Loading field map
Behavioral game theory is the study of how people actually make strategic decisions, in contrast to classical game theory, which specifies how perfectly rational agents should decide. It sits at the intersection of economics and psychology, using experimental methods to test, refine, and sometimes replace the assumptions of traditional game theory. The field asks a deceptively simple question: when real people interact in situations where outcomes depend on the choices of all involved, what do they do, and why?
Classical game theory, developed in the mid-twentieth century, models strategic interaction through concepts like Nash equilibrium—a set of strategies where no player can improve their payoff by unilaterally changing their own choice. This framework is elegant and powerful, but it rests on strong assumptions: players are self-interested, perfectly rational, have common knowledge of the game's structure, and can perform arbitrarily complex calculations. Behavioral game theory does not reject this framework outright. Instead, it treats it as a benchmark—a precise null hypothesis against which actual behavior can be measured.
The field's central questions follow from this stance. When do people actually reach equilibrium, and when do they systematically deviate? What cognitive shortcuts or social preferences explain those deviations? How do people learn to play strategically over time, and does that learning converge to what theory predicts? And crucially, can we build better models—models that predict what people will do, not just what they should do—by incorporating psychological realism?
The roots of behavioral game theory lie in the experimental economics movement of the mid-twentieth century. Early researchers, most prominently Vernon Smith in the 1960s, began testing economic theories in controlled laboratory settings. At first, the focus was on markets, where results often aligned with competitive equilibrium predictions. But when researchers turned to simple two-person games, the picture changed.
The ultimatum game, introduced in the early 1980s, became the field's signature demonstration. In this game, one player proposes a division of a sum of money; the second player can accept or reject. If the second rejects, both get nothing. Classical game theory predicts the proposer offers the smallest possible amount, and the responder accepts, since something is better than nothing. The experimental result was strikingly different: proposers typically offered 40–50% of the stake, and responders frequently rejected offers below 20–30%, even though rejection meant sacrificing real money. This single game crystallized the challenge: either people are not purely self-interested, or they are not perfectly rational, or both.
The 1980s and 1990s saw the field consolidate. Researchers systematically documented deviations from equilibrium predictions across a range of games—public goods games, prisoner's dilemmas, bargaining games, and coordination games. Two broad explanations emerged. One emphasized social preferences: people care not only about their own payoffs but also about fairness, reciprocity, and equity. The other emphasized bounded rationality: people lack the cognitive capacity to compute equilibria and instead use simple heuristics or rules of thumb. These two strands—one motivational, one cognitive—remain the field's twin pillars.
The social preferences approach explains deviations from self-interested behavior by positing that people have preferences over outcomes beyond their own material payoff. The most influential formalization is inequality aversion, developed by Ernst Fehr and Klaus Schmidt in 1999. Their model assumes that people dislike unequal outcomes—both when they are disadvantaged and, to a lesser extent, when they are advantaged. This single assumption elegantly explains the ultimatum game results: responders reject low offers because the utility loss from inequality outweighs the material gain, and proposers anticipate this and offer fairer splits.
A complementary model, developed by Matthew Rabin in 1993, emphasizes reciprocity: people respond kindly to perceived kindness and harshly to perceived hostility, even at personal cost. This approach captures the dynamic, intention-based nature of fairness—people care not just about outcomes but about the intentions behind them. A low offer in the ultimatum game is not merely unequal; it is mean, and rejection is a form of punishment.
These models have been remarkably successful at organizing experimental data across many games. They predict, for example, why people contribute to public goods despite incentives to free-ride, and why they punish free-riders even when punishment is costly. However, the approach has limits. It treats social preferences as stable individual traits, but experimental evidence shows that behavior varies substantially with context, framing, and the identity of the other player. People are not uniformly fair or reciprocal; they are conditional and situation-dependent. Moreover, the models are often ex post—they fit observed data well but have limited power to predict behavior in novel games without additional assumptions.
The bounded rationality approach, by contrast, focuses on cognitive constraints. Its central insight is that people do not solve games from first principles; they learn, adapt, and use simplified mental models. The most developed framework here is reinforcement learning, borrowed from psychology: players repeat strategies that yielded good outcomes and abandon those that yielded poor ones. This approach predicts that behavior will converge to equilibrium in many games, but slowly and with substantial noise along the way.
A more sophisticated variant is belief learning, where players form expectations about what others will do based on past observations and choose best responses to those beliefs. The experience-weighted attraction (EWA) model, developed by Colin Camerer and Teck-Hua Ho, unifies reinforcement and belief learning into a single framework, showing that both processes operate simultaneously and that their relative weight varies across individuals and games.
A different strand of bounded rationality research emphasizes level-k reasoning. In this model, players are classified by the depth of their strategic thinking. A level-0 player chooses randomly or heuristically; a level-1 player best-responds to level-0; a level-2 player best-responds to level-1; and so on. This approach has been particularly successful in explaining behavior in games like the "beauty contest," where players must guess a number that is a fraction of the average of all guesses. Classical game theory predicts a unique equilibrium at zero, but experimental subjects typically guess numbers in the 20–40 range, consistent with level-1 or level-2 reasoning. The level-k model captures the crucial fact that people do not assume others are fully rational; they assume others are somewhat less sophisticated than themselves.
The bounded rationality approach has its own limits. It is often descriptive rather than predictive—the models can fit data well but require knowing which learning rule or which level of reasoning a given player will use. It also struggles to explain the persistence of non-equilibrium behavior in games where learning should eventually converge. And it does not easily accommodate social preferences; a purely cognitive model cannot explain why people reject positive offers in the ultimatum game.
It is tempting to see social preferences and bounded rationality as competing explanations, and some researchers do frame them that way. But the more accurate picture is one of complementarity. Real behavior in strategic settings is shaped by both motivational and cognitive factors, often simultaneously. A person might reject an unfair offer because they dislike inequality and because they have not fully calculated the expected value of acceptance. A proposer might offer a fair split because they anticipate the responder's anger and because they lack the strategic sophistication to exploit the situation.
Modern behavioral game theory increasingly integrates both strands. Models of social learning combine reinforcement with social preferences, allowing players to learn not only what works but also what feels fair. Models of cognitive hierarchy incorporate social preferences by allowing players at different levels of reasoning to have different utility functions. The field has also expanded to include emotions—anger, guilt, pride—as direct drivers of behavior, and identity—the social categories people belong to and the norms attached to those categories.
Contemporary behavioral game theory is characterized by several ongoing developments. First, there is a strong emphasis on external validity: researchers increasingly ask whether laboratory findings generalize to field settings, real markets, and policy interventions. This has led to a productive exchange with development economics, where behavioral insights inform the design of institutions, contracts, and incentives.
Second, the field has become more computational. Advances in machine learning and artificial intelligence have produced new tools for modeling strategic behavior, including models that learn from large datasets of human play. These models are often more accurate than traditional theory at predicting behavior, though they are less interpretable and offer less insight into the underlying psychological mechanisms.
Third, there is growing attention to heterogeneity. Early behavioral game theory often reported average behavior, obscuring substantial individual differences. Modern research uses larger samples and statistical techniques to identify distinct behavioral types—for example, people who are consistently fair-minded, consistently strategic, or consistently erratic. This type-based approach has improved prediction and has connected behavioral game theory to personality psychology and individual differences research.
Fourth, the field has engaged with neuroeconomics, using brain imaging and physiological measures to understand the neural basis of strategic decision-making. This work has shown, for example, that unfair offers activate brain regions associated with disgust and pain, and that costly punishment activates reward-related regions. Neuroeconomic evidence does not replace behavioral models, but it constrains them by revealing the psychological processes that generate behavior.
Finally, behavioral game theory has become a normative tool. Insights from the field are used to design mechanisms—auctions, matching markets, voting systems—that are robust to real human behavior rather than idealized rationality. This application-oriented work, sometimes called behavioral mechanism design, represents the field's practical payoff: a better understanding of how people actually behave allows institutions to be built that work with human nature rather than against it.
The field's durability rests on its methodological commitment to empirical testing. Behavioral game theory does not offer a single grand theory of strategic behavior; it offers a set of tools—experimental designs, statistical methods, and formal models—for discovering how people actually decide. Its findings are often surprising, sometimes unsettling, and always informative. The result is a picture of human strategic behavior that is far richer than classical game theory alone, and far more accurate.