Human judgment systematically depends on the context in which information is presented, ordered, or framed, even when the underlying factual content is held constant. This feature, known as contextuality, means that the same question can elicit different responses depending on which other questions are asked before it, how options are grouped, or what reference points are made salient. Traditional probabilistic models that assume stable underlying preferences and beliefs struggle to accommodate these reversals and incompatibilities without introducing ad hoc parameters. Quantum cognition approaches propose that these context effects are not mere noise or bias but signatures of an underlying contextual structure in cognitive states, analogous to quantum contextuality in physical systems.
In many experiments, judgments violate classical assumptions of joint probability distributions. For instance, the response to a question about a person’s honesty depends on whether participants were first asked about the person’s kindness or their competence, and the joint pattern of responses cannot be reconciled with a single, context-independent set of probabilities. Such order effects imply that mental states are transformed by the act of measurement—here, the act of asking a question—so that subsequent judgments are made from a different cognitive state than they would have been otherwise. This mirrors how in quantum theory, the outcome probabilities for a measurement depend on which other measurements have been or could be performed on the system.
Contextuality in human judgment appears in a wide range of domains, including probability estimation, moral evaluation, social perception, and economic choice. In probability estimation, people often give different likelihood judgments for the same event when it is part of a disjunction, a conjunction, or a conditional statement, even when the logical relationships among these descriptions are explicit. In moral evaluation, the perceived wrongness of an action can increase or decrease depending on what other morally salient scenarios are concurrently considered. In social perception, judgments of traits such as warmth, intelligence, and dominance can interfere with each other, such that attending to one trait changes how evidence for another is encoded and used.
One hallmark of this form of contextuality is the presence of incompatible questions—pairs of judgments that cannot be assigned simultaneous, well-defined values without contradiction. When people are asked about intentions and outcomes in moral dilemmas, for example, the order in which these aspects are probed can generate different attributions of blame or praise. Attempts to assign a single, context-free moral evaluation that would reproduce all observed responses across different orders and framings fail, just as one cannot assign a single set of predetermined values to all measurements in a contextual quantum system. This suggests that mental states may be more fruitfully modeled as context-dependent superpositions rather than as fixed points in a classical state space.
Quantum cognition models capture these phenomena by representing cognitive states as vectors in abstract spaces where different questions correspond to different bases. A given state can yield different outcome probabilities depending on which basis is used to “measure” it, reflecting the influence of perspective, framing, and task demands. Interference terms can account for constructive or destructive combination of evidence, explaining why adding an option or presenting an additional piece of information can paradoxically increase or decrease the judged probability of a target outcome. Contextuality thus becomes a structural property of the cognitive architecture, not merely an artifact of inattention or error.
In decision-making, contextuality is visible in phenomena such as preference reversals and attraction effects. The presence of a dominated alternative can shift preference between two main options, and the same pairwise choice can flip when embedded in different choice sets. Classical models require unstable utilities or shifting error terms to address this, whereas contextual models treat the act of comparison itself as a transformation of the cognitive state. Each comparison context selects and weights attributes differently, so that the probabilities of choosing each option depend on which other options are simultaneously evaluated.
Contextuality also shows up in temporal judgments, where earlier queries shape the cognitive landscape on which later judgments are formed. When people repeatedly evaluate risk, fairness, or trustworthiness, previous responses can prime, inhibit, or reconfigure latent associations. This is more than simple anchoring: the logical relations among successive questions can create interference patterns, producing non-linear shifts in response distributions that are inconsistent with additive noise models. Here, contextuality reveals itself as a dynamic property of evolving cognitive states rather than a static bias.
The notion of contextuality is closely tied to the idea that mental representations are not fully specified until they are queried. Before a question is asked, a person may hold only a partially resolved stance on an issue, combining multiple incompatible considerations. The question acts as an operator that collapses this indeterminate state into a particular response while simultaneously preparing the mind for subsequent queries. This helps explain why people can be remarkably consistent within a given question order yet diverge sharply when the order is rearranged. What appears as inconsistency from a classical standpoint can be understood as lawful, context-sensitive updating of an inherently indeterminate cognitive state.
Importantly, contextuality does not imply irrationality in the sense of random or arbitrary behavior. Instead, it suggests that rationality is conditional on the cognitive context in which information is processed and judgments are requested. A person’s responses can follow coherent patterns within each context, even as no single global assignment of beliefs and preferences fits all contexts simultaneously. This reframes many so-called cognitive biases as consequences of contextual structure rather than as simple deviations from normative probabilistic rules.
Seen through this lens, human judgment can be interpreted as the outcome of neural inference under constraints of limited resources, noisy information, and shifting task demands. If the bayesian brain hypothesis is correct and the brain continuously maintains and updates priors about the world, then contextuality reflects how these priors are reweighted and reorganized when different questions probe different aspects of the same latent generative model. The observed context effects in responses can then be viewed as signatures of how internal models are sliced, rotated, and collapsed by the cognitive operations that a particular environment, query, or decision problem imposes.
Mathematical frameworks for contextual cognitive states
To formalize contextuality in human judgment, mathematical frameworks from quantum theory are repurposed to describe how cognitive states depend on the questions that are asked and the order in which they are posed. Instead of treating beliefs and preferences as fixed points in a classical probability space, these models represent cognitive states as vectors in a Hilbert space, an abstract vector space equipped with an inner product. Questions, cues, and decision tasks are represented as operators or bases acting on this space. The probabilities of different responses are then derived from the squared magnitudes of projection amplitudes, in close analogy with quantum cognition models of measurement and collapse.
In this geometrical picture, a person’s latent stance on an issue corresponds to a state vector that does not yet commit to a specific answer. Each question defines a measurement basis, a set of mutually exclusive response options that span the relevant subspace. When a question is asked, the current cognitive state is projected onto this basis, and the outcome is selected according to the Born rule: the probability of each response is given by the squared length of the component of the state along the corresponding basis vector. This rule naturally handles probabilistic responses without requiring hidden random noise; instead, uncertainty is encoded as superposition of incompatible dispositions within the state itself.
Crucially, different questions often correspond to incompatible bases. Asking first about a person’s honesty and then about their competence is modeled as projecting onto one basis and then onto another that is rotated relative to the first. Because projections in Hilbert space do not generally commute, the sequence of measurements matters: measuring honesty then competence yields a different final state than measuring competence then honesty. This non-commutativity mathematically captures order effects in human judgment that violate classical models where joint probabilities are presumed to be independent of the sequence of evaluation.
The operators representing questions can be defined as projection operators or more general Hermitian operators whose eigenvectors correspond to possible responses and whose eigenvalues represent response categories or intensities. When the cognitive system is queried, the measurement operator acts on the current state, producing both an outcome and an updated state. The updated state is no longer the original superposition but an eigenstate or mixture of eigenstates compatible with the observed answer. Subsequent questions act on this updated state, so their outcome probabilities reflect the history of prior questions. In this way, the formalism embeds a dynamic model of question-induced cognitive change directly into the algebraic structure.
Beyond simple binary or discrete-choice questions, contextual cognitive states can also be modeled using positive operator-valued measures, which generalize the notion of measurement beyond sharp projectors. This allows for “fuzzy” queries in which different response options partially overlap in meaning, as is common in everyday language and social judgment. The associated operators can capture graded membership, ambiguous categories, and overlapping mental constructs in ways that classical set-theoretic models struggle to represent without contradictions or arbitrary thresholds.
Interference is a central feature of this formalism. When multiple cognitive paths to a judgment exist, such as different lines of reasoning favoring or disfavoring an option, these paths can be represented by probability amplitudes that combine either constructively or destructively. The overall probability of a response is then determined not just by summing independent contributions but by summing amplitudes and taking their squared magnitude, which introduces cross terms that encode interference. This explains empirical cases where presenting an additional scenario or attribute changes the perceived likelihood of an outcome in a non-additive way, reflecting interference among underlying considerations rather than simple aggregation.
Another family of mathematical frameworks adapts contextuality-by-default and sheaf-theoretic approaches from the foundations of quantum theory. Instead of starting from Hilbert spaces, these frameworks begin by specifying random variables that are indexed not only by content (which question is asked) but also by context (what other questions are asked alongside it, in what order, or under which framing). For each context, one can observe joint distributions of responses to the questions presented together, but across contexts, certain combinations of variables are never jointly observable. The central problem is whether there exists a single global joint distribution over all variables that agrees with each observed contextual distribution. Failure of such a global embedding is a formal signature of contextuality.
In this probabilistic approach, contextuality is operationalized by comparing observed patterns to constraints imposed by classical probability theory. If all contextual distributions can be glued together into one global distribution, the system is noncontextual; each question is assumed to tap into a pre-existing, context-free property. If no such global distribution exists, the system is contextual, meaning that no assignment of stable hidden values can reproduce all observed correlations simultaneously. This analysis does not require physical quantum mechanics, yet it captures the same structural impossibility that motivated quantum contextuality theorems in physics.
For cognitive systems, the sheaf-theoretic perspective treats experimental conditions—such as different question orders, framing manipulations, or choice sets—as distinct contexts. Each context is associated with an observed empirical distribution of judgments. The mathematical machinery then assesses whether there is a global assignment of latent cognitive states and response tendencies that could generate all these distributions while respecting the rules of classical probability. When such a global model fails to exist, one can quantify the degree of contextuality, providing a principled measure of how deeply context pervades the system’s responses.
The Hilbert-space and contextuality-by-default frameworks are complementary. The former provides a compact geometric representation and can generate specific experimental predictions about interference, order effects, and violation of classical inequalities. The latter offers a representation-agnostic test of contextuality that can be applied to behavioral data without assuming any particular cognitive architecture. When both approaches converge on the same set of phenomena, the case strengthens that cognitive contextuality is not an artifact of model choice but a genuine structural feature of neural inference.
Bayesian perspectives can be integrated into these quantum-like formalisms by recasting priors and likelihoods as components of a contextual state rather than as fixed scalar probabilities. Instead of a single prior distribution over hypotheses, the cognitive system is endowed with a prior state vector that encodes overlapping, partially incompatible hypotheses. Measurement operators derived from sensory evidence, questions, or decision tasks select different “slices” of this prior state. In place of classical Bayes’ rule, state update rules involve unitary evolution followed by projection, so that updating is shaped not only by the data but also by the measurement context in which data are interpreted.
This integration maintains the core intuition of the bayesian brain hypothesis—that the brain performs approximate probabilistic inference—while relaxing the assumption that all relevant probabilities can be embedded in a single Kolmogorovian space. Under this view, neural inference operates in a high-dimensional representational space in which different tasks correspond to different effective bases. Learning can be modeled as continuous transformations (unitary or approximately unitary) that reshape the geometry of this space, aligning it with the statistical structure of the environment while preserving the contextual nature of cognitive states.
Time evolution of contextual cognitive states is typically described with unitary operators, analogous to how quantum systems evolve between measurements. In the cognitive setting, unitary transformations capture internal processing such as mental simulation, re-interpretation of evidence, or re-framing of a problem before an explicit response is made. These transformations rotate the state vector within the Hilbert space, changing the potential outcome probabilities for future questions even in the absence of new external data. This provides a mathematical handle on phenomena like spontaneous reappraisal, delayed preference shifts, and the impact of internal rumination on later choices.
More general dynamical laws, such as Lindblad-type equations, have been proposed to model open cognitive systems that interact continuously with their environment and with noise. These equations allow for decoherence, in which superposed cognitive states gradually lose their interference potential, effectively approximating classical mixtures. In everyday decision-making, decoherence can be interpreted as the stabilization of a particular viewpoint or habit after repeated queries or experiences. Once a person has repeatedly answered similar questions or made similar choices, their cognitive state may evolve toward an eigenstate of the corresponding measurement operators, reducing contextuality for those specific judgments while leaving other domains more indeterminate.
At a more abstract level, contextual frameworks can also be expressed using generalized probabilistic theories, which characterize the state space, measurements, and transformations of a system without committing to specific Hilbert-space structures. Cognitive states are elements of a convex set, measurements are affine maps that yield outcome probabilities, and transformations are positive maps preserving normalization. Quantum cognition is then a special case of a broader family of contextual models. This perspective is useful for comparing different mathematical assumptions about cognition—such as dimensionality, type of interference, and composition rules—and for identifying which empirical patterns uniquely point to quantum-like structure versus other forms of contextual probabilistic reasoning.
These mathematical tools jointly provide a rich language for describing how human judgments depend on context, order, and framing while remaining formally coherent. Instead of forcing behavioral data into a single global probability space or treating violations of classical axioms as errors, they allow contextuality to be built into the foundations of the model. Cognitive states become relational entities, defined not in isolation but through the set of potential questions and tasks that can be posed to the system, and the geometry of the underlying state space captures the lawful ways in which these relations shape observable responses.
Cognitive priors as quantum-like probability amplitudes
Within a quantum-like framework, cognitive priors are not merely scalar probabilities attached to isolated hypotheses but structured states that encode overlapping tendencies, partial commitments, and mutually incompatible perspectives. Instead of assigning a single numerical prior to each hypothesis within one global sample space, the system is characterized by a vector of probability amplitudes whose squared magnitudes yield classical priors only when a particular question is posed. This representation treats priors as latent dispositions that depend on the measurement basis defined by the task, thereby embedding contextuality directly into the prior structure itself.
Consider a person’s belief about whether a new technology is “safe.” In a classical Bayesian model, one might posit a prior probability that the technology is safe and then update this belief using evidence. In a quantum-like model, the cognitive prior is a superposition of several interpretive stances—trust in scientific institutions, fear of unknown risks, comparison to past technologies—each corresponding to different bases of evaluation. The amplitude structure encodes how these stances coexist before a particular question directs attention to one or another. When the person is asked, “How safe is this compared to existing alternatives?” the prior amplitudes are projected into a comparative-safety basis, yielding context-specific prior probabilities that differ from those that would have emerged if the question had emphasized, say, regulatory oversight or long-term environmental impact.
This shift from scalar priors to amplitude-based priors allows the same latent state to support multiple, incompatible prior distributions depending on the query. In standard Bayesian modeling, such context dependence seems to violate coherence: the agent appears to change priors arbitrarily when the framing changes. In a quantum cognition model, the apparent incoherence reflects the geometry of the state space. Different questions correspond to different bases, and the prior amplitudes relative to one basis need not be simultaneously realizable as priors in another. The resulting pattern of responses can still satisfy internal consistency constraints within each context, even though no single global prior distribution reconciles all of them.
Mathematically, a cognitive prior state can be represented by a normalized vector in a Hilbert space whose components in a chosen basis correspond to complex amplitudes over interpretive categories or hypotheses. The squared modulus of each component gives the prior probability associated with that hypothesis in that basis, but the relative phases between components capture information that is invisible in purely classical priors. These phases give rise to interference effects when evidence or questions combine different reasoning paths. For example, the judged probability that a politician is “honest and competent” can differ from what would be predicted by multiplying separate prior probabilities for honesty and competence, because the joint question probes a different superposition of trait concepts than the individual questions do.
From this perspective, learning and experience are not simply changes in scalar priors but deformations of the amplitude structure. When a person acquires new information, the cognitive state undergoes a transformation that can both shift the magnitudes of amplitudes and alter their phases. Evidence that is framed in a way that reconciles previously conflicting considerations may reduce destructive interference and increase the effective prior for a coherent narrative. Conversely, evidence that highlights contradictions can increase destructive interference, lowering the net prior for a simple, unified interpretation and making ambivalence or indeterminacy more likely.
Interference among cognitive priors becomes especially relevant in scenarios where multiple narratives or explanatory schemes compete. Suppose an individual evaluates whether a health intervention is effective. One narrative emphasizes statistical trial results; another stresses anecdotal reports; a third foregrounds institutional trust or distrust. In a quantum-like representation, each narrative corresponds to a subspace or basis, and the prior state may be a superposition across them. When a question or piece of evidence activates more than one narrative simultaneously, the associated amplitude paths can interfere. If the phases are aligned, the narratives support each other, effectively elevating the judged likelihood of effectiveness beyond what a simple weighted average of priors would suggest. If the phases are opposed, the narratives undercut each other, yielding judgments that look overly skeptical given the individual strands of evidence.
Interference also offers a structured account of well-documented conjunction and disjunction fallacies. Classic experiments show that people sometimes assign a higher probability to a specific conjunctive scenario (for example, a bank teller who is active in a feminist movement) than to a more general one (a bank teller). In the amplitude framework, the general category and the conjunctive scenario correspond to different projections of the same prior state. The prior amplitude for the conjunctive narrative can benefit from constructive interference among traits and stereotypes, while the projection onto the more generic category can suffer from destructive interference among less coherent or less salient features. The resulting pattern of judged probabilities no longer appears as a simple logical error but as an outcome of the geometry of cognitive priors in a structured state space.
Crucially, the notion of priors as quantum-like amplitudes meshes naturally with the bayesian brain hypothesis while modifying its architectural assumptions. The brain is still engaged in continuous prediction and neural inference about hidden causes, but the underlying representation of uncertainty is enriched. Instead of a single Kolmogorovian probability distribution, the system maintains a context-sensitive prior state that stores both magnitudes and phases, encoding how different internal models relate and potentially interfere. Prediction errors then act as measurement-like events that partially collapse this state, reweighing and rephasing priors in a manner that depends on the task, attentional set, and framing of the incoming data.
This richer representation sheds light on how prior expectations can be highly resistant to change in some contexts yet remarkably labile in others. When a prior state is close to an eigenstate of the measurement operators associated with a particular task—for instance, a deeply entrenched political identity relative to partisan cues—questions in that domain will elicit highly stable prior probabilities with little interference. By contrast, in domains where the prior is a delicate superposition of partially compatible models—such as assessing risks of emerging technologies—different framings can rotate the effective measurement basis, revealing different components of the prior state and thereby producing large shifts in expressed beliefs without large changes in underlying evidence.
The amplitude formalism also illuminates how implicit and explicit priors can diverge. Implicit priors, shaped by long-term associative learning and cultural immersion, may determine the deep geometry of the cognitive state space. Explicit priors, elicited by direct questioning or introspection, correspond to projections of this state onto specific bases defined by language and conscious reasoning. Because different elicitation methods implement different measurement operators, the same underlying state can yield divergent explicit priors across tasks that appear superficially similar. This helps to explain why surveys, behavioral tasks, and implicit association measures often produce systematically different “prior beliefs” for the same participants.
Within this framework, updating rules generalize classical Bayes’ rule. A piece of evidence is represented as an operator that acts on the prior state, sometimes modeled as a unitary transformation followed by a projection in the measurement basis defined by the query. The unitary part encodes context-sensitive preprocessing: how the evidence is interpreted, which associations it activates, and how it reshapes the relationships among latent hypotheses before a commitment is made. The projection step then selects a specific outcome, generating revised priors for subsequent questions. Because different evidence contexts correspond to different operators, the same raw data can induce distinct updates depending on how they are presented or which comparisons they invite.
This generalized updating scheme naturally produces path dependence. The order in which evidence is encountered, and the sequence of questions posed, jointly determine the trajectory of the prior state through the cognitive state space. Once certain projections have been made—such as publicly endorsing a position or repeatedly answering similar questions—the state may be driven closer to eigenstates associated with those commitments. Later evidence, especially if framed in compatible bases, will then have a reduced capacity to induce large rotations of the state. In contrast, early evidence that is framed in a way that emphasizes conflict or ambiguity can preserve superposition and interference, keeping priors flexible and sensitive to subsequent contextual shifts.
Viewing priors as quantum-like probability amplitudes also clarifies how contextuality can coexist with pragmatic rationality. From a classical vantage point, shifting priors across contexts appears as a violation of coherence axioms. Yet if the mind’s representational substrate is inherently contextual, there may be no single “true” set of priors to which all observable judgments should conform. Instead, rationality may be better understood as local coherence within a given context, together with efficient transitions between contexts given the organism’s goals, constraints, and resource limits. The amplitude representation makes this explicit: what counts as a prior is always relative to a measurement basis, and different bases are related by lawful geometric transformations rather than arbitrary inconsistency.
Modeling cognitive priors as amplitudes opens the door to systematic exploration of how education, persuasion, and social communication reshape the geometry of belief. Persuasive messages can be thought of as designed sequences of operators that move a prior state from one region of the state space to another, not only changing the magnitudes of specific beliefs but also altering interference patterns among them. Narratives that align previously orthogonal considerations can generate constructive interference, amplifying the perceived plausibility of targeted conclusions. Conversely, introducing dissonant frames or highlighting inconsistencies can induce destructive interference, weakening prior commitments even when surface-level probabilities appear only modestly affected. In this way, quantum-like modeling offers a principled vocabulary for understanding how complex patterns of messaging and experience can sculpt the deep structure of cognitive priors over time.
Empirical studies of contextual effects in decision-making
Empirical research on decision-making has uncovered a dense landscape of context effects that resist explanation within classical probability theory yet align naturally with quantum cognition models. One core line of work focuses on question order effects in attitude surveys. For example, when respondents are asked to evaluate the honesty and then the competence of a political figure, the distribution of answers differs systematically from when the competence question comes first. Crucially, the joint pattern across all possible orders cannot be embedded into a single context-independent joint distribution: a hallmark of contextuality. Quantum-like models capture this by representing each question as a measurement in a different basis, with non-commuting operators yielding order-dependent projections of the underlying cognitive state.
These order effects are not limited to social judgment. Experiments in probabilistic reasoning show that the assessed likelihood of events can change depending on whether questions about related events are posed beforehand. For instance, participants may judge the probability of a stock market crash differently depending on whether they were first asked about political instability, inflation, or technological disruption. Each preceding question selectively activates particular subspaces of the belief state, re-weighting and re-phasing the priors that govern subsequent answers. Empirically, this manifests as systematic, directionally consistent shifts rather than random noise, supporting the view that internal states evolve according to structured, context-sensitive dynamics.
Another rich empirical domain concerns violations of classical rationality axioms in economic choice. Preference reversals, attraction effects, and compromise effects show that the inclusion or framing of options can flip preferences between the same pair of alternatives. In a typical attraction effect paradigm, people prefer option A over option B in a binary choice, but when a third, dominated option A′ is added, preference shifts toward A. Classical utility theory struggles to account for why a clearly inferior option would increase the attractiveness of a related one. In a quantum cognition framework, the expanded choice set is understood as a different measurement context, with the additional option altering the comparison basis and inducing interference among attribute-based evaluations. Empirical choice frequencies can often be fit by quantum-like models using a small number of geometrical parameters representing rotations and phases in the decision space.
Disjunction and conjunction fallacies offer further evidence of contextual structure in judgment. In classic experiments, participants rate a specific conjunctive scenario (“Linda is a bank teller and active in the feminist movement”) as more probable than a more general one (“Linda is a bank teller”), contrary to classical probability rules. Variants of this paradigm have been replicated across different storylines and domains, including medical diagnoses, legal judgments, and everyday forecasting. Quantum-like models reproduce these patterns by representing stereotypical narratives as states that are closer, in Hilbert-space geometry, to the conjunctive description than to the generic category. Empirical data reveal interference-like signatures: probability judgments for combined scenarios deviate from the sums and products one would expect if the underlying beliefs were governed by a single Kolmogorovian distribution.
Experimental studies of disjunction effects in decision under uncertainty provide another compelling test bed. In the well-known two-stage gambling paradigm, many participants prefer to accept a second gamble if they are told they won the first, and also if they are told they lost the first, yet decline the second gamble when they do not know the outcome. This violates the classical “sure-thing principle,” which states that if an option is preferred in each possible state of the world, it should be preferred when the state is unknown. Quantum-like models explain this via interference between cognitive paths corresponding to different imagined outcomes. The unknown-outcome condition allows these paths to superpose, generating destructive interference that suppresses the propensity to gamble. Empirical parameter estimates from such studies quantify the strength of interference and link it to individual differences in risk attitudes and ambiguity tolerance.
Survey research has systematically investigated attitude polarization and response inconsistency across frames, revealing patterns incompatible with a single, stable set of preferences. For example, support for a public policy can depend strongly on whether it is framed in terms of gains versus losses, or as an extension versus a repeal. Within each frame, responses can appear coherent and transitive, but across frames they become cyclic or intransitive. Large-scale studies using multiple framing conditions have attempted to reconstruct a global preference ordering, only to find no consistent ranking that fits all contexts. Contextuality-by-default analysis of these datasets often detects violations of noncontextual inequalities, suggesting that no classical joint distribution over latent preferences can reproduce the observed pattern. Quantum-like models, in contrast, treat each frame as a distinct measurement that rotates the evaluative basis, making frame-sensitive judgments an expected outcome of the underlying geometry.
Empirical work on moral judgment provides analogous evidence. Experiments manipulating the order in which intentions, outcomes, and social norms are probed show strong sequence-dependent attributions of blame and praise. For instance, asking about harmful outcomes before highlighting benign intentions tends to increase condemnation relative to the reverse order. These effects persist even when participants are explicitly instructed to consider all relevant information, indicating that they are not mere artifacts of inattention. Quantum-like models interpret each query as an operator that partially collapses a morally ambivalent state—one that superposes conflicting considerations such as harm, fairness, loyalty, and authority—into a more determinate stance. Subsequent questions then act on this updated state, yielding the characteristic path dependence seen in empirical moral evaluation.
In intertemporal choice experiments, where participants choose between smaller-sooner and larger-later rewards, contextual manipulations such as time framing (“in 12 months” versus “one year from now”), presence of intermediate options, or prior tasks involving waiting can alter discount rates in systematic but non-classical ways. For example, presenting multiple delayed options can make the far-future payoff relatively more attractive than when it is evaluated in isolation, a kind of temporal attraction effect. When modeled with quantum-like dynamics, these phenomena emerge from rotations along a time-preference axis in the underlying state space, with different contexts emphasizing immediacy, growth, or self-control dimensions. Measured response patterns, including reversals and non-monotonic discounting, frequently violate the assumptions of stationary exponential or hyperbolic discounting, but can be captured by context-sensitive transformations of a single latent decision state.
Empirical studies of probability judgments in dynamic environments further highlight the role of contextuality. In sequential prediction tasks—such as forecasting the next move in a market, the next event in a narrative, or the next action of an opponent—participants’ forecasts depend strongly on the sequence of prior queries and feedback. Even when the statistical structure of the environment is stationary, humans often exhibit path-dependent estimation patterns, including hysteresis and primacy–recency trade-offs. Under a bayesian brain interpretation, these phenomena are traditionally modeled as miscalibrated updates or selective attention. Quantum-like models instead represent the time course of neural inference as unitary-like evolution punctuated by measurement events (questions, feedback), which steer the belief state along different trajectories. Empirical fits often reveal that the effective state dimensionality needed to explain these trajectories is relatively low, consistent with a geometric representation of a small number of latent hypotheses rather than an unwieldy high-dimensional classical distribution.
Another important empirical line involves so-called “answering without knowing” paradigms, where participants exhibit above-chance performance on forced-choice tasks despite reporting no conscious knowledge. For example, in some perceptual or implicit learning tasks, participants choose the correct option more often than chance while claiming to guess randomly. From a classical vantage point, this suggests the existence of unconscious probabilities that have not yet surfaced in explicit reports. Quantum cognition approaches sharpen this distinction by treating implicit and explicit responses as measurements in different bases on the same underlying state. Experiments that alternate implicit and explicit probes reveal order effects: implicit performance changes after explicit questioning, and vice versa. This mutual influence is naturally explained as sequential projections on incompatible observables within a single state space.
Group decision-making provides yet another domain where contextuality has been empirically documented. Studies of deliberating juries, committees, and crowds show that aggregate judgments depend strongly on the sequence and framing of arguments, the order of individual contributions, and the presence of salient exemplars. Group polarization experiments demonstrate that discussion often leads to more extreme positions than initial private judgments would predict. Rather than attributing this solely to social influence or conformity, some researchers have modeled the group as a composite system whose joint state encodes superpositions of members’ stances. Deliberation is then cast as a sequence of interactions (unitary-like transformations) and measurements (votes, public statements) that entangle and gradually decohere individual states. Empirical patterns such as path dependency of outcomes, non-averaging aggregation, and sudden shifts after “pivotal” contributions align with such contextual, quantum-like representations.
More formal tests of contextuality in cognition adapt tools from the foundations of quantum theory, such as Bell-type inequalities and contextuality inequalities. Experiments have constructed cognitive analogs of these tests by designing tasks in which participants make judgments about combinations of features or questions that are never all observed together in a single context. For instance, respondents might evaluate pairs of traits of a public figure under different pairings—honest versus friendly, friendly versus competent, competent versus decisive—without ever evaluating all four simultaneously. Classical models impose constraints on the correlations among these pairwise judgments; violations indicate contextuality. Multiple empirical studies have reported such violations, even after carefully controlling for memory, fatigue, and response biases, suggesting that the observed contextuality is structural rather than an experimental artifact.
Longitudinal and learning studies provide insight into how contextual effects evolve over time. When participants are repeatedly exposed to similar decision problems, some initially strong order and framing effects diminish, while others become more entrenched. For example, training in statistical reasoning can reduce certain conjunction fallacies, but framing effects in moral and political judgments often remain robust. From a quantum-like standpoint, this pattern indicates selective decoherence: repeated measurement in the same basis stabilizes certain dimensions of the cognitive state, pushing them toward eigenstates that exhibit little contextuality for those queries. Other dimensions remain superposed and context-sensitive, particularly where competing narratives or values are chronically activated. Empirical tracking of these shifts reveals that learning does not simply converge toward a classical, noncontextual state; rather, it sculpts the geometry of contextuality, making some subspaces more classical while preserving interference in others.
Neuroscientific studies, though still preliminary, provide converging evidence for context-sensitive neural dynamics consistent with a quantum-like description at the computational level, without claiming physical quantum processes in the brain. Electrophysiological and fMRI work has shown that patterns of neural activation representing choices or beliefs depend sharply on task context, prior questions, and framing, even when stimulus content is held constant. Representational similarity analyses reveal that the metric structure of neural state spaces is not fixed: the same options can occupy different relative positions depending on which comparisons are currently salient. This dynamic remapping mirrors the rotation of bases in Hilbert-space models. Moreover, prediction error signals in fronto-striatal circuits are modulated by context in ways that standard reinforcement learning models fail to capture but that align with state-dependent measurement operators acting on structured priors.
Together, these empirical strands—spanning individual judgment, economic choice, moral evaluation, group deliberation, learning, and neural representation—depict a decision-making system pervaded by contextuality. Rather than revealing a collection of unrelated biases, the data suggest a unified pattern: judgments and choices are produced by a cognitive architecture whose states and updates are intrinsically context-dependent. Quantum cognition models, with their emphasis on superposition, interference, and non-commuting measurements, provide a compact, generative account of these findings, turning what appear as violations of classical rationality into lawful consequences of how beliefs and preferences are represented and transformed in the mind.
Implications for models of rationality and belief updating
Recasting rationality within a contextual, quantum cognition framework changes both the normative benchmarks and the descriptive models used to understand belief updating. Classical rational choice theory presumes that agents possess a single, globally coherent probability distribution over states of the world and a stable utility function over outcomes. Under this assumption, principles such as the law of total probability, the sure-thing principle, and dynamic consistency specify how rational agents should update beliefs and preferences in light of new information. Contextuality challenges the feasibility of this picture at the representational level: if cognitive states are inherently basis-dependent, there may be no context-independent joint distribution that captures all observed attitudes and choices simultaneously.
Within a contextual framework, rationality becomes local rather than global. Local coherence requires that, for any given context—defined by a set of questions, options, or frames—the agent’s responses satisfy internal consistency constraints, such as transitivity within that context and adherence to a suitable probability calculus defined on the relevant subspace. However, there is no requirement that these context-specific distributions glue together into a single Kolmogorovian structure. Apparent violations of global consistency, such as preference reversals across frames or order-dependent judgments, then cease to be straightforward indicators of irrationality. Instead, they reveal that the cognitive architecture implements a form of bounded rationality adapted to a world where queries, tasks, and informational structures change rapidly and cannot be anticipated in a single, static model.
This reconceptualization has direct implications for how belief updating is modeled. In classical Bayesianism, updating is governed by Bayes’ rule applied to a fixed hypothesis space with stable priors and likelihoods. In a quantum-like model, the hypothesis space is encoded by a Hilbert space, and belief states are represented by vectors or density operators that change both through unitary-like transformations (internal processing, mental simulation) and through measurement-like projections (responses to questions, observation of outcomes). The update process depends on which operator is applied, not only on the data per se. Two pieces of evidence that are informationally equivalent in a classical sense can produce different updates if they are framed so as to correspond to different measurement operators. Belief updating thus becomes inherently context-sensitive, with history-dependent trajectories that reflect the sequence and structure of cognitive operations.
From the perspective of the bayesian brain hypothesis, this means that neural inference operates with priors that are themselves contextual states rather than static probability tables. The brain approximates Bayesian updating, but the “probabilities” it manipulates are derived from the squared amplitudes of a higher-dimensional state and are always defined relative to a measurement basis determined by current goals, attention, and task demands. Prediction errors then act as signals that not only adjust scalar priors but also reshape the geometry of the state space, altering which dimensions are aligned or orthogonal and how interference among competing narratives unfolds. A prediction that is violated in one context may lead to a different reconfiguration of the cognitive state than an equivalent violation in another context, because the underlying operators that encode the evidence differ.
This geometric view offers a new lens on phenomena traditionally classified as biases in belief updating. Confirmation bias, for instance, can be modeled as a tendency for evidence operators to be aligned with existing eigenstates of belief, minimizing rotations in the state space and preserving interference patterns that favor entrenched narratives. Evidence framed in a dissonant basis, by contrast, induces larger rotations, potentially triggering destructive interference that reduces the effective weight of old commitments. Rather than treating confirmation bias as a sheer failure of Bayesian updating, the contextual model interprets it as a resource-rational strategy: maintaining stability in frequently encountered contexts while limiting costly reconfiguration of deeply entrenched subspaces of belief.
Similarly, phenomena such as belief polarization and motivated reasoning can be understood as outcomes of context-dependent trajectories in the cognitive state space. When individuals with different initial states are exposed to the same sequence of evidence, the operators representing that evidence act on distinct regions of the space, leading to divergent rotations and projections. The same data can therefore increase the separation between states rather than bringing them closer, as classical Bayesian convergence theorems would predict. From a contextual standpoint, this divergence does not necessarily reflect irrational data processing; it can result from lawful interference patterns given the initial geometry of priors and the contextual structure of evidence presentation.
Dynamic consistency—the requirement that plans made now remain preferred as time unfolds and information arrives—is also transformed under contextuality. In classical models, dynamic inconsistency emerges when discount functions are non-exponential or when agents fail to anticipate how future preferences will evolve. In the quantum-like framework, dynamic inconsistency may arise even when discounting is structurally simple, because future decision contexts correspond to different measurement bases than those envisioned at planning time. The state vector evolves between planning and execution, and the operators representing later decisions need not commute with those used in earlier deliberations. An agent may therefore anticipate a certain preference ordering in the abstract but, when the moment of choice arrives under a different contextual configuration, the actual measurement yields a different outcome. Rationality, in this setting, may emphasize flexible meta-policies for managing context shifts rather than strict commitment to time-invariant plans.
These considerations impact normative decision theory. Traditional prescriptions—such as maximizing expected utility under a single prior or under a set of subjective probabilities—implicitly assume that all relevant questions can be embedded in a joint probability space. Contextuality suggests that any normative theory aiming to be psychologically realistic must either relax this assumption or treat violations as inevitable costs of biological and cognitive constraints. One possibility is to define “contextual rationality” as adherence to quantum-like coherence axioms: for each context, probabilities must follow the Born rule with respect to some underlying state, and transitions between contexts must be describable by admissible transformations (unitary-like or completely positive maps). Under this regime, violations of classical axioms such as the law of total probability or the sure-thing principle are not inherently irrational as long as they can be derived from a single, well-structured state space endowed with a family of context-specific observables.
In practical modeling, this means that belief updating rules must explicitly encode measurement dependence. For instance, a decision support system informed by quantum cognition would not merely ask for a user’s prior probability about a policy’s success; it would also model the order, wording, and comparative structure of questions as operators that shape the expressed priors. Recommender systems, forecasting platforms, and deliberative tools could then be designed to manage the sequence of queries in ways that encourage constructive interference among relevant considerations and reduce destructive interference that yields unstable or erratic judgments. Normatively, one might aim to design “context-stabilizing” sequences that lead users toward eigenstates that are robust across nearby contexts, thereby improving calibration without demanding impossible global consistency.
Contextuality also reframes the interpretation of revealed preferences. In classical economics, observed choices are taken as evidence of underlying utilities and priors, assumed to be stable across decision problems. In a contextual framework, each choice reveals only a projection of the underlying state onto the basis defined by the particular task. No single choice or finite set of choices uniquely identifies the full state, because different tasks probe different combinations of latent dimensions. Econometric models that attempt to infer stable utilities and beliefs from behavior may therefore be underdetermined in principle, regardless of data volume. A more appropriate goal is to recover the structure of the state space and the family of relevant measurement operators, identifying which subspaces are relatively stable and which are highly context-sensitive, rather than insisting on a context-free utility function.
Belief updating in social and institutional environments further illustrates the implications of contextual rationality. Institutions such as courts, scientific communities, and markets impose standardized contexts—rules of evidence, protocols of debate, market formats—that effectively fix certain measurement bases. By constraining how information can be presented, what comparisons are permissible, and which outcomes count as legitimate, these institutions reduce the space of allowable operators and thereby stabilize belief updating across individuals. Contextuality is not eliminated but channeled: within the institutional context, local coherence is enhanced, and interference patterns that would otherwise produce idiosyncratic judgments are dampened. Rationality at the institutional level thus involves designing contexts that guide individual neural inference toward predictable, collectively functional equilibria, without requiring that every agent maintain a single context-independent belief system.
At the computational level, the notion that neural inference implements a contextual, amplitude-based representation of uncertainty opens possibilities for new algorithmic models. Instead of approximate Bayesian updating over large classical state spaces, cognitive architectures could be modeled as performing low-dimensional projections in a structured space with interference. From this standpoint, “rational” updating is efficient updating: using context to select a small number of relevant dimensions in which to compute predictions and make choices, while tolerating global inconsistencies across distant contexts as a trade-off for tractability. What appears as irrational framing sensitivity from a classical perspective can then be reinterpreted as a side effect of an optimization criterion that prioritizes computational economy and adaptability over universal coherence.
These insights have consequences for how interventions aimed at improving reasoning and decision-making are conceived. Traditional debiasing strategies often attempt to teach individuals to conform more closely to classical Bayesian and expected-utility norms, encouraging invariance to framing and order. In a contextual framework, more promising strategies might focus on meta-cognitive skills: recognizing when context is likely to induce large rotations in belief space; learning to compare judgments across contexts; and deliberately constructing sequences of self-questioning that counteract unhelpful interference. Training could aim to help individuals approximate a form of second-order rationality, in which they understand their own contextuality and can strategically manage it, rather than striving to achieve an impossible context-free coherence.
The contextual and quantum-like view suggests that belief updating in humans is not merely about adjusting numbers in a probability table but about actively reorganizing the geometry of cognitive representations. Rationality, in this sense, involves maintaining a state space whose structure supports flexible reconfiguration in response to changing tasks, while preserving enough stability to sustain identity, long-term projects, and coordination with others. Contextuality is thus not simply a defect in human reasoning; it is a structural feature that enables the brain to perform prediction and decision-making under severe informational and computational constraints. Models of rationality that ignore this feature risk mischaracterizing both the limitations and the strengths of human belief updating.
