Satisficing

In the AI Story

Simon coined 'satisfice' — combining 'satisfy' and 'suffice' — because the concept it described had no name. Optimization had a name. Maximization had a name. The thing human beings actually do when facing complex decisions had been rendered invisible by a discipline so captured by the fiction of optimization that the empirical reality of decision-making escaped its vocabulary.

The mechanism is structural rather than cultural. Satisficing is not a failure of rationality; it is the rational response to the conditions of bounded rationality. Given finite time, finite attention, and finite computational capacity, sequential evaluation against a threshold is more efficient than exhaustive comparison. The inefficiency appears only when the threshold is poorly calibrated — too high for the available resources, or too low for the stakes of the decision.

The threshold's responsiveness to search cost connects satisficing theory to every technological transition that has reduced generation cost. The printing press reduced the cost of producing text; publishers' thresholds for which manuscripts deserved production rose. The spreadsheet reduced the cost of financial models; analysts' thresholds for which scenarios deserved modeling rose. The word processor reduced the cost of revision; writers' thresholds for which drafts deserved acceptance rose. Each transition produced both better output and greater cognitive demand on the evaluator.

AI accelerates this dynamic to a qualitatively new regime. When implementation cost drops to the duration of a conversation, the satisficing threshold undergoes a phase transition — the cost of the next alternative approaches zero while the cost of evaluating it remains constant. The result is the specific cognitive exhaustion that the Berkeley study documented: workers satisficing at a perpetually rising threshold, driven by the rational logic of a cost curve that makes continued search feel efficient while depleting the attentional resources evaluation requires.

Origin

The 1956 paper 'Rational Choice and the Structure of the Environment' formalized satisficing as a mathematical alternative to expected utility theory. Simon demonstrated that under realistic assumptions about information availability and computational capacity, satisficing often produces outcomes indistinguishable from optimization while requiring a fraction of the cognitive resources. The framework was controversial among neoclassical economists and foundational for the behavioral tradition that followed.

The concept's durability across seven decades reflects its descriptive accuracy. Every empirical study of actual decision-making — in corporations, in households, in governments, in professional practice — has confirmed that real agents satisfice rather than optimize. The question has never been whether satisficing describes behavior. The question has been whether the description deserved the theoretical attention that optimization received. Simon's answer, vindicated over time, was yes.

Key Ideas

Sequential search. Alternatives are evaluated one at a time rather than compared simultaneously across the full option space.

Threshold termination. The search ends when an alternative meets a minimum acceptability criterion, not when the best alternative has been identified.

Adaptive threshold. The acceptability criterion adjusts to the cost of continued search — rising when alternatives are cheap to generate, falling when they are expensive.

Phase transition at zero cost. When generation cost approaches zero, the threshold rises without natural bound, until it outpaces the evaluative capacity that would determine whether the threshold has actually been met.

Evaluation lags generation. The cognitive resources required to assess each alternative remain fixed while the rate at which alternatives become available expands without limit.

Debates & Critiques

The central contemporary debate concerns whether satisficing in the AI age remains rational or becomes pathological. Traditional satisficing assumes that search cost is meaningful — that the agent stops searching because continuing is expensive. When search cost approaches zero, the agent has no rational reason to stop at any particular threshold, which produces the pattern Byung-Chul Han diagnoses as self-exploitation. Simon's framework offers a more precise account: the behavior is not irrational but satisficing on a cost curve that has been radically reshaped, with the dysfunction emerging from the mismatch between generation cost (approaching zero) and evaluation cost (unchanged).