Robert Axelrod and William D. Hamilton (1981), Science 211(4489): 1390–1396, 27 March 1981 — DOI-10.1126/science.7466396
Key takeaways
- With a probability w that two individuals meet again, cooperation based on reciprocity can be robust (TIT FOR TAT won both of Axelrod’s tournaments and took over an ecological simulation), stable (no mutant does better against an established TIT FOR TAT if w is large enough) and initially viable (via kinship or clustering), although always-defect is stable for every w.
- “The gear wheels of social evolution have a ratchet”: a nice strategy that is stable cannot be invaded even by a cluster, whereas a cluster of TIT FOR TAT can invade always-defect.
- Effective retaliation needs individual recognition, or a substitute (permanent pairing, a fixed meeting place, territoriality), and a high enough w; falling w (illness, ageing) should make partners defect.
Standing in the record
Filed on 2026-10-02 at the owner’s request, as one of the evolutionary sources of morality-as-cooperation. NOTE-034’s lineage paragraph does not name it: it names Trivers for reciprocity, and Curry 2016 (LIT-503) cites Axelrod’s 1984 book, not this paper, beside Trivers. It is filed because it is the formal core of both.
NOTE-411 is the close reading of 2026-10-02, and it placed the work: Active. It turns Trivers’s verbal model (LIT-510) into an ESS analysis and adds the question of how reciprocity gets started, answering it with Hamilton’s kinship (LIT-497) and with clustering. For morality-as-cooperation it grounds the exchange domain, and it shows why that domain needs retaliation as well as giving. TIT FOR TAT’s three properties (never first to defect, provocable, forgiving after one retaliation) are behaviours, not moral judgements; the paper does not discuss morality.
Two cautions for anyone citing it for stability. The proof shows that no strategy does better than TIT FOR TAT against TIT FOR TAT; always cooperate does exactly as well, so TIT FOR TAT is not an ESS in Maynard Smith and Price’s strict sense (LIT-505, p. 17). Boyd & Lorberbaum, Nature 327:58–59 (1987), DOI-10.1038/327058a0, showed that no pure strategy is evolutionarily stable in the repeated Prisoner’s Dilemma; not read here. And Nowak (LIT-501) reports that with errors TIT FOR TAT’s performance declines and generous variants and win-stay, lose-shift replace it.
The applications include speculative ones the authors label as such: chronic and acute phases of infection, cancer onset, and Down’s syndrome as defection between homologous chromosomes in oogenesis (“pure speculation”).
Its continuation probability w is the “shadow of the future” that Fitouchi, André & Baumard’s moral disciplining (LIT-101) needs: their reading (NOTE-132, assumption A1) cites Axelrod for the claim that reciprocal cooperation requires delayed gratification.
Topic. natural-sciences first: it is evolutionary biology, applied
from bacteria to primates. mathematics for the game theory.
complex-systems for its account of how cooperation arises in a
population of strategies (the tournament’s ecology, clustering, the
ratchet). Not probabilistic-modeling (its blurb is Bayesian and
statistical models) and not moral-psychology.
Anthology. No instruction for machine-learning practice. Iterated Prisoner’s Dilemma tournaments are used in multi-agent learning research, but the anthology’s literature.d holds no such link (grepped for “Prisoner’s Dilemma” and “tit for tat”), so it is not flagged.