Loading field map
How do people actually reason and make decisions, and how should they? This tension between normative ideals and descriptive reality has driven the study of reasoning and decision making since its modern origins. Early frameworks borrowed from economics and logic to specify what rational choice and valid inference ought to look like. Later work, often from psychology, revealed systematic departures from those ideals and proposed alternative accounts of the cognitive machinery behind judgment and choice. The history of the subfield is a series of competing answers to a single question: what kind of theory can capture both the power and the fallibility of human inference?
The first major frameworks came from outside cognitive science proper. Expected Utility Theory (1944–1970), developed by John von Neumann and Oskar Morgenstern, provided a mathematical definition of rational choice under risk. It assumed that a decision maker assigns numerical utilities to outcomes, multiplies each by its objective probability, and selects the option with the highest expected value. The theory was normative: it specified what a perfectly rational agent would do, not what people actually do. Subjective Expected Utility (1954–1970), formulated by Leonard Savage, extended the logic to situations where probabilities are not given objectively. Here the decision maker holds subjective beliefs (personal probabilities) and still maximizes expected utility. Together, these frameworks set the benchmark against which all later descriptive theories would be measured. They treated reasoning as a matter of applying formal rules to well-defined representations—an assumption that the cognitive revolution would soon challenge.
As cognitive science emerged in the 1960s, researchers began asking what mental processes underlie deductive reasoning. Mental Logic (1960–1990) proposed that people possess a mental analogue of a formal proof system: a set of inference rules (like modus ponens) that operate on logical forms extracted from sentences. Reasoning was seen as a syntactic manipulation of abstract symbols. This framework aligned naturally with the classical computationalism of the era, which treated cognition as symbol processing. But critics argued that Mental Logic could not explain why people find some logically valid inferences easy and others hard, nor why content and context so strongly influence reasoning.
Mental Models (1983–Present), introduced by Philip Johnson-Laird, offered a direct alternative. Instead of applying syntactic rules, reasoners construct mental models of the possibilities described by premises. A conclusion is accepted if it holds in all models of the premises; reasoning errors arise when people fail to consider all possible models. This framework shifted the explanatory focus from rules to representations: reasoning is about exploring the space of possibilities, not manipulating logical forms. Mental Models coexists with Mental Logic as a live competitor, and it has been extended to domains beyond deduction, including probabilistic and causal reasoning.
The most dramatic challenge to normative frameworks came from Daniel Kahneman and Amos Tversky. Their Heuristics and Biases program (1974–Present) showed that people routinely rely on mental shortcuts—representativeness, availability, anchoring—that produce systematic errors. Judgment deviates from statistical norms in predictable ways, not because of random noise but because the heuristics are efficient but imperfect. This work did not merely catalog errors; it argued that the human mind uses different cognitive machinery than the one assumed by Expected Utility Theory. The program remains active, especially in behavioral economics and applied judgment research.
Prospect Theory (1979–Present), also by Kahneman and Tversky, directly replaced Expected Utility Theory as a descriptive model of choice under risk. Its core claims are striking: people evaluate outcomes relative to a reference point (gains vs. losses), are loss-averse (losses hurt more than equivalent gains please), and overweight small probabilities while underweighting moderate and large ones. The value function is S-shaped, not linear. Prospect Theory did not reject the idea of utility; it transformed the utility function to fit actual behavior. It remains the dominant descriptive framework for risky choice and has absorbed many findings from Heuristics and Biases into a unified model.
By the 1990s, a synthesis began to take shape. Dual-Process Theories (1990–Present) propose that reasoning and decision making involve two distinct types of processing: Type 1, which is fast, automatic, and intuitive, and Type 2, which is slow, deliberate, and analytical. This framework absorbed the Heuristics and Biases findings by locating heuristics in Type 1 processing, while preserving a role for normative reasoning in Type 2. It also provided a bridge to older frameworks: Mental Logic and Mental Models could be seen as accounts of Type 2 reasoning, while Prospect Theory described the output of Type 1. Dual-Process Theories remain widely influential, though critics argue that the two-system dichotomy oversimplifies the diversity of cognitive processes.
At the same time, a very different reaction to laboratory-based bias research emerged. Naturalistic Decision Making (1989–Present) studies how experts make decisions in real-world settings—firefighters, nurses, military commanders—under time pressure, uncertainty, and shifting goals. Its flagship model, the Recognition-Primed Decision Model (1993–Present), developed by Gary Klein, argues that experts do not compare options; they recognize a situation as familiar, retrieve a typical course of action, and mentally simulate whether it will work. If it will not, they adjust. This model challenges the assumption that good decision making requires deliberation or explicit trade-offs. Naturalistic Decision Making coexists with Dual-Process Theories: both acknowledge fast, intuitive judgment, but they interpret it differently. For NDM, intuition is pattern recognition built on experience, not a bias-prone shortcut.
While psychologists debated heuristics and intuition, cognitive modelers built computational frameworks that aimed to unify reasoning, memory, and decision making. SOAR (1987–Present), developed by John Laird, Allen Newell, and Paul Rosenbloom, is a symbolic cognitive architecture based on production rules and a universal subgoaling mechanism. It treats all cognition as problem solving within a state space, using a single set of mechanisms for everything from routine skill to novel reasoning. ACT-R (1993–Present), created by John Anderson, is another production-system architecture but with a stronger emphasis on memory processes and rational analysis. ACT-R assumes that cognition is optimized for the statistical structure of the environment—a claim that aligns it with Bayesian approaches. Both SOAR and ACT-R remain active, but they have diverged: SOAR emphasizes architectural uniformity and learning from experience, while ACT-R integrates declarative and procedural memory with a rational (Bayesian) account of retrieval and decision making.
Connectionist Models (1986–Present) offered a fundamentally different computational vision. Instead of symbolic rules, connectionist networks use distributed representations and learning through weight adjustment. Reasoning and decision making emerge from the activation patterns of many simple units, not from the manipulation of discrete symbols. Connectionist models challenged both Mental Logic and production-system architectures by arguing that rule-like behavior can arise from subsymbolic processes. They have been especially successful in modeling pattern recognition, categorization, and intuitive judgment—the very domains where Heuristics and Biases found systematic errors. Connectionism remains a live tradition, though it has been partly absorbed into deep learning.
Bayesian Cognitive Science (1990–Present) and Rational Analysis (1990–Present) represent a different kind of reintegration. Both argue that human cognition can be understood as approximately optimal given the structure of the environment and the costs of computation. Rational Analysis, developed by John Anderson, treats cognitive mechanisms as adaptations to the statistical regularities of the world: memory retrieval, for example, follows a rational (Bayesian) calculus. Bayesian Cognitive Science goes further, modeling reasoning and decision making as Bayesian inference over probabilistic generative models. This framework reframes many of the Heuristics and Biases findings: biases are not errors but rational responses to uncertainty and limited data. For instance, the representativeness heuristic can be seen as a form of Bayesian inference with a small sample. Bayesian Cognitive Science and Rational Analysis overlap substantially, but Bayesian Cognitive Science places greater emphasis on explicit probabilistic models and hierarchical structure. Both remain highly active and have become dominant in computational cognitive science.
Embodied and Enactive Approaches (1990–Present) contest a premise shared by nearly all the frameworks above: that reasoning and decision making are internal, representation-driven processes. Instead, these approaches argue that cognition is shaped by the body, the environment, and action. Reasoning is not just in the head; it is distributed across brain, body, and world. Decision making is not a matter of computing utilities but of engaging with affordances in a situation. This framework challenges the very idea of a central reasoning module and questions whether formal models (whether symbolic, connectionist, or Bayesian) capture the situated, dynamic nature of real-world cognition. Embodied and enactive approaches remain a minority position in reasoning and decision making research, but they have influenced human–computer interaction, robotics, and ecological psychology.
Today, no single framework dominates reasoning and decision making. The leading active programs—Dual-Process Theories, Prospect Theory, Bayesian Cognitive Science, ACT-R, and Naturalistic Decision Making—agree on several points: human cognition is not a straightforward implementation of normative logic or utility maximization; context, experience, and cognitive constraints matter deeply; and any adequate theory must account for both fast intuitive processes and slower deliberative ones. But they disagree sharply on what intuition is. For Dual-Process Theories, intuition is automatic Type 1 processing that can be biased. For Naturalistic Decision Making, intuition is expert pattern recognition. For Bayesian Cognitive Science, intuition is approximate Bayesian inference. For Connectionist Models, intuition is the output of a trained network. These disagreements are not merely terminological; they lead to different predictions, different experimental paradigms, and different applications. The field remains pluralist, with each framework best suited to particular questions—normative benchmarks for policy, heuristics for behavioral economics, architectures for cognitive modeling, and naturalistic studies for training and expertise. The central tension between normative ideals and descriptive reality has not been resolved; it has been deepened and enriched by the very diversity of frameworks that now populate the subfield.