The classical theory of revealed preference rests on a deceptively simple axiom: choices expose pre-existing valuations. When an agent selects A over B, we infer that U(A) > U(B), treating the choice as a diagnostic instrument peering into a stable utility function. This framework, formalized by Samuelson and refined through decades of decision theory, treats preferences as antecedent causes and choices as their observable consequences.

Yet a growing body of neuroeconomic and psychological evidence disrupts this unidirectional causality. Choices do not merely reveal preferences—they construct them. The act of selecting an option triggers cascading neural and cognitive processes that retroactively inflate the chosen alternative's subjective value while devaluing the rejected one. What began as a marginal preference difference becomes, post-decision, a chasm.

This bidirectionality carries profound implications. If preferences are partially endogenous to the choice process itself, then revealed preference theory conflates two distinct signals: what the agent valued before choosing and what the choice produced. For welfare economics, policy design, and computational models of decision-making, this distinction is not merely philosophical. It suggests that our formal apparatus may be measuring an artifact of its own instrumentation.

Since Brehm's 1956 free-choice paradigm, experimenters have documented a robust phenomenon: after choosing between two similarly-rated options, subjects subsequently rate the chosen alternative higher and the rejected one lower. This spreading of alternatives occurs even when initial ratings suggested near-indifference, and it persists across timeframes ranging from minutes to years.

The effect's magnitude is non-trivial. Meta-analyses report spreading effects of 0.4 to 0.8 standard deviations on post-choice valuation measures, comparable in size to substantial experimental manipulations of intrinsic value. Critically, this spreading emerges even in blind-choice paradigms where subjects cannot recall which option they selected, ruling out simple self-perception explanations.

Sharot and colleagues demonstrated the neural signature of this effect using fMRI, showing that caudate nucleus activity during rating increases for chosen items post-decision, tracking the elevated subjective value. The brain does not simply report a stable utility—it recomputes valuation in light of the commitment already made.

This poses a fundamental identification problem for revealed preference theory. If we observe choice C at time t and measure preference P at time t+1, we cannot cleanly attribute P to the antecedent preferences that produced C. The measurement is contaminated by the choice itself, creating what econometricians would recognize as a form of endogeneity bias baked into the very structure of preference elicitation.

The theoretical implication is that utility functions, as inferred from choice data, are not invariant objects being sampled but rather dynamic constructs being sculpted. Each observation perturbs the underlying quantity we seek to measure.

Takeaway

Preferences are not static objects that choices reveal—they are dynamic constructions that choices actively sculpt. Every decision is simultaneously an act of measurement and modification.

The mechanisms underlying choice-induced preference change operate at multiple levels of analysis. At the psychological level, cognitive dissonance theory posits that inconsistency between behavior (having chosen A) and cognition (A and B are equally attractive) generates aversive arousal, motivating attitude change to restore consonance. But the neural evidence suggests processes more fundamental than post-hoc rationalization.

Izuma and colleagues traced the effect to activity in the posterior medial frontal cortex during the choice itself, with dorsolateral prefrontal engagement mediating the subsequent revaluation. This suggests a predictive coding architecture: once a commitment is made, downstream systems update their value priors to align with the action, minimizing prediction error between behavior and internal valuation.

From a computational perspective, this can be modeled as a Bayesian updating process where the choice itself serves as evidence about the agent's own preferences. The agent, uncertain about their true utility parameters, treats their behavior as informative data—a form of self-directed inference sometimes formalized in models of introspective learning.

Reinforcement learning frameworks offer complementary insight. The chosen option, by virtue of being sampled, receives updated value estimates through experience, while the rejected option's value remains untested and often decays. This asymmetric information updating creates a structural bias toward the chosen alternative, independent of any dissonance-reduction motive.

The convergence of these mechanisms—dissonance, predictive coding, self-inference, and asymmetric learning—suggests that preference entrenchment is not a peripheral quirk but a fundamental feature of how choosing systems maintain coherent behavior over time.

Takeaway

Commitment is not merely a psychological attitude but a computational necessity: agents that fail to consolidate their choices cannot maintain the behavioral coherence required for coordinated action.

If choices construct preferences, the epistemic foundations of applied welfare economics require reconsideration. The standard approach infers consumer surplus and welfare gains from observed choices, assuming these reveal pre-existing utility. But when the choice environment shapes what is chosen and subsequently valued, welfare inferences risk circularity.

Consider default effects in retirement savings or organ donation. If defaults not only alter behavior but also reshape subsequent preferences—such that individuals come to prefer whatever default they were assigned—then evaluating welfare by post-choice preferences legitimizes any manipulation of the choice architecture. The libertarian paternalism debate hinges partly on this issue.

Bernheim and Rangel have proposed behavioral welfare economics frameworks that distinguish between choices made under conditions favorable versus unfavorable to preference expression. But operationalizing this distinction requires theoretical criteria for which choice contexts produce authentic versus constructed preferences—a demarcation that remains philosophically contested.

For computational modelers, the finding demands reformulation of utility as a state variable rather than a fixed parameter. Models incorporating dynamic preference updating—such as those in dual-process frameworks or drift-diffusion models with post-decisional consolidation—better capture observed behavior but complicate parameter identification substantially.

The methodological upshot is not that revealed preference theory should be abandoned but that its inferences require humility. What we call preferences may be better understood as equilibrium states of an ongoing process of choice and consolidation, meaningful within their temporal context but not the timeless utilities the theory once presumed.

Takeaway

Welfare analysis based on revealed preferences may be measuring what choices have produced rather than what agents antecedently wanted—a distinction that fundamentally alters the normative weight of behavioral data.

The bidirectional relationship between choice and preference dissolves the neat separation between what agents want and what they do. Choices function simultaneously as diagnostic instruments and constructive acts, revealing prior valuations while manufacturing new ones through commitment mechanisms operating at psychological, neural, and computational levels.

This does not render decision theory incoherent, but it does require sophistication about what our models represent. The utility functions we infer are not photographs of stable mental states but rather dynamic equilibria between antecedent dispositions and the shaping force of decisions themselves.

For researchers, the imperative is methodological: designing paradigms that distinguish antecedent from constructed preference, and models that treat valuation as endogenous to the choice process. For theorists, it is conceptual: recognizing that the boundary between description and prescription, between measuring preferences and creating them, is more porous than classical frameworks assumed.