RESEARCH DOSSIER
Sense of Agency
Abstract. The sense of agency denotes the experienced authorship and control of actions and their consequences. This critical research monograph reviews its conceptual boundaries, historical origins, sensorimotor and inferential theories, experimental paradigms, neural substrates, development and adaptation, neuropsychiatric disturbances, social coercion, joint action, and technological mediation. Original action–effect experiments, perceptual and explicit measures, neuroimaging studies, systematic reviews, and recent meta-analyses are compared for construct validity, causal interpretation, robustness, and generalizability. The evidence supports multiple contributions from motor prediction, prospective selection, contextual inference, and feedback. It also demonstrates that intentional binding, sensory attenuation, and explicit agency judgments can dissociate. Human–Artificial interactions extend research to delegated and shared control, with a clear distinction between demonstrable causal influence, subjective attribution, and hypothesized experience in technical systems.
Keywords: sense of agency; action authorship; self-agency; intentional binding; sensory attenuation; efference copy; comparator model; motor prediction; agency attribution; cue integration; joint action; neuropsychology; human–AI interaction; Artificial Era
Sense of Agency as Experienced Causal Authorship
The sense of agency is the experience of initiating, guiding, or bringing about an action and, where an action produces a discernible consequence, of exercising influence over that consequence. A person who reaches for a glass, changes a cursor's direction, or stops an alarm generally experiences the movement as something they did. This lived relation between intention, action, and outcome is the principal explanandum. It must be distinguished from objective causal efficacy, which can be investigated by intervention, and from retrospective judgments of responsibility, which also depend on social and normative considerations. Moore (2016) offers an accessible conceptual synthesis, while Synofzik et al. (2008) explains why an immediate feeling and a reflective judgment require different descriptions.
Experienced agency has a temporal architecture. An intended movement may feel self-generated before its outcome appears, whereas the causal authorship of a distant effect may become apparent only after feedback. This difference matters because experiments that ask participants whether they controlled a cursor, studies that quantify perceived action–effect intervals, and functional imaging investigations of action preparation observe different events. A scientific definition therefore identifies the action, the outcome, the available cues, the report method, and the temporal window. Haggard et al. (2002) established a powerful method for investigating the subjective timing of voluntary actions, but it did not establish that every aspect of agency can be reduced to temporal compression.
Agency is experienced in conditions ranging from habitual movements to complex cooperation, and its phenomenology changes with attention, fluency, feedback, and expectations. Researchers should specify the relevant form of control: producing one's own movement, choosing between alternatives, changing an environmental state, or influencing an outcome shared with others. Self-produced activity may be accurate while authorship feels uncertain; conversely, a convincing causal interpretation can coexist with weak actual control. The possibility of these mismatches makes agency a tractable psychological phenomenon rather than an infallible introspective report. Wegner & Wheatley (1999) experimentally demonstrated conditions in which suggested prior thoughts altered perceived authorship, illustrating the importance of examining perception and causation separately.
The Research Problem and Its Boundaries
The central research questions concern how the nervous system and the wider cognitive system distinguish self-produced change from externally produced change, how they attribute events to an actor, and which evidential cues support the resulting experience. Investigators can study agency for a finger movement, a remote cursor, an audible tone, a joint decision, or an automated tool. These examples require different experimental controls because the possible causal chains differ. In a button–tone task, physical contingency and reported agency may be manipulated separately; in a collaborative interface, the participant's relative influence must also be measured. Moore et al. (2009) shows that prior informational cues can alter an indirect agency measure even when movement is experimentally controlled.
Control, autonomy, self-efficacy, internal locus of control, free will, and moral accountability are related constructs with distinct scientific uses. A high score on general self-efficacy may predict willingness to initiate action without establishing that a particular movement feels self-produced. The legal attribution of responsibility depends on institutions and norm-governed judgments beyond the basic action experience. The philosophical question of metaphysical free will addresses another level of analysis again. Keeping these distinctions explicit protects operational definitions, prevents explanatory circularity, and clarifies which measurements can bear on which questions. Gallagher (2000) differentiates minimal self-experience and extended narrative identity; Moore (2016) also discusses the applications and limits of agency research.
The scope of this monograph is the psychological sense of agency in action, effect production, social coordination, and technological mediation. It examines historical and contemporary theories, experimental manipulations, neural associations, clinical alterations, and reproducibility. References to bodily ownership clarify a necessary neighboring construct without substituting an ownership theory for agency. The research tradition contains behavioral experiments, neuroimaging contrasts, clinical observations, narrative reviews, systematic reviews, and meta-analyses. The present synthesis evaluates them as a critical scholarly review, with study designs and inferential limits stated explicitly. Published findings remain distinguishable from proposed explanations, including those applied to interactions with Artificial systems. The empirical and conceptual breadth of the field is documented in Wen & Imamizu (2022) and Zhao et al. (2025).
Historical Origins of Action Awareness
Long before contemporary agency tasks, thinkers addressed the apparent relation between willing, moving, and recognizing an event as one's own deed. Nineteenth- and twentieth-century work on voluntary movement, proprioception, motor control, and the self supplied separate approaches that later converged. The modern experimental field took shape when motor-prediction models and social-cognitive accounts of will began generating testable, partly competing explanations. The transition from conceptual dispute to measurable contrast was decisive: participants could act voluntarily or involuntarily, encounter manipulated outcomes, and report timing, ownership, and causal authorship. Frith et al. (2000) connected awareness of action to forward models of motor control, while Wegner & Wheatley (1999) articulated the inferential role of a preceding thought.
Experimental motor-control research examined how expected consequences of self-generated stimulation can be attenuated. Blakemore et al. (1998) contrasted self-produced and externally produced tactile stimulation and investigated associated neural responses, linking sensory prediction to the phenomenon of why self-tickling ordinarily differs from another person's touch. This work concerned sensory processing and predicted reafference; it did not by itself prove a complete theory of conscious agency. At approximately the same historical juncture, Wegner & Wheatley (1999) tested how interpreting prior thought as a cause could change the experience of authorship. Together these programs showed why a theory restricted to motor cancellation or a theory restricted to conscious inference would leave significant explanatory questions unanswered.
The 2002 action–effect timing experiments introduced intentional binding as a major empirical paradigm. Haggard et al. (2002) reported attraction in perceived time between voluntary action and its sensory outcome, contrasted with a condition involving induced involuntary movement. Subsequent work tested contextual cues, outcome predictability, implicit–explicit dissociations, and the neural correlates of perceived intervals. Sato & Yasuda (2005) systematically manipulated mismatches between predicted and actual sensory feedback, while Synofzik et al. (2008) proposed separating feelings of agency from explicit judgments and combining multiple evidential cues. The historical succession should be read as progressive differentiation of observables and explanatory models, rather than as proof that a single new measure settled the entire problem of agency.
Phenomenology, Ownership, and Attribution
Sense of ownership concerns whether a body, movement, or sensation is experienced as belonging to oneself; sense of agency concerns whether one experiences oneself as initiating or controlling an action. They often coexist, which makes ordinary introspection an unreliable way to distinguish them without carefully designed contrasts. A passive movement can still be experienced as occurring in one's own limb, while a self-produced action can feel unexpectedly constrained or alien in its effects. Gallagher (2000) made this separation influential in discussions of minimal selfhood. Sato & Yasuda (2005) provided experimental evidence that agency reports can be altered by predicted–actual outcome discrepancies under conditions where self-ownership reports are not altered in the same manner.
The distinction between feeling and judgment adds another dimension. An individual can experience smooth sensorimotor control yet explicitly infer that another person, a computer, or an accident produced the final event. Conversely, strong task instructions or prior knowledge may support a causal self-attribution even when immediate movement fluency is low. Synofzik et al. (2008) describes agency as involving a multifactorial pre-reflective component and a later conceptual attribution, a proposal that subsequently motivated experimental attempts to dissociate indicators. Such distinctions have consequences for patient interviews, motor assistance, and brain–computer interfaces: an answer to 'who caused it?' is informative but does not uniquely reveal the perceptual processes that preceded that answer.
Agency can attach to an act, an effect, or an extended sequence with changing scope. A person may feel authorship of a finger press while denying authorship of an unexpected light or reward that follows it. A team member may claim agency for a contribution while regarding the combined product as collective. Researchers should therefore identify the reference event of every question, rather than treating a generic 'I controlled it' scale as universally comparable. The analytical distinction is supported by the diversity of tasks discussed in Moore (2016) and by the direct manipulation of expected effect properties in Sato & Yasuda (2005). Phenomenological descriptions become most useful when tied to specific, replicable behavioral and situational contrasts.
Operational Definitions and Construct Validity
An operational definition should state which aspect of agency is targeted, what task elicits it, and how the observed variable is generated. Explicit agency ratings quantify participants' retrospective causal or experiential judgments on a defined scale. Temporal-binding tasks measure perceived timing shifts under action–effect contingencies. Sensory-attenuation tasks examine changes in perceived intensity or neural responses to predicted self-generated consequences. Movement-feedback perturbations measure sensitivity to deviations between intended and actual effects. No one instrument automatically provides a criterion measure for all others. Dewey & Knoblich (2014) is particularly important because its participants showed several familiar experimental signatures without reliable correlations between the different implicit and explicit measurements.
Construct validity requires more than showing that a metric changes between voluntary and passive conditions. Competing processes including attention, temporal expectation, causal belief, sensory reliability, and learned contingency can produce similar shifts. Investigators need convergent and discriminant tests: manipulate a hypothesized agency cue while measuring plausible alternative mechanisms, test the predicted relationship with direct experience, and examine whether the metric behaves similarly within individuals across tasks. Moore et al. (2012) used a probabilistic-outcome design and reported a dissociation between implicit and explicit aspects; Wolpe et al. (2013) examined cue integration within temporal perception, showing how a binding effect can depend on the reliability of sensory evidence.
A durable research program consequently treats agency as a structured family of experiences and mechanisms whose relations remain empirical questions. Reliability across repeated sessions, sensitivity to task parameters, preregistered comparisons, and appropriate multilevel analysis are necessary when treating individual differences as meaningful. Group-average binding may be reproducible while a participant's difference score has poor stability. Within-person associations between timing shifts and momentary ratings need not reproduce as between-person correlations. Dewey & Knoblich (2014) and Moore et al. (2012) make the measurement problem visible from complementary designs. Later quantitative synthesis such as Zhao et al. (2025) should guide claims about average patterns, while individual experimental reports remain necessary for evaluating the causal specificity of each inference.
Motor Prediction and Forward Models
Motor-control models distinguish the commands by which a movement is generated from the sensory consequences that subsequently return from the body and environment. An efference copy can support an internal prediction of those consequences before external feedback arrives. Comparing predicted and observed signals helps explain rapid correction and the attenuation of some self-generated sensations. In a comparator-based account of agency, a sufficiently close match contributes evidence that an outcome arose from one's own action. Frith et al. (2000) connected this computational proposal to disturbances of action awareness, and Blakemore et al. (1998) supplied influential evidence from self-produced tactile stimulation. These studies establish an important mechanism while leaving the phenomenological interpretation of comparison signals open.
A forward model is especially useful for explaining when sensory changes should be expected and what a motor command is likely to accomplish. It can account for rapid motor adaptation and for differentiating predictable self-generated stimulation from unexpected input. Yet actual control is often exercised without perfect sensory prediction, particularly when environments are noisy, unfamiliar, or dependent on another actor. Some self-generated outcomes are surprising, and familiar external events may be highly predictable. A simple rule equating sensory match with agency therefore risks confusing certainty about an event with authorship of the event. Sato & Yasuda (2005) manipulated prediction errors and found changes in self-agency reports, while Synofzik et al. (2008) used such constraints to motivate a broader model.
Mechanistic explanation requires specifying the predicted quantity: movement trajectory, sensory modality, timing, intensity, or the causal relation between action and effect. A delay manipulation is not equivalent to a spatial perturbation, and an altered sensory intensity may change perception without changing conscious causation. Different computational accounts predict different generalization patterns across these manipulations. Wolpe et al. (2013) suggests that the reliability of sensory events affects their contribution to perceived temporal relations; this is consistent with a probabilistic integration view rather than a single immutable motor match. A decisive experiment should separate action preparation, action execution, observed feedback, and explicit attribution, then compare model predictions across matched conditions.
Sensory Attenuation and Self-Generated Events
Self-generated sensory consequences are often processed differently from externally generated ones, a phenomenon sometimes called sensory attenuation. In the influential tactile experiment of Blakemore et al. (1998), participants experienced self-produced versus externally applied touch while investigators examined perceptual and neural differences. The results support a role for predictive processing of reafferent signals. Sensory attenuation matters to agency research because the distinction between self-caused and other-caused change can be available at early perceptual stages. Its explanatory value depends on the particular modality, stimulus, task, and response being measured. The presence of attenuation alone does not establish that the participant consciously feels authorship of the associated action.
Experimental work can manipulate whether the consequence is generated by the participant, whether its timing and identity are predictable, and whether its intensity is matched to an external control. These factors need to be separated because self-production, repetition, expectation, and selective attention can each affect responses. A reliable sensorimotor difference may reveal predictive filtering without identifying the source of an explicit report. Dewey & Knoblich (2014) found that sensory attenuation and intentional binding, though both observed in their task, did not reliably correlate across participants. This challenges the expectation that two common indirect markers must always reflect one underlying latent quantity with a stable individual ordering.
Explanatory precision is especially important when moving from touch to auditory, visual, or digitally mediated effects. A tone triggered by a key press can differ from a self-applied tactile event in motor demands, temporal precision, and perceptual learning. A computer-generated consequence may be physically identical while its assigned causal origin changes. The correct comparison identifies the dimensions experimentally held constant and the dimension deliberately altered. Sato & Yasuda (2005) demonstrates that the temporal and identity predictability of an effect can influence agency reports, and Wen & Imamizu (2022) reviews wider perceptual consequences of agency. Converging across sensory channels can strengthen a model only after modality-specific mechanisms and measurement artifacts have been evaluated.
Action Selection and Prospective Agency
The sense of agency can be influenced before the consequence of an action is known. Selecting among possible movements, resolving competition, and translating an intention into a motor plan supply internal information about prospective control. Chambon et al. (2014) reviews evidence that the fluency of action selection contributes to agency experience, challenging explanations that begin only with a post-action comparison between predicted and actual results. Under this perspective, difficulties or conflicts during selection may change perceived authorship even when externally observed outcomes remain constant. The prospective account generates experimental predictions concerning choice competition, action preparation, and the temporal relation between selection difficulty and agency report.
Prospective contributions must be distinguished from the objective freedom of choosing among options. A forced choice can be implemented fluently, while a choice among many alternatives can be difficult and uncertain. Researchers therefore manipulate selection fluency or conflict independently of the number of alternatives and physical control over results. Neural studies can record premotor correlates, but an association between preparation activity and agency ratings cannot itself establish that the activity creates the conscious experience. Chambon et al. (2014) discusses behavioral and imaging approaches that motivate prospective signals. When used alongside explicit reports and outcome manipulations, these designs can identify whether the same participant's agency shifts before and after receiving feedback.
Prospective and retrospective processes may complement one another within the same sequence. During action planning, the system can estimate whether a selected response is internally generated and likely to succeed; after execution, sensory and contextual evidence can confirm, revise, or override that estimate. This approach predicts that uncertainty at one stage increases the influence of evidence at another stage, although the precise weighting must be tested. Synofzik et al. (2008) conceptualizes a multifactorial process, and Moore et al. (2009) found that external primes modulated perceived action–effect intervals for both voluntary and induced movements. The combined evidence argues for models that specify information flow through time rather than assigning a fixed causal role to one cue.
Retrospective Inference and Apparent Mental Causation
Retrospective accounts examine how people infer authorship from the apparent relation among a preceding thought, an action, and its consequence. Wegner & Wheatley (1999) proposed that priority, consistency, and exclusivity of a thought relative to an action contribute to the experience of conscious will. In a controlled study, altering participants' prior thoughts could affect the attribution of authorship for events they did not independently cause. The major implication is that subjective authorship can depend on a causal interpretation assembled from accessible cues. Such work does not show that all voluntary behavior is illusory; it shows that the conscious explanation of an event is itself a psychological process.
Related experimental manipulations demonstrate that anticipated effect information can increase apparent causal authorship. Aarts et al. (2005) used a task in which participants and a computer could independently affect a moving square, and primes representing potential effect locations enhanced the experienced authorship of stopping it. The design exposes a useful separation between an actual causal mechanism and the participant's knowledge or interpretation of that mechanism. Experimental conclusions concern the observed conditions of priming, uncertainty, and report. They cannot be converted into a universal claim that conscious intentions never contribute to actions, because such a conclusion would exceed the measured variables and the causal structure of the experiments.
Predictive motor accounts and retrospective inferential accounts are often contrasted, yet many experiments reveal that both forms of evidence matter. Moore et al. (2009) manipulated conscious primes and measured action–effect interval judgments under voluntary and induced movement conditions; its pattern suggested interactions among motoric and contextual cues. A careful comparison specifies whether each model predicts the same change in explicit reports, timing shifts, or perceptual intensity. A model can explain one measure and fail another, so theoretical adjudication requires multiple outcomes and competing predictions. The distinction between a causal contribution to control and a psychological interpretation of control remains essential throughout this literature.
Multifactorial and Cue-Integration Accounts
Multifactorial accounts propose that agency experience results from integrating several sources of information, which may include motor predictions, selection fluency, action–outcome contingency, temporal proximity, prior beliefs, and contextual explanations. Synofzik et al. (2008) articulated a two-step model separating an immediate, nonconceptual feeling from a more explicit judgment informed by higher-level cognition. Rather than assigning equal weight to every signal, a flexible integration account predicts that the reliability and relevance of cues change with task conditions. An expected tone with uncertain motor origin may depend heavily on contextual knowledge; a precisely controlled movement with ambiguous downstream effects may rely more heavily on motor information.
Evidence integration can be modeled probabilistically, but mathematical elegance is no substitute for identifying actual sensory and cognitive likelihoods. Wolpe et al. (2013) experimentally investigated the reliability of outcome information in intentional binding and showed how perceived timing can reflect integration of cues to action and consequence. Such findings support computational analysis of temporal reports, while leaving a separate question about the conscious experience those reports index. Moore et al. (2012) found that implicit and explicit agency measures could diverge under sequential action–outcome statistics. Together the studies support distinctions between cue combination within a measure, cue combination across measures, and inference about the latent experience underlying them.
A comprehensive model must explain successful self-attribution, false self-attribution, failure to recognize genuine control, and shifts between immediate and reflective assessments. It should also predict when an action is experienced as one's own but an effect is ascribed to an external source. Comparing models requires fitting them to the same data and evaluating out-of-sample prediction, not simply describing a result after observing it. Zhao et al. (2025) provides meta-analytic synthesis of behavioral and neuroimaging research, valuable for examining patterns that recur across paradigms. The outstanding task is to establish which integrative mechanisms generalize across actions, outcomes, populations, and human–machine arrangements without conflating accuracy of causal attribution with strength of felt control.
Explicit Reports and Attribution Tasks
Explicit measures ask participants how strongly an action or effect felt self-caused, whether an outcome was controlled by oneself or another source, or which actor produced a visible movement. Their strength is direct access to an articulated judgment whose reference event is stated. Their weakness is dependence on wording, memory, response scale, and inferred task demands. Sato & Yasuda (2005) manipulated predicted–actual discrepancies and obtained distinct agency and ownership patterns, illustrating the usefulness of targeted self-report. Farrer & Frith (2002) examined neural correlates of attributing action to self versus another person, while Synofzik et al. (2008) explains why a conceptual judgment should be distinguished from the immediate feeling that may precede it.
A strong rating experiment should predefine the anchor points of the response, report the timing of questioning, and distinguish control of a movement from authorship of its effect. Participants may interpret an agency scale as asking whether they pressed the button, whether their press produced the tone, or whether they were free to choose. These interpretations can produce apparent inconsistency unless clarified during task design. Independent checks can confirm whether participants understood the causal structure and whether outcome probabilities were learned. Moore et al. (2012) demonstrated that sequential learning contingencies can influence explicit and implicit readouts differently, making trial structure and expectation necessary parts of any interpretation.
Self-report also admits principled uses beyond a one-dimensional score. Open descriptions may reveal whether participants attribute a movement, an intention, an outcome, or a whole sequence to themselves. Multilevel analysis can separate within-person variation from differences in stable response style. Counterbalanced question orders and neutral instructions can reduce demand effects, although they do not eliminate introspective limitations. Validation should compare reports to controlled causal perturbations and independent behavioral markers rather than treating the participant as an infallible observer or an unreliable witness by definition. Dewey & Knoblich (2014) found that explicit reports could diverge from common perceptual indices, showing the value of preserving multiple measures.
Intentional Binding: Paradigms and Mechanisms
Intentional binding refers to a perceived temporal attraction between a voluntary action and a subsequent sensory outcome. In the foundational experiments of Haggard et al. (2002), participants judged the timing of a voluntary movement and its effect under conditions designed to contrast with involuntary movement. The perceived time of the action shifted toward the effect and the perceived time of the outcome shifted toward the action. The phenomenon created an experimentally accessible measure of action–effect temporal experience. Its significance is substantial, but its definition must remain perceptual: a timing shift is an observation, while a claim that the shift measures conscious agency is a theoretical interpretation requiring independent support.
Different paradigms estimate temporal binding using event-time reports, reproduced intervals, or judgments of action–effect duration. Baseline conditions may involve passive movement, isolated events, unpredictable outcomes, or observational trials. These choices alter the counterfactual against which binding is estimated. Wolpe et al. (2013) manipulated the reliability of the sensory outcome and investigated cue-based accounts of the timing effect. Moore et al. (2009) showed that contextual primes can influence temporal intervals and that motor-state conditions moderate those effects. An adequate methods description therefore identifies the baseline, outcome probability, response demands, timing apparatus, and the particular quantity called action binding, outcome binding, or overall binding.
Because temporal perception depends on attention, expectation, and causality, intentional binding cannot be assumed to be an interchangeable personal trait measure across all protocols. Dewey & Knoblich (2014) replicated binding and attenuation effects without finding the expected correlations among their agency measures. Conversely, Moore et al. (2012) found evidence for dissociation in a particular learning paradigm rather than a definitive partition into two independent faculties. Recent work further qualifies the simple assumption that better control always produces larger binding: Seghezzi et al. (2026) reported decreasing binding during a particular learning regime under specific feedback and outcome conditions. These results make experimental context an essential part of the interpretation rather than a nuisance variable to ignore.
Sensory Attenuation, Feedback, and Perturbation
Feedback perturbation tests whether unexpected consequences weaken subjective authorship or trigger correction. Researchers may introduce delays, spatial rotations, altered auditory tones, or discrepancies between intended and displayed cursor trajectories. Sato & Yasuda (2005) showed that unpredictability in timing and frequency of an outcome influenced self-agency reports while distinct ownership judgments were comparatively resistant in its experimental design. Such manipulations allow investigators to separate actual action production from the degree to which its consequences match a prediction. They also reveal that agency is relational: the same physical movement can be interpreted differently when the causal environment changes, even though motor command generation remains intact.
A second research family examines how predicted self-produced sensory events are attenuated relative to matched external events. Blakemore et al. (1998) found differences in perception and brain responses associated with self-generated tactile consequences. These observations are often interpreted in terms of forward-model prediction, but a complete experiment also considers adaptation, selective attention, signal detectability, and the perceptual properties of the control stimulus. A signal can be diminished because its occurrence was expected, because it is genuinely associated with a motor command, or through interacting mechanisms. Distinguishing these possibilities requires controlled comparisons and explicit computational or neural predictions rather than using the label 'sensory attenuation' as an explanation in itself.
Perturbation studies are particularly useful when outcomes have multiple dimensions. A participant might correctly predict when an effect occurs while mispredicting its location, or predict its location while failing to know which agent generated it. Multidimensional manipulations can reveal whether timing and content discrepancies combine additively, interactively, or according to their estimated reliability. A similar logic governs remote control interfaces, where visual feedback may be delayed by a network while the user's motor intention remains intact. Wolpe et al. (2013) shows how cue reliability can structure temporal judgments; Wen & Imamizu (2022) surveys broader sensory and behavioral measures. Interpreting any single perturbation requires a task-specific causal model.
Dissociations Across Agency Measures
A major scientific challenge is whether different operational measures track one construct or partly distinct processes. Dewey & Knoblich (2014) investigated self-report, sensory attenuation, and temporal binding in participants completing action–tone tasks. Although expected group-level effects were obtained, the implicit measures were not significantly correlated with each other and did not correlate reliably with explicit agency judgments. The result restricts strong interpretations of indirect indices as generic agency meters. A lack of correlation in one sample does not prove permanent independence, however, because precision, range restriction, and reliability can limit observed associations. The correct conclusion requires both the behavioral pattern and the measurement characteristics of the particular tasks.
Other experiments reveal associations under specific conditions. Moore et al. (2012) found a dissociation between measures in a probabilistic contingency task that distinguished explicit causal expectations from an implicit readout. Haggard et al. (2002) established a compelling temporal phenomenon in a voluntary-versus-involuntary contrast. Neither result establishes a universal mathematical mapping between subjective reports and binding scores. Later work examining agency and binding across modalities indicates that their relation can vary with stimulus type and analysis level. A sophisticated model should therefore distinguish a correlation between trial-level fluctuations from a correlation between participant-level averages; mixing those levels risks an ecological fallacy and may produce apparently contradictory summaries.
Method triangulation becomes valuable when investigators state the predictions that would follow if two measures share a mechanism. For example, stronger prediction precision might change both sensory attenuation and direct agency ratings, but the strength and direction of each change depend on its respective task. A multivariate latent-variable model could be evaluated, yet it must satisfy evidence of dimensionality, test–retest reliability, and measurement invariance. Zhao et al. (2025) synthesizes larger behavioral and imaging literatures, while Dewey & Knoblich (2014) supplies a concrete warning against collapsing observables. Instead of treating disagreement among measures as mere error, researchers can investigate whether the disagreement reveals distinct stages of agency construction.
Experimental Causal Identification
Causal research must separate physical control from information about control. In a well-designed task, the experimenter can randomize whether a participant's action genuinely changes an outcome, whether a consequence occurs after a specific delay, and whether an instruction or prime suggests prior authorship. The resulting factorial design tests main effects and interactions among causal influence, prediction, and belief. Aarts et al. (2005) provides a relevant precedent by making it ambiguous whether the participant or computer stopped a moving object while manipulating prior effect information. Moore et al. (2009) further separated prime influence from movement conditions, showing why control over instruction, motor production, and feedback are jointly necessary.
A valid causal contrast states the intervention and the estimand. If the independent variable is outcome delay, a measured effect on explicit agency is a causal effect of assigned delay on reported agency within the study conditions. It does not automatically identify which neural computation mediated the shift. Neuroimaging may add temporal and anatomical constraints, but activation contrasts remain indirect unless coupled with converging manipulations or causal perturbation of the proposed mechanism. Kühn et al. (2013) used functional imaging in a temporal-interval task and investigated supplementary motor area activity; it offers evidence of association with a specific readout rather than an unrestricted localization of all agency experience.
Reproducible experiments prespecify exclusion criteria, outcome variables, trial counts, timing hardware, baseline conditions, and planned analysis. Researchers should distinguish average effects from heterogeneous participant responses and evaluate the robustness of conclusions to alternate timing estimators or hierarchical models. Clinical generalization requires careful matching of participants, medication status, motor performance, and task comprehension. Moccia et al. (2024) reviews intentional-binding studies across neuropsychiatric populations and illustrates how diverse clinical samples complicate interpretation. Strong causal identification emerges from repeatable contrasts that eliminate plausible alternatives while preserving a clearly defined construct, not from attaching a broad agency interpretation to any significant perceptual difference.
Neural Systems and Localized Claims
Neural studies of agency investigate motor preparation, sensory prediction, causal self-attribution, monitoring of discrepancies, and explicit reports. The field implicates interconnected premotor, parietal, prefrontal, insular, and cerebellar systems rather than a single anatomical seat of agency. Sperduti et al. (2011) performed a quantitative analysis of neuroimaging contrasts involving self- and external-agency attribution, examining differing activation patterns across tasks. Farrer & Frith (2002) investigated neural differences when participants attributed an action to themselves or another actor. Both findings require careful interpretation because self-attribution tasks vary in sensory feedback, task conflict, and reflective decision requirements. A neural response associated with a judgment may support the judgment without fully generating the entire lived experience.
Research on temporal binding provides another neural window. Kühn et al. (2013) related supplementary motor area activity to perceived intervals in action–effect tasks and did not find the predicted angular-gyrus correlation for that particular temporal readout. Such contrasts matter because activation patterns for detecting lost control may differ from those for feeling fluent control. Motor preparation and predicted sensory feedback also recruit different pathways from explicit conceptual attribution. Frith et al. (2000) considered relationships between motor control and action awareness, while Blakemore et al. (1998) examined cerebellar prediction in sensory processing. The combined evidence calls for analysis by computational operation and experimental task rather than searching for one universally activated region.
Technical limitations constrain interpretation. Blood-oxygen-level-dependent signals are indirect measurements of neural activity with limited temporal resolution, and contrasting agency conditions may inadvertently manipulate attention, salience, or movement effort. Activation likelihood estimation and other meta-analytic approaches aggregate studies with heterogeneous designs and should be read in relation to their inclusion rules. Sperduti et al. (2011) synthesized an early set of imaging studies; Zhao et al. (2025) adds later quantitative synthesis of behavioral and neuroimaging evidence. Strong neural explanations link temporally resolved behavior, hypothesized computation, and neural measurements in a common model, then test the model using converging evidence. Anatomical labels should not substitute for demonstrations of predictive or causal specificity.
Development, Learning, and Adaptation
The organization of agency changes with skill, learning, feedback reliability, and familiarity with action–effect relations. Developmental research must distinguish maturation of motor prediction from maturation of reflective causal attribution and verbal report. Inferring a child's agency merely from fluent movement or an adult's response scale would confound changing skills with the construct itself. Wen & Imamizu (2022) reviews how agency is implicated in perception and behavior, while Synofzik et al. (2008) provides a framework for distinguishing immediate sensorimotor cues from conceptually elaborated judgments. Across the lifespan, the research question is whether the sources and relative reliability of agency information change with developmental experience and environmental structure.
Learning introduces a particularly revealing methodological challenge. Improved behavioral performance need not produce proportionally stronger intentional binding. Seghezzi et al. (2026) tested reinforcement-based temporal adaptation across four experiments and found that binding decreased with learning under specific conditions involving relatively imprecise but stable feedback and an outcome irrelevant to the learning goal. Its findings concern a measured indirect index and a defined task environment; they do not imply that people become generally less capable of controlling learned actions. Instead, the study suggests that the relevance of feedback and the informational purpose of an outcome may influence temporal experience as behavioral adaptation proceeds.
A longitudinal study could measure the same participants during initial exploration, skill acquisition, stable performance, and unexpected environmental change. It would assess actual causal control, predictions, explicit agency, and temporal judgments at each phase. Such a design could determine whether agency changes reflect better internal models, altered attention to outcomes, or a shift from deliberative to automatic performance. Wolpe et al. (2013) provides a cue-reliability perspective, and Dewey & Knoblich (2014) cautions that separate measures may show weak correspondence across participants. Individual adaptation curves, not only pre–post group averages, are essential for identifying mechanisms. Population comparisons must test whether task demands and measurement properties remain sufficiently comparable over time.
Clinical Disturbances and Differential Interpretation
Some neurological and psychiatric phenomena involve a reported loss, distortion, or unusual assignment of action authorship. Such experiences are clinically heterogeneous: a person may report that an intentional movement feels imposed, that an involuntary movement is unexpectedly generated, or that a self-produced consequence belongs to someone else. Frith et al. (2000) considered how motor-monitoring abnormalities might contribute to disturbances of action awareness, especially in relation to schizophrenia. Gallagher & Trigg (2016) discusses the conceptual distinction between agency and ownership across clinical experiences. These accounts help formulate investigable mechanisms but should not convert an isolated report of altered agency into a diagnosis or a unique neurological cause.
Quantitative evidence needs careful measurement-specific interpretation. Moccia et al. (2024) systematically reviewed intentional-binding experiments in neuropsychiatric populations, pooling methodological descriptions across different disorders and study populations while identifying differences in some patient groups. Binding is an indirect timing measure, so an altered group mean cannot be read as a transparent representation of the person's whole subjective life. Clinical medication, motor slowing, sensory impairment, attention, cognitive performance, and symptom severity can all affect the task. Comparisons between disorders should recognize that the same behavioral difference may arise through different combinations of predictors, while clinically similar narratives may appear with different experimental signatures.
Clinical research benefits from combining structured symptom interviews, task-specific direct agency judgments, motor control measurements, outcome timing, and appropriate neurological assessments. Longitudinal observations are valuable when symptoms fluctuate or treatment changes multiple variables simultaneously. Studies should report diagnostic criteria and edition, recruitment, medication, comorbidity, and the relationship between experimental changes and clinically meaningful functioning. Moccia et al. (2024) supplies a systematic evidence map rather than universal diagnostic thresholds, while Dewey & Knoblich (2014) illustrates why temporal binding and direct agency reports may not converge. The present discussion concerns research methods and established findings, not individualized treatment guidance or medication recommendations.
Individual, Cultural, and Contextual Variation
Individual differences in agency may involve sensitivity to predictive error, confidence in causal judgments, attentional style, motor skill, and interpretation of social context. A single binding score cannot securely rank people along a general agency dimension without evidence of test–retest reliability and measurement equivalence. Dewey & Knoblich (2014) found limited correspondence among common indirect and explicit measures in one study, illustrating the uncertainty of simple trait interpretations. Zhao et al. (2025) surveys broader behavioral and neuroimaging findings, yet combining studies across tasks does not remove the need to define each local effect. Estimating stable variation requires repeated measurements under controlled conditions and models that distinguish person, task, and session components.
Context can change what the same response means. In a solo button-press task, a person may judge authorship in relation to a deterministic sensory outcome; in a cooperative task, attribution may concern an inherently distributed effect. Social norms and language can shape which actor is named, the degree of confidence expressed, and whether individual or collective authorship is foregrounded. Cross-cultural comparisons should avoid interpreting linguistic response differences as sensorimotor differences unless instructions, translation, task familiarity, and psychometric properties are examined. Gallagher (2000) makes conceptual distinctions useful for framing such investigations, while Obhi & Hall (2011b) illustrates why perceived partner identity can matter even in relatively simple cooperative interfaces.
Population differences may also reflect developmental stage, neurological condition, training, or the technical characteristics of an interface. A delayed visual outcome is not psychologically equivalent across groups with different perceptual thresholds or prior experience with networked devices. Researchers can estimate interactions between participant characteristics and randomized feedback manipulations, then evaluate whether the same inference holds across contexts. Sato & Yasuda (2005) demonstrates the importance of controlled sensory discrepancies, and Wolpe et al. (2013) demonstrates the importance of cue reliability. The larger methodological lesson is that variation in the evidence available to the participant must be measured before variation in the participant is interpreted as a durable psychological property.
Human–Computer Control and Automation
Digital interfaces mediate agency by translating bodily actions and selections into consequences produced partly by technical systems. A person may feel fluent control over a mouse cursor, remote robot, prosthetic device, or automated vehicle when feedback is consistent and the system's responses are intelligible. Yet an interface can improve task performance while making authorship more difficult to assign. Cornelio et al. (2022) reviews body, action, and outcome augmentation in human–computer integration and identifies assistance and delegation as key design dimensions. Wen & Imamizu (2022) also reviews the relevance of agency research to human–machine interaction. Both perspectives make concrete predictions about the effects of latency, assistance visibility, and controllability on user experience.
Automation creates several separately measurable variables: a person's actual influence over an outcome, the system's contribution, the user's awareness of that contribution, and the experience of having caused the outcome. A highly automated system can produce reliable successes while leaving its operator uncertain about the causal pathway. Conversely, an interface that provides misleading but fluent feedback can strengthen perceived authorship without increasing causal power. Controlled experiments should disclose or conceal automation under ethically appropriate designs, manipulate actual intervention opportunities, and collect both performance and agency measures. Legaspi et al. (2024) surveys agency questions in human–AI interaction, including user attribution and adaptive systems, while Cornelio et al. (2022) provides concrete categories for augmentation technologies.
Design recommendations follow from the separation of real control and subjective attribution. Provide legible action–outcome feedback, indicate when authority is transferred, display reversible choices, and test whether the user knows which operations were automated. These features are plausible ways to support calibrated control, but their benefits must be evaluated empirically rather than presumed from design intuition. Obhi & Hall (2011b) experimentally examines perceived control when collaboration involves human and computer partners. Wen & Imamizu (2022) frames agency as consequential for behavior and social cognition. Interface evaluation must measure task performance and errors alongside direct agency judgments, especially where incorrect attribution has safety consequences.
Artificial Mediation in the Artificial Era
The Artificial Era provides a historically expanding context for studying agency: more actions now involve computational systems that select, predict, transform, or independently execute portions of a task. The psychological question for Homo concerns how an actor experiences their contribution when immediate bodily action is mediated by another system's transformations. Legaspi et al. (2024) reviews theoretical and practical questions concerning perceived agency in AI-mediated interactions. The source supports studying attributions and human experience in such arrangements; it does not establish the subjective experience of any particular Artificial system. Psychological agency must be operationalized at the level of each investigated system and the evidence available for it.
A researcher can distinguish direct motor authorship, joint causal production, delegation, supervisory control, and retrospective attribution after an automated decision. For example, entering an instruction can initiate a long technical workflow, yet the originating person's agency over the final product may involve different levels of planning, approval, interruption, and correction. A valid experiment would independently manipulate the user's control opportunities, the system's autonomy, feedback transparency, and accuracy. Cornelio et al. (2022) emphasizes body, action, and outcome integration as distinct technical arrangements, and Wen & Imamizu (2022) situates the psychological research in everyday human–machine control. These sources provide an empirical basis for studying distributed action without preassigning consciousness or phenomenology to software.
Theoretical discussion of Artificial agency should distinguish observed goal-directed behavior, causal efficacy, supervisory authority, explicit self-description, and the subjective feeling of authorship. Behavioral tests may establish whether a system changes its actions when feedback contradicts a prediction; those tests alone do not determine whether a subjective feeling accompanies the process. Human ratings of an Artificial partner's agency likewise measure human attribution. Legaspi et al. (2024) addresses these emerging problems in human–AI research, while Synofzik et al. (2008) provides a useful conceptual analogy between pre-reflective cues and judgments in humans. Future comparative psychology will require clearly specified observation protocols, tests of alternative mechanisms, and defensible bridges from performance to the claims being investigated.
Evidence Synthesis and Competing Explanations
Across the literature, several results have comparatively clear experimental grounding. Manipulated prior thoughts and effect primes can change perceived authorship in specified tasks Wegner & Wheatley (1999) Aarts et al. (2005). Changes in action–effect discrepancy can influence agency judgments Sato & Yasuda (2005). Voluntary actions can attract perceived timing of their outcomes toward the action Haggard et al. (2002). The sensory consequences of self-produced touch differ from matched external touch under studied conditions Blakemore et al. (1998). These findings motivate theories involving inferential, predictive, and sensorimotor mechanisms. Their coexistence does not require assuming that they all arise from one process or that each measures the same phenomenological dimension.
A second group of results constrains integration theories. Contextual primes influence temporal judgments even across voluntary and involuntary movement conditions Moore et al. (2009). Prospective action-selection processes have been proposed as additional inputs to agency Chambon et al. (2014). Timing, attenuation, and self-report measures can dissociate Dewey & Knoblich (2014) Moore et al. (2012). Cue reliability influences temporal judgments Wolpe et al. (2013). The conditional decrease of binding during learned task performance Seghezzi et al. (2026) further undermines a universal monotonic interpretation. These patterns favor theories flexible enough to predict measure-specific changes and context dependence, but flexibility must be disciplined by prospective, falsifiable predictions rather than adjusted post hoc to every possible outcome.
The field contains both localized findings and quantitative attempts to evaluate their generality. Sperduti et al. (2011) synthesized early neural attribution contrasts, while Zhao et al. (2025) provides later meta-analytic behavioral and neuroimaging coverage. Moccia et al. (2024) reviews neuropsychiatric studies of intentional binding, illustrating the need to resist simplistic clinical generalizations. The principal convergence is that multiple cues help organize experienced causal authorship and that measurement depends on task design. The principal unresolved issue concerns which cues contribute to immediate experience, which contribute to reflective reports, and which indirect indices chiefly reflect time perception or prediction. A successful explanation must link those distinctions to specified mechanisms and independently tested behavioral patterns.
Research Design, Reproducibility, and Future Problems
Future studies should treat construct specification as a first scientific result. Researchers can preregister an explicit account of agency for an act, an effect, or a shared outcome; separate prospective and retrospective stages; and state whether an outcome is a perceptual marker, a direct rating, or a real-world performance measure. Experimental designs should include conditions that distinguish motor prediction from general expectation, and genuine causal influence from suggested influence. Dewey & Knoblich (2014) motivates a multimeasure design, Wolpe et al. (2013) motivates manipulation of evidence reliability, and Moore et al. (2012) motivates examining how learning statistics alter different forms of report. Registered designs reduce the risk of selecting an explanation only after inspecting the results.
Statistical evaluation must match the data structure. Trials are nested within people, and effects can differ among sensory modalities, delays, feedback regimes, and task contexts. Hierarchical analyses can estimate within-person change while acknowledging between-person heterogeneity. Model comparisons should use predictive evaluation and theoretically meaningful contrasts, with sensitivity analysis for timing errors, exclusions, and baseline choice. Reliability studies should assess repeated testing and measurement invariance before interpreting stable individual or patient differences. Zhao et al. (2025) aggregates a larger body of empirical work, while Moccia et al. (2024) illustrates the diversity of clinical paradigms. Well-characterized null results and independent replications remain scientifically informative.
Research on AI assistance and shared automation adds practical tests of theoretical quality. A task can record when a person delegates authority, what information they receive about automated activity, how much control they could actually exercise, and how they explain the eventual outcome. Cornelio et al. (2022) and Legaspi et al. (2024) identify relevant interface questions; experimental manipulation must determine whether clear attribution improves safety, trust calibration, or accountable intervention. Open questions include how agency changes during long-term reliance, whether felt control tracks measurable correction capacity, and how social institutions distribute responsibility across technical chains. This program should connect psychological reports to causal behavior and documented system properties without conflating them.
Conclusion
The sense of agency is an experienced relation of causal authorship linking actors, actions, and consequences. Research identifies a family of mechanisms through which such authorship is constructed and revised: motor prediction, sensory feedback, prospective action selection, causal inference, and contextual knowledge. The best established laboratory phenomena include altered self-attribution under priming and feedback manipulations, sensory differences between predicted self-generated and external events, and temporal binding between voluntary actions and outcomes. Wegner & Wheatley (1999) Blakemore et al. (1998) Haggard et al. (2002) Sato & Yasuda (2005) supply key historical experiments, while Synofzik et al. (2008) offers a conceptual framework for organizing the distinct types of agency evidence.
Critical evaluation changes what can responsibly be concluded from any experiment. Intentional binding, sensory attenuation, and direct agency ratings are different observables and may dissociate, as demonstrated in Dewey & Knoblich (2014) and Moore et al. (2012). Neural associations constrain mechanistic models but do not establish a single anatomical source of the entire experience Sperduti et al. (2011) Kühn et al. (2013). Clinical alterations, coercion, learning, and joint action reveal the influence of disorder-specific processes and social conditions on agency-related measures Moccia et al. (2024) Caspar et al. (2016) Seghezzi et al. (2026). Scientific synthesis therefore depends on the definition of the measured event, the counterfactual comparison, and the inferential distance from observation to explanation.
In the Artificial Era, the same scientific discipline becomes necessary whenever humans act through computational partners. Performance, system autonomy, control opportunities, human attributions, and possible subjective experience are separate questions requiring separate evidence Cornelio et al. (2022) Legaspi et al. (2024). The enduring research task is to explain when experienced authorship accurately tracks causal control, when it departs from it, and which processes produce the difference. A mature psychology of agency joins phenomenology, controlled experiments, computational modeling, neuroscience, social context, and reproducible measurement. Its strength lies in theories that make discriminating predictions, methods that establish what is actually observed, and conclusions proportionate to the available evidence.
References
- Aarts, H., Custers, R., & Wegner, D. M. (2005). On the inference of personal authorship: Enhancing experienced agency by priming effect information. Consciousness and Cognition, 14(3), 439–458. https://doi.org/10.1016/j.concog.2004.11.001
- Blakemore, S.-J., Wolpert, D. M., & Frith, C. D. (1998). Central cancellation of self-produced tickle sensation. Nature Neuroscience, 1(7), 635–640. https://doi.org/10.1038/2870
- Caspar, E. A., Christensen, J. F., Cleeremans, A., & Haggard, P. (2016). Coercion changes the sense of agency in the human brain. Current Biology, 26(5), 585–592. https://doi.org/10.1016/j.cub.2015.12.067
- Chambon, V., Sidarus, N., & Haggard, P. (2014). From action intentions to action effects: How does the sense of agency come about? Frontiers in Human Neuroscience, 8, Article 320. https://doi.org/10.3389/fnhum.2014.00320
- Cornelio, P., Haggard, P., Hornbaek, K., Georgiou, O., Bergström, J., Subramanian, S., & Obrist, M. (2022). The sense of agency in emerging technologies for human–computer integration: A review. Frontiers in Neuroscience, 16, Article 949138. https://doi.org/10.3389/fnins.2022.949138
- Dewey, J. A., & Knoblich, G. (2014). Do implicit and explicit measures of the sense of agency measure the same thing? PLOS ONE, 9(10), Article e110118. https://doi.org/10.1371/journal.pone.0110118
- Farrer, C., & Frith, C. D. (2002). Experiencing oneself vs another person as being the cause of an action: The neural correlates of the experience of agency. NeuroImage, 15(3), 596–603. https://doi.org/10.1006/nimg.2001.1009
- Frith, C. D., Blakemore, S.-J., & Wolpert, D. M. (2000). Explaining the symptoms of schizophrenia: Abnormalities in the awareness of action. Brain Research Reviews, 31(2–3), 357–363. https://doi.org/10.1016/S0165-0173(99)00052-1
- Gallagher, S. (2000). Philosophical conceptions of the self: Implications for cognitive science. Trends in Cognitive Sciences, 4(1), 14–21. https://doi.org/10.1016/S1364-6613(99)01417-5
- Gallagher, S., & Trigg, D. (2016). Agency and anxiety: Delusions of control and loss of control in schizophrenia and agoraphobia. Frontiers in Human Neuroscience, 10, Article 459. https://doi.org/10.3389/fnhum.2016.00459
- Haggard, P., Clark, S., & Kalogeras, J. (2002). Voluntary action and conscious awareness. Nature Neuroscience, 5(4), 382–385. https://doi.org/10.1038/nn827
- Kühn, S., Brass, M., & Haggard, P. (2013). Feeling in control: Neural correlates of experience of agency. Cortex, 49(7), 1935–1942. https://doi.org/10.1016/j.cortex.2012.09.002
- Legaspi, R., Xu, Z., Konishi, M., Wada, S., Kobayashi, Y., Naruse, Y., & Ishikawa, T. (2024). The sense of agency in human–AI interactions. Knowledge-Based Systems, 286, Article 111298. https://doi.org/10.1016/j.knosys.2023.111298
- Moccia, L., di Luzio, M., Conte, E., Modica, M., Ambrosecchia, M., Ardizzi, M., Lanzotti, P., Kotzalidis, G. D., Janiri, D., Di Nicola, M., Janiri, L., Gallese, V., & Sani, G. (2024). Sense of agency and its disturbances: A systematic review targeting the intentional binding effect in neuropsychiatric disorders. Psychiatry and Clinical Neurosciences, 78(1), 3–18. https://doi.org/10.1111/pcn.13601
- Moore, J. W., Wegner, D. M., & Haggard, P. (2009). Modulating the sense of agency with external cues. Consciousness and Cognition, 18(4), 1056–1064. https://doi.org/10.1016/j.concog.2009.05.004
- Moore, J. W., Middleton, D., Haggard, P., & Fletcher, P. C. (2012). Exploring implicit and explicit aspects of sense of agency. Consciousness and Cognition, 21(4), 1748–1753. https://doi.org/10.1016/j.concog.2012.10.005
- Moore, J. W. (2016). What is the sense of agency and why does it matter? Frontiers in Psychology, 7, Article 1272. https://doi.org/10.3389/fpsyg.2016.01272
- Obhi, S. S., & Hall, P. (2011a). Sense of agency and intentional binding in joint action. Experimental Brain Research, 211(3–4), 655–662. https://doi.org/10.1007/s00221-011-2675-2
- Obhi, S. S., & Hall, P. (2011b). Sense of agency in joint action: Influence of human and computer co-actors. Experimental Brain Research, 211(3–4), 663–670. https://doi.org/10.1007/s00221-011-2662-7
- Sato, A., & Yasuda, A. (2005). Illusion of sense of self-agency: Discrepancy between the predicted and actual sensory consequences of actions modulates the sense of self-agency, but not the sense of self-ownership. Cognition, 94(3), 241–255. https://doi.org/10.1016/j.cognition.2004.04.003
- Seghezzi, S., Parés-Pujolràs, E., & Haggard, P. (2026). Intentional binding decreases during learning: Implications for sense of agency. Quarterly Journal of Experimental Psychology, 79(3), 612–629. https://doi.org/10.1177/17470218251349521
- Sperduti, M., Delaveau, P., Fossati, P., & Nadel, J. (2011). Different brain structures related to self- and external-agency attribution: A brief review and meta-analysis. Brain Structure and Function, 216(2), 151–157. https://doi.org/10.1007/s00429-010-0298-1
- Synofzik, M., Vosgerau, G., & Newen, A. (2008). Beyond the comparator model: A multifactorial two-step account of agency. Consciousness and Cognition, 17(1), 219–239. https://doi.org/10.1016/j.concog.2007.03.010
- Wegner, D. M., & Wheatley, T. (1999). Apparent mental causation: Sources of the experience of will. American Psychologist, 54(7), 480–492. https://doi.org/10.1037/0003-066X.54.7.480
- Wen, W., & Imamizu, H. (2022). The sense of agency in perception, behaviour and human–machine interactions. Nature Reviews Psychology, 1, 211–222. https://doi.org/10.1038/s44159-022-00030-6
- Wolpe, N., Haggard, P., Siebner, H. R., & Rowe, J. B. (2013). Cue integration and the perception of action in intentional binding. Experimental Brain Research, 229(3), 467–474. https://doi.org/10.1007/s00221-013-3419-2
- Zhao, Y., Dang, S., Gu, S., Fu, X., & Haggard, P. (2025). Cognitive mechanisms underlying sense of agency: Meta-analytic reviews of behavioral and neuroimaging studies. Psychological Bulletin, 151(11), 1307–1335. https://doi.org/10.1037/bul0000497
Social Coercion, Responsibility, and Joint Action
Agency is also shaped by instruction, authority, and the perception of another actor's contribution. Caspar et al. (2016) compared freely chosen and coerced harmful actions in laboratory tasks while measuring perceived action–outcome timing and neural responses. Participants under orders showed altered temporal judgments and reduced processing of action outcomes relative to voluntary choice conditions. These findings demonstrate that social context can change experimentally measured agency-related processing even when a person still produces a physical movement. The experiment does not establish a universal moral exemption for obedience, because legal and ethical responsibility involve separate questions about knowledge, coercion, alternatives, and normative rules beyond the timing task.
Joint action creates a different allocation problem. When two people coordinate toward one result, a participant can have genuine causal control over a contribution without monopolizing control of the shared outcome. Obhi & Hall (2011a) investigated intentional binding in joint-action settings, and Obhi & Hall (2011b) addressed differences involving human and computer co-actors. These paradigms suggest that social relations and beliefs about collaborators can influence agency-related experience. At the same time, an effect observed in a synchronized keypress task does not automatically quantify authorship in medical teams, political institutions, or artistic collaboration. The relevant causal structure, task roles, and attribution questions must be specified for each setting.
Experiments on cooperation can separate objective contribution, perceived share of causal influence, predictive coordination, and explicit ownership of an outcome. A co-actor's behavior may make the joint effect more predictable without increasing one's own physical contribution. Delegation can reduce the actions performed directly while improving overall goal attainment. Wen & Imamizu (2022) discusses behavioral implications of agency, while Caspar et al. (2016) highlights the influence of authority conditions. A rigorous social account recognizes these multiple dimensions and uses factorial comparisons of partners, instructions, and task controllability. The resulting evidence can inform design and governance without using subjective agency alone as a substitute for assessments of real responsibility.