Key Takeaways
Use experiments to make game theory describe actual strategic behavior
A map, not a portrait. A game specifies who acts, what choices they have, what they know, and how outcomes are valued. Nash equilibrium predicts choices from the assumption that nobody can improve by changing alone, given what others do. Behavioral game theory tests that mathematical prediction against how people actually behave, then adds emotion, limited calculation, beliefs, and learning where needed.
Evidence improves the model. Camerer reports hundreds of controlled experiments. People rarely begin at equilibrium, but in many games their choices move toward it with experience. The aim is not to disprove game theory. Experiments supply the observations that theory needs, much as a wind tunnel tests an aircraft design before it is used in practice.
The book's central contribution is methodological: strategic behavior becomes an empirical question rather than a debate over whose intuition sounds more rational. Simple laboratory games isolate causes that everyday settings mix together, such as incentives, information, and fairness. That precision is powerful, but it does not automatically establish that students playing short sessions represent firms, families, or governments. The strongest use of the method is therefore iterative: isolate a mechanism, test it under controlled conditions, then check whether it survives richer settings.
Treat fairness as a payoff people may defend at real cost
Rejection can be deliberate. In an ultimatum game, one player proposes how to divide a sum and the other can accept or leave both with nothing. With $10 at stake, offers commonly approach an even split, while offers of $2 or less are rejected about half the time. This costly punishment is called negative reciprocity: repaying unfair treatment by harming the person who offered it, even at a cost to oneself. Dictator games remove the rejection option and show that some generosity remains, but much less.
Norms vary by culture. In some small-scale societies, low offers were common and accepted, closely matching self-interest predictions. Other communities made unusually generous offers. Across the studied societies, more market integration and cooperative production were associated with more equal sharing.
These results change what counts as a rational response. Rejecting money may be irrational under a utility function that values only personal cash, yet sensible if people also value fair treatment or want to punish exploitation. The distinction matters for interpreting behavior, not merely labeling it. At the same time, rejection alone cannot identify the motive: anger at unfairness, a preference for equality, and beliefs about future encounters can produce similar choices. The experiments make those motives measurable by changing the game around them.
Expect a few reasoning steps, not perfect chains of logic
Reasoning has a ceiling. In a beauty-contest game, players choose numbers from 0 to 100, and the closest number to a fraction of the average wins. If the target is two-thirds of the average, a player who guesses an average of 50 should choose about 33. If they expect others to make that calculation, they should choose about 22. The theoretical equilibrium is zero, but choosing zero is usually a poor response to actual first-round choices.
Most stop early. Across many subject pools, including game theorists, executives, and business-news readers, first choices typically reflect one or two steps of reasoning. Iterated dominance gives a related measure: people commonly avoid their own dominated choices but hesitate to assume everyone else will do so.
The useful lesson is calibration, not contempt for human reasoning. A person choosing zero in the beauty contest may have solved the formal game and still lose because the other players have not. The right action depends on the reasoning likely in the room, not the reasoning available on paper. This also explains why training matters: subjects taught backward induction can use it, but do not reliably apply it spontaneously. Strategic skill is partly a learned procedure, not simply a measure of intelligence.
Treat mixed-strategy predictions as useful averages, not literal coin flips
Randomness is hard to imitate. In a mixed-strategy equilibrium, players choose among actions with probabilities that make opponents indifferent. Laboratory subjects usually choose with frequencies reasonably close to those probabilities, although statistical tests often detect meaningful deviations. They also alternate actions more than genuinely random sequences would.
Experts can learn the pattern. Data from major tennis matches show that players win at similar rates when serving to the left or right, as equilibrium predicts. A study of soccer penalty kicks found a similar balance between kick direction and goalkeeper movement. These cases suggest that experienced professionals can approach strategic unpredictability, even when laboratory novices do not randomize cleanly.
A population-level prediction need not describe each individual mind. Players may use imperfect personal rules whose aggregate behavior looks like equilibrium mixing, or they may learn to avoid patterns that opponents can exploit. That distinction matters for applications: a policy based on average frequencies may work even if individuals do not consciously randomize. But average fit can conceal exploitable habits in particular players, which is why field evidence should examine sequences and responses, not just overall percentages.
Model learning, because people rarely begin at equilibrium
Experience changes choices. In repeated games, behavior often moves toward equilibrium, but the route depends on what players observe and remember. Reinforcement learning strengthens actions that paid off. Belief learning adjusts expectations about what others will do. Camerer and Ho's experience-weighted attraction model combines both: players learn from received payoffs and give some weight to payoffs they could have received from unchosen actions.
No single rule fits everything. In games with mixed equilibria, simple reinforcement can perform respectably. In coordination games, models that track others' behavior often do better. Across many tested games, the hybrid approach is generally more robust than either pure version, though its best-fitting parameters vary with the environment.
Learning models explain an important fact that equilibrium alone cannot: the path matters. A group may reach a good outcome, a bad one, or neither, depending on early feedback and how much attention players pay to missed opportunities. The hybrid model is valuable because it treats adaptation as a combination of memory and counterfactual thought. Its flexibility also creates a risk of overfitting, so out-of-sample prediction and comparisons across different games are essential tests.
Small starting differences can lock groups into unequal outcomes
History can split a group. In the continental-divide game, seven players chose numbers whose payoffs depended on the group's median. Starting medians at 7 or below led groups toward an equilibrium near 3; starting medians at 8 or above led them toward 12. Across ten groups, five went each way, although the low equilibrium paid about 60 cents per player and the high one about $1.12.
Coordination is not automatic. This path dependence also appears in weak-link games, where the group's payoff depends on its least ambitious choice. Large groups often drift toward the least efficient outcome, while small, stable groups can coordinate on better ones. A focal point, a choice that seems especially noticeable or shared, can help people coordinate, but it does not guarantee an efficient result.
The striking point is that a small initial difference can become a durable distributional fact. Once people respond to what has already happened, the history that created an outcome also helps justify its continuation. This is a useful warning for organizations and policy: coordination may not repair itself just because everyone would benefit from a better convention. Still, a controlled game demonstrates a mechanism, not proof that every real-world regional or industrial divide has the same cause.
Anchor bargaining around fairness, because self-serving stories prolong costly delay
Fairness has rival interpretations. In unstructured bargaining, people often gravitate toward focal divisions such as an equal split. When each side has a different idea of what is fair, their beliefs can become self-serving. In a legal-case experiment, plaintiffs and defendants differed by about $18,500 in their estimates of the likely judgment. Only 72 percent of pairs settled, taking an average of 3.75 periods.
Perspective changes outcomes. When subjects learned the case before learning which side they represented, 94 percent settled in 2.51 periods. Asking bargainers to list weaknesses in their own case also helped: 96 percent settled, in 2.39 periods. In structured bargaining, offers tend to fall between equal divisions and the game-theoretic prediction, and players often fail to look far enough ahead to apply backward induction.
The intervention works by making the other side's case easier to imagine before self-interest hardens into conviction. That is more precise than telling people to be reasonable: it changes the information they attend to. The result also suggests why settlements can fail even when both sides would benefit. They may disagree not only about their preferences, but about the facts and what an impartial decision-maker would do. Weakness-listing helps, though a party that becomes more realistic may also concede too much if only the other side stays biased.
Trust can sustain effort, while penalties can destroy goodwill
Reciprocity can solve hidden effort. In gift-exchange labor experiments, firms offered wages and workers chose how much effort to provide, which firms could not enforce contractually. Self-interest predicts minimum wages and minimal effort. Instead, firms often paid above-market wages, and workers responded with greater effort. The pattern appeared even when workers and firms were rematched, so it could not be explained simply by reputation with the same partner.
Control can backfire. Adding fines for detected shirking reduced workers' effort and weakened the response to generous wages. A later stage allowing firms to reward or punish workers improved effort again. The evidence suggests that explicit penalties can crowd out reciprocal motivation, while credible rewards and sanctions can sometimes strengthen it.
The experiments reveal that incentives are not just prices attached to behavior. They also communicate how one side views the other. A generous wage can invite reciprocity; a fine can imply distrust and change the relationship it is meant to regulate. This does not mean goodwill should replace contracts everywhere. The experiments use particular tasks, payoffs, and subject pools. The practical implication is narrower: test whether a control system complements an existing norm or undermines it.
Competition and uncertainty change whether fairness can be enforced
Structure limits expression. When several proposers competed to make an ultimatum offer to one responder, offers quickly rose toward almost the entire sum. Competition left proposers little ability to enforce equal sharing without losing the deal. When responders did not know how large the sum was, they accepted less, since a low offer might have come from a small pool. Proposers took advantage of that uncertainty.
Fairness is not simply absent. These results suggest that people can care about fair treatment yet behave self-interestedly when competition, limited information, or lack of leverage prevents them from expressing that preference. In markets, concerns about unfair outcomes may instead appear through protest, regulation, or law, as in rules against price gouging after disasters.
This is a useful distinction between a preference and the ability to act on it. Competitive institutions can make generous behavior costly, while opacity makes it difficult to know whether another party has acted unfairly in the first place. Observed self-interest therefore does not prove that fairness is irrelevant. The challenge is to avoid turning that interpretation into an unfalsifiable excuse: evidence about information, competition, and the costs of refusing a deal must be measured, not presumed.
Treat reputation as learned evidence, not instant certainty
Signals invite inference. In signaling games, one player observes private information and chooses an action that another player can see. Education, for example, can signal ability if high-ability people find it more worthwhile than low-ability people do. Experiments show that people sometimes separate by type or learn to pool on the same signal, but they do not consistently make the subtle off-equilibrium inferences required by the most refined theories.
History shapes belief. In repeated trust and entry games, behavior broadly followed reputation predictions: lenders became cautious after default, and entrants reacted to incumbent behavior. Yet rates were often different from equilibrium values. In limit-pricing games, firms sometimes revealed their types unintentionally. Players do reason strategically, but their beliefs also reflect what they have actually observed.
The most persuasive account combines inference with memory. Formal refinements ask what a perfectly reasoning observer should believe after an unexpected action; experimental players often rely on the record built by earlier actions instead. That is not a minor technical detail. A history-sensitive belief can sustain a convention that pure logic would reject, or delay the outcome that logic predicts. Models of reputation become stronger when they explain both the inference and the learning process that produces it.
Test institutions with people before trusting elegant predictions
Design needs behavioral evidence. In government spectrum auctions, theorists and regulators used experimental test runs to expose weaknesses in proposed rules before companies bid for valuable rights. The final designs were informed by likely bidder behavior, not copied mechanically from mathematical theory. Similar logic applies to contracts, market rules, and organizational incentives: a design that works under perfect rationality may fail when people misunderstand it or learn slowly.
Generalize with care. Many experiments use students, short sessions, and modest stakes. The source text reports that raising stakes often changes behavior less than expected, but calls for more work with teams, communication, complex environments, and overlapping generations. Experiments are a practical test bed, not a substitute for field evidence.
The strongest policy lesson is that institutions are behavioral technologies. Rules alter what participants can observe, what they can credibly threaten, and how quickly they learn from mistakes. Testing a design before deployment is therefore part of the theory, not an afterthought. Laboratory control also has limits: it can reveal causal mechanisms cleanly, but cannot certify that the same mechanism dominates in a workplace, a diplomatic negotiation, or a national market.
Analysis
The book's deepest argument is not that people are irrational, but that rationality is only one layer in strategic behavior. Outcomes depend on at least three separable forces: what players value, what they believe others will do, and how they learn or calculate. A rejection in bargaining may reflect a social preference; a low number in a beauty contest may reflect a reasonable estimate of limited reasoning; and a failed coordination effort may reflect path dependence rather than disagreement about the best outcome. Treating every deviation from Nash equilibrium as the same kind of mistake erases these distinctions.
The experimental method is especially powerful because game theory is unusually sensitive to details. A change in who moves first, what information is visible, or whether players compete can alter the prediction. Controlled variation can therefore identify causal mechanisms that casual observation confounds. Yet the same precision creates a translation problem: a stripped-down game isolates a mechanism precisely by omitting relationships, institutions, and history. The solution is not to abandon laboratory work, but to connect it systematically to field settings and richer experiments.
The strongest models in this account preserve strategic interaction while loosening perfect calculation and pure self-interest. That combination is more informative than either extreme: assuming people maximize only their own money, or describing every choice as unique psychology. The next conceptual advance depends on integrating preferences, beliefs, learning, and mental representations without making the model so flexible that it can explain anything after the fact. Behavioral game theory earns its name when it predicts new behavior, not simply when it explains old surprises.
Glossary
Behavioral game theory
Game theory informed by behaviorAn approach that combines the mathematical structure of games with evidence about how people actually choose. It retains strategic reasoning but adds empirically grounded features such as social preferences, emotion, limited calculation, uncertainty about others, and learning. Its purpose is to improve predictions and explanations, not to reject game theory's formal logic.
Nash equilibrium
Mutual best-response outcomeA set of strategies in which each player's choice is a best response to the choices of the others. No player can improve their payoff by changing strategy alone, assuming everyone else keeps their strategy. A game may have several Nash equilibria, and the concept by itself does not say which one people will reach.
Social preferences
Payoffs that include othersPreferences in which a person's utility depends on more than their own monetary payoff. They may care about equality, another person's welfare, fair treatment, or reciprocating kindness and unfairness. These preferences help explain why people give money away, reject profitable offers, or repay trust even when doing so reduces their personal earnings.
Negative reciprocity
Costly punishment of unfairnessA tendency to respond to perceived unfair treatment by harming the person responsible, even when the punishment also costs the responder. In an ultimatum game, a responder may reject a low offer and leave both players with nothing. The behavior is evidence of a preference about treatment, not necessarily a failure to understand the monetary consequences.
Mixed-strategy equilibrium
Equilibrium using probabilitiesAn equilibrium in which players choose among actions with specified probabilities rather than selecting one action every time. The probabilities make opponents indifferent among the actions they use, preventing a predictable pattern from being exploited. Experiments often find population frequencies near these predictions, even when individual choices are not genuinely random.
Iterated dominance
Repeated deletion of inferior choicesA reasoning procedure that removes strategies that are always worse than another choice, then repeats the process after accounting for the choices already removed. In some games, repeated deletion identifies a unique outcome. Experiments use these games to estimate how many levels of reasoning people perform about what others will choose.
Experience-weighted attraction
Learning from outcomes and alternativesA learning model in which the appeal of each strategy changes with experience. Players give special weight to the payoff they actually received, while also learning from payoffs they could have received by choosing differently. This combines the central ideas of reinforcement learning and belief learning, and can represent either approach under particular parameter settings.
Focal point
Noticeable coordination choiceA choice that seems especially prominent or likely to be recognized by others, making it a natural candidate for coordination when several outcomes are equally rewarding. Distinctive labels, familiar conventions, equal divisions, and salient positions can all make a strategy focal. Focality helps explain matching, but does not guarantee that players will select the most efficient outcome.
Risk-dominance
Equilibrium with lower joint riskA criterion for comparing equilibria that favors the outcome with smaller losses when a player chooses it but another player deviates. In a stag hunt, the safer equilibrium can be risk-dominant even though a different equilibrium gives everyone more. The concept captures why fear of miscoordination can outweigh the appeal of a larger shared payoff.
Signaling
Action conveying hidden informationA strategic action that communicates information about a player's privately known type to someone who cannot observe it directly. Education can signal ability, and a warranty can signal confidence in product quality. A signal is credible when it is worthwhile for some types to send but too costly for other types to imitate.
Intuitive criterion
Belief test for unexpected actionsA refinement for signaling games that restricts beliefs about who made an unexpected choice. If one type could not benefit from deviating from an equilibrium but another type could, the receiver should regard the possible beneficiary as the more plausible deviator. Experiments show that this logical filter does not always describe how people use observed history.
Sequential equilibrium
Consistent strategies and beliefsAn equilibrium for games with ordered moves that requires players to choose best responses and to hold beliefs consistent with observed actions wherever Bayesian updating is possible. It also specifies beliefs after actions that the equilibrium predicts should never occur. This makes it a demanding prediction about both behavior and how players interpret unexpected events.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.