{
  "id": "prisoners-dilemma",
  "title": "Prisoner's Dilemma",
  "lang": "en",
  "source_lang": "en",
  "summary": "The prisoner's dilemma is a strategic game in which individually dominant choices produce an outcome worse for both players than mutual cooperation.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Prisoners' dilemma",
    "PD"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/prisoners-dilemma/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/prisoners-dilemma/",
      "json": "https://aiwiki.page/zh/prisoners-dilemma.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The prisoner's dilemma is a two-player game in game theory in which each player has an individual incentive to choose an action that leaves both players worse off than another available outcome. In its standard form, each player chooses between cooperation and defection. Defection yields a higher payoff regardless of the other player's choice, yet mutual cooperation gives both players higher payoffs than mutual defection. The dilemma therefore separates individual strategic incentives from collective advantage.\n\nFirst formulated in 1950, the game became a central model in economics, political science, evolutionary biology, and the study of social behavior. Its importance lies not in a claim that people always behave selfishly, but in its precise representation of circumstances in which mutually beneficial cooperation is vulnerable to unilateral exploitation. Repeated interaction, reputation, institutions, population structure, and preferences concerning others can alter its predictions, although each does so through a different mechanism."
    },
    {
      "heading": "Payoff structure and defining conditions",
      "text": "The standard prisoner's dilemma is a simultaneous, symmetric game with two players and two available actions: cooperate, denoted \\(C\\), and defect, denoted \\(D\\). “Simultaneous” means that neither player observes the other's current choice before choosing; their physical actions need not occur at exactly the same instant. The consequences are represented by a payoff matrix, with the first entry in each cell belonging to the row player:\n\nRow player's action | Column player cooperates | Column player defects\nCooperate | \\((R,R)\\) | \\((S,T)\\)\nDefect | \\((T,S)\\) | \\((P,P)\\)\n\nThe conventional symbols identify the reward \\(R\\) for mutual cooperation, the temptation \\(T\\) to defect against a cooperator, the punishment \\(P\\) for mutual defection, and the sucker's payoff \\(S\\) received by an exploited cooperator. The defining ordering is\n\n\\[T>R>P>S.\\]\n\nFor example, \\(T=5\\), \\(R=3\\), \\(P=1\\), and \\(S=0\\) produce a prisoner's dilemma. Against cooperation, defection gives 5 rather than 3; against defection, it gives 1 rather than 0. Nevertheless, mutual cooperation gives each player 3, exceeding the payoff of 1 from mutual defection. The dilemma is thus not merely a situation in which cooperation is risky. Defection is better for an individual under either possible action of the other player.\n\nMany treatments, particularly of repeated games, also impose\n\n\\[2R>T+S.\\]\n\nThis additional condition ensures that mutual cooperation produces a greater combined payoff than an outcome in which one player exploits the other. It also makes sustained mutual cooperation preferable, in average payoff, to taking turns exploiting one another. It is not required for the basic dominance argument, and definitions differ over whether to include it.\n\nPayoffs represent the players' preferences, not necessarily money. Larger values may indicate profit, reproductive success, or shorter prison sentences after an appropriate conversion. When randomized outcomes are considered, numerical payoffs normally have an expected-utility interpretation. Positive affine transformations of an individual player's payoffs preserve that player's strategic incentives.\n\nThe standard game is not a zero-sum game: one player's gain need not equal the other's loss, and different outcomes can produce different total payoffs. Symmetry is also not essential to the underlying conflict. An asymmetric version remains a prisoner's dilemma if each player strictly prefers defection against either action while both prefer mutual cooperation to mutual defection."
    },
    {
      "heading": "Origins and intellectual development",
      "text": "The modern game originated at the RAND Corporation in 1950 through work by Merrill Flood and Melvin Dresher. Their investigations belonged to an early period of game theory in which researchers were exploring whether mathematical solution concepts adequately described actual decisions. The broader foundations had been established by John von Neumann and Oskar Morgenstern in Theory of Games and Economic Behavior in 1944. The new example was especially striking because its individually compelling choice produced a jointly inferior result.\n\nAlbert W. Tucker supplied the prison story and the name in 1950. In the familiar narrative, two suspects are questioned separately. Each may remain silent or betray the other. If both remain silent, both receive relatively light punishment; if one betrays a silent partner, the betrayer receives favorable treatment while the other receives a severe sentence; if both betray, both receive an intermediate punishment. Versions use different sentences, but the precise numbers are secondary to the ordering of preferences. Cooperation refers to cooperation between the suspects, not cooperation with investigators.\n\nThe game appeared alongside the development of Nash equilibrium, introduced by John Forbes Nash Jr. in 1950–1951. This concept identifies a combination of strategies from which no player can profitably deviate alone. In the prisoner's dilemma, it gives a particularly clear result: mutual defection is the unique equilibrium, despite being worse for both players than mutual cooperation. The example helped distinguish strategic stability from collective desirability.\n\nAnatol Rapoport and Albert Chammah's 1965 book Prisoner's Dilemma consolidated theoretical and experimental work on the game. During the following decades, it became an important model for bargaining, conflict, and the emergence of cooperation. Robert Trivers's 1971 account of reciprocal altruism connected repeated exchanges with evolutionary explanations of helping behavior. These developments did not make cooperation automatic; they identified conditions under which a short-term cost could yield a longer-term return.\n\nRobert Axelrod's computer tournaments, reported in 1980, brought the repeated game to a wider audience. His 1984 book The Evolution of Cooperation presented repeated interaction as a framework for examining cooperation without centralized enforcement. A 1981 paper by Axelrod and William D. Hamilton connected tournament results with evolutionary reasoning.\n\nThis history contains two distinct research questions. One asks what strategically rational players should do under specified assumptions. The other asks how people, organisms, or adaptive programs actually behave. The prisoner's dilemma became influential partly because a single, compact payoff structure could support both inquiries without making them identical."
    },
    {
      "heading": "Dominant strategies and the one-shot prediction",
      "text": "In the one-shot game, defection is a strictly dominant strategy: it gives a higher payoff than cooperation for every possible action of the other player. If the other player cooperates, \\(T>R\\); if the other defects, \\(P>S\\). A player maximizing the specified payoff therefore does not need to predict the opponent's action to choose defection.\n\nThe same argument applies when the opponent is believed to cooperate with probability \\(q\\). The expected payoffs from the two actions are\n\n\\[U(C)=qR+(1-q)S\\]\n\nand\n\n\\[U(D)=qT+(1-q)P.\\]\n\nTheir difference is\n\n\\[U(D)-U(C)=q(T-R)+(1-q)(P-S)>0\\]\n\nfor every \\(q\\in[0,1]\\). Consequently, uncertainty alone cannot make cooperation payoff-maximizing in the standard one-shot game. Even complete confidence that the other player will cooperate leaves defection strictly preferable.\n\nMutual defection, \\((D,D)\\), is therefore the unique Nash equilibrium. There is no additional equilibrium involving a mixed strategy, because an action that is strictly dominated cannot receive positive probability in equilibrium. Allowing a mediator to send nonbinding private recommendations also cannot sustain cooperation as a correlated equilibrium of this unmodified game: whenever cooperation is recommended, defection still gives the recipient a higher payoff.\n\nYet \\((D,D)\\) is not Pareto efficient. Mutual cooperation makes both players strictly better off, moving each from \\(P\\) to \\(R\\). This is the exact sense in which the equilibrium is collectively inferior. It does not establish that every equilibrium is socially undesirable, nor that maximizing total payoff is always the appropriate ethical criterion. It establishes a particular conflict between unilateral incentives and an available mutual improvement.\n\nNonbinding communication does not by itself change the dominance calculation. A promise to cooperate, if it neither changes payoffs nor supplies enforceable commitments, leaves the same incentive to defect. Communication can nevertheless matter experimentally by changing beliefs about the social situation, making norms salient, or revealing preferences not captured by the original material-payoff model.\n\nThe prediction is conditional rather than a universal claim about rational behavior. It assumes that the listed payoffs adequately describe what each player values and that the choice has no strategically relevant consequences outside the game. Concern for another player's welfare, aversion to violating a promise, or an effect on future reputation may change the actual utility structure. Such changes require a revised model rather than a different solution to the same payoff matrix."
    },
    {
      "heading": "Repetition, backward induction, and future incentives",
      "text": "A repeated prisoner's dilemma allows the same players to interact over several rounds and condition current behavior on earlier actions. The available strategies are therefore not simply “cooperate” and “defect,” but complete rules specifying actions after every possible history. A strategy might cooperate initially and continue only while the other player has cooperated. This dependence on history creates incentives absent from the isolated game.\n\nIf the number of rounds is finite, known to both players, and the usual rationality assumptions hold throughout, backward induction predicts defection in every round. In the last round there is no future reward or punishment, so both defect. Because the last-round outcome is then fixed, the preceding round has no effect on it, and the same argument moves backward to the beginning. This conclusion depends on the end date and relevant strategic assumptions being sufficiently shared; it does not describe every finite interaction observed in practice.\n\nAn indefinitely repeated game has no known final round. Let \\(\\delta\\), with \\(0\\leq\\delta<1\\), be the discount factor applied to future payoffs. It may represent impatience, a probability that the relationship continues, or a combination of both. Consider grim trigger: cooperate until a defection occurs, then defect forever. Against an opponent using this strategy, continued cooperation yields\n\n\\[V_C=\\frac{R}{1-\\delta},\\]\n\nwhereas a defection followed by permanent mutual defection yields\n\n\\[V_D=T+\\frac{\\delta P}{1-\\delta}.\\]\n\nCooperation is preferable when\n\n\\[\\delta\\geq\\frac{T-R}{T-P}.\\]\n\nFor \\(T=5\\), \\(R=3\\), and \\(P=1\\), the threshold is \\(1/2\\). The threat is credible in the punishment phase because defection remains a best response to defection.\n\nUnder these conditions, grim trigger can support a subgame-perfect equilibrium, meaning that the prescribed strategies remain equilibrium strategies after every possible history. However, repetition does not select cooperation uniquely. Always defecting is still an equilibrium. The folk theorem describes a broader phenomenon: under suitable conditions and sufficiently patient players, many feasible payoff combinations above the relevant security levels can be supported by repeated-game equilibria.\n\nMonitoring is crucial. The preceding calculation assumes that actions are observed accurately. If an accidental failure is mistaken for deliberate defection, a permanent punishment can destroy an otherwise valuable relationship. If monitoring is private rather than shared, players may disagree about whether a deviation occurred. Repeated interaction therefore creates a possible basis for cooperation, but its effectiveness depends on patience, information, and the design of responses to apparent violations."
    },
    {
      "heading": "Conditional strategies, tournaments, and noise",
      "text": "A prominent repeated-game strategy is tit for tat: cooperate in the first round, then copy the opponent's previous action. It rewards cooperation with cooperation and answers defection with defection. Unlike grim trigger, it can return to cooperation immediately after the opponent does. Its simplicity made it important both as a theoretical example and as a computer strategy.\n\nAnatol Rapoport submitted tit for tat to Axelrod's first computer tournament, which included 14 submitted programs. A second tournament attracted 62 entries. Tit for tat achieved the highest overall score in both. The tournaments evaluated accumulated performance against a particular collection of opponents; they did not prove that tit for tat is optimal against every strategy. A strategy's success depends on the population it encounters, the length of interactions, the payoff values, and the presence of errors.\n\nTit for tat also illustrates a difference between doing well and defeating an opponent. In a direct encounter it cannot obtain a higher cumulative payoff than the opponent by exploiting cooperation: it never defects first and generally restores equality after retaliation. Its tournament success instead arose from avoiding unnecessary conflict with cooperative opponents while limiting exploitation by others. An always-defect strategy can defeat a cooperator in their bilateral comparison yet perform poorly in a population where mutually cooperative relationships generate larger returns.\n\nErrors expose an important weakness. If two tit-for-tat players intend to cooperate but one accidentally defects, the other retaliates in the next round. The first then retaliates against that response. Without another error or a forgiving modification, they can alternate between cooperation and defection indefinitely. Thus, the ability to punish is not sufficient for robust cooperation; a strategy must also recover from misunderstandings.\n\nVariants address this problem differently. Generous tit for tat sometimes cooperates after an opponent's defection, interrupting retaliatory cycles at the cost of greater vulnerability to exploitation. Win–stay, lose–shift, also known as Pavlov, repeats its previous action after a favorable payoff and switches after an unfavorable one. In the common version, \\(T\\) and \\(R\\) count as favorable, while \\(P\\) and \\(S\\) count as unfavorable. Two such players can recover mutual cooperation after an isolated implementation error through a brief sequence that includes mutual defection.\n\nNoise can affect actions, observations, or both. With independently randomized, finite-memory strategies, the interaction can often be represented as a Markov chain over recent outcomes. Long-run performance is then calculated from its stationary distribution. This approach makes it possible to distinguish intended cooperation from realized cooperation and to assess how apparently minor error rates alter the success of punishment, forgiveness, and exploitation."
    },
    {
      "heading": "Evolutionary dynamics and population structure",
      "text": "In evolutionary game theory, strategies need not result from conscious calculation. They may be inherited behavioral tendencies, learned rules, or programs copied because they perform well. Payoff affects reproductive success or imitation, and the main question becomes whether cooperative behavior can increase or persist within a population.\n\nFor a large, well-mixed population playing the one-shot game, let \\(x\\) be the proportion of cooperators. Their expected payoffs are\n\n\\[\\pi_C=xR+(1-x)S, \\qquad \\pi_D=xT+(1-x)P.\\]\n\nThe replicator equation for the cooperator frequency is\n\n\\[\\dot{x}=x(1-x)(\\pi_C-\\pi_D).\\]\n\nBecause \\(T>R\\) and \\(P>S\\), the difference \\(\\pi_C-\\pi_D\\) is negative for every population composition. Consequently, the frequency of cooperation decreases from every interior starting point. Merely placing the one-shot game in an evolutionary setting does not remove its incentive structure.\n\nRepeated encounters can change which strategies compete. A population of conditional cooperators may resist unconditional defectors when future interactions make exploitation sufficiently costly. This is different from making unconditional cooperation stable in the original one-shot matrix: the competing objects are now complete repeated-game strategies. Mutation, population size, and the possibility of neutral invasion can all affect their long-run distribution.\n\nAnother mechanism is assortment: cooperators interact disproportionately with cooperators. A useful special case is the donation game, in which cooperation costs the actor \\(c>0\\) and gives the recipient a benefit \\(b>c\\). Its payoffs are\n\n\\[T=b,\\qquad R=b-c,\\qquad P=0,\\qquad S=-c.\\]\n\nIn a well-mixed population, defection avoids the cost while retaining any received benefit. Under appropriate assortment, however, the benefits of cooperation fall more often on other cooperators. Kin-selection models express related reasoning through Hamilton's rule, \\(rb>c\\), although the interpretation of relatedness \\(r\\) depends on the population model.\n\nSpatial and network models make interaction patterns explicit using graphs. Players occupy vertices and interact with neighbors, so cooperative clusters can protect their members from repeated exploitation by outsiders. A widely studied weak-selection result for donation games on large regular graphs under death–birth updating gives the approximate condition\n\n\\[\\frac{b}{c}>k,\\]\n\nwhere \\(k\\) is the number of neighbors. This is not a universal rule: changing the update procedure, graph structure, selection strength, or relationship between competition and interaction can change the result.\n\nEvolutionary “success” means success under specified population dynamics. It does not imply foresight, moral superiority, or maximization of group welfare. The framework is valuable because it identifies mechanisms—such as repetition and assortment—that alter which behaviors receive the greater return."
    },
    {
      "heading": "Experimental evidence and social preferences",
      "text": "Experiments use the prisoner's dilemma to compare strategic predictions with observed choices. Typical designs assign participants monetary payoffs, specify whether partners remain the same, and control what participants know about the duration of interaction. These details matter because an apparently similar choice can represent a one-shot encounter, a repeated relationship, or participation in a larger environment with reputational consequences.\n\nA recurring finding is that some participants cooperate even in anonymous one-shot games. Cooperation rates vary with payoff magnitudes, instructions, participant populations, and beliefs about the other player. This variation prevents a single percentage from serving as a general behavioral constant. It also makes the game useful for examining which institutional and informational conditions change behavior.\n\nOne explanation is that monetary earnings do not exhaust participants' preferences. A person may value the other player's payoff, dislike unequal outcomes, or regard cooperation as an obligation. For illustration, suppose a player's utility is their own monetary payoff plus a weight \\(\\alpha\\) on the other player's payoff. In the donation game, switching from defection to cooperation changes that utility by\n\n\\[-c+\\alpha b.\\]\n\nIf \\(\\alpha b>c\\), cooperation is individually preferable under this utility function. The monetary payoff matrix still has the form of a prisoner's dilemma, but the utility matrix governing choice does not. This distinction prevents cooperative behavior from being labeled irrational solely because it sacrifices money.\n\nBeliefs, comprehension, and social interpretation also matter. Participants may misunderstand dominance, expect a concealed later encounter, or treat an anonymous interaction as a social relationship. Framing an action as “cooperation” rather than presenting neutral labels may invoke norms that are absent from the formal payoff table. Careful experimental design therefore separates material incentives from wording, instructions, and information about other participants.\n\nIn repeated experiments, cooperation often depends on the expected duration of the relationship and the attractiveness of exploiting a cooperator. Known finite endings can produce declining cooperation near the final rounds, although behavior does not invariably match the complete backward-induction prediction. Learning, uncertainty about an opponent's strategy, and a mixture of cooperative and opportunistic participants can sustain earlier cooperation.\n\nInterpretation also requires appropriate statistical treatment. Multiple decisions by one participant are not automatically independent observations, and players' outcomes within a session may be interdependent. Experiments involving fixed partners, random rematching, or group-level treatments create different units of variation. Observed cooperation consequently provides evidence about behavior under a specified design, not a direct measurement of a universal disposition to cooperate."
    },
    {
      "heading": "Applications and institutional responses",
      "text": "The prisoner's dilemma applies when a proposed situation has the relevant incentive ordering, not merely when parties disagree. A valid application must identify the players, their feasible actions, what each values, and why defection benefits each player regardless of the other's action. It must also establish that mutual cooperation is preferable to mutual defection. Without these elements, the analogy may conceal a coordination problem, a bargaining problem, or a conflict with a different strategic structure.\n\nA classic economic application concerns firms attempting to maintain high prices. If both restrict output, each may earn more than under aggressive competition, while either may gain by secretly undercutting the other. Mutual undercutting can then be the equilibrium. However, cooperation between firms can harm consumers. The game's term “cooperation” describes conduct within the modeled pair; it does not establish that the conduct benefits everyone affected. Competition law illustrates why the interests of omitted parties matter.\n\nEnvironmental problems can exhibit similar incentives. Reducing emissions may impose a local cost while distributing benefits broadly, encouraging each jurisdiction to rely on reductions made elsewhere. Models of climate change often extend this reasoning to many players, unequal costs, uncertain damages, and repeated negotiations. The two-player prisoner's dilemma isolates one possible incentive conflict but does not capture the full structure of international environmental agreements.\n\nInstitutions can change the game rather than merely ask players to behave differently within it. Suppose an enforceable rule imposes an expected penalty \\(f\\) whenever a player defects. Defection's payoffs become \\(T-f\\) and \\(P-f\\). If\n\n\\[f>\\max(T-R,\\;P-S),\\]\n\ncooperation becomes strictly preferable against either action. Actual enforcement also involves detection probabilities, administrative costs, and the incentives of those responsible for imposing penalties. An announced sanction that is not credible may leave the original incentives largely unchanged.\n\nContracts, deposits, monitoring systems, and reputation mechanisms operate through related but distinct channels. A contract can make future transfers depend on verified behavior. A deposit can make breach costly before it occurs. Public records can cause current defection to reduce access to future partners. Reputation therefore connects present actions with later opportunities, even when the same two individuals do not meet again.\n\nEnforcement itself may become a collective-action problem if everyone prefers that others bear its cost. Institutions must then explain not only why participants cooperate, but also why monitoring and punishment are maintained. This leads beyond the original two-player matrix to models of governance, repeated participation, and incentives among multiple roles."
    },
    {
      "heading": "Related games, advanced extensions, and limits",
      "text": "The prisoner's dilemma is one member of a broader family of social dilemmas, but nearby games have importantly different predictions. In a stag hunt, mutual cooperation and mutual defection can both be equilibria; cooperation is attractive when the other player cooperates, but risky when the other defects. Its central issue is assurance. In chicken, or the hawk–dove game, each player prefers yielding when the other escalates and escalating when the other yields. Mutual escalation is especially costly. Neither game has defection as a strictly dominant action in the prisoner's-dilemma sense.\n\nMany-person public-goods games generalize the conflict between private cost and shared benefit. If each contribution costs its contributor more than the contributor's private share of the resulting benefit, withholding contributions may be individually dominant even though universal contribution raises everyone's payoff. Threshold technologies, unequal benefits, and nonlinear returns can produce quite different strategic structures. Consequently, a resource dispute or free-rider problem should not automatically be classified as an \\(n\\)-person prisoner's dilemma.\n\nAdvanced repeated-game analysis includes zero-determinant strategies, introduced by William H. Press and Freeman J. Dyson in 2012. In the standard memory-one formulation, a player's cooperation probabilities depend on the preceding outcome: \\(CC\\), \\(CD\\), \\(DC\\), or \\(DD\\). Certain choices of these probabilities can impose a linear relationship between long-run expected payoffs,\n\n\\[a\\pi_X+b\\pi_Y+c=0.\\]\n\nSome such strategies enforce extortionate payoff relationships, linking gains above a baseline in an unequal ratio. These results concern expected payoffs under specified repeated-game conditions. They do not establish that an extortionate player must receive a high absolute payoff, that opponents will cooperate, or that extortion is evolutionarily stable in every population.\n\nThe game also has limits as an empirical explanation. Observing mutual defection does not establish that the players faced a prisoner's dilemma, because many other games produce the same visible outcome. Conversely, observing cooperation does not by itself refute the one-shot prediction: the relevant interaction may be repeated, preferences may include social concerns, or commitments may have altered the available incentives. Inferring the underlying game requires information about counterfactual payoffs and choices, not just recorded outcomes.\n\nFinally, the game distinguishes strategic analysis from moral evaluation. “Cooperate” and “defect” are labels for actions relative to the specified payoff structure, not universal judgments about right and wrong. The original suspects' story, competition between firms, and environmental cooperation involve different legal and ethical considerations. A precise application retains the mathematical insight while identifying whose interests, which consequences, and what institutional constraints lie outside the model."
    }
  ],
  "links": [
    "game-theory",
    "economics",
    "matrix",
    "expected-utility-theory",
    "zero-sum-game",
    "john-von-neumann",
    "nash-equilibrium",
    "john-forbes-nash",
    "probability",
    "expected-value",
    "mixed-strategy",
    "pareto-efficiency",
    "backward-induction",
    "subgame-perfect-equilibrium",
    "markov-chain",
    "stationary-distribution",
    "graph-theory",
    "experimental-design",
    "climate-change"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}