To treat conscious experience as probabilistic inference is to see each moment of awareness as the brain’s best guess about the hidden causes of sensory input. Rather than passively recording the world, the nervous system is constantly engaged in a process of hypothesis testing, where internal models compete to explain what is happening both outside and inside the organism. On this view, the conscious now is not a simple readout of raw data; it is the most probable interpretation, given the brain’s current model of the world and the noisy, partial signals arriving from the senses.
This perspective draws on the idea of the bayesian brain. In Bayesian terms, the brain maintains prior beliefs—priors—about how the world is structured: what objects exist, how they persist, which events tend to follow others, and what bodily states are likely in different situations. Incoming sensory data function as evidence that updates these priors into posterior beliefs. Moment-to-moment experience reflects those posterior beliefs, not the evidence alone. Perception, in this sense, is a form of ongoing Bayesian inference, where the content of consciousness tracks the most probable causes of sensations rather than their exact physical details.
Because the brain’s sensory channels are noisy and delayed, the system cannot directly access the world as it is at any given instant. The signals from light, sound, touch, and the body arrive with differing latencies and uncertainties. To reconcile these delays, the brain engages in temporal integration, combining information over short windows of time to infer what is happening “now.” During this process, predictions about how the world is likely to evolve fill in gaps and smooth discontinuities in the raw input. The resulting experiential moment is therefore an inference over a temporally extended set of signals, compressed into what feels like a unified present.
Probabilistic inference is especially evident in cases where sensory data are ambiguous or incomplete. The brain uses its priors to resolve competing interpretations: a shadow becomes either a dent or a bump; an indistinct sound becomes speech in a familiar language; a vague bodily sensation becomes either hunger, anxiety, or illness depending on context. These interpretations do not arise from a single bottom-up pathway but from a negotiation between top-down predictions and bottom-up evidence. The content that wins this negotiation is what becomes consciously experienced, while alternative possibilities are suppressed or remain in the background of awareness as vague feelings of uncertainty.
In this framework, prediction is not an optional extra added on top of perception; it is the primary mechanism by which perception operates. The brain constantly projects forward what it expects to encounter—visual contours, familiar voices, bodily feedback from movement—and then compares those expectations with actual input. Deviations from expectation, or prediction errors, are used to adjust the internal model. Conscious experience can be seen as the running summary of these model-based predictions constrained by the prediction errors that cannot be ignored. When the world conforms to expectations, awareness may feel smooth and continuous. When prediction errors are large or surprising, experience feels vivid, disjointed, or even disruptive, as when an unexpected loud sound abruptly captures attention.
Because probabilistic inference is hierarchical, the brain does not maintain just one level of prediction but many, ranging from low-level expectations about edges and tones to high-level beliefs about people, intentions, and abstract concepts. Each level passes predictions downward and receives prediction errors upward. Conscious experience seems to be most tightly linked to intermediate and higher levels of this hierarchy, where perceptual hypotheses are relatively stable yet still sensitive to context and meaning. What one consciously experiences as a coherent scene or narrative emerges from the interaction of these layered predictions, rather than from the raw flux of sensory data.
Uncertainty plays a central role in shaping what reaches awareness. In a Bayesian system, beliefs are not only about what is likely to be true but also about how confident the system should be. The brain estimates the precision of both its priors and its incoming signals. When priors are assigned high precision relative to the data, experience is dominated by expectation: illusions become more robust, habits of interpretation more entrenched, and new evidence more easily discounted. When sensory evidence is given high precision, perception becomes more data-driven, and surprising stimuli can rapidly reshape what is experienced. Conscious experience thus reflects a dynamic balance between prior-driven prediction and evidence-driven correction, modulated by estimates of precision at each moment.
Seeing experience as probabilistic inference also clarifies why consciousness is both rich and selective. At any time, the brain entertains many possible interpretations of the same sensory stream, but only a subset achieves high posterior probability given the current priors and evidence. These “winning” interpretations form the content of awareness. Others remain latent as alternative hypotheses, influencing behavior and emotional tone without entering full consciousness. The apparent unity of the conscious now masks an underlying competition among probabilistic models, each striving to best account for the data within the constraints of the system’s prior structure.
This inferential view naturally extends to interoception, the perception of internal bodily states. Signals from the heart, lungs, gut, and immune system are likewise noisy and ambiguous. The brain infers their causes—stress, safety, fatigue, threat—using priors shaped by past bodily states and social learning. The resulting feelings, from calmness to anxiety, can be understood as probabilistic judgments about the organism’s internal conditions and prospects. Conscious affect and mood therefore arise as interpretations of bodily evidence, constrained by expectations about what such evidence typically means, rather than as direct, transparent readouts of physiological status.
By construing each moment of awareness as a probabilistic best guess, this approach treats the conscious now as a bridge between what has been sensed recently and what is expected to unfold next. The brain uses its recent history of inputs, weighted by their estimated reliability, to update its model and generate predictions that shape the very experience of the present. In this way, consciousness does not merely reside at an instant; it sits at the evolving boundary where prior beliefs, current evidence, and anticipated futures are continuously reconciled through Bayesian inference.
Temporal binding and the construction of the present moment
The experience of a single, unified moment depends on the way the nervous system binds together events that, in physical time, are slightly misaligned. Light from a flash reaches the visual cortex later than the vibrations from the associated sound reach the auditory cortex, and both arrive later than signals from the skin or muscles. Yet in experience, the flash, the bang, and the bodily jolt can appear to occur simultaneously. To construct this apparent simultaneity, the bayesian brain performs a form of temporal integration, pooling sensory evidence over a short time window and assigning it to a common cause when doing so yields a more probable overall interpretation of the situation.
This temporal binding is not a mere averaging of arrival times but a probabilistic inference about when an external event most likely occurred. The system relies on learned priors about the typical delays between different sensory modalities—how long it usually takes for sound to follow light at various distances, how quickly bodily feedback follows voluntary movement, how internal signals lag behind external events. Given these priors and the noisy, delayed input streams, the brain estimates a best-fitting global timeline on which events are arranged. The conscious now is experienced at that inferred point in the constructed timeline, not at the raw, physical arrival time of any single signal.
Laboratory phenomena such as the flash-lag effect illustrate how this inferential process reshapes temporal experience. When a moving object and a briefly flashed object are physically aligned, the moving one is often seen as being ahead. A straightforward recording system would not generate such a distortion. But if the perceptual system is constantly generating prediction-based estimates of where objects will be a fraction of a second in the future, then the moving object’s perceived position can be pulled forward in time. What is experienced as present location partly reflects an extrapolation beyond the last available sensory sample. The conscious now, on this view, is a bridge between recent evidence and near-future states implied by the brain’s dynamical model of the world.
Similar considerations apply in the “color phi” phenomenon, where two differently colored flashes at slightly different positions can give rise to the experience of a single object moving between them and changing color mid-trajectory. The change of color is subjectively located between the first and second flash, even though the second flash has not yet occurred at the moment when, physically, the intermediate motion would have taken place. This suggests that the brain does not simply lock in the content of the present at the earliest possible moment. Instead, it waits just long enough to gather a small slice of future evidence and then retrospectively assigns a coherent trajectory and transformation to that slice. The felt present is thus shaped by a short reach into the immediate future, as later information is folded back into an interpretation of what was happening “just now.”
This retrospective component is not a literal retrocausality in the physical sense but a statistical reallocation of events within the internal timeline. The system gathers data over a brief temporal window, performs Bayesian inference over that window, and then issues a unified percept that feels instantaneous. From the inside, it seems as though perception is tracking the world moment by moment, but in fact the conscious now lags behind physical time just enough to allow this integration and correction. The result is a present that is slightly thick—spread across tens or hundreds of milliseconds—yet experienced as a single, indivisible moment.
The need to coordinate action places strict constraints on how wide this temporal integration window can be. If it were too long, behavior would be out of sync with rapidly changing circumstances; if too short, the system could not effectively bind related events into a coherent now. The brain appears to strike a compromise, using windows whose effective width varies with context and task demands. When fine motor control is critical, such as catching a ball, the system may weight the most recent evidence more heavily and rely on rapid prediction of trajectories. When interpreting ambiguous social cues, it may use a broader window, integrating facial expressions, tone of voice, and bodily signals over a longer span to infer a stable emotional state or intention.
Temporal binding is also shaped by expectations about agency and causation. In intentional binding experiments, people perceive the interval between their own voluntary action and its outcome as shorter than the same interval between an involuntary movement and an outcome. The motor system generates a prediction about the expected sensory consequences of a chosen action, including their approximate timing. When the outcome matches this prediction, the brain infers a strong causal link and compresses the perceived temporal gap. This compression can be understood as a Bayesian adjustment: given a prior that intentional actions promptly cause their effects, the subjective timeline is warped so that cause and effect appear closer together, reinforcing the sense of agency.
Conversely, when events violate strong temporal or causal expectations, they are likely to be experienced as asynchronous or disjointed, even when physical timing differences are small. A word that arrives slightly off-beat in a familiar rhythm, or a visual cue that appears late relative to a predicted impact, can jump out as “out of time.” In probabilistic terms, prediction errors in the temporal domain signal that the current generative model is misaligned with the data. The system must then either update its priors about timing relationships or treat the event as an anomaly. In both cases, the disturbance is often felt as a disruption in the smooth flow of the present moment.
On this account, the construction of the present is less like reading off a timestamp and more like stitching together a narrative segment. Within a brief temporal window, sensory events, motor commands, and internal signals are assigned roles—cause, effect, background, context—according to the model that best explains their order and delays. Temporal binding is the process by which these assignments are negotiated so that the world appears as a set of synchronized happenings rather than as a jumble of asynchronies. The felt unity of a clap, a spoken word, or a quick gesture is the phenomenological correlate of the brain’s success in inferring a common temporal structure underlying disparate signals.
This stitching process extends to the domain of memory and expectation. The boundaries of what feels like “just now” flex depending on how memory traces and predictions bleed into current experience. When immersed in a conversation or a piece of music, the immediate past—the previous phrase, the unfinished melody—remains partially active, serving as a constraint on how the current input is interpreted. At the same time, predictions about what is about to be said or played reach forward into the next beats. The conscious now is suspended between these two directions: it is defined by a region of time where prior context has not yet fully decayed and anticipated continuation has not yet been fully confirmed or disconfirmed.
Evidence from neurophysiology supports this picture of a temporally extended present. Neural assemblies in sensory and association areas show patterns of activity that integrate information over successive tens or hundreds of milliseconds, with higher-order areas often exhibiting longer integration timescales than primary sensory cortices. These dynamics enable the system to maintain a partially updated representation of the recent past while incorporating new input. In this sense, the conscious now is not anchored to a single neuronal event but arises from the ongoing evolution of distributed networks that pool evidence over time while constantly adjusting their predictions about what is currently happening.
To understand the present as a constructed result of temporal binding is therefore to see it as a moving inference window, sliding over the stream of sensory and internal events. At each position of the window, the brain leverages priors about typical delays, causal chains, and action consequences, along with the precision of incoming signals, to determine which events belong together in one coherent now. The boundaries of that window shift with context, attention, and uncertainty, but its function remains the same: to provide an operationally useful, probabilistically grounded estimate of what is happening in time that can guide perception, action, and learning.
Predictive processing and the continuity of self
Within the predictive processing framework, the continuity of self emerges from the same machinery that constructs the conscious now: hierarchical generative models performing Bayesian inference over time. Rather than being a static essence, the self is an ongoing prediction—an inferred center of perspective, control, and preference that the bayesian brain posits to explain patterns in bodily sensations, actions, and social feedback. Just as the system infers stable objects behind changing sensory input, it infers a relatively stable subject behind fluctuating thoughts, feelings, and perceptions. The apparent persistence of this subject across moments is the result of temporally extended inference, where priors about personal identity and bodily continuity constrain the interpretation of incoming evidence.
In this setting, predictive processing operates on multiple nested timescales. Fast levels track immediate sensory details, like the shape of a hand or the pitch of a voice, while slower levels encode enduring regularities such as one’s typical posture, one’s characteristic voice, and one’s recurring habits and goals. The continuity of self is tied to these slower, higher levels of the hierarchy, where predictions unfold over seconds, minutes, hours, and even years. These levels generate expectations about what kinds of experiences “I” tend to have, what “I” am likely to do next, and how “I” will feel in various contexts. Moment-to-moment experience is then interpreted against these high-level priors, so that the shifting content of consciousness is constantly tagged as belonging to the same underlying self.
Temporally deep models are crucial here. A system that could only predict the next few hundred milliseconds would be able to coordinate movement and bind sensory events, but it would lack any coherent narrative about who is acting and why. By learning long-range statistical regularities—how one’s own body typically changes with age, how one’s social roles evolve, how one’s preferences tend to remain stable or drift—the brain constructs a generative model that reaches far beyond the immediate present. The conscious now then functions as a bayesian bridge between these long-horizon predictions and the current stream of sensory and interoceptive data, allowing the organism to feel like the same entity that had past experiences and will have future ones.
Autobiographical memory can be viewed as part of this predictive architecture rather than as a mere archive of stored snapshots. Memories are reconstructed as inferences about what must have happened, given the current generative model of the self and fragments of stored traces. When you recall a past event, the brain does not replay a perfect recording; it runs a constrained simulation that fits current priors about your character, your relationships, and your history. This simulation updates those priors in subtle ways, reinforcing some aspects of self and weakening others. Over time, the remembered past is gradually reshaped to become more consistent with present identity and expectations about the future, thereby tightening the probabilistic coupling between different temporal segments of the self-model.
From this angle, the sense of “I was there then, and I am here now” is the experiential face of successful temporal integration within the self-model. The system continuously checks whether its current states and actions are consistent with the predicted trajectory of the same agent that existed moments ago. When there is a close match, the inference that “this is still me” is strongly reinforced. When there are mismatches—dramatic mood shifts, dissociative episodes, or radical changes in belief—prediction errors arise at higher levels of the hierarchy, signaling that the existing self-model may no longer adequately explain the data. The resulting instability can be felt as fragmentation, estrangement, or a break in the continuity of identity.
Interoception plays a central role in securing this continuity. Predictive models of the body’s internal milieu—heart rate, breathing patterns, gut activity, hormonal states—provide a relatively slow-changing, high-reliability anchor for self-inference. Even when external circumstances fluctuate rapidly, the organism can track a familiar signature of bodily states, using it as evidence for the persistence of a single embodied self. When interoceptive prediction and regulation are working smoothly, there is a background sense of bodily ownership and presence that underwrites more abstract aspects of identity. When these processes are disrupted, as in certain anxiety disorders, depersonalization, or chronic pain conditions, the felt continuity of self can erode, as prediction errors about bodily states propagate upward and destabilize higher-level self-beliefs.
This predictive view also reframes the relationship between self and action. The self is not only a perceptual construct but also a control construct: a hypothesized locus of agency that helps explain why certain patterns of movement and decision-making occur. Generative models at intermediate and high levels encode policies—structured expectations about how actions lead to outcomes over time. When the organism acts in ways that conform to these policies and outcomes match predictions, the inference that “I am the author of this sequence” gains strength. The continuity of self thus depends on the continuity of predicted action-outcome chains, stitched together across successive nows into extended projects, habits, and life plans.
Breakdowns in this stitching process illuminate its normal operation. In some forms of anosognosia, individuals deny paralysis of a limb despite overwhelming evidence. One interpretation in predictive processing terms is that high-level priors about bodily integrity and agency are assigned such high precision that conflicting sensory prediction errors are down-weighted or reinterpreted. The self-model insists on a continuous, intact body and a coherent repertoire of actions, even at the cost of distorting current experience. Conversely, in certain depersonalization or dissociative states, priors about the unity and authorship of experience lose precision, allowing fragmentary sensations and thoughts to be experienced as alien or disconnected from “me.” In both cases, the continuity of self is modulated by how the system allocates precision between priors and incoming evidence.
Language and narrative add another layer of temporal depth to the predictive self. Verbal models allow the organism to encode and rehearse story-like structures with beginnings, middles, and anticipated endings. These stories are not passive descriptions but generative scripts that shape expectation and interpretation. By telling and retelling stories about “who I am,” “what has happened to me,” and “where I am going,” the brain trains its own high-level priors over identity. These narrative priors then influence how new events are perceived and remembered: successes may be assimilated into a story of competence, failures into a story of inadequacy, depending on which narrative currently dominates. The conscious now is continuously evaluated against these narrative templates, reinforcing a sense of continuity whenever new experiences can be smoothly woven into the existing plot.
Even rapid fluctuations in mood and thought content can be accommodated by this framework, because predictive models can treat short-term variability as noise or as expected oscillation around a more stable core. A person may feel angry one moment and calm the next, distracted now and focused later, without experiencing a rupture of identity, because higher-level self-priors predict that such fluctuations are typical for them. Only when deviations exceed the expected range—prolonged mania in someone who expects emotional balance, or a sudden inability to recognize familiar surroundings in someone who expects reliable memory—do high-level prediction errors rise to consciousness as crises of self and continuity.
Importantly, continuity does not require perfect stability in content but stability in the structure of prediction. The parameters of the generative model—beliefs about one’s traits, capacities, and social roles—can drift gradually while the system preserves a meta-level expectation that “I am the same one whose parameters are drifting.” This meta-level expectation, itself a prior in a higher stratum of the hierarchy, ensures that even recognized change is interpreted as change of a single persisting subject. The self is thus not an unchanging core but an evolving model that treats its own evolution as the history of one agent, rather than as a sequence of unrelated owners of experience.
Seen in this way, the continuity of self is an emergent property of temporally deep, hierarchically organized predictive processing. The brain’s generative models must span enough time to make sense of extended patterns in bodily states, interactions, and outcomes, and they must assign those patterns to a single, persisting source. The conscious now is the active interface where these temporally extended self-predictions meet the latest evidence, allowing each new moment to be both a fresh inference and a continuation of an ongoing personal narrative. The stability we feel is the felt success of these models in explaining away prediction errors over time, maintaining a coherent center of experience even as its contents continually change.
Attention, uncertainty, and the dynamics of the now
Attention, in this framework, is best understood as the dynamic allocation of precision within the bayesian brain. Rather than being a spotlight that illuminates some contents and leaves others in the dark, it is a control mechanism that selectively increases or decreases the weight of particular prediction errors relative to competing priors. When attention is directed toward a stimulus, the system assigns higher precision to the associated sensory channels and their errors. As a result, those signals exert more influence on updating the internal model and on shaping what is experienced in the conscious now. When precision is lowered, even salient signals can be absorbed by existing expectations with minimal experiential impact.
This way of thinking turns attention into a form of uncertainty management. Precision is, in effect, an estimate of inverse uncertainty: the more precise a signal is deemed to be, the less uncertain the system considers it. The brain must constantly infer both the likely causes of its inputs and how reliable those inputs are. When environmental conditions are stable and predictable, it can safely rely on strong priors, keeping attention narrowly focused on task-relevant cues while down-weighting unexpected but likely irrelevant deviations. When conditions are volatile or ambiguous, it is adaptive to broaden attention, increase the precision of a wider range of sensory inputs, and allow prediction errors to drive more substantial model revision.
The dynamics of the present moment emerge from this continuous precision-weighting. At any given instant, the felt texture of the conscious now—whether it is sharp and vivid, or diffuse and hazy—depends on how the system is balancing the precision of priors against the precision of sensory evidence. A highly focused state, such as deep concentration on a demanding task, corresponds to a configuration in which a specific subset of priors and associated prediction errors dominate the inferential process. Background events are effectively marginalized: their errors are treated as low-precision noise and so rarely gain access to awareness. By contrast, in states of fatigue, anxiety, or sensory overload, precision estimates may become dysregulated, leading to a noisy, unstable present in which minor stimuli intrude disproportionately and it becomes difficult to sustain a coherent line of thought or action.
Temporal integration, which underlies the construction of a thick present, is also modulated by attention. The brain does not use a fixed temporal window for all processing; instead, it flexibly adjusts the duration and overlap of its integration windows according to current estimates of uncertainty and task demands. When attention narrows onto a rapidly changing stimulus—for example, tracking a fast-moving object—the system shortens effective temporal integration at the relevant levels, giving more weight to the most recent samples and predictions while discounting older evidence. When attention broadens to encompass slower, more contextual information—such as monitoring the emotional climate of a conversation—the system can rely on longer integration windows, smoothing over transient fluctuations and emphasizing more stable patterns.
This flexibility in temporal integration helps explain why the subjective pace and granularity of the present can change so dramatically. During high-adrenaline episodes, time may feel slowed and segmented, as if each moment were expanded. One interpretation is that attention has sharply increased the precision of short-timescale prediction errors, leading the system to process and store more detailed snapshots within each physical second. In quieter, low-arousal states, precision may be redistributed toward slower processes, producing an experience of time that feels smoother and less densely populated with discrete events. In both cases, the objective clock remains the same; what changes is how the brain allocates precision across temporal scales, thus reshaping the felt dynamics of the now.
Uncertainty about action outcomes is another key driver of attentional deployment. When an agent is unsure how its actions will affect the world, or when prediction errors about outcomes are unexpectedly large, attention tends to be drawn to the relevant sensory modalities and contextual cues. This can be understood as a strategy for reducing uncertainty: by selectively increasing the precision of outcome-related signals, the system can learn more effectively which policies are reliable. The conscious now becomes populated with those aspects of the environment most informative for improving future predictions, while less informative or redundant features recede into the background.
Conversely, in highly overlearned situations where outcomes are extremely predictable, attention can be relaxed or diverted elsewhere. Walking on a familiar path, for example, involves priors and motor policies so entrenched that prediction errors about each step are minimal and low-precision; the brain effectively runs an automatic generative script. This frees attentional resources to track other, more uncertain streams—planning an upcoming conversation, scanning for unusual sounds, or mind-wandering. The immediate sensorimotor details of walking are barely represented in the conscious now because their prediction errors contribute little new information to the model.
The precision of interoceptive signals plays a similarly important role in shaping the dynamics of the present. When the system judges bodily signals as highly informative—say, under threat or physiological disturbance—it increases their precision, pulling them into the center of awareness. Heartbeat, breathing, and muscle tension then color the entire field of experience, narrowing attention and biasing interpretation toward threat-related priors. When interoceptive noise is low and regulation is effective, precision can be shifted outward, allowing attention to rest more fully on external tasks and social cues. The conscious now, in this sense, reflects not only what is happening in the environment but also how the brain is weighting evidence from within the body.
Disorders of attention can be reframed within this precision-allocation scheme. In attention-deficit presentations, for instance, it is plausible that the system struggles to stabilize precision on task-relevant prediction errors. Competing stimuli and internal thoughts repeatedly gain transient precision bumps, fragmenting experience into a succession of partially formed nows that do not cohere into extended engagement. In anxiety and hypervigilance, by contrast, precision may be chronically over-allocated to possible threat signals and to high-level catastrophic priors. The result is a present that feels both constricted and unstable: the system is certain that something might be wrong but uncertain exactly what, so attention is repeatedly yanked across potential sources of danger.
These phenomena highlight that attention is not merely a filter on finished percepts but an intrinsic component of the inferential process that constructs them. By tuning precision, the system determines which signals will drive updating at each level of the hierarchy and over which temporal spans. The conscious now is the emergent product of this tuning: a continuously renegotiated balance between entrenched expectations and fresh evidence, between brief and extended temporal integration, between internal and external sources of information. Shifts in attention correspond to shifts in this balance, altering not just what is noticed but how the very flow of experience is segmented and stabilized.
Social contexts introduce additional layers of complexity. When interacting with others, the system must track not only the physical environment but also inferred mental states, intentions, and norms—high-level hidden causes with substantial uncertainty. Attention is often pulled toward socially informative cues such as gaze direction, tone of voice, or subtle facial movements. Here, precision is dynamically redistributed toward signals that reduce uncertainty about others’ beliefs and future actions. The conscious now becomes densely populated with socially salient micro-events—raised eyebrows, pauses, hesitations—that might be ignored in asocial contexts. In this way, the dynamics of attention mirror the shifting structure of the generative model as it places greater or lesser emphasis on social prediction.
Importantly, the allocation of precision is itself subject to learning. Over time, the system forms higher-order priors about which cues are typically reliable, which contexts are safe or dangerous, and which tasks warrant sustained focus. These meta-priors about uncertainty guide future attentional strategies, often outside of awareness. A person who has repeatedly encountered unpredictable punishment, for example, may learn to assign high precision to ambiguous threat cues and low precision to signals of safety, thereby stabilizing a chronically tense and narrow present. Another individual, raised in reliably supportive contexts, may develop priors that favor broader, more exploratory attention, yielding a present that feels more open and less dominated by potential dangers.
Practices often described as “training attention” can be interpreted as deliberate attempts to recalibrate these precision hierarchies. Mindfulness exercises, for instance, involve sustaining attention on simple, ongoing sensations—breath, bodily contact, ambient sounds—while reducing the automatic precision boosts given to intrusive thoughts and evaluative narratives. Over time, this may weaken certain high-level priors and strengthen precision allocation to low-level, moment-to-moment sensory evidence. The conscious now then becomes less entangled with elaborate predictions about the future or reconstructions of the past and more anchored in the immediate flow of experience, even though that flow remains inferentially constructed.
Across these examples, the central role of uncertainty becomes clear. The bayesian brain must continuously decide not only what to believe about the world and the body, but also how confident to be in those beliefs, over what temporal horizons, and in which contexts. Attention is the mechanism by which these confidence estimates are enacted in real time, sculpting which prediction errors matter and how long their influence persists. The dynamics of the now—its clarity, stability, speed, and scope—are the phenomenological footprint of this ongoing negotiation between priors, evidence, and the ever-shifting mapping of precision onto the layered processes of perception, action, and interpretation.
Implications for cognition, agency, and free will
Understanding cognition through the lens of probabilistic inference reshapes familiar notions of thinking, deciding, and acting. Cognitive processes become the operation of a hierarchically organized generative model that continually updates beliefs in light of prediction errors. Deliberation, problem-solving, and imagination are not add-ons to perception; they are extensions of the same inferential machinery that constructs the conscious now. When you weigh options, simulate outcomes, or reinterpret a past event, the bayesian brain is using its model to explore counterfactual trajectories—possible futures and alternative histories—and to assess their plausibility given current priors and goals. Cognition is thus deeply future-oriented and model-based: it operates by projecting ahead, comparing different predicted paths, and revising higher-level beliefs about what is worth pursuing or avoiding.
On this view, agency is likewise reinterpreted. Rather than a mysterious faculty that steps in from outside the causal chain, agency is an emergent property of how the generative model organizes its policies—structured expectations about sequences of actions and their outcomes. The system does not merely predict what will happen; it predicts what will happen if it acts in certain ways. Policies that reliably reduce uncertainty, maintain bodily integrity, and satisfy learned preferences become entrenched, so that acting in line with them feels natural and self-generated. The experience of “I am doing this” corresponds to the successful enactment of a predicted action-outcome sequence, where prediction errors remain within expected bounds and reinforce high-level beliefs about being an efficacious agent.
From inside this process, agency is felt as the capacity to select among alternatives—different possible policies—and to bring about one rather than another. In inferential terms, this selection is governed by expected value and expected uncertainty reduction. The system evaluates potential policies not just by their anticipated sensory consequences but also by how well they align with long-term priors about identity, goals, and norms. Choosing to speak up in a meeting, for instance, is not only a motor decision; it is the adoption of a policy that has been judged probable to yield outcomes compatible with one’s self-model and social expectations. When the ensuing events match predictions closely enough, the sense of authorship over the action and its consequences is strengthened.
This does not render agency illusory, but it does relocate it. Agency is not a simple, atomistic power to initiate actions ex nihilo; it is the organized capacity of a temporally deep model to anticipate the effects of its own interventions and to steer itself along trajectories that it deems valuable. The conscious feeling of being able to do otherwise is rooted in access to multiple competing policy predictions. When the brain can simulate divergent courses of action and assign meaningful probabilities and utilities to them, the space of “could have” and “still might” options becomes phenomenologically rich. The more flexible and well-calibrated these simulations are, the more robust the sense of agency.
This reframing directly intersects with debates about free will. Traditional, non-inferential pictures often pit free will against causal determination, as if any lawful regularity in behavior must undermine genuine choice. Within a bayesian framework, however, the very capacity to choose depends on having a coherent, causally structured model of oneself and the environment. Free will, insofar as it is a psychological reality, does not require breaks in causality; it requires a system that can represent alternative futures, evaluate them against its priors and preferences, and adjust its behavior accordingly. The organism is not free from causes; it is free in virtue of being a cause that can learn, predict, and self-modify.
The perceived gap between “I could have done otherwise” and “I inevitably did what I did, given prior causes” narrows when we consider how temporally extended inference works. When an action is taken, it reflects not only immediate stimuli but the accumulated shaping of priors over development, culture, and previous choices. Yet the same mechanism that entrenches these priors also allows them to change. New evidence, new social feedback, and new internal states can drive prediction errors at high levels, leading to revisions of preferences, values, and characteristic policies. Over time, the space of available futures that the model can realistically entertain and implement is reshaped by the very actions it selects. In this sense, agents participate in the ongoing construction of the constraints that will guide their later choices.
Responsibility and moral agency can thus be understood in probabilistic terms. To hold someone responsible is, among other things, to treat their behavior as emerging from a relatively stable, temporally deep self-model that can learn from consequences. Sanctions, praise, and negotiation all act as structured feedback that updates the agent’s priors about which policies lead to which social outcomes. When responsibility is attributed, we implicitly assume that the person’s generative model is capable of reconfiguring its policies under new evidence—that is, that their future predictions about action-outcome patterns can be altered. A purely reflexive or rigidly conditioned system, by contrast, would not be an appropriate target of moral appraisal because its policy space is too limited or inflexible to meaningfully incorporate such feedback.
This perspective also sheds light on feelings of compulsion and loss of control. In addictive behaviors or certain compulsive disorders, some policies acquire such entrenched priors and such high expected value in the model that alternative policies are systematically underweighted, even when their long-term outcomes are judged better in abstract terms. The agent might explicitly endorse a desire to act differently while still finding themselves repeatedly enacting the same harmful script. The resulting experience is that of diminished freedom: the space of live, action-guiding predictions has narrowed, and high-level intentions fail to propagate into effective policy selection at lower levels. Therapeutic interventions can be viewed as attempts to alter these entrenched priors and reward structures so that alternative policies regain inferential and motivational weight.
Another consequence of this framework concerns how we think about rationality and bias. Cognitive biases become systematic patterns in how priors are shaped and updated, and in how precision is allocated to different sources of evidence. Confirmation bias, for example, arises when priors about beliefs or identities are given disproportionately high precision, such that disconfirming prediction errors are down-weighted or dismissed. From the inside, actions remain experienced as freely chosen; from the outside, we can see that the generative model’s rigidity has effectively narrowed the corridor of plausible futures. Efforts to “debias” cognition can then be conceived as attempts to adjust precision hierarchies and encourage exposure to evidence that challenges overly rigid priors.
Because the conscious now functions as a bridge between recent evidence and anticipated futures, it is also where many experiences relevant to free will are felt most acutely: hesitation, temptation, resolve, regret. Hesitation reflects the coexistence of competing policy predictions whose relative expected values and uncertainties have not yet been sufficiently discriminated. Temptation arises when a short-term, high-reward policy competes with a longer-term, more abstract policy anchored in identity or values. Resolve is the stabilization of a particular high-level policy and its propagation down the hierarchy so that lower-level systems align their micro-actions accordingly. Regret emerges when later evidence reveals that a chosen policy generated worse-than-expected outcomes, prompting revision of priors about similar future situations.
Importantly, these experiences are not epiphenomenal glosses on an otherwise blind mechanism; they are integral aspects of how the inferential system monitors its own uncertainty and calibrates policy selection. The felt tension of a difficult choice tracks the degree of residual uncertainty between competing future trajectories. The relief of having decided corresponds to the collapse of that uncertainty as one policy gains dominant posterior probability. Over time, the system learns to recognize patterns of internal conflict and to adjust its priors or its search strategies, making similar decisions easier or harder depending on what has been learned. In this way, subjective experiences associated with freedom and responsibility play a functional role in refining the generative model.
This account also has implications for how we think about collective agency. Groups, institutions, and cultures can be seen as higher-level generative structures that shape individual priors, define available policies, and constrain which futures are afforded or even conceivable. Social norms, laws, and narratives function as shared priors on what actions are acceptable or likely to succeed. An individual’s felt freedom is therefore partly determined by the breadth and diversity of policies that their social environment renders plausible and practicable. When social structures systematically restrict policy exploration—through coercion, deprivation, or rigid role expectations—the effective policy space shrinks, and with it the scope of lived agency, even if local moment-to-moment choices still feel voluntary.
Conversely, environments that foster exploration and learning expand the range of policies that the generative model can safely consider. Education, supportive relationships, and material security reduce certain forms of uncertainty while enabling the system to tolerate and learn from others. This broadens the set of futures that can be meaningfully represented and pursued. In these contexts, the experience of freedom deepens not because causality is suspended, but because the inferential machinery has richer models, more flexible priors, and access to a wider landscape of predicted trajectories that can be enacted and revised.
