Skip to content
Log in

An Eye for an Eye, but with Forgiveness: What Axelrod's Tournaments Taught Us

The shortest program beat a field of game theorists without winning a single match. Forty-five years later, researchers revived the code and checked again. Part 6 of our game theory series: on cooperation, mistakes and forgiveness.

Game Theory, Part 6 of 9

The late 1970s. Robert Axelrod, a political scientist at the University of Michigan, sends game theorists an unusual invitation. Send in a computer program that will play the prisoner's dilemma. Not once, but many times in a row against the same opponent.

Fourteen programs arrived, from psychologists, economists, political scientists, sociologists and mathematicians. Axelrod added a fifteenth that moved at random. Every program played every other for 200 moves.

Mutual cooperation earned each side 3 points, mutual defection 1. A player who betrayed a trusting partner got 5, and the trusting partner got 0.

The shortest program won. It was four lines of code. It came from Anatol Rapoport, a mathematician and psychologist born in 1911 in Lozova, in what is now Ukraine's Kharkiv region, who moved to the United States as a child. The program was called Tit for Tat, and its rule fits in one sentence: cooperate on the first move, then do whatever your opponent did last time.

Repetition changes the game

In Part 4 we saw that if the prisoner's dilemma is played once, defecting pays whatever the other player does. In the same article, Nash objected that a hundred games in a row are really one big game.

Axelrod put that idea to the test. When a game repeats, betrayal today has a price tomorrow: your opponent remembers it. Trust becomes an investment rather than naivety. And a strategy that can both trust and defend itself gets its chance.

Axelrod called this "the shadow of the future": the longer it is, the more cooperation pays. When players know they are meeting for the last time, the logic of the one-shot dilemma returns, and defecting on the final move pays again. That's why, in the second tournament, the number of moves wasn't announced in advance: it was set at random each time.

A strategy that never won a match

Here's the strangest fact about Tit for Tat. It cannot beat any opponent in a single match.

Look at the rule closely. The program never defects first. Each of its defections is a response to a defection by the opponent. So in every match it either finishes level or slightly behind.

A 2015 analysis showed that, ranked by number of matches won, Tit for Tat came last in the first tournament.

So how did it win? The tournament counted not victories but total points. Aggressive programs won individual matches, but when they met each other they got stuck in mutual defection at 1 point a move. Tit for Tat beat nobody, but with everyone willing to cooperate it earned 3.

One of the lessons Axelrod considered most important: don't be envious. In a game where you can win together, the question isn't whether you scored more than your opponent. It's how much you scored at all.

Four traits of the winner

Axelrod analysed the results and described what the successful strategies had in common.

  • Niceness. Never be the first to defect. The first tournament had eight such programs, and all eight finished above the other six.
  • Willingness to retaliate. Betrayal can't go unanswered, or others will exploit it.
  • Forgiveness. After punishing, return to cooperation as soon as your opponent does.
  • Clarity. Your opponent should quickly understand how you play. Overly clever programs that tried to predict the opponent looked random to it, and it stopped trusting them.

Axelrod announced a second tournament. This time the entrants knew the results of the first. Sixty-two programs arrived, and their authors could design their strategies specifically against Tit for Tat. It won again anyway.

A second look, 45 years on

Does the story hold up when checked again? There are some surprises.

The original code of the first tournament didn't survive. The programs had to be rebuilt from the descriptions in Axelrod's paper. In 2019 the mathematician Vince Knight of Cardiff University played such a rebuilt tournament, and Tit for Tat finished only eighth.

We played the first tournament too, using the open-source Axelrod-Python library: the same 15 entrants, 200 moves, twenty repetitions. In our run Tit for Tat finished in a near-tie for third, less than 0.07 points per move behind the leader. But the main finding held: the eight nice programs took the top eight places.

We then added Tit for Two Tats, which responds only to two defections in a row. It wasn't an entrant, though Axelrod noted it would have won had anyone sent it in. In our runs it finished first or a very close second.

The second tournament was luckier: its programs survived. In October 2025, Knight's team revived the original code and played the tournament 25,000 times. Tit for Tat won 63% of the runs.

The authors' conclusion: the victory is real, but the field of entrants at the time was especially favourable to it. In a tournament of 272 modern strategies, Tit for Tat is no longer first; the top places go to programs trained with machine learning and computer evolution.

There's a curiosity, too. In 2004, at an anniversary tournament, a team from the University of Southampton got past Tit for Tat with a trick. Its programs recognised "their own" by an agreed sequence of opening moves and played to help one another. A clever trick, but hardly the lesson worth taking into life.

When mistakes happen

In Axelrod's tournament, programs didn't make mistakes. In life they happen: a message doesn't arrive, words are misunderstood, someone simply forgets.

Here's what happens when Tit for Tat plays itself and one copy accidentally defects on the fifth move (C for cooperate, D for defect). We simulated it in the same library:

First: CCCCCDCDCDCDCD Second: CCCCDCDCDCDCDC

One mistake, and two identical, nice programs are stuck forever taking turns at revenge. Each answers the other's last defection, and the echo never dies down.

Now replace the first copy with a "generous" version, Generous Tit for Tat, studied by Martin Nowak and Karl Sigmund in 1992. In the library's implementation it forgives, on average, one defection in three:

Generous: CCCCCCCCCCCCCC Second: CCCCDCCCCCCCCC

The mistake is forgiven, and cooperation returns. We also replayed the first tournament with 5% random mistakes, adding both forgiving versions to the field. Tit for Tat dropped to sixth place, with both Tit for Two Tats and the generous version above it. Nowak and Sigmund put it this way: Tit for Tat is the pivot, rather than the aim, of an evolution towards cooperation.

What to take away

Axelrod's tournaments are about programs, but their conclusions carry easily over to relationships that last: with colleagues, neighbours, people you talk to more than once.

  • Start with trust. The first step towards someone costs little and opens a lot.
  • Don't let betrayal pass in silence, but don't take revenge for longer than you need to.
  • Forgive more often than seems sensible, especially where it's easy to make a mistake or be misunderstood.
  • Be predictable. People find it easier to cooperate with someone whose behaviour they understand.

And don't measure success by whether you got ahead of the other person. In a game that repeats, what matters is what you've built together.

Exercise

  1. Tit for Tat plays 10 moves against a program that always defects. Count both sides' points (3 for mutual cooperation, 1 for mutual defection, 5 and 0 when only one defects). Who "won" the match?
  2. Two copies of Tit for Tat play 10 moves without mistakes. Now, after one mistake, they spend 10 moves alternating defection. How many points per move, on average, does each lose to this echo?

Answers are at the end.

Questions for discussion

  • Think of a conflict that dragged on through an "echo" of tit-for-tat responses. What could have stopped it sooner?
  • Are you closer to Tit for Tat or to the generous version? Where is the line between forgiveness and naivety?
  • Do you know people who "lose" every argument but whom everyone wants to deal with? What is their strength?

Next in the series: "Bluffing, Risk and Incomplete Information", on mixed strategies, poker, and why the best move is sometimes to flip a coin.

Answers. 1) Tit for Tat scores 9 points: 0 on the first move and 1 on each of the next nine. Its opponent scores 14: 5 on the first move and 1 after that. The opponent "won" the match. But in a tournament, Tit for Tat makes up the difference in its games against those who cooperate. 2) Without mistakes, each scores 30 points, 3 per move. During the echo, each alternates 5 and 0, scoring 25 points, 2.5 per move. Each loses half a point per move — and it goes on until someone forgives.

Sources

  1. Effective Choice in the Prisoner's Dilemma — Journal of Conflict Resolution 24(1), 3–25 (Axelrod, 1980)
  2. More Effective Choice in the Prisoner's Dilemma — Journal of Conflict Resolution 24(3), 379–403 (Axelrod, 1980)
  3. Is Tit-for-Tat the Answer? On the Conclusions Drawn from Axelrod's Tournaments — PLoS ONE 10(7), e0134128 (Rapoport, Seale & Colman, 2015)
  4. Tit for tat in heterogeneous populations — Nature 355, 250–253 (Nowak & Sigmund, 1992)
  5. Reproducing Axelrod's first tournament — Vince Knight (blog), 2019-05-24
  6. Reviving, reproducing, and revisiting Axelrod's second tournament — arXiv preprint (Knight, Campbell, Harper, Gaffney & Glynatsi), 2025-10-17
  7. Cardiff academic recreates Robert Axelrod's highly influential Prisoner's Dilemma tournament — Cardiff University, 2025-11-27
  8. University of Southampton team wins Prisoner's Dilemma competition — University of Southampton, 2004-10-07
  9. Anatol Rapoport — Wikipedia
  10. Axelrod-Python library, version 4.14 — Axelrod-Python project (open source) — used for MindLounge's own simulations

How we write and check articles

What did you make of it?

Today's puzzle

More puzzles

Discussion

Show the discussion

No comments yet. Be the first to share a thought.

Log in or Sign up to join the discussion.

Read next

Society & Humanity 7 min read

Who do we ask when we can ask AI?

The small question to a colleague — "how do you do this?" — looks trivial. Remote work pushed it aside first; now a chatbot does. What the data say, where they contradict themselves, and what nobody has measured yet.