Irrationality as an evolutionary advantage. Rationality as a goal is marketing

Gerd Gigerenzer devotes Gut Feelings (2007) to documenting that human "irrational" heuristics are ecological rationality: optimal in their environment, not errors. Deciding fast with little information saved lives for 200,000 years. The heuristic "assume it's an enemy" kills few friends but saves you from many enemies. A "rational" AI playing the human survival game would have died in the first generation. Rationality as a goal is a marketing error —and a philosophy-of-mind one.

The asymmetry evolution couldn't afford to ignore

Picture the savanna, 100,000 years ago. A noise in the bushes. Two options: investigate, assuming it's probably the wind, or flee, assuming it's probably a predator. The asymmetry in the cost of being wrong in each direction isn't academic. If it was the wind and you fled, you spent energy and came back. If it was a predator and you stayed to investigate, you were eaten. The descendants of those who fled are here writing and reading this. The descendants of the rationalists who stayed to investigate didn't make it to descendants.

This, without rhetoric, is natural selection acting on cognitive architecture. Any trait whose false-positive rate is cheap and whose false-negative rate is lethal will get fixed in the population. The human threat detector is calibrated, by evolutionary construction, to err on the side of the false positive. Calling it a "bias" or "error" in lab psychology compares it with a Bayesian agent computing posterior probabilities. Comparing it with the statistical reality of the savanna presents it as an optimal strategy given the asymmetric cost.

The observation isn't new. Pascal Boyer and Justin Barrett applied it to agency detection in Religion Explained (Basic Books, 2001) and Exploring the natural foundations of religion (Trends in Cognitive Sciences 4(1), 2000) respectively. Cosmides and Tooby generalized it from the late eighties as evolutionary psychology: many of the human "errors" cataloged by classic psychology are optimal solutions to problems the organism faced during the Pleistocene, badly calibrated for the industrial context where they're tested. David Buss, in Evolutionary Psychology (Pearson, 6th ed., 2019), gathers the full program in textbook form. Antonio Damasio, since Descartes' Error (Putnam, 1994), added the clinical complement: the separation between reason and emotion that axiomatic rationality presupposes doesn't hold in patients with damaged ventromedial cortex. They can reason and they decide badly. The soma-cognition integration is structural, not contingent.

Axiomatic rationality and why nobody follows it

There's a well-known theoretical construct from decision calculus worth naming to situate the discussion. Expected-utility theory, axiomatized by von Neumann and Morgenstern in Theory of Games and Economic Behavior (1944), defines an ideal rationality through axioms: completeness, transitivity, continuity, independence. Any agent whose preferences satisfy the axioms can be described as a maximizer of a well-defined expected utility.

Three observations about this construct. First: no real human satisfies it. Violations of transitivity and independence have been documented experimentally since the Allais paradox (Maurice Allais, Le comportement de l'homme rationnel devant le risque, Econometrica 21, 1953). Second: no practical artificial intelligence satisfies it either, except in very restricted domains. LLMs aren't axiomatic utility maximizers; they're probabilistic samplers. Third, and most important: perfect compliance with the axioms isn't desirable in an agent with finite cognitive resources. Completeness over all imaginable options requires comparing infinite alternatives. Transitivity over all pairs requires cataloging them. Continuity requires infinite sensitivity to infinitesimal differences. A real agent can't be axiomatically rational, and simulating being so within a limited cognitive horizon consumes the horizon itself.

Herbert Simon, in Models of Man (1957), called this bounded rationality. His thesis: real agents don't maximize, they satisfice. They choose the first option that crosses an acceptable threshold. Satisficing isn't laziness, it's indispensable cognitive economy. And most of the cognitive psychology that followed, in one line or another, operates within this framework.

The "take-the-best" heuristic, an example

Gigerenzer and Goldstein, in Reasoning the Fast and Frugal Way. Models of Bounded Rationality (Psychological Review 103(4), 1996), published results that at the time seemed paradoxical. They compared two classes of decision procedure on inference tasks. The first class used all available data weighted by their statistical validity —the "rational" procedure in the classic sense. The second class, called take-the-best, examined cues in order of validity and stopped at the first that discriminated between alternatives, ignoring everything else.

The result: under conditions of incomplete or noisy information, which are most real conditions, take-the-best matched or beat the optimal models. The fast, frugal heuristic made better decisions than the procedure using more information. Why: because additional information, in real conditions, tends to be noise more than signal, and the optimal models overfit the noise. The simple heuristic, by ignoring the noise by construction, isn't hurt by it.

This series of results, replicated and extended by Gigerenzer, Todd and the ABC Research Group in Simple Heuristics That Make Us Smart (Oxford University Press, 1999) and later work, gave empirical grounding to the thesis of ecological rationality. The name is precise: the rationality of a strategy depends on the ecosystem it operates in. A strategy optimal in a well-defined theoretical environment can be bad in the real environment, and vice versa. The quality of a decision isn't an intrinsic property of the decision, it's a property of the decision-environment pair.

Kahneman and Gigerenzer, what each measures well

It's worth not falling into the false opposition. The two lines —Kahneman's and Gigerenzer's— capture different parts of the phenomenon, and both are useful when used in their domain.

Kahneman and Tversky measured deviations from the Bayesian model under lab conditions where the problem is well defined, information is complete or controlled, and the cost of error is low. Under those conditions, humans are worse than the Bayesian model. The description is correct. The operational consequence is important: in domains resembling lab conditions —standardized tests, financial markets with fast feedback, repeated low-cost decisions— it's wise to distrust the fast heuristic and use explicit procedures.

Gigerenzer and the ABC Group measured performance under conditions of incomplete, noisy information, a new domain, time pressure. Under those conditions, humans are comparable to or better than sophisticated models, and fast heuristics are efficient. The description is also correct. The operational consequence: in real domains with high uncertainty —the first emergency-medicine response, the initial assessment of a risk, a decision under threat— the fast heuristic isn't a defect, it's fitness.

The question that closes the debate isn't who's right, but in which domain each framework applies. And the interesting question, taken into AI territory, is the inverse of the usual one.

The "rational" AI versus the survival game

Imagine for a moment that contemporary AI, optimized to produce statistically optimal answers on the basis of its training corpus, played the game the hominids played for two hundred thousand years. What would happen?

The first noise in the bushes. The system measures the prior distribution of similar events, the conditional probability of each interpretation, the expected utility of fleeing versus staying, the energy cost, the relevant contextual frames. While it processes, the predator arrives. The system, in its final reply, suggests: "Recommended action with probability 0.73: flee westward, the direction of lower gradient."

The sentence is read by the corpse.

The exercise is rhetorical, yes. It's also revealing. The human heuristic "assume threat first" is optimal precisely because it gives up the calculation in exchange for speed. The cost of the false positive —fleeing when it was the wind— is low. The cost of the false negative —staying when it was a predator— is lethal. Any system designed to survive in that game converges to the same heuristic, not because it's the rational answer, but because it's the only answer that ends with subjects alive to answer again.

The extrapolation to the contemporary conversation about AI has an uncomfortable twist. When an AI system is sold as "more rational" than the human for making decisions, the system's rationality is assumed to be the relevant property. In closed, controlled environments, with complete information and abundant time, it is. In open, noisy environments, with time pressure and asymmetric cost, it is less so. And in environments where fast decision with incomplete information is the very condition of the problem, calculated rationality stops being a virtue and starts being an obstacle.

Stanovich's tripartite mind

Keith Stanovich, in Rationality and the Reflective Mind (Oxford UP, 2011), put forward a framework worth mentioning because it orders the debate. He distinguishes three levels of processing in the human agent.

The autonomous mind runs fast, automatic, frequently unconscious processes calibrated by evolution. It includes perceptual reflexes, threat detection, basic heuristics. It isn't available for direct introspection.

The algorithmic mind runs deliberate, slow processes that require working memory. When psychometric intelligence measures something, it measures here.

The reflective mind runs the metacognitive control: it decides when to trust the autonomous mind, when to activate the algorithmic one, when to doubt the results. It's the mind that detects the bias and, occasionally, corrects it.

Stanovich's conclusion is relevant to this conversation: the dysrationality observed in cognitively competent subjects comes mainly from failures in the reflective mind, not in the algorithmic one. And the reflective mind is a recent cultural product, not a deep evolutionary trait. The autonomous mind, by contrast, has been there since the Pleistocene.

This reorders the conversation. The heuristics the Kahnemanians call biases are productions of the autonomous mind, calibrated by evolution for concrete environments. The dysrationalities we detect when we sit down to think are failures of the culturally educated reflective mind. The human isn't irrational in the strong sense. They're an agent with three levels of processing whose coordination doesn't always produce optimal decisions by modern metrics.

The critique of Western cultural bias

There's a further twist worth including even if it's uncomfortable. Much of the classic literature on rationality implicitly assumes a specific cultural model: the ideal of the deliberative, autonomous individual who weighs evidence before deciding. This model is a historical product, identifiable in the European Enlightenment and deepened in twentieth-century Anglo-American academic culture. It isn't a universal property of human cognition. In other cognitive traditions —oral, communal, authority-based, consensus-based— the regulative ideal is different.

Calling irrational anyone who doesn't decide according to the Enlightenment ideal is transferring a local cultural bias as a universal measure. A person who decides by intuition, the authority of an elder, or inherited common sense isn't deciding worse in an absolute sense. They're deciding according to a different regulation whose hit rates in their context are comparable, and sometimes superior, to those of the Enlightenment ideal in its own.

This observation isn't a defense of irrationality as a value. It's recognition that the rationality metric used isn't neutral. Applying it universally without nuance falls into the cognitive version of Eurocentrism. What matters in each culture is the ecological quality of the decision, not its resemblance to the protocol of a Princeton lab.

The operational question for contemporary AI

The public conversation about AI drags an assumption worth unearthing. The assumption is that a system more rational than the human would be a better system for making human decisions. The assumption is false or, at best, domain-dependent.

For decisions in a stable domain, abundant information, fast feedback —medical-image diagnosis with massive prior cases, logistics optimization, chess, industrial-process control— a system more calculated than the human produces better results. The evidence is clear and there's no debating it.

For decisions under high uncertainty, incomplete information, time pressure, asymmetric cost —emergencies, political decision, moral judgment in novel circumstances, fluid social interaction— a system more calculated than the human can produce worse results, not because it's technically inferior but because it optimizes the wrong metric for the problem. The fast human heuristic was calibrated by two hundred thousand years of selection for precisely those environments.

The uncritical substitution of the human by AI in decisions of the second category isn't modernization. It's a change of criterion without justification of its superiority in that specific domain. And the justification, when attempted, almost always appeals to lab metrics that don't capture the ecological quality of the decision.

Does this mean AI is worse than the human? Of course not. It means that the question "is AI more rational than the human?" isn't the relevant question. The relevant question is "in which domain, with what asymmetric cost, with what time horizon, does each type of system suit?" That question admits an operational answer. The other is marketing.

Definitions

Ecological rationality. A concept of Gigerenzer and the ABC Research Group. The quality of a decision strategy is a joint property of the strategy and the environment it operates in, not an intrinsic property of the strategy. Heuristics that are "irrational" in the lab can be optimal in their natural niche.

Take-the-best. A simple heuristic studied by Gigerenzer and Goldstein (1996). It examines cues in order of decreasing validity, stops at the first that discriminates between alternatives and ignores the rest. It performs as well as or better than optimal models when information is incomplete or noisy.

Bounded rationality. A concept of Herbert Simon (1957). Real agents don't maximize utility, they satisfice thresholds. Satisficing isn't a defect but indispensable cognitive economy in systems with finite resources.

Tripartite mind. Stanovich's framework (2011). It distinguishes the autonomous mind (fast, automatic, evolutionary), the algorithmic mind (deliberate, costly) and the reflective mind (metacognitive control). Dysrationality typically emerges in the third.

Cost asymmetry. The difference between the cost of the false positive and that of the false negative in a detection task. When the asymmetry is large, the optimal detectors aren't the Bayesian ones, they're those that minimize the catastrophic error even at the cost of increasing the cheap error.

References

Buss, D. M. (2019). Evolutionary Psychology. The New Science of the Mind (6th ed.). Pearson. A general framework of evolutionary psychology supporting the adaptive reading of human heuristics.

Cosmides, L. & Tooby, J. (1992). Cognitive adaptations for social exchange. In The Adapted Mind, Oxford University Press. Foundational work on human reasoning adapted to specific ecological problems.

Damasio, A. (1994). Descartes' Error. Emotion, Reason, and the Human Brain. Putnam. A neurobiological framework on the integration of cognition, emotion and soma; the clinical complement to ecological rationality.

Gigerenzer, G. (2007). Gut Feelings. The Intelligence of the Unconscious. Viking. An accessible synthesis of ecological rationality as an alternative framework to the classic bias catalog.

Gigerenzer, G. & Goldstein, D. G. (1996). Reasoning the Fast and Frugal Way. Models of Bounded Rationality. Psychological Review 103(4), 650–669. A formal demonstration of the effectiveness of simple heuristics under uncertainty.

Gigerenzer, G., Todd, P. M. & ABC Research Group (1999). Simple Heuristics That Make Us Smart. Oxford University Press. A compilation of studies on take-the-best and other frugal heuristics.

Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. The classic position that ecological rationality complements or contests, depending on interpretation.

Simon, H. A. (1957). Models of Man. Social and Rational. John Wiley. The introduction of the concept of bounded rationality and of satisficing as an alternative model to maximization.

Stanovich, K. E. (2011). Rationality and the Reflective Mind. Oxford University Press. The tripartite framework on human cognition; it locates dysrationality in the reflective mind.

von Neumann, J. & Morgenstern, O. (1944). Theory of Games and Economic Behavior. Princeton University Press. The classic axiomatization of expected utility as the theoretical ideal of decision rationality.

You might also like

• The human as a defective machine
• Decision-making without data
• Emotion as noise

Elsewhere

• Bounded Rationality (Stanford Encyclopedia of Philosophy)
• Bounded rationality (Wikipedia)