thrives. With some mistakes, forgiving strategies do better. With too many, trust collapses and cheaters win.

A further lesson is that unconditional cooperators are not the winners. In a population of pure cooperators, a few cheaters thrive and spread. Cooperation is protected by strategies that are willing to cooperate and willing to retaliate.

The explainer is an accurate popular summary of well-established results. Its closing suggestions about modern social trust are the author's interpretation, reasonable but not tested by the simulation itself.

So Is It Best Not to Be Selfish?

The popular conclusion, "in the end it's best not to be selfish", is partly right. It needs to be stated precisely.

In a one-shot interaction with a stranger you will never meet again, and with nothing else at stake, defection is the self-interested choice. The mathematics is clear.

In repeated interactions, reciprocal cooperation usually does better than selfishness, provided that:

1. the chance of meeting again is high enough ($w > c/b$);
2. mutual cooperation is rewarding enough compared with the temptation to cheat;
3. mistakes and misunderstandings are fairly rare, or strategies are forgiving enough to recover from them;
4. cooperators are able and willing to respond to betrayal.

Under these conditions, the strategies that win are not selfless. They are nice, retaliatory, forgiving, and clear. They cooperate first, refuse to be exploited, and do not hold grudges forever.

So a more accurate version of the conclusion is: in long relationships, being reliably cooperative, while refusing to be a pushover, is usually in your own interest.

Limitations and Open Problems

• Simplification. Real interactions have more than two choices, unequal power, varying payoffs, and communication. Most models leave these out.
• Many equilibria. In repeated games, many outcomes can be stable, including permanent mutual defection. The theory says cooperation can be stable, not that it will appear. ^[11]
• No single best strategy. Which strategy wins depends on who else is in the population. Tit for Tat, Generous Tit for Tat, and Win-Stay, Lose-Shift each do best in some environments and poorly in others. ^[8,9]
• From model to society. The step from computer tournaments to claims about politics or modern distrust involves interpretation. It is suggestive, not proven.

The Conditions for Cooperation

The prisoner's dilemma explains why cooperation is fragile: in any single moment, betrayal is tempting. The research on repeated games explains why cooperation exists anyway: when relationships last, when their rewards are real, and when mistakes can be forgiven, reciprocal strategies outperform exploitation.

This turns a moral question into a structural one. Whether people cooperate depends not only on their character but on the situation they are in: how often they meet, how clearly they communicate, and whether betrayal has consequences. Changing those conditions can change behavior.

Across many models and experiments, reciprocity can sustain cooperation among self-interested agents. Scaling that result to a society is possible, not guaranteed. The practical problem is how to create the long horizons, clear signals and tolerance for error in which trust has time to grow.

References

[1] Case, N. (2017). The Evolution of Trust. Interactive guide, released into the public domain (CC0).
https://ncase.me/trust/

[2] Flood, M. M. (1952). Some Experimental Games. RAND Corporation, Research Memorandum RM-789-1.
https://www.rand.org/pubs/research_memoranda/RM789-1.html

[3] Kuhn, S. (1997, rev. 2025). "Prisoner's Dilemma." Stanford Encyclopedia of Philosophy.
https://plato.stanford.edu/entries/prisoner-dilemma/

[4] Axelrod, R. (1980). "Effective choice in the prisoner's dilemma." Journal of Conflict Resolution, 24(1), 3–25.
https://doi.org/10.1177/002200278002400101

[5] Axelrod, R., & Hamilton, W. D. (1981). "The evolution of cooperation." Science, 211(4489), 1390–1396.
https://doi.org/10.1126/science.7466396

[6] Trivers, R. L. (1971). "The evolution of reciprocal altruism." The Quarterly Review of Biology, 46(1), 35–57.
https://doi.org/10.1086/406755

[7] Molander, P. (1985). "The optimal level of generosity in a selfish, uncertain environment." Journal of Conflict Resolution, 29(4), 611–618.
https://doi.org/10.1177/0022002785029004004

[8] Nowak, M. A., & Sigmund, K. (1992). "Tit for tat in heterogeneous populations." Nature, 355, 250–253.
https://doi.org/10.1038/355250a0

[9] Nowak, M. A., & Sigmund, K. (1993). "A strategy of win-stay, lose-shift that outperforms tit-for-tat in the Prisoner's Dilemma game." Nature, 364, 56–58.
https://doi.org/10.1038/364056a0

[10] Nowak, M. A. (2006). "Five rules for the evolution of cooperation." Science, 314(5805), 1560–1563.
https://doi.org/10.1126/science.1133755

[11] Dal Bó, P., & Fréchette, G. R. (2011). "The evolution of cooperation in infinitely repeated games: Experimental evidence." American Economic Review, 101(1), 411–429.
https://doi.org/10.1257/aer.101.1.411

[12] Case, N. (2017). "'The Evolution of Trust' is out!" Nicky's Blog.
https://blog.ncase.me/the-evolution-of-trust-is-out/

[13] Axelrod, R. (1980). "More effective choice in the prisoner's dilemma." Journal of Conflict Resolution, 24(3), 379–403.
https://doi.org/10.1177/002200278002400301