Bounded Rationality
Bounded rationality is Herbert Simon's foundational contribution to the social sciences: the recognition that real decision-makers — human or institutional — cannot optimize because the informational and computational requirements of optimization exceed the capacity of any biological mind. Instead, they satisfice, searching sequentially through alternatives until one clears a threshold of acceptability, then stopping. The insight earned Simon the 1978 Nobel Prize in Economics and dissolved the fiction of Homo economicus that had powered a century of economic theory. Its architectural implication is that every institution, hierarchy, and organizational form ever built is at bottom a device for managing cognitive limits. The concept has acquired new urgency in the AI age: the bounds Simon identified — information, computation, time — have been dramatically relaxed, while a fourth bound he also named, attention, remains intact and has become the binding constraint.
In the AI Story
Simon's 1955 paper 'A Behavioral Model of Rational Choice' introduced bounded rationality as a corrective to the prevailing rational-choice framework. The argument was not that humans are irrational but that rationality operates within bounds imposed by limited information, limited cognitive capacity, and limited time. Given these constraints, decision-makers cannot evaluate all alternatives simultaneously; they must search through them one at a time, using heuristics to direct the search.
The concept reshaped economics, organizational theory, and the design of institutions. If human rationality is bounded, then organizational structures exist not as expressions of authority or efficiency in the abstract but as decomposition architectures that break complex decisions into sub-decisions manageable by bounded minds working in coordination. The division of labor, the reporting chain, the standard operating procedure — all are prosthetics for cognitive limitation.
The AI age produces a specific test of bounded rationality's durability. Three of the four bounds Simon identified — information, computation, and time — have been substantially relaxed by large language models. The fourth, attention, has not. This asymmetric relaxation does not dissolve bounded rationality but relocates its binding constraint from generation to evaluation, producing the specific cognitive pathology of AI-augmented work: unbounded output meeting bounded judgment.
The framework connects directly to Kahneman and Tversky's heuristics and biases program, which mapped the specific shortcuts that bounded minds deploy when facing decisions beyond their computational capacity. Where Simon described the structural fact of boundedness, the later behavioral program cataloged its systematic consequences.
Origin
Simon arrived at bounded rationality from political science rather than economics, having studied municipal administration and organizational decision-making firsthand. He observed that corporate executives, chess grandmasters, and government officials alike did not conform to the optimizing model economists assumed. They considered handfuls of plausible alternatives, selected the first that met their criteria, and moved on. The pattern was so pervasive that Simon concluded the optimizing model was describing a species that did not exist.
The 1955 paper formalized the observation into a theory that Simon developed across six decades — through work on satisficing, organizational design, human problem-solving, and the science of the artificial. Each subsequent domain extended the implications of the original insight rather than replacing it.
Key Ideas
Rationality has bounds. Not because minds are irrational, but because the informational and computational requirements of optimization exceed biological capacity.
Search replaces optimization. Bounded agents evaluate alternatives sequentially rather than simultaneously, terminating when a threshold is met.
Institutions are cognitive prosthetics. Organizational structures exist to decompose decisions too complex for any single bounded mind into sub-decisions manageable by coordinated bounded minds.
The bounds determine the design. Every well-functioning system is designed around the cognitive characteristics of the agents who will operate within it; systems designed for idealized agents produce overload and failure.
AI relocates bounds; it does not remove them. The relaxation of information, computation, and time constraints does not dissolve bounded rationality — it shifts the binding constraint to attention and evaluation.
Debates & Critiques
Critics have argued that bounded rationality is a transitional concept awaiting replacement by more specific accounts of cognitive limitation (behavioral economics, dual-process theory, ecological rationality). Simon himself anticipated this and welcomed specification — bounded rationality was never intended as a final theory but as a corrective to the optimizing fiction. The AI age has prompted a secondary debate: whether the expansion of information, computation, and time constraints constitutes a transcendence of bounded rationality or merely a relocation of its binding constraint. This volume argues the second reading.
Appears in the Orange Pill Cycle
Further reading
- Simon, A Behavioral Model of Rational Choice (1955)
- Simon, Models of Bounded Rationality (1982)
- Gerd Gigerenzer and Reinhard Selten, eds., Bounded Rationality: The Adaptive Toolbox (2001)
- Daniel Kahneman, Thinking, Fast and Slow (2011)
- Hugo Mercier and Dan Sperber, The Enigma of Reason (2017)