From a Bayesian perspective, perception and action are often modeled as processes of probabilistic inference, in which the brain combines sensory evidence with prior expectations to generate a coherent interpretation of the world. Traditional formulations assume that these priors encode knowledge about how causes typically precede and generate effects. However, when we consider the temporal aspects of agency, it becomes necessary to entertain a broader class of priors that are defined not only over states of the world, but over temporal relations themselves, including the possibility that information about later events can modulate inferences about earlier ones. These temporally extended expectations can be described as retrocausal priors: structured beliefs that allow the bayesian brain to use information arriving after an event to refine, or even reshape, the inferred causes leading up to that event.
In such models, the distinction between forward-looking and backward-looking components of inference is fluid. Perception of an action’s outcome is typically framed as the result of prediction: motor commands generate efference copies that are used to anticipate sensory consequences, and these predictions serve as priors against which incoming sensory evidence is compared. Yet the same computational machinery can, in principle, operate in reverse through postdiction, where later sensory signals update the inferred state of earlier moments. Retrocausal priors formalize this by assigning probabilities not just to what will happen next, but to how subsequent evidence should revise the interpretation of what has already occurred, including the perceived onset of intention and the sense of having caused an outcome.
In a standard Bayesian model of perception and action, the generative model specifies how hidden causes produce observable data over time. Extending this model to include retrocausal priors means that the joint probability distribution spans entire trajectories rather than discrete, ordered events. The brain does not merely ask, “Given my current state and motor command, what sensory input should I expect?” It also implicitly encodes, “Given the sensory data I will receive over a short temporal window, what was the most plausible configuration of intentions, movements, and external states that produced them?” This shift reframes agency as an inference over spatiotemporal patterns, constrained by expectations that link earlier internal states with later environmental feedback.
Retrocausal priors are particularly relevant for explaining phenomena in which the subjective timing of intention and action deviates from their physical order. The bayesian brain can maintain flexible temporal hypotheses, weighting them by both internal predictions and externally observed consequences. When an outcome closely matches a predicted effect of a plausible action, retrocausal inference can “pull” the perceived time of intention closer to that outcome, reinforcing the sense of agency. Here, the prior is not merely about which outcome is likely, but also about how tightly linked, in time, intentions and outcomes should be if they belong to the same causal chain.
Formally, these priors can be implemented as probability distributions over temporal lags and causal link strengths. For example, the model might encode that, under normal circumstances, voluntary actions precede their sensory consequences by a characteristic delay, with some variability. When sensory feedback is noisy or ambiguous, the prior over this delay can dominate, effectively snapping the perceived timing of events into a canonical pattern that maximizes overall coherence. When new evidence arrives—such as an unexpectedly rapid or delayed outcome—the inference process can retroactively adjust the estimated onset of intention and the perceived duration of the action to preserve a high posterior probability for a unified, agent-caused episode.
This framework also helps explain how the sense of agency remains robust despite delays, distortions, or interruptions in sensorimotor processing. If the generative model expects that self-produced outcomes will conform to particular statistical regularities—such as matching predicted sensory features or satisfying higher-level goals—then retrocausal priors can bias inference toward attributing ambiguous outcomes to one’s own actions whenever they fit these patterns, even if the temporal order is imperfect. In this way, retrocausality at the level of neural inference does not violate physical causation; instead, it reflects the brain’s strategy of fitting noisy data into a consistent causal narrative, sometimes by “editing” the temporal structure of that narrative after additional evidence becomes available.
Integrating retrocausal priors into Bayesian models also clarifies the relationship between motor control and conscious awareness of intention. Motor planning and execution can be understood as largely forward-generative processes, but the conscious experience of deciding and acting may be constructed in part through retroactive interpretation. Within a hierarchical generative model, higher levels encode abstract intentions and goals, while lower levels handle detailed kinematics and sensory predictions. When an action and its outcome unfold, evidence is propagated upward and downward across the hierarchy, allowing later-confirmed success or failure to reshape the inferred timing and strength of prior intentions. Retrocausal priors at higher levels bias this reconstruction toward narratives in which intentions reliably precede and cause successful outcomes, supporting a stable sense of being the author of one’s actions.
These ideas can be extended to encompass learning. Each episode of action and feedback updates the parameters of the generative model, including priors over how intentions, actions, and outcomes are temporally and causally related. Repeated experiences in which predictions are confirmed tighten these priors, making future retrocausal inferences more confident and less tolerant of temporal deviations. Conversely, environments with unpredictable delays or outcomes may lead to broader, more flexible priors, weakening the automatic attribution of agency. Over developmental timescales, this iterative updating sculpts an individualized temporal template of what “normal” voluntary action feels like, such that departures from this template—whether due to external perturbations or internal disruptions—can be experienced as a loss or distortion of agency.
Importantly, retrocausal priors need not be consciously accessible. They operate as implicit constraints on neural inference, shaping both perceptual timelines and the probability of attributing causation to oneself versus external sources. The resulting phenomenology—such as feeling that one intended to act earlier than one actually did—is a byproduct of the system’s attempt to minimize prediction error across a temporally extended window. By embedding these priors within a unified Bayesian framework, it becomes possible to describe agency not as a simple readout of motor commands, but as an emergent property of ongoing probabilistic negotiations between prior expectations and sensory evidence, distributed over both past and future moments within the inferential horizon.
Temporal structure of agency: from prediction to postdiction
The experience of agency is not confined to a single instant; it is distributed over a temporal interval in which prediction and postdiction are intertwined. At the outset, prospective processes dominate: motor systems generate forward models of anticipated sensory consequences, and higher-order states encode intentions and goals that are projected into the near future. This anticipatory stage involves both explicit planning and implicit calibration of expected delays between action and outcome. Yet the eventual feeling that one has acted voluntarily is not simply the readout of these early predictions. Instead, the bayesian brain retrospectively evaluates the unfolding sensorimotor sequence, using later sensory outcomes and contextual information to refine the timing, strength, and even the existence of the prior intention.
Within this temporal framework, prediction supplies an initial scaffold for agency. When a voluntary command is prepared, efference copies of motor signals generate immediate expectations about which body parts will move, what proprioceptive feedback will arise, and which environmental changes are likely to follow. These predictions are time-stamped, specifying not only what should happen but also when it should occur relative to the initiating intention. This structure allows the system to preemptively attenuate self-produced sensations and to prioritize unpredicted events for further processing. Crucially, however, this forward-looking organization does not close the inferential loop; it opens a window in which subsequent evidence can confirm, disconfirm, or reshape the prior narrative of having willed the action.
Postdiction exploits the fact that sensory information often arrives with variable delays and noise. After an action unfolds, the nervous system has access to a richer dataset: proprioceptive traces, visual and auditory feedback, interoceptive changes, and external consequences. Neural inference uses this temporally extended evidence to reconstruct the most coherent interpretation of what the agent was doing and intending at earlier moments. For example, if an outcome strongly matches a predicted effect of a plausible action, the posterior inference may shift the perceived onset of intention backward in time, aligning it more tightly with the observed consequence. Conversely, if outcomes are surprising or inconsistent with prior expectations, the reconstructed timeline may weaken or delay the felt intention, giving rise to experiences of diminished or disrupted agency.
This reconstruction is constrained by temporal priors that encode typical delays between intention, movement, and outcome. The system expects certain characteristic intervals: decisions are usually followed quickly by motor execution, and environmental effects occur within a limited window after the relevant action. When actual timings deviate from these expectations, postdictive processes adjust perceived onset and duration of internal events to minimize global prediction error. For instance, a slightly delayed outcome that still fits the expected sensory profile may be “pulled” closer in subjective time to the initiating movement, preserving a compact, causally plausible episode. In contrast, outcomes that fall outside the learned temporal window may be segregated into a different causal stream, and the corresponding retrospective narrative may omit or downplay a prior intention.
Temporal binding illustrates this interplay between prediction and postdiction. When people perform an action that yields a predictable outcome, the perceived time of the action often shifts forward, while the perceived time of the outcome shifts backward, compressing the interval between them. From a retrocausal perspective, this binding reflects a bidirectional negotiation: prospective models specify that a self-caused event should be followed by a matching consequence within a given delay, and postdictive inference, upon observing such a consequence, re-centers both action and outcome around a unified causal episode. The result is a temporally compact representation that supports a strong sense of agency, even though the subjective timestamps have been altered by later evidence.
Importantly, the temporal structure of agency operates across multiple scales. At the sub-second level, fine-grained sensorimotor delays are adjusted to maintain coherence between movement and immediate sensory feedback. At the scale of several seconds, higher-level inferences integrate sequences of actions and outcomes into goal-directed episodes, potentially reassigning individual events to different intentions depending on how the sequence unfolds. Over even longer intervals, the system can retrospectively reinterpret past decisions in light of outcomes that arise well after the original choice, as when success or failure leads one to feel that an earlier intention was more or less decisive than it seemed at the time. Across all these scales, postdiction complements prediction, continually reshaping the perceived temporal profile of intentions and actions.
This dynamic is particularly relevant for understanding the relation between consciousness and action. The subjective feeling of “now deciding” may not coincide with the earliest neural events that bias future behavior. Instead, conscious intention can function as a temporally smoothed estimate, distilled from both predictive signals leading up to movement and postdictive signals that confirm or revise the initial plan. When an action produces expected consequences, retrospective amplification of intention contributes to a vivid sense of authorship. When outcomes are mismatched or externally driven, the same inferential machinery can down-regulate the experienced contribution of will, making the action feel automatic, coerced, or alien. Thus, conscious agency is a temporally extended construct, sustained by ongoing negotiations between forward models and backward-looking evaluation.
The temporal flexibility of agency also explains why subtle manipulations of delay and contingency can profoundly alter subjective experience without changing the objective order of events. Small, experimentally imposed shifts between action and feedback can weaken or strengthen the attribution of causation, not because the physical sequence has changed in any deep sense, but because the probabilistic mapping between internal timestamps and external outcomes has been recalibrated. When feedback is artificially delayed beyond the window encoded by temporal priors, postdiction struggles to integrate the outcome into the same causal chain, and the action may be reinterpreted as ineffective or merely coincidental. Conversely, when outcomes are made to appear synchronously with or even slightly before the movement, retrocausal inference can still knit them into a coherent episode, provided they match the predicted sensory profile and fit the broader context.
This framing highlights that the sense of agency is better conceived as a property of temporally distributed neural inference than as a simple function of instantaneous neural events. Prediction sets up an initial hypothesis about how the near future will unfold, while postdiction monitors the unfolding stream and continuously edits the temporal narrative of what one intended and did. Retrocausality, in this context, does not imply violations of physical time, but rather expresses the brain’s ability to allocate credit and responsibility across a moving temporal horizon. The subjective “present” of agency is thus a constructed interval that integrates past signals and near-future evidence into a single, coherent experience of having initiated and controlled action.
Neural mechanisms supporting retrocausal inference
Understanding how the nervous system could implement retrocausal inference requires examining how neural circuits represent time, causation, and agency within a distributed architecture. Rather than encoding a fixed timeline of events, many brain systems appear to maintain temporally extended activity patterns that integrate information from both recent past and near-future expectations. These patterns provide a substrate for inference that is not strictly locked to the present moment, but instead spans a window over which sensory evidence and motor predictions are jointly evaluated. Within this window, the bayesian brain can update its interpretation of earlier neural states in light of later inputs, effectively allowing retrocausal priors to reshape the perceived sequence of intention, action, and outcome.
At the core of this capacity are recurrent and feedback connections that run both within and between cortical and subcortical regions. Hierarchical generative models of perception and action map naturally onto the laminar and columnar organization of cortex, where higher-order areas send top-down signals encoding predictions about lower-level activity. These predictions include not only what features should be present, but when they should occur relative to planned movements and goals. Crucially, bottom-up signals do not merely report current sensory states; they convey temporally structured prediction errors that can be used to revise both ongoing expectations and the inferred history of internal states that gave rise to the current moment. In this way, sensory feedback that arrives after a movement can still alter the inferred timing and strength of the intention that preceded it, as long as the relevant neural states remain partially active or accessible within the recurrent circuitry.
Several candidate systems are especially relevant for supporting this kind of temporally extended neural inference. The prefrontal cortex, particularly medial and dorsolateral regions, is implicated in representing abstract intentions, goals, and action plans over extended timescales. Activity in these regions often precedes overt movement, but it also persists and evolves after outcomes are observed, reflecting reevaluation and updating of internal models. The supplementary motor area and pre-supplementary motor area are central to initiating voluntary actions and are frequently associated with the subjective experience of deciding to move. Their dense bidirectional connections with prefrontal and parietal regions position them as hubs where higher-level intentions, motor programs, and retrospective evaluations converge, enabling both prospective planning and postdictive reconstruction of agency.
Parietal cortex, particularly the inferior parietal lobule and temporoparietal junction, plays a key role in integrating multisensory information and attributing actions to self versus others. Neural responses in these regions are sensitive to discrepancies between predicted and actual sensory consequences of movement, and they have been linked to the temporal binding of actions and outcomes. When visual, proprioceptive, or auditory signals match efference-based predictions, parietal circuits support a compact, unified representation of the event as self-caused. When mismatches arise, these same circuits can reassign causation, sometimes leading to the experience that movements or outcomes are externally generated. The capacity to adjust causal attribution based on post hoc evidence suggests that parietal networks maintain a temporally flexible encoding of who did what, and when, that can be revised retroactively.
Subcortical structures, especially the basal ganglia and cerebellum, contribute essential predictive and corrective components to this architecture. The cerebellum has long been associated with forward models of motor control: it predicts the sensory consequences of outgoing motor commands and uses errors to refine future predictions. Yet cerebellar circuits also process delayed feedback and can adjust internal timing estimates when outcomes arrive earlier or later than expected. This makes them well suited for recalibrating temporal priors that govern the expected intervals between intention, movement, and outcome. The basal ganglia, through their role in action selection and reinforcement learning, encode the probability that particular actions will lead to particular outcomes. Dopaminergic signals broadcasting prediction error not only drive future learning but can also modulate current and recent neural activity, reshaping the perceived salience and intentionality of actions that have just occurred.
These systems are embedded in a broader network that supports metacognitive access to agency. Prefrontal-parietal loops, often implicated in conscious awareness, appear to integrate information about performance, confidence, and context into a higher-order representation of “I am doing this.” Retrocausality in this network does not require that late-arriving signals physically alter past neural events; instead, it relies on the fact that the conscious narrative of action is constructed from distributed traces that remain malleable for some period. As visual, auditory, and proprioceptive evidence accumulates, higher-order representations in prefrontal and parietal cortex can reweigh earlier motor and premotor signals, effectively editing the experienced onset and strength of intention. Consciousness of agency thus emerges as a post hoc yet timely summary of an episode, rather than a direct readout of its initiating neural cause.
Oscillatory dynamics and neural synchrony provide an additional mechanism for implementing temporally extended inference. Coordinated oscillations, particularly in beta and gamma bands, support communication between distant regions involved in motor planning, sensory processing, and evaluation. Changes in phase relationships between these oscillations can realign the temporal integration windows across networks, effectively shifting which signals are grouped into a single causal episode. When action-related and outcome-related activity become more tightly synchronized, the brain can bind them into a unified event, even if the objective time lag is variable. Conversely, desynchronization may promote segregation of events into separate streams, weakening the sense that an outcome belongs to a prior intention. Through such dynamic reconfiguration, neural oscillations help tune the temporal boundaries within which retrocausal inference operates.
Importantly, many neural mechanisms that support prediction over short timescales naturally enable postdiction when considered over slightly longer intervals. For example, predictive coding models posit that higher cortical levels maintain beliefs about causes that evolve slowly relative to the rapid fluctuations of sensory input. These slowly varying beliefs can be updated not only by current prediction errors but also by those arising a short time later, as new evidence continues to arrive. Because the same latent states are responsible for generating a segment of the data stream, late errors can adjust the inferred trajectory of those states, implicitly altering what the system takes itself to have been representing or intending just moments before. In this sense, retrocausal priors are simply an extension of standard prediction machinery along the temporal axis of the generative model.
Temporal working memory and eligibility traces, as studied in reinforcement learning and synaptic plasticity, further illuminate how the nervous system keeps recent events “alive” for retroactive credit assignment. Neurons and synapses often exhibit decaying traces of recent activity, which can be strengthened or weakened by reinforcement signals that arrive later. These mechanisms allow outcomes to retroactively modify the significance of antecedent states and actions, both at the level of synaptic weights and at the level of network activity patterns. The same eligibility-like processes can be harnessed to update the neural representation of intention and control: when an outcome strongly confirms a predicted consequence, neuromodulatory signals can selectively enhance the salience of the preceding intentional state, making it more likely to be encoded in conscious memory as a decisive act of will. When outcomes disconfirm expectations, the corresponding traces may be weakened or recoded as guesses, impulses, or externally driven events.
The temporal resolution of different neural systems also shapes the window over which retrocausal inference can operate. Fast sensory pathways, such as the visual magnocellular stream, provide quick but coarse feedback about movement consequences, while slower, more detailed pathways deliver refined information. Motor and premotor regions can maintain preparatory activity long enough to be updated by both fast and slow feedback channels, allowing multiple “passes” of evaluation. Interoceptive systems, including insular and cingulate cortices, track bodily states and affective responses that often unfold over seconds, such as effort, success, or failure. These slower signals can retroactively color the perceived voluntariness and significance of recent actions, embedding them within an affectively laden narrative that either reinforces or diminishes the sense of agency.
Disruptions to any of these mechanisms provide convergent evidence for their role in retrocausal processing. Lesions to parietal regions, for instance, can lead to disorders in which patients misattribute externally generated movements to themselves or fail to recognize their own actions, suggesting a breakdown in the integration of predictive and postdictive information. Abnormalities in cerebellar or basal ganglia function are associated with altered timing of actions and feedback, as seen in Parkinson’s disease or cerebellar ataxia, often accompanied by changes in perceived control. In schizophrenia and related conditions, aberrant dopaminergic signaling and disrupted connectivity within fronto-parietal networks have been linked to experiences of thought insertion, passivity phenomena, and impaired temporal binding. These pathologies can be interpreted as failures of the neural machinery that normally maintains coherent, retroactively editable narratives of agency.
Neuroimaging and electrophysiological studies of intentional action also reveal signatures consistent with retrocausal inference. Readiness potentials and build-up activity in motor areas sometimes precede the reported time of conscious intention, suggesting that the subjective decision may be a delayed reconstruction from earlier unconscious processes. At the same time, the perceived timing of intention can be shifted by manipulating outcomes or contextual cues that occur after movement onset, indicating that higher-order networks revise the temporal label of intention based on post hoc information. Such findings align with the idea that the nervous system implements a temporally smeared representation of decision and action, subject to continuous revision as new evidence arrives.
On this view, neural mechanisms supporting retrocausal inference do not require any exotic reversal of physical causation. They rely instead on the brain’s intrinsic architecture for maintaining and updating distributed, time-sensitive representations. Recurrent circuits, hierarchical predictive coding, eligibility traces, oscillatory synchronization, and neuromodulatory feedback collectively allow later events to influence how earlier states are encoded in memory and consciousness. The sense of agency arises from this ongoing negotiation, as the nervous system settles on a coherent account of what it intended, did, and caused over a short but elastic temporal horizon. Retrocausality, in this neurobiological sense, is simply the flip side of prediction: given that the same generative model governs both past and future segments of experience, revisions based on later evidence inevitably propagate backward through the inferred history of the agent’s own actions.
Experimental paradigms probing retrocausal agency
Experimental work on the sense of agency has increasingly exploited the temporal flexibility of perception to probe how the bayesian brain integrates prediction and postdiction. One influential approach centers on intentional binding, in which participants perform simple voluntary actions—such as pressing a key—that trigger sensory outcomes like tones or visual flashes after a short delay. When individuals initiate the action themselves, they typically judge the time of the action as later and the time of the outcome as earlier, compressing the perceived interval between them. When the same movements are triggered externally, or when outcomes are decoupled from actions, this binding effect is reduced or absent. These findings are naturally interpreted within a framework of retrocausal priors: when an outcome matches a predicted consequence within an expected temporal window, postdictive processes shift subjective timestamps to produce a compact, high-probability narrative of self-caused action.
Variants of the intentional binding paradigm manipulate the reliability and structure of action–outcome contingencies to test how flexible these priors are. In some experiments, outcomes occur with only partial probability following a voluntary action, or with variable delays drawn from different distributions. When the action–outcome relationship is strong and consistent, temporal binding is robust; when the mapping is noisy or inconsistent, binding weakens, and participants report a diminished sense of having caused the outcome. By systematically altering the statistics of the environment, researchers can observe how the nervous system recalibrates its temporal priors and how this recalibration changes both objective measures of binding and subjective reports of agency. The pattern suggests that retrocausal inference is actively tuned by experience, with learned expectations about delays and contingencies guiding how later evidence reshapes the perceived onset of intention and the felt authorship of events.
Another family of paradigms explores retrocausality through spatial and identity manipulations of feedback, often using virtual reality or mirror setups. Participants may control a virtual limb or cursor while the experimenter systematically distorts the spatial trajectory or timing of the visual feedback. If these perturbations remain within a certain tolerance—consistent with the participant’s internal model of sensorimotor noise—people still claim ownership of the movement and experience a stable sense of agency. However, once distortions exceed that tolerance, the feedback is increasingly experienced as alien or externally controlled. Retrocausal priors come into play as the system evaluates, after the fact, whether the observed sensory consequences could plausibly have arisen from the prior motor command, given both spatial and temporal expectations. When the match is deemed sufficient, postdiction pulls the experience of intention toward the outcome; when the mismatch is too large, the narrative of self-causation is weakened or broken.
Temporal illusions of decision-making provide a complementary window into how retrospective processes help construct the experience of will. Libet-style experiments, in which participants make spontaneous finger movements while observing a rapidly rotating clock, reveal that the reported time of conscious intention often lags behind neural readiness potentials recorded from motor areas. Yet the reported time of intention can be modulated by manipulations occurring after movement onset, such as the presence or absence of an outcome or the experimenter’s feedback about whether the action met a criterion. When outcomes are manipulated to appear more or less congruent with the participant’s putative goal, the remembered time of deciding can shift accordingly, indicating that the experience of “when I chose” is not a fixed readout of early neural events but a construct shaped retroactively by later evidence. Experimental paradigms that systematically vary these post-movement cues show that agency and even the timing of volition are sensitive to retrocausal updates.
Closely related are paradigms on choice blindness, where participants make a decision—for example, choosing between two faces or options—and then, through a sleight-of-hand manipulation, are shown the alternative as if it were their own choice. Many individuals accept this substitution and proceed to justify the “chosen” option, providing reasons that fit the new outcome. This suggests that the cognitive system, confronted with a mismatch between earlier intention and later evidence, can retroactively reconstruct its own prior state to preserve a coherent narrative of agency. The mechanisms underlying this reconstruction likely overlap with those that support retrocausal neural inference at shorter timescales, but here they operate over several seconds, reshaping not only the perceived timing but the very content of intention in light of present circumstances.
Experiments involving delayed or reordered feedback offer more direct probes of the temporal windows within which retrocausal priors operate. In some studies, participants move a joystick or tap a key, and the resulting cursor movement or tone is presented with systematically manipulated delays, sometimes even appearing slightly before the registered movement due to experimental trickery. Within limits, participants still report a sense of control and may even experience the action as having occurred earlier than it did, effectively “backdating” their intention to preserve causality. When delays exceed a critical threshold, however, the action and outcome become perceptually disjoint; temporal binding collapses, and agency ratings drop. By charting these thresholds for different sensory modalities and task contexts, researchers infer the temporal extent over which the bayesian brain is willing to let later evidence influence the inferred timing of prior intentional states.
Studies of sensory attenuation complement these timing-based paradigms by focusing on the perceived intensity of outcomes rather than their timing. When a person voluntarily generates a sensory event, such as pressing a button that produces a tone or a tactile stimulus, the resulting sensation is commonly perceived as less intense than an equivalent externally produced event. This attenuation is often interpreted as evidence for predictive motor control: efference copies dampen responses to expected self-produced stimuli. However, experiments that introduce uncertainty or delay into the action–outcome mapping reveal that attenuation can be modulated by post hoc evaluations. If an outcome is later judged to be self-caused, even when its timing or occurrence was initially ambiguous, perceived intensity can be retrospectively downregulated, consistent with a retroactive application of the “self-generated” label. Conversely, if outcomes are reclassified as externally generated, attenuation diminishes. Here, retrocausality emerges in the reassignment of sensory events to self or other after they have been perceived.
Illusions of ownership and agency in virtual and augmented reality further demonstrate how temporal and causal expectations are jointly exploited to probe retrocausal processes. In the rubber hand illusion and its variants, synchronous stroking of a visible fake hand and a participant’s hidden real hand can induce a feeling that the fake hand is part of one’s body. When the fake hand is also made to move in correspondence with the participant’s voluntary movements, a sense of agency over the artificial limb often follows. Researchers can introduce small temporal offsets between real and virtual movements, or introduce unexpected outcomes linked to the virtual limb, to test when and how participants revise their judgments about ownership and control. As long as action–outcome correlations stay within the bounds set by temporal priors, participants maintain both ownership and agency; when these bounds are violated, the system retroactively withdraws these attributions, evidencing the role of postdictive evaluation in sustaining embodied agency.
Clinical populations provide naturally occurring tests of how disruptions to priors, prediction, and postdiction affect agency. In patients with schizophrenia, paradigms that involve intentional binding, sensory attenuation, or virtual-reality–based agency manipulations often show attenuated binding, abnormal timing judgments, or paradoxical attributions of self to externally controlled events. For example, when asked to judge the timing of actions and outcomes, some patients display reduced compression or even reversal of the usual binding pattern, hinting at altered retrocausal inference. In passivity phenomena, where individuals experience their own movements as controlled by external agents, experimental tasks that manipulate feedback timing can elicit exaggerated shifts in agency judgments, suggesting that the temporal priors that normally stabilize the sense of self-causation are weakened or distorted. These findings support the idea that disorders of agency may arise from atypical tuning of the inferential machinery that links earlier internal states to later sensory evidence.
Pharmacological and neuromodulatory interventions have been used to establish causal links between specific neural processes and retrocausal aspects of agency. For instance, altering dopaminergic function with psychostimulants or dopamine antagonists can modify intentional binding and the perceived linkage between action and outcome, consistent with dopamine’s role in prediction-error signaling and credit assignment. Noninvasive brain stimulation techniques, such as transcranial magnetic stimulation applied to parietal or prefrontal regions, can transiently perturb temporal binding, timing judgments, and self–other attribution. When stimulation is applied at carefully chosen intervals relative to action and outcome, it can disproportionately affect postdictive components of processing, revealing how late-arriving neural signals contribute to updating the timing and authorship of events that have already unfolded in objective time.
Advanced neuroimaging and electrophysiological paradigms bring these behavioral and phenomenological findings into contact with underlying neural dynamics. In many tasks, time-resolved functional MRI or EEG/MEG recordings track how predictive signals before action and evaluative signals after feedback jointly shape the experience of agency. For example, increased coupling between motor regions and sensory cortices during successful action–outcome episodes correlates with stronger intentional binding and higher agency ratings, whereas weakened or delayed coupling accompanies diminished agency. Decoding analyses sometimes show that post-outcome brain states can predict participants’ later reports of when they decided or how strongly they intended to act, over and above what can be inferred from pre-movement activity alone. These results underscore that consciousness of agency is not simply read off from preparatory neural activity, but is co-determined by later evaluative processes that engage retrocausal inference over a temporally extended window.
Together, these experimental paradigms exploit controlled manipulations of timing, contingency, and context to reveal how the sense of agency depends on the brain’s ability to integrate information across past and future within a coherent generative model. By varying the statistical relationship between actions and outcomes, researchers probe the limits of temporal binding, the recalibration of priors, and the conditions under which postdictive processes revise the interpretation of earlier mental states. Across behavioral, clinical, and neurophysiological studies, a convergent picture emerges: agency is not a static property of individual events but a dynamically constructed inference, shaped by the brain’s ongoing efforts to reconcile predictive models with the unfolding stream of sensory evidence, even when that requires retrocausal adjustments to the perceived timing and content of intention.
Implications for free will and cognitive architecture
Taking retrocausal priors seriously alters how debates about free will are framed. In place of a picture in which an isolated, instantaneous decision causes action in a simple forward chain, the bayesian brain is better described as maintaining a temporally extended hypothesis about what it is doing, why, and with what consequences. Intentions, movements, and outcomes are all treated as noisy data points within this hypothesis, updated both prospectively and retrospectively. Within this framework, the experience of free will is not the direct imprint of a single initiating neural event, but an emergent property of ongoing neural inference that integrates pre-action predictions with post-action evidence to construct a self-consistent narrative of agency.
This shift has immediate implications for how we interpret classical challenges from neuroscience to free will, such as the finding that readiness potentials precede reported conscious decisions. If conscious intention is partly a postdictive estimate, then early motor-related activity need not undermine the authenticity of felt choice. Instead, the system may use preparatory signals as one ingredient among many in reconstructing what “I decided” at a slightly later moment, after enough information is available to support a coherent account. Retrocausal priors formalize this: the brain expects that successful, goal-congruent outcomes should be preceded by appropriately timed intentions, and it therefore biases its later reconstruction of mental chronology toward patterns in which the agent appears as an effective initiator. On this view, the gap between neural precursors and reported intention reflects the temporal dynamics of inference, not a simple revelation that free will is illusory.
At the same time, this perspective does constrain naive libertarian conceptions of free will. If the contents and timing of conscious intention are shaped by retroactive evaluation of outcomes, then the felt sense of having freely chosen cannot be taken as an infallible guide to the underlying causal process. The architecture of agency is built to favor internally coherent explanations over veridical access to microstructural neural events. When outcomes are fortuitously aligned with existing goals and predictive models, postdiction can render them “meant all along,” even if the initiating neural dynamics were partially stochastic or influenced by unnoticed biases. Retrocausality thus supports compatibilist accounts in which free will is grounded not in metaphysical indeterminism, but in the organism’s capacity to maintain and refine a temporally integrated model of itself as a reliable source of controlled action.
Within such compatibilist views, what matters is that the agent’s actions flow from an internal, hierarchically organized system of goals, values, and policies, and that this system can flexibly adapt to evidence over time. Retrocausal priors and postdictive processes are not enemies of free will but enablers: they allow the agent to assimilate noisy, temporally scattered events into coherent projects, to learn from delayed feedback, and to revise self-understanding in light of outcomes. The sense of “I could have done otherwise” can then be interpreted as a counterfactual judgment within the generative model: given my higher-level dispositions and the range of possible trajectories they license, alternative actions were within the space of policies the system could have selected. Neural inference, operating over an extended window, supports both actual and counterfactual narratives, updating their probabilities as new evidence arrives.
This has important consequences for moral and legal conceptions of responsibility. If the sense of having willed an action is partly constructed after the fact, questions of culpability cannot be settled simply by appealing to the agent’s immediate introspective report. However, the relevant notion of responsibility can be anchored in the stability and structure of the generative model itself: how consistent are the priors and policies that shaped the action? How robustly does the agent’s architecture support long-range prediction, self-monitoring, and retrospective correction? An individual whose neural systems reliably integrate prediction and postdiction to maintain coherent control across contexts may justifiably be treated as a responsible agent, even if their conscious timeline of decision-making is reconstructed. Conversely, when pathology or developmental disruption fractures these integrative capacities, responsibility is diminished not because retrocausality undermines agency in principle, but because the underlying architecture can no longer sustain a stable model of self-authored action.
Considering cognitive architecture more directly, retrocausal priors suggest that systems supporting agency must be organized around temporally deep generative models. Rather than a pipeline from perception to decision to action, the architecture resembles a multi-layered loop in which higher levels encode slow-changing goals and identity-defining narratives, while lower levels handle fast sensorimotor contingencies. Prediction flows downward as provisional commitments about what will happen and what the agent will do; postdiction flows upward as a stream of prediction errors that may retroactively revise both the interpretation of recent events and the configuration of long-term priors. This bidirectional traffic means that agency is distributed across the hierarchy: local sensorimotor routines contribute to control in the moment, while abstract self-models determine how these routines are grouped into longer-term projects and are later remembered as intentional or accidental.
One consequence of this architecture is that consciousness of agency likely arises at intermediate levels of the hierarchy, where temporal resolution and abstraction are balanced. Very low-level sensory and motor processes operate too quickly and automatically to be introspectively accessible, while very high-level identity narratives evolve too slowly to capture the fine structure of single actions. The conscious feeling of deciding and acting emerges instead where action policies, action–outcome contingencies, and context converge over windows of hundreds of milliseconds to a few seconds. Retrocausal inference at this mesoscale level allows late-arriving feedback to adjust not only the perceived timing of decisions but also their qualitative framing—for example, as deliberate versus impulsive, coerced versus voluntary. As these intermediate representations are broadcast across networks involved in working memory and report, they form the substrate of explicit judgments about free will and responsibility.
Retrocausal priors also imply that the architecture of selfhood is inherently narrative and revisable. The brain does not merely store static snapshots of past choices; it maintains compressive, story-like summaries that can be updated when new evidence changes the inferred meaning of earlier events. Free will, in this sense, is intertwined with the capacity to curate and revise one’s own history of agency. When outcomes reveal that prior policies are maladaptive or misaligned with higher-order goals, the system can retroactively reinterpret those episodes as mistakes, lapses, or coerced actions, adjusting priors on future behavior accordingly. This narrative plasticity is central to rehabilitation, learning, and moral growth. Far from undermining freedom, retrocausality enables agents to treat their own past as material for reorganization, aligning subsequent choices more closely with endorsed values.
At a more computational level, adopting retrocausal inference as a design principle leads to cognitive architectures that emphasize credit assignment over extended horizons. Hierarchical predictive coding, model-based reinforcement learning, and control-theoretic approaches all converge on the need to maintain eligibility traces linking earlier internal states to delayed outcomes. Embedding these traces within a temporally deep generative model effectively implements retrocausal priors: the system assumes that certain constellations of internal activity and environmental context are the typical precursors of particular outcomes, and it uses later evidence to refine these assumptions. In artificial agents, such architectures would naturally support phenomena analogous to temporal binding and post hoc reinterpretation of decisions, raising further questions about how agency and free will should be attributed to non-biological systems that share similar inferential dynamics.
Recognizing the centrality of retrocausality also helps to unify disparate cognitive functions under a single architectural principle. Memory, prediction, imagination, and agency all become facets of the same temporally deep modeling capacity. Episodic memory is not a passive archive but an active reconstruction that reuses the generative model to simulate past events; prediction uses the same model to simulate possible futures; imagination explores counterfactual trajectories; and agency selects among these trajectories while continuously revising their probability given new evidence. Free will, on this picture, is not a mysterious faculty added on top of cognition, but the name we give to the organism’s ability to treat its own internal simulations as actionable options, and to commit to some trajectories rather than others, while remaining capable of revising the interpretation of those commitments after the fact.
This framework invites a reframing of philosophical skepticism about agency. Many arguments that conclude that free will is an illusion presuppose a simplistic, forward-only causal picture in which either the conscious intention is the initial mover or it is epiphenomenal. Once temporal inference and retrocausal priors are acknowledged, this dichotomy becomes less compelling. Conscious intention is neither an omnipotent initiator nor a powerless afterthought; it is a dynamically updated estimate, constructed within a richly structured architecture that spans past and future. The integrity of free will, understood compatibilistically, depends on how this architecture functions: whether it supports coherent, context-sensitive, and self-correcting patterns of control. Retrocausality therefore shifts attention from metaphysical questions about timing to empirical and engineering questions about what kinds of neural and computational organizations can sustain robust forms of agency in a temporally uncertain world.
