Every act of seeing, hearing, or touching is shaped by expectations that were not chosen in the moment but sedimented through a long history of learning and evolution. These expectations, or priors, rarely operate in isolation. Instead, they form dense networks of mutual constraint: beliefs about the shape of objects are tied to beliefs about lighting, beliefs about other people’s goals are tied to beliefs about social norms, and beliefs about what is likely to happen next are tied to beliefs about what has just happened. In perceptual inference, this tangle of dependencies means that updating one belief silently tugs on many others, even when we are not aware that anything has changed. The result is a form of entanglement among priors, where no single expectation can be adjusted without subtly reshaping the web of expectations that surrounds it.
Within a Bayesian inference framework, perception can be modeled as a process of combining sensory evidence with priors to generate the most probable interpretation of the world. In simple textbook examples, these priors are often depicted as independent: one prior for orientation, another for motion, another for color. But actual experience shows that the brain does not maintain such a clean separation. The prior that “light usually comes from above” is entangled with priors about object convexity and concavity. A shaded pattern on a flat surface is interpreted as a bump or a dent depending on this structured set of expectations, not on an isolated assumption about illumination. Similarly, priors about object continuity over time interact with priors about occlusion, so that a person walking behind a pillar is seen as persisting even when momentarily invisible. Entangled priors thus encode not just frequencies of isolated features, but the lawful ways in which features co-occur in the environment.
Neural coding studies support this picture of interdependent expectations. In early sensory areas, neurons are tuned to basic features such as orientation or spatial frequency, yet their responses are modulated by context in ways that reflect higher-level predictions. When a contour is likely to continue behind an occluder, neurons representing that invisible segment can still show enhanced activity, indicating that the system’s expectation shapes local activity patterns. This influence does not arise from a single prior about edges but from an ensemble of priors about object boundaries, surface continuity, and lighting conditions. Recurrent connections and feedback projections distribute these expectations across multiple levels of the sensory hierarchy, effectively implementing a network of entangled priors that jointly constrain interpretation.
Classic perceptual illusions illustrate how this entanglement can override raw sensory input. In the hollow-mask illusion, a concave mask of a face is often experienced as convex when viewed from a distance. The prior that faces are typically convex is not an isolated rule; it is tightly linked to priors about lighting, shading, and typical viewing conditions. When these linked expectations are activated together, they warp the interpretation of the incoming data in a coordinated way, turning a physically concave object into a compelling experience of convexity. The illusion persists even when observers know that the mask is hollow, showing that the entangled structure of priors can stabilize an interpretation against explicit contrary knowledge.
Social perception offers another domain where entangled priors are evident. Inferring another person’s intention from their actions depends on expectations about bodily movement, typical goals, cultural conventions, and emotional expressions. These expectations mutually inform each other: a subtle shift in facial expression can change the interpreted intention behind an identical movement, because priors about emotion, context, and goal-directed behavior are jointly activated. When such priors become biased—through stereotypes or limited exposure—the entire network of entangled expectations can tilt social perception in systematic ways. Correcting these distortions is difficult precisely because no single belief is wrong in isolation; it is the joint configuration of priors that is misaligned with the actual distribution of social behavior.
Entangled priors are especially important when perception must integrate information across sensory modalities. The brain often resolves conflicts between vision and audition by favoring the modality with higher expected reliability in that context, but this decision is not based on separate, rigid priors for each sense. Instead, it reflects a joint prior over multimodal combinations: what kinds of audiovisual patterns typically come from a single cause. The ventriloquist effect, where a sound is experienced as coming from the location of a moving mouth, depends on an entangled prior that links visual mouth movements and speech sounds as originating from the same source. This expectation reaches into both visual and auditory systems, influencing how ambiguous stimuli in each channel are interpreted and fused.
Another important aspect of entangled priors involves their dependence on time. The brain expects the world to change in lawful ways: objects move smoothly, causes precede effects, and scenes do not flicker into radically different configurations without intermediate steps. Priors about temporal continuity interact with priors about object identity and causal structure. For instance, when an object disappears at one location and a very similar object appears shortly afterward at a nearby location, the system tends to infer that it is the same object moving rather than a new one popping into existence. This inference draws jointly on expectations about motion smoothness, object persistence, and the typical temporal spacing of events, and it is implemented through entangled priors that span both space and time.
Motor predictions also reveal how deeply intertwined priors can be. When planning a movement, the brain recruits expectations about one’s own body dynamics, environmental resistance, and the sensory consequences of action. These priors shape perception even before movement begins; the expected feel of lifting an object or the anticipated path of a reaching motion biases how its weight, distance, and position are experienced. Corollary discharge and efference copy mechanisms send predictions about incoming sensory feedback to sensory regions, attenuating expected reafferent signals and emphasizing unexpected deviations. Because these predictions depend simultaneously on priors about the body, the environment, and sensorimotor contingencies, altering any one component—such as learning to use a tool—reshapes the whole network of entangled expectations that govern both action and perception.
Developmental trajectories further reveal the entangled nature of priors. Infants do not acquire expectations about object permanence, physical support, or agentive behavior as separate modules. Instead, emerging priors about how objects behave under gravity are linked to priors about solidity and continuity, while expectations about agents are tied to patterns of contingent interaction, gaze direction, and goal-directed reach. Violations of one type of expectation often provoke surprise across multiple domains—for instance, an object that passes through a solid barrier challenges priors about solidity, support, and temporal continuity at once. As learning progresses, these initially diffuse expectations crystallize into more sharply tuned, yet still interdependent, priors that guide mature perception.
Pathological alterations in experience can be viewed through the lens of entangled priors as well. In conditions such as schizophrenia, it has been proposed that the balance between prior expectations and sensory evidence is disrupted. If certain priors are weakened while others remain strong, the resulting experience can fragment: irrelevant coincidences are imbued with salience, while stable features of the environment become uncertain. Because priors about self-generated action, bodily signals, and external events are normally intertwined, a disturbance in their coordination can give rise to the feeling that one’s own thoughts or actions are externally controlled. This perspective frames altered experiences not merely as deficits in sensory processing, but as changes in the structure and strength of entangled priors that organize perceptual inference.
Viewed as a whole, perceptual inference operates not on a set of isolated assumptions but on a richly structured fabric of expectations. Each prior is a thread that gains function only in relation to the others, and the brain’s predictive machinery continually weaves and reweaves this fabric as environments, tasks, and bodies change. The entanglement of priors allows the system to exploit regularities that span features, modalities, individuals, and time, generating experiences that are coherent and actionable even in the face of noisy, incomplete, and ambiguous sensory inputs.
Bayesian structure of experience
If perception is understood as a process of probabilistic inference, then experience itself can be seen as the unfolding of a continuously updated posterior distribution over possible worlds. On this view, the mind is not passively illuminated by sensory input; it is actively engaged in bayesian inference, using priors and likelihoods to construct a best-guess model at every moment. What is given in experience is not a raw sample of the world, but the current estimate of what most likely exists and is happening, given both incoming data and a deep reservoir of structured expectations. Qualities such as stability, coherence, and persistence over time are not intrinsic properties of sensations; they emerge from the brain’s ongoing negotiation between top-down prediction and bottom-up evidence.
This bayesian structure becomes clearer when we distinguish between three layers: the generative model that encodes how causes give rise to sensory inputs, the inference process that updates beliefs about causes, and the resulting phenomenal scene that appears to us as “what there is.” The generative model is embodied in neural circuitry: patterns of connectivity, synaptic strengths, and dynamical rules specify which configurations of the world tend to produce which patterns of sensory activity. Inference is realized by the dynamics of neural populations that adjust their activity to minimize prediction error, effectively moving from prior expectations to posterior beliefs. The phenomenal scene is what it is like to occupy a particular point in this high-dimensional belief space. Experience, on this view, is not a transparent mirror of the external world, but a lived cross-section of the current posterior—an internal map that carries the imprint of both the constraints of the environment and the entangled priors of the organism.
Within this framework, entanglement among priors is not a side effect but a core structural feature. The generative model is built around conditional dependencies: objects tend to cast shadows, sounds tend to have sources, actions tend to have goals, and bodies tend to obey approximate physical laws. These dependencies are encoded as links between variables, so that the prior over the entire scene does not factor into independent components but instead forms a joint distribution with rich correlations. As a result, a small update in one part of the model—say, revising the estimate of light direction—propagates through the network, changing expectations about shape, color constancy, and even material properties. The experienced scene shifts accordingly, often without any conscious awareness that a cascade of probabilistic adjustments has taken place beneath the surface.
A key implication is that what we call “objects” and “events” are not primitive inputs but inferred summaries that optimize the trade-off between complexity and predictive power. From a bayesian perspective, an object is a cluster of variables whose joint behavior can be captured by a relatively compact set of priors and likelihoods, yielding powerful predictions across sensory modalities and time. The mind carves the world into such clusters because they support efficient entanglement: priors about an object’s typical trajectory, its likely textures, its characteristic sounds, and its functional affordances can all be linked within a shared generative scaffold. When the posterior assigns high probability to such a cluster, the result in experience is the seeming presence of a distinct thing, persisting and acting in a world of other things.
Temporal structure is woven into this picture at a fundamental level. Priors are not only about what goes with what, but about what tends to follow what. The generative model specifies transition probabilities: how hidden states evolve over time, which sequences of events are likely, and which changes are improbable or even ruled out. Experience of temporal flow can then be viewed as the progressive updating of beliefs about these evolving states, guided by prediction and correction. When you watch a glass tipping over, your experience of it falling, shattering, and scattering shards across the floor is not a series of static images but a smooth trajectory through a space of predictions and error signals. The subjective sense of continuity derives from the fact that the posterior at each moment is heavily constrained by priors about lawful temporal evolution, so that successive experiential snapshots are tightly coupled by shared expectations.
This coupling becomes especially vivid in moments of surprise. Prediction error acts like a spotlight that illuminates the underlying bayesian machinery. When the world behaves in a way that drastically violates priors—an object suddenly vanishes, a voice comes from an impossible direction, or a familiar sequence abruptly breaks—the posterior must adjust rapidly, and the resulting experience is marked by disorientation or shock. Such moments reveal that ordinary, unremarkable experience is characterized by a near-constant cancellation of prediction error: the world seems stable because the generative model is usually good enough that incoming data fall within expected ranges. The calm transparency of everyday perception hides the fact that each instant of experience is a negotiated compromise between the inertia of priors and the corrective force of new evidence.
The same architecture that supports stability also allows for ambiguity and multistability in experience. In ambiguous stimuli like the Necker cube or certain auditory illusions, the sensory likelihood function is consistent with multiple distinct clusters of latent causes, each supported by different constellations of priors. Because the generative model encodes these alternatives as separate high-probability basins in belief space, the posterior can occupy one basin for a while, then jump to another without any change in the stimulus. The resulting alternation in phenomenal content—the cube flipping orientation, the melody reorganizing—reflects the dynamics of bayesian inference under entangled priors. Experience here is not determined once and for all by the sensory input; it is a temporally extended negotiation in which different prior-driven hypotheses take turns dominating the posterior.
Social and conceptual experience exhibit the same structure. When understanding a sentence, for instance, the brain uses a hierarchical generative model that links phonemes to words, words to syntactic structures, and structures to meanings and communicative intentions. Priors about grammar, typical word co-occurrences, and conversational norms are all entangled, so that expectations at higher levels constrain interpretation at lower ones. As each word unfolds in time, the posterior over possible parse trees and meanings shifts, narrowing or broadening based on both the incoming sound and the current web of priors. The felt immediacy of comprehension—“just understanding” what is being said—is the experiential correlate of this rapidly evolving posterior, in which multiple candidate interpretations are implicitly entertained, weighted, and pruned.
This bayesian organization extends to the experience of one’s own body and agency. The generative model includes parameters for limb positions, muscle forces, tactile feedback, and visual appearance, along with priors about typical sensorimotor contingencies. When an action is initiated, predictions are sent forward about the sensory consequences of that action, and the posterior over bodily states is updated as feedback arrives. The sense that “I am moving my arm” corresponds to a posterior in which predicted and actual feedback match closely under a self-generated action hypothesis. When predictions and evidence diverge beyond what the priors can accommodate—due to delays, perturbations, or altered feedback—the posterior shifts, and the sense of agency can weaken or fracture. The structure of bodily and agentive experience is thus tightly linked to how generative models and entangled priors encode the expected coupling between intention, movement, and sensation over time.
Even relatively abstract aspects of experience can be framed in this way. Feelings of familiarity, confidence, and uncertainty arise from properties of the posterior distribution itself. A sharply peaked posterior, in which one hypothesis dominates, tends to be accompanied by a sense of clarity or certainty, while a broad posterior, distributing probability mass across many competing hypotheses, manifests phenomenally as ambiguity or doubt. When the posterior is bimodal or multimodal, the mind may oscillate between distinct experiential contents, or it may sustain a vague, unresolved experience that never fully collapses into a single interpretation. What we introspect as “not being sure” is not a mere commentary layered on top of perception; it is the lived expression of the underlying bayesian geometry of belief.
One consequence of viewing experience as structured by bayesian inference is that it blurs the boundary between perception and cognition. If both are implemented by the same kind of generative modeling and inference machinery, differing mainly in time scale, level of abstraction, and the sensory modalities engaged, then the distinction between “seeing that something is the case” and “thinking that something is the case” becomes a matter of degree. Perceptual experience sits at a point in the hierarchy where priors are tightly constrained by long-term statistics of the environment and by the immediate sensorium, while more cognitive states occupy regions where priors are shaped by language, symbolic reasoning, and social communication. Yet both are posterior beliefs in a shared probabilistic architecture, and their interaction—how expectations informed by deliberation can reshape what is seen or heard—emerges naturally from the entanglement of priors across levels.
This probabilistic perspective also changes how spatial organization in experience can be understood. The apparent layout of a scene—the arrangement of surfaces, distances, and occlusions—is not merely read off from depth cues; it is a solution to a high-dimensional optimization problem, constrained by priors over three-dimensional structure, illumination, material properties, and typical viewing conditions. The visual field appears as a continuous, metrically structured space because the generative model favors world configurations that can be embedded in such a space and that support coherent prediction across eye movements and head turns. Spatial experience thus inherits its geometry from the constraints of a model designed to generate stable, lawful sensory patterns, rather than from any direct access to an independently existing Euclidean arena.
Undergirding all of this is the notion that experience at any moment is shaped not only by current sensory input but by a thick temporal context encoded in the brain’s evolving priors. What has just happened leaves traces in the form of adapted neural responses, updated transition probabilities, and altered expectations about what is likely to come next. These traces act as short-lived priors that bias interpretation in the immediate future, while more slowly changing structural priors encode enduring regularities of the environment and the organism’s own body. The flow of experience over time therefore reflects a layered interplay between fast, context-sensitive adjustments and slower, deeply entrenched assumptions, all of which are entangled within a single bayesian framework that strives to keep prediction error within manageable bounds.
Geometry of representational spaces
To understand how entangled priors shape experience, it helps to think in terms of geometry rather than mere lists of variables. A generative model can be seen as defining a high-dimensional space of possible latent causes, where each point corresponds to a particular configuration of the world as the system takes it to be. Priors, likelihoods, and posteriors then define structures within this space: regions of high probability, valleys and ridges of prediction error, and trajectories along which beliefs tend to move over time. Experience at any moment corresponds to occupying one of these regions under the constraints imposed by neural coding, bodily dynamics, and environmental statistics.
In this geometric picture, entangled priors manifest as curvature and correlation structure. If priors over two features are independent, the corresponding dimensions of the representational space are orthogonal and can be varied without affecting each other. But when priors are entangled—when, for instance, certain shapes typically co-occur with certain textures or motions—the high-probability regions become elongated, bent, or twisted submanifolds rather than simple axis-aligned boxes. Moving along such a manifold changes several feature dimensions together, reflecting the fact that the system treats those feature combinations as natural and others as unlikely or even impossible. Perception thus tends to settle into these curved ridges of probability, and the felt regularities of the world mirror the geometry of these entangled structures.
Neural population activity offers a concrete realization of this geometry. At any instant, activity across many neurons can be treated as a point in a high-dimensional firing-rate space. Over time, perception and action correspond to trajectories through this space, carved out by the dynamics of recurrent connectivity and ongoing prediction-error minimization. Patterns that recur across contexts—such as the recognition of a particular object or phoneme—appear as attractors or low-dimensional subspaces, where many different inputs funnel into similar population states. Entangled priors are encoded in the way these attractors are arranged and connected: nearby or overlapping regions reflect expectations that certain configurations share causes, while distant or separated regions encode sharp distinctions that the system maintains between categories, scenes, or bodily states.
Recurrent and hierarchical circuitry further shape this representational geometry. Feedback connections carry top-down predictions that bias the trajectory of neural activity toward regions consistent with prior expectations, while feedforward inputs push activity toward configurations that better match the current sensory evidence. The compromise between these forces defines not just which region of state space is occupied, but also how the space itself is effectively warped. Where priors are strong and precise, the state space becomes sharply channeled, with narrow valleys that guide activity along stereotyped paths. Where priors are weak or diffuse, the space is flatter and more open, allowing for wider excursions and more variable experiences under similar inputs. This dynamic shaping of the geometry is an expression of ongoing bayesian inference realized in the physics of neural networks.
One way to visualize this is to imagine a low-dimensional projection of the representational space—say, a two-dimensional map that captures the most important modes of variation relevant to a task. Within this map, different perceptual hypotheses appear as basins of attraction. Ambiguous stimuli lie near the boundaries between basins, so that small perturbations in input or internal noise can nudge the system from one basin to another, yielding bistable or multistable experience. The fact that such switches are not uniform but biased toward certain sequences or dwell times indicates that the underlying geometry is not symmetric: entangled priors sculpt the energy landscape so that some transitions are easier than others, and some sequences of interpretations are more natural than others even when the stimulus remains fixed.
Crucially, this geometry is not restricted to low-level sensory features; it spans multiple hierarchical levels. At higher levels, variables encode abstract properties such as object identity, causal structure, or social intention. These higher-level variables impose additional constraints on the geometry of lower-level representations, effectively bundling diverse sensory features into coherent trajectories. When you recognize a face or a familiar environment, the system rapidly moves into a high-level attractor corresponding to that identity or place. This attractor then shapes the lower-level representational space, tightening or relaxing couplings among colors, contours, sounds, and bodily responses in a way that reflects the entrenched priors associated with that high-level state. The subjective impression of a stable, unified object or situation arises from traversing a coordinated path through this multi-layered space.
Spatial experience itself can be treated as emerging from a specialized geometry of representational states. Neural systems supporting navigation, such as place cells and grid cells, exemplify a mapping between physical layout and internal geometry: different locations in the environment correspond to distinct patterns of activity, and transitions between locations map to continuous trajectories in neural space. Yet this mapping is not a simple Euclidean copy of external space; it is distorted by the organism’s movement statistics, goals, and constraints. Frequently traversed paths are represented with higher resolution and stronger connectivity, while rarely visited regions may be compressed or fragmented. The entanglement of priors about typical routes, landmarks, and affordances effectively reshapes the internal spatial manifold, and the felt ease or difficulty of imagining, planning, or remembering certain movements reflects this non-uniform geometry.
Representational spaces also encode temporal structure. States are not only arranged according to similarity at a given moment, but also linked by transition probabilities that specify which states tend to follow which others. This gives rise to a kind of spatiotemporal manifold, where paths correspond to expected event sequences. When you anticipate the continuation of a melody or the next word in a sentence, you are implicitly moving along well-trodden paths in this manifold, guided by priors about temporal regularities. Entanglement across time means that certain present configurations pull strongly toward specific future ones, while others open onto a wide range of possibilities. The asymmetry between past and future in experience is encoded in this directed structure of connections, where prediction flows along particular channels that have been carved by learning and embodied interaction.
At finer scales, the geometry of representational spaces can be characterized in terms of manifolds corresponding to categories, concepts, or skill repertoires. For instance, images of a particular object under varying viewpoints, lighting, and occlusion tend to lie on a low-dimensional manifold embedded in the high-dimensional space of visual inputs. Learning to recognize that object can be understood as shaping neural representations so that these varied inputs are mapped into a compact, smooth region in neural state space. Entangled priors ensure that this manifold respects not just visual similarity, but also functional and contextual regularities: cups that can be grasped and used for drinking will be grouped together despite surface differences, while visually similar but functionally distinct items occupy separate manifolds. Perceptual invariance is thus tied to the curvature and connectivity of these learned subspaces.
The same logic applies to motor control. Possible postures and movements of the body form a vast configuration space that is heavily constrained by biomechanics, joint limits, and typical task demands. Motor cortex and associated regions appear to encode low-dimensional manifolds within this space, corresponding to coordinated synergies and commonly used trajectories. Entangled priors about the body and environment shape these manifolds so that certain combinations of muscle activations are easy to generate and control, while others are effectively inaccessible except under special training or pathological conditions. The felt naturalness or awkwardness of a movement, and the fluency with which a skill can be executed, reflect how the trajectory required by the task aligns or conflicts with the pre-existing geometry of motor representations.
Representational geometry also clarifies how context can radically reconfigure experience without large changes in sensory input. Context can be viewed as selecting or modulating a subspace within the overall representational manifold, effectively rotating or stretching the space so that different dimensions become salient. In a noisy crowd, priors about a conversation partner’s voice define a subspace in which that voice is amplified and separated from the background. In a demanding visual search task, task-relevant features such as color or orientation define axes along which distances are magnified, making small differences easier to detect. Neural coding studies show that attention and task demands can reshape population codes so that discriminability increases along behaviorally important directions, another expression of how priors and goals sculpt the geometry in which perception unfolds.
Even subtle experiential qualities such as similarity and difference can be grounded in this framework. Psychological similarity between two stimuli or concepts corresponds to the distance between their neural representations along behaviorally relevant dimensions. Entangled priors determine which dimensions matter for similarity judgments: objects that are functionally interchangeable may be represented as near each other even if they differ visually, while visually similar objects with different roles may be separated. As learning proceeds, the representational space can be re-metrized so that distinctions that matter more for prediction and action are stretched, and irrelevant variations are compressed. Shifts in judgment, categorization, and intuitive generalization thus trace back to smooth yet far-reaching deformations in the underlying geometry.
This geometric viewpoint naturally accommodates uncertainty and graded belief. A posterior distribution over latent causes can be represented as a cloud or region in state space rather than a single point. Narrow, sharply peaked distributions correspond to tight clusters in this space, while broad or multimodal distributions occupy larger or disjoint regions. The subjective sense of confidence or ambiguity aligns with the spread and structure of these regions: when belief is concentrated in a small neighborhood, experience feels determinate; when it is dispersed or split across distant basins, experience is correspondingly vague or unstable. Entanglement ensures that uncertainty in one dimension ripples through others, expanding or contracting the effective volume of the experiential manifold in coordinated ways.
The geometry of representational spaces is not static. It is continually reshaped by learning, adaptation, and ongoing experience. Each episode of prediction and correction slightly adjusts synaptic weights, thereby altering distances, angles, and curvatures in the high-dimensional space. Regularities that are repeatedly encountered carve deep channels that guide future trajectories, while rarely used pathways fade or become harder to access. As a result, the long-term history of interaction with the world is inscribed in the very shape of the space within which new experiences are realized. Entangled priors thus appear not only as statistical constraints but as geometric facts about how representations are arranged, connected, and allowed to move, grounding the structure of experience in the evolving topology of neural state space.
Learning, plasticity, and shifting priors
Learning reshapes the web of expectations that underwrites perception, but it does so in a way that preserves and often deepens entanglement among priors. Each new experience is not merely a data point that updates a single belief; it is an event that reverberates across a network of conditional dependencies. When a child learns that certain animals are dangerous, for instance, the update does not remain confined to a label for that animal. It reshapes expectations about typical environments in which the animal appears, characteristic sounds and movements, likely outcomes of different actions, and appropriate emotional and bodily responses. The new information inserts itself into an existing lattice of priors, bending multiple strands at once rather than attaching as an isolated fact.
From the perspective of bayesian inference, this process involves adjusting both the parameters and the structure of the generative model. Parameter learning refines the numerical values that govern how likely certain patterns are—how heavy a typical mug feels, how often people smile when they are content. Structural learning goes further, modifying which variables are linked at all—discovering, for example, that a certain tone of voice reliably predicts sarcasm, or that particular colors tend to signal ripeness. Because these changes occur in a model whose variables are already interconnected, any local update propagates to other expectations that share those variables. Learning that a particular person is trustworthy does not only affect priors about their future behavior; it also shifts priors about the reliability of testimony in general, the weight given to others’ recommendations, and even the perceived riskiness of environments associated with that person.
Neurobiologically, plasticity mechanisms provide the substrate for this continual reshaping. Synaptic strengths change through experience-dependent processes such as Hebbian learning, spike-timing-dependent plasticity, and neuromodulator-gated adaptations. These mechanisms operate in networks rich with recurrent and feedback connections, meaning that when a synapse is strengthened in one pathway, the pattern of activity it supports interacts with many others. A visual neuron that strengthens its response to a certain contour in the presence of a particular motion signal is not simply encoding a new standalone feature; it is embodying an updated joint prior over shape and motion. Over time, many such adjustments warp the effective geometry of neural state space, tightening some correlations, loosening others, and thereby altering the texture of future experience.
Experience-dependent plasticity is especially prominent during sensitive and critical periods, when the brain is unusually malleable to particular classes of regularities. Early visual deprivation, for example, can permanently alter orientation-selective maps and depth perception. But what is modified is not just the sensitivity to light or contrast; it is the entangled configuration of priors about contours, surfaces, and three-dimensional layout. A child who does not receive typical binocular input does not simply lack a specific depth cue; their entire ensemble of priors about spatial arrangement, object boundaries, and motion parallax becomes differently organized. Later learning can compensate to some extent, but it must operate on a neural architecture that has already consolidated an atypical set of entangled expectations.
Learning in adulthood reveals that priors remain pliable, though often in more domain-specific and context-dependent ways. Acquiring a new skill—such as playing a musical instrument, speaking a foreign language, or navigating a novel city—requires building new manifolds in representational space and weaving them into the existing fabric. When learning to play the violin, for instance, the system must develop precise priors linking finger positions, bow pressure, and the resulting timbre and pitch. These sensorimotor contingencies soon become entangled with emotional and social priors: expectations about audience response, typical performance contexts, and the internal feeling-tone of successful execution. The novice’s early, effortful attempts correspond to a posterior that ranges widely across possibilities; with practice, the posterior sharpens along narrow channels that embody well-learned couplings, yielding the subjective sense of fluency and automaticity.
Habit formation illustrates how repeated prediction and confirmation can crystallize priors into rigid patterns of experience. When a behavior reliably leads to a particular outcome, and that outcome consistently reduces prediction error, the system converges on a compact set of expectations that bypasses more exploratory inference. Over time, contextual nuances that originally modulated the behavior may be suppressed or ignored, as the prior for the habitual response grows strong enough to dominate moderate discrepancies in sensory input. This rigidity is not simply a matter of a single overlearned association; it reflects a broad entanglement across cues, actions, and consequences that narrows the experiential pathways available in a given situation. Breaking a habit, therefore, is not just about inhibiting a single response; it requires restructuring a network of priors that define what feels natural or inevitable in that context.
Conversely, certain learning regimes deliberately destabilize entrenched priors to make room for new ones. Exposure to diverse and conflicting examples can prevent premature closure of category boundaries, keeping representational manifolds more flexible. In perceptual learning paradigms, where participants learn to discriminate barely distinguishable stimuli, improvements do not only sharpen a single sensory dimension; they reweight the contribution of multiple cues and adjust correlations between them. Training to distinguish phonetically similar speech sounds in a new language, for instance, requires not only tuning the auditory system to fine-grained acoustic differences, but also weakening previously dominant priors that lump those sounds into a single category. The resulting entanglement differs from that of native speakers, which can manifest in persistent accent or subtle differences in comprehension under noise, even after years of practice.
An important driver of plasticity is the structure of prediction errors over time. When discrepancies between expected and observed inputs are small, they can be absorbed by minor parameter tweaks within the existing model. When errors are large, persistent, and systematically patterned, they exert pressure for more radical reorganization. This distinction mirrors the difference between fine-tuning an entangled network of priors and revising the network’s architecture. For example, encountering occasional exceptions to a social rule may lead to slightly more diffuse expectations about that rule’s applicability, whereas sustained immersion in a different culture can prompt the formation of qualitatively new priors about politeness, hierarchy, and interpersonal distance. These new priors then entangle with older ones about trust, risk, and cooperation, sometimes generating internal tension as competing networks vie for dominance in borderline situations.
Plasticity can also operate at the level of meta-priors: expectations about how stable or revisable other priors are. Some individuals and developmental phases are characterized by strong meta-priors favoring stability: the assumption that the world’s regularities, once learned, will rarely change. In such cases, surprising experiences may be explained away as noise or anomalies, leaving deep generative structures untouched. Other individuals or contexts cultivate meta-priors that favor change: the expectation that categories are fluid, norms are negotiable, and causal relations may shift. Here, similar surprises more readily trigger restructuring of the underlying model. This higher-order entanglement means that learning rates and degrees of openness to evidence are themselves learned and can vary across domains, subtly steering the trajectory of perceptual and conceptual development.
Emotion and motivation modulate plasticity by selectively amplifying or dampening certain prediction errors. Neuromodulatory systems such as dopamine, norepinephrine, and serotonin adjust synaptic learning rules in a context-dependent way, effectively assigning different weights to different experiences. Rewarding or threatening events are tagged as particularly informative, encouraging updates to priors that govern similar future situations. Because emotional contexts are themselves defined by broad constellations of cues and internal states, the resulting changes tend to be widely entangled. A single traumatic episode, for instance, can reshape priors about safety, self-efficacy, bodily vulnerability, and the intentions of others, not as isolated beliefs but as a coordinated shift in the entire landscape of expectations. Subsequent experiences are then filtered through this altered landscape, often reinforcing the new configuration unless explicitly counteracted.
Attentional control plays a complementary role by determining which parts of the input are sampled and thus which prediction errors are even available to drive learning. Attention can be seen as a policy for selecting data, guided by existing priors about what matters. If attention is habitually drawn to confirming cues and away from disconfirming ones, the system may rarely experience errors large enough to challenge its current model. This self-reinforcing loop deepens entanglement along familiar pathways while leaving alternative manifolds underdeveloped. In contrast, deliberate efforts to widen attentional scope or to seek out disconfirming evidence expose the system to a richer pattern of discrepancies, enabling more balanced updates. Over time, these policies themselves become entrenched as priors about where informative signals are likely to be found.
Sensorimotor adaptation experiments provide a controlled window into how quickly and flexibly priors can shift. When participants wear prism glasses that displace the visual field, their reaching movements initially miss the target, producing large prediction errors between expected and actual hand positions. With practice, the motor system retunes, adjusting priors about the mapping between visual and proprioceptive signals so that accurate reaches are restored. Once the prisms are removed, a reverse aftereffect is observed, revealing that the newly entangled priors have temporarily become dominant. These dynamics illustrate that learning operates not only on isolated mappings but on joint distributions over multiple modalities and time, rapidly reconfiguring how the body and environment are jointly represented.
Long-term expertise shows how extensive training can sculpt highly specialized yet robust configurations of priors. A radiologist, for instance, acquires refined expectations about the statistical structure of medical images: typical textures, shapes, and contrasts that signal pathology versus normal variation. These priors are deeply entangled with conceptual knowledge about disease processes, patient histories, and technological artifacts. As a result, the expert’s experience of a scan is not that of ambiguous grayscale patches, but of a richly organized scene where certain regions “pop out” as suspicious and others fade into the background of normalcy. This transformed phenomenology reflects millions of micro-updates to the generative model, each subtly altering how features co-occur, how strongly they are linked to potential diagnoses, and how they unfold over the temporal course of reading a case.
Maladaptive learning illustrates that plasticity does not inherently lead toward more accurate or beneficial priors. In chronic pain, for example, repeated nociceptive input can entrench expectations that certain movements or contexts will be painful, even after tissue damage has healed. Sensory and affective systems become tuned so that innocuous stimuli are interpreted through a high-gain prior of threat, and neural circuits amplify prediction-consistent patterns while discounting counterevidence. The result is a self-sustaining entanglement among bodily sensations, motor avoidance, and negative affect that stabilizes a distorted experiential world. Similar mechanisms may underlie anxiety disorders, addiction, and certain forms of bias, where learning has carved deep channels that keep posterior beliefs—and thus lived experience—circulating within a narrow, maladaptive portion of the broader hypothesis space.
Rehabilitation, psychotherapy, and educational interventions can be understood as attempts to deliberately reshape entangled priors by engineering sequences of prediction and feedback. Graduated exposure therapies, for instance, systematically confront a person with situations that have been deemed threatening, under conditions designed to violate their strongest fear-related expectations. Repeated experiences of safety in these contexts generate large but structured prediction errors, which, if integrated, can loosen the coupling between previously linked cues and catastrophic outcomes. Crucially, what changes is not only a single fear association but the broader network of priors about controllability, self-competence, and the generalizability of risk. Effective interventions thus recognize that learning targets an interconnected system, not a single node.
Across development and throughout life, the shifting of priors is therefore not a process of erasing and rewriting isolated entries in a mental ledger. It is a continuous, path-dependent reconfiguration of a high-dimensional, entangled structure. Each new episode of prediction and correction slightly alters the curvature of representational manifolds, the depth of attractor basins, and the connectivity between them. Some regions become smoother and more integrated, supporting rapid, confident inferences; others remain rough, fragmented, or sparsely explored. Time, in this sense, is not just an external dimension along which experiences are laid out; it is the medium through which the very shape of the experiential state space is sculpted. Learning and plasticity are the processes by which that sculpture is ongoingly revised, ensuring that the priors that quietly govern perception remain, for better or worse, attuned to the patterns of a changing world.
Implications for consciousness and phenomenology
Thinking about consciousness in terms of entangled priors reframes what it means for an experience to be “mine” and to unfold the way it does. If moment-to-moment phenomenology is identified with the current posterior of a deeply structured generative model, then the distinctive feel of a conscious state reflects which parts of that model are currently engaged, how strongly priors are constraining interpretation, and how widely prediction errors are being broadcast. The unity of consciousness—the way a multitude of features, thoughts, and bodily sensations appear as aspects of a single scene—emerges from the fact that these elements are not inferred independently. They are bound together by shared causes and co-modulated by entangled priors, so that updating one part of the model has coordinated consequences elsewhere. The resulting global coherence is not imposed from outside experience; it is the lived manifestation of a system whose internal dependencies ensure that its best-guess world hangs together.
This view suggests that the so-called “binding problem” is not primarily a matter of synchronizing separate feature maps, but of maintaining a joint posterior over latent causes that explains away diverse sensory inputs. When the system infers that a particular cluster of hidden variables underlies a pattern of color, motion, sound, and touch, these features are experienced as belonging to a single object or event. The sense that “this is one thing I am aware of” corresponds to occupying a region of belief space where those features share a common explanation under strongly entangled priors. Failures of binding—such as illusory conjunctions or certain neuropsychological syndromes—can be interpreted as cases where the joint distribution fragmentizes, so that features that should be tightly coupled are instead assigned to separate causes, or vice versa. Consciousness here tracks not raw data, but the structural organization of inference.
The entanglement of priors also reshapes questions about the access we have to our own mental states. Introspection does not reveal the machinery of bayesian inference; it presents its outputs in a format that is sparse, categorical, and often overconfident. Yet the confidence or doubt that accompanies a percept, memory, or decision is not arbitrary. It reflects properties of the posterior distribution such as its precision, stability under small perturbations, and sensitivity to additional evidence. When experience feels vivid and certain, the underlying posterior is sharply peaked and strongly constrained by entrenched priors. When it feels dim, ambiguous, or “as if it could be otherwise,” the posterior is broader or multimodal, and prediction errors are tugging at several competing hypotheses. Phenomenology thus contains a metacognitive aspect built into it: the felt “how sure this seems” is the experiential shadow of the geometry and spread of belief.
Temporal aspects of consciousness can likewise be reinterpreted in terms of prediction and updating. The sense that experience flows, rather than consisting of isolated snapshots, arises from the way priors about time structure how incoming data are integrated. The generative model encodes not only what configurations are likely, but how they tend to evolve; the posterior at any instant already embeds expectations about the near future and traces of just-updated beliefs about the immediate past. Phenomenologically, this yields a “specious present” in which what is consciously given includes anticipations and retentions as well as raw sensations. The fact that we can be startled by an event that completes a pattern—such as a punchline, a sudden chord resolution, or an expected impact—shows that consciousness is continuously leaning into the future, with violations of temporal priors registering as particularly salient alterations of the experiential stream.
Subjective time can speed up, slow down, or fracture when the patterning of predictions changes. In highly predictable situations—routine tasks, monotonous environments—priors capture most of what will occur, and only small, easily absorbed prediction errors arise. Reports that “time flew by” can be understood as reflecting a regime in which little new structure was added to the generative model; the trajectory through representational space followed a well-worn path with minimal branching. In contrast, during novel, emotionally charged, or threatening episodes, large, structured errors demand extensive updating of priors. The richness and density of these updates correspond to an expanded record in memory and a retrospective sense that time was elongated. The phenomenology of duration is thus linked to the rate at which entangled priors are being revised by unexpected information.
Diversity in conscious experience across individuals and cultures can be framed as diversity in the learned structure of entangled priors. People inhabiting different social worlds, speaking different languages, or practicing different skills occupy distinct regions of representational space that have been curved and channeled by their histories. These differences are not confined to explicit beliefs; they extend into what feels salient, what categories seem natural, what emotional tones attach to particular cues, and what possibilities spontaneously present themselves in imagination. From this standpoint, there is no universal “view from nowhere”; every stream of consciousness is a locally coherent exploration of a world carved out by particular priors. Mutual understanding requires, at least in part, learning enough about another’s generative model to approximate how their entanglements make certain experiences feel obvious, threatening, sacred, or invisible.
Emotion itself can be seen as an aspect of phenomenology that expresses how prediction and control are faring relative to entrenched expectations. When outcomes match priors in domains that matter for survival or social standing, affect tends to be neutral or mildly positive; the system is operating within its comfort zone, and prediction errors are low. When events systematically overshoot or undershoot expectations—when rewards are unexpectedly large, threats unexpectedly absent, or losses unexpectedly severe—neuromodulatory systems signal these mismatches in ways that reshape both learning and conscious feeling. Joy, relief, disappointment, and dread are not mere overlays on a neutral picture of the world; they are signals about how successfully the entangled network of priors is maintaining grip on a volatile environment. The felt “valence” of an experience is deeply tied to whether prediction error is being reduced or amplified in domains weighted as important.
On this picture, disorders of consciousness and altered states can be interpreted as large-scale reorganizations or breakdowns in the normal pattern of entangled inference. Under anesthesia or deep sleep, feedback connectivity and long-range synchronization are reduced, limiting the integration of priors across cortical and subcortical regions. The system still performs local predictions, but these do not coalesce into a globally coherent posterior that would support rich, reportable experience. In psychedelic states, by contrast, certain high-level priors—particularly those that stabilize self-representation and ordinary causal structure—may be weakened or rendered more flexible. This loosening allows bottom-up signals and normally suppressed hypotheses to influence the posterior more strongly, generating vivid, labile experiences that feel both highly meaningful and difficult to integrate with everyday categories. The reported dissolution of ego boundaries, synesthetic blending of modalities, and altered sense of time can be seen as phenomenological signatures of an inference system operating under partially relaxed constraints.
The self, in this framework, is not a separate substance but a special configuration of entangled priors that organizes bodily, interoceptive, and social information. Expectations about where the body is, what it can do, how it typically feels, and how others will respond to it are woven into a dense web that stabilizes a center of gravity for perception and action. The sense of ownership—“this is happening to me”—arises when incoming signals are successfully explained by this self-model: predicted and observed interoceptive, proprioceptive, and exteroceptive cues align under a hypothesis that attributes them to a single persisting agent. When this alignment is perturbed, as in out-of-body experiences, depersonalization, or certain illusions of body ownership, the boundaries of the self can shift or blur. Consciousness, on this view, always occurs from some standpoint defined by an active self-model, but the exact shape and rigidity of that model depend on how tightly its priors are entangled with the rest of the generative structure.
This standpoint-dependence connects to phenomenological traditions that emphasize intentionality—the aboutness of experience. If every conscious state corresponds to a posterior over possible worlds as oriented by a self-model, then experience is inherently relational: it presents a world as it appears to someone with particular capacities, needs, and histories. The world is not simply there; it shows up as affordant, inviting, threatening, or irrelevant depending on how it fits into the agent’s ongoing projects encoded in their priors. Entanglement ensures that these projects are not localized; expectations about what matters suffuse perception at every level, affecting which features stand out, which patterns are recognized, and which possibilities feel live. Phenomenology thus cannot be reduced to a catalog of neutral qualia; it is a structured field of significance shaped by predictive engagement with a world of potential action.
Because entangled priors reach across sensory, cognitive, emotional, and motor domains, any attempt to understand or modify consciousness must take into account their systemic character. Efforts to cultivate particular forms of awareness—through meditation, contemplative practice, or cognitive training—can be interpreted as systematic attempts to adjust the balance between priors and sensory evidence, to alter meta-priors about the stability of beliefs, or to attenuate habitual chains of prediction that ordinarily dominate experience. Practices that emphasize sustained, nonreactive attention to present-moment sensation, for example, may gradually weaken coarse, evaluative priors that instantly categorize stimuli as good or bad, relevant or irrelevant. As these fast, affect-laden predictions lose some of their grip, finer-grained aspects of experience become accessible, and the felt structure of the experiential field shifts. The resulting reports of spaciousness, clarity, or reduced identification with passing thoughts can thus be seen as phenomenological descriptions of an inference system in which certain entanglements have been loosened and others strengthened.
Conceiving of consciousness in terms of bayesian inference and neural coding has implications for debates about its physical basis. Rather than seeking a single “consciousness center” or a privileged neural correlate, this perspective points toward patterns of large-scale integration and complexity: regimes in which many specialized subsystems contribute to a shared posterior in a way that is both globally coherent and locally differentiated. Entangled priors are central to such regimes because they specify which combinations of activity across areas count as plausible world-states and which do not. When these priors are strong and well-calibrated, the system can generate a rich, stable experiential world that is nonetheless sensitive to surprise. The physical signatures of such functioning—particular forms of connectivity, rhythmic coordination, and information flow—do not explain consciousness apart from inference; they are the dynamically realized scaffold that allows an organism to sustain, from moment to moment, the structured, world-involving phenomenology that we call being conscious.
