For decades, decision theory operated under a tidy assumption: emotion was noise, a perturbation in the signal of rational choice. The expected utility framework bequeathed by von Neumann and Morgenstern treated affect as either irrelevant or pathological—something to be controlled for, corrected, or transcended. The agent computed; the agent did not feel.

This dichotomy has collapsed under empirical pressure. Patients with ventromedial prefrontal lesions retain intact logical reasoning yet make catastrophic life decisions. Healthy subjects systematically integrate momentary affect into intertemporal trade-offs. Neuroeconomic recordings show value signals in the orbitofrontal cortex that are inseparable from interoceptive states. The data demand a richer formalism—one in which emotion is not the antithesis of reason but a distinct computational input to it.

What follows examines three threads in this reformulation: affect-as-information theories that treat feelings as data about value, the analytic distinction between integral and incidental affect, and Damasio's somatic marker hypothesis as a candidate mechanism. Together they suggest a decision architecture in which the calculus itself includes affective variables, weighted and combined according to principles we are only beginning to specify mathematically. The question is no longer whether emotions enter the decision calculus, but how we should formally represent their entry.

The affect-as-information framework, developed initially by Schwarz and Clore and later formalized within Bayesian decision models, treats emotional states as signals carrying probabilistic information about the value of options. Rather than corrupting utility computation, affect provides a low-dimensional summary statistic of high-dimensional evaluative input.

Formally, we can represent the agent's posterior expected utility as E[U|x, a], where x denotes cognitively appraised features of the option and a denotes the affective state. The affective channel functions as an additional likelihood term, updating value estimates much as any other observation would. Crucially, this is not dual-process romanticism—it is conditional inference.

Empirical work supports this formulation. Pham and colleagues demonstrated that decisions made under instructions to consult feelings show greater coherence and predictive validity for experiential outcomes than purely deliberative judgments. Affect appears to integrate information that explicit cognition cannot easily verbalize, particularly in domains involving aggregate consequences, gestalt evaluations, or temporally distributed outcomes.

The framework dovetails with reinforcement learning accounts in which mood functions as a momentum term over reward prediction errors. Eldar and colleagues showed that mood biases learning rates in ways consistent with treating affect as a running estimate of environmental favorability. Under this view, an emotion is a compressed sufficient statistic, not an irrational impulse.

What this reframing buys us is theoretical parsimony. We need not posit a separate emotional system competing with a rational one. We need only a unified inferential agent whose value function incorporates affective variables as legitimate evidence, with appropriate priors and update rules.

Takeaway

Treat your feelings as data, not verdicts. They are probabilistic signals about value—informative when calibrated, misleading when their source is misattributed.

Not all affect is created equal. The distinction between integral and incidental affect, sharpened by Loewenstein and Lerner, partitions emotional inputs by their causal relation to the choice at hand. Integral affect is generated by the decision options themselves—the dread one feels contemplating a risky surgery, the anticipated pleasure of a vacation. Incidental affect arises from sources unrelated to the decision yet bleeds into it: residual anger from morning traffic shaping an afternoon negotiation.

From a normative standpoint, integral affect is informationally legitimate. It encodes value-relevant features of the options. Incidental affect, by contrast, constitutes a misattribution—a violation of the procedural invariance axiom that decisions should not depend on irrelevant context. Yet behaviorally, the two are difficult to separate. Lerner and Keltner's appraisal-tendency framework shows that incidental fear induces risk aversion across unrelated domains, while incidental anger produces risk seeking and optimism.

This poses a formal challenge. If the agent cannot distinguish the source of an affective signal, how should the decision calculus handle it? One proposal models the system as performing source monitoring with imperfect resolution, applying a discount factor to affect of uncertain provenance. When attention is drawn to the irrelevant source—the classic Schwarz and Clore weather-mood study—the misattribution effect attenuates.

The distinction has substantial implications for both descriptive accuracy and prescriptive design. Choice architectures that decouple affective context from decision moments—cooling-off periods, structured deliberation protocols—exploit the integral-incidental split to reduce noise. Markets, courts, and clinics implicitly assume agents can perform this disentangling, often without warrant.

The deeper theoretical point is that emotion's epistemic status depends on its referential structure. The same physiological state can be diagnostic or noisy depending on what it is about. Decision theory must therefore track not only the magnitude of affect but its intentional object.

Takeaway

An emotion's value as evidence depends entirely on whether it points at the decision in front of you. The feeling itself does not announce its source.

Damasio's somatic marker hypothesis offers perhaps the most influential mechanistic proposal for affect's role in choice. The claim is that decision options become tagged with bodily states—visceral, autonomic, and musculoskeletal signals—accumulated through prior experience. These markers, processed through ventromedial prefrontal and insular circuits, bias deliberation toward advantageous options before explicit cognitive evaluation completes.

The Iowa Gambling Task provided the foundational evidence. Healthy participants developed anticipatory skin conductance responses to disadvantageous decks before they could verbally articulate the contingencies. Patients with ventromedial prefrontal damage failed to develop these markers and continued selecting from disadvantageous decks even when they could describe the optimal strategy. The dissociation between knowing and choosing implicated somatic feedback as causally efficacious in selection.

Subsequent work has been more equivocal. Maia and McClelland argued that participants often possess explicit knowledge earlier than initially reported, weakening claims about preconscious somatic guidance. Replications have produced mixed effect sizes, and the interpretation of skin conductance as a pure value signal—rather than an arousal correlate—remains contested. The hypothesis as originally stated may be too strong.

Yet the underlying intuition has survived in modified form. Interoceptive accuracy correlates with decision quality on tasks involving uncertainty. Insular cortex activity tracks risk and ambiguity in ways consistent with bodily state monitoring. The hypothesis pointed correctly at something real: the brain treats the body as part of its evidential apparatus, and decision computations recruit afferent signals about physiological state.

The contemporary synthesis is less dramatic but more defensible. Somatic signals are one class of input to value computation, weighted alongside semantic, episodic, and counterfactual information. They are neither the secret engine of rationality nor a primitive holdover, but a channel whose contribution depends on context, calibration, and the agent's interoceptive resolution.

Takeaway

The body keeps a ledger the conscious mind cannot fully read. Whether to trust it depends not on whether it speaks, but on whether it has been well-trained.

The reason-versus-emotion dichotomy was always a philosophical artifact, not an empirical finding. What replaces it is more demanding: a decision theory in which affect is a structured input with formal properties, sources, and weights. The mathematics is incomplete, but the architecture is becoming visible.

Three principles emerge. Emotions function as information about value, subject to inferential update. Their epistemic status depends on whether they refer to the decision at hand or leak in from elsewhere. And their substrate includes bodily signals whose contribution is real but contextually calibrated rather than universally diagnostic.

The agent we are constructing in our theories is neither the bloodless calculator of classical economics nor the passion-driven actor of romantic counter-narrative. It is an integrative system that uses every signal it can—including the felt ones—and whose rationality is measured not by emotional absence but by appropriate weighting. The calculus has always included emotion. We are only now writing the equations.