You press a key and the letter appears.
Most of the time there is no philosophical interval between the movement and the result. You do not consciously calculate whether the finger movement caused the mark on the screen. The action feels like yours, and the consequence feels linked to what you just did.
Now add a half-second delay.
Or let the cursor move slightly differently from the direction you intended.
Or imagine that an automated system silently corrects the result before you see it.
The question becomes noticeable: Did I do that?
That question belongs to the sense of agency.
Definition
The sense of agency is the experience of oneself as the initiator or controller of one's actions and, through them, of events in the world.
It is a component of self-consciousness, but it is not identical to voluntary movement, intention, responsibility or body ownership.
A person can make a voluntary movement without experiencing complete control over its consequence. A person can experience an outcome as self-caused even when the causal chain is noisy or partially mediated. A person can own a hand without controlling its movement.
Research therefore treats agency as a constructed experience informed by several cues rather than as a direct readout of motor commands.
Why agency feels immediate
Action normally generates a dense stream of predictive and sensory information.
Before movement, the motor system prepares commands and expected consequences. During and after movement, proprioceptive, visual, tactile and auditory signals report what happened. The environment supplies feedback. Goals and intentions supply context.
When these sources align, agency is usually effortless.
When they diverge, agency can become uncertain.
A light switches on after your hand moves toward the control. Did you cause it, or did someone else trigger it remotely? A document changes after you type. Was that your edit, an autocorrection or an AI rewrite? Modern interfaces make these attribution problems increasingly ordinary.
Comparator models and retrospective inference
One influential tradition explains agency through comparisons between predicted and actual sensory consequences of action.
If the result resembles what the motor system predicted, the event is more likely to feel self-caused. If there is a mismatch, agency can weaken.
This family of accounts has substantial explanatory power, especially for action-effect congruence.
But it is not the whole story.
People also infer agency retrospectively from contextual cues. If an outcome matches a goal, occurs at the expected time and no plausible external cause is visible, it may feel more self-produced. Conversely, an unexpected outcome can weaken agency even when the motor command was entirely one's own.
Nicole David, Albert Newen and Kai Vogeley argued in 2008 for a multifactorial, multilevel account precisely because no single cue explains the full phenomenon.
Prospective agency begins before the outcome
Agency is not only reconstructed after seeing what happened.
Valerian Chambon and Patrick Haggard showed that the fluency of selecting an action can influence the sense of control. When action selection is facilitated, control can feel stronger even before outcome comparison has done all the explanatory work.
This introduces a prospective component.
The system has information about how readily an action became available, how coherent the intention was and how fluently selection unfolded.
The distinction matters because it prevents agency from being reduced to a post hoc story.
Part of the feeling of control can arise from the process of choosing itself.
Agency is not body ownership
Body ownership asks whether a body or body part feels like mine.
Agency asks whether I am causing an action or effect.
The two usually coincide in everyday movement, which makes them easy to confuse.
Experimental work can separate them. A rubber or virtual hand may be experienced as owned under synchronous multisensory stimulation even without active movement. Conversely, a person can control an external tool or cursor without experiencing it as part of the body.
This distinction is particularly important for prosthetics, robotics, avatars and human-computer interaction.
A successful interface may increase control without increasing ownership, or vice versa.
Agency is not moral or legal responsibility
The phrase "sense of agency" can sound morally loaded.
In experimental psychology, it is primarily a perceptual and cognitive construct.
Feeling that one caused an event is not the same as being legally responsible for it. Failing to feel agency does not prove absence of causal involvement. Feeling agency does not prove sole authorship.
This boundary becomes especially important when automated systems share control.
A driver may feel in charge of a semi-automated vehicle while important decisions are being made by software. An editor may feel authorship over text that has been heavily rewritten by a model. Conversely, a person may underestimate how much their initial instruction shaped a system's result.
Phenomenology and causal responsibility can diverge.
Measuring agency
Researchers use several approaches.
The most direct is self-report: participants rate how much control or authorship they felt over an action or outcome.
Other paradigms manipulate action-effect delays, congruence or predictability and measure changes in agency judgments.
A widely used implicit measure has been intentional binding, the apparent compression of perceived time between a voluntary action and its outcome.
Because voluntary actions and their effects can seem temporally closer together, intentional binding was often interpreted as an implicit signature of agency.
That interpretation is now contested.
The intentional-binding problem
A 2024 study directly compared voluntary movement, passive movement and external sensory triggers while matching predictability more carefully than many earlier designs.
It found no evidence that temporal compression was specific to voluntary action. Instead, binding appeared to arise from sensory integration and temporal prediction more generally.
The implication is important.
An effect can be robust while its interpretation is wrong.
If temporal binding occurs without intentional action, it cannot by itself serve as a clean measure of agency.
This does not erase decades of agency research. It strengthens the case for using multiple measures and for separating the subjective construct from any one proxy.
Responsibility feelings and binding can diverge
Other recent work has similarly shown that explicit feelings of responsibility and temporal binding do not necessarily move together.
That divergence reinforces a general methodological rule: agency is not whatever an agency task happens to measure.
A task can capture temporal prediction, causal inference, motor fluency or demand characteristics alongside agency.
The more important the conclusion, the more carefully the proxy needs to be defended.
Agency and prediction
Prediction remains central even if no single comparator model is sufficient.
Actions normally have expected consequences. Learning strengthens those expectations. When a system makes outcomes predictable, agency often becomes easier to attribute.
But predictability can also produce false confidence.
A highly automated tool can learn a user's patterns so well that its output feels like an extension of intention even when the system contributed substantial transformations.
That possibility is scientifically and socially important.
Sense of agency is a phenomenological variable. It is not a perfect audit trail of causation.
Joint action and shared agency
Human action is often collective.
Two people lift a table. A team writes a report. A surgical group coordinates a procedure. A band improvises.
Researchers have explored whether there is a distinct "we-agency" or shared sense of control in joint action.
Evidence suggests that coordination, prediction and shared goals can shape agency judgments, but the field remains conceptually unsettled. It is not always clear whether "we-agency" is a distinct phenomenological state or an attribution built from individual agency plus social inference.
The uncertainty matters for institutions.
Collective work can produce outcomes for which everyone feels partly responsible and no one feels fully responsible. Agency research can illuminate that experience without deciding the normative question of who should be accountable.
Agency under automation
Artificial intelligence makes agency attribution newly practical.
When a system completes a sentence, ranks options, selects a route, filters applicants or proposes a decision, human and machine contributions become intertwined.
A user may initiate the process but not understand how the result was generated. They may approve an output without having produced its reasoning. They may experience strong agency because the interaction is fluent and responsive.
This creates an important distinction between felt control, effective control, causal contribution, understanding and responsibility.
Agency research does not tell us how governance should assign accountability. It does tell us that subjective control cannot safely be treated as proof that meaningful control existed.
Agency at different levels
Agency can be asked at more than one level.
At the motor level, did I initiate this movement?
At the effect level, did my movement cause this sound, visual change or object movement?
At the goal level, did the outcome correspond to what I intended?
At the authorship level, do I regard the resulting product as something I made?
At the social level, was this outcome mine, ours, or imposed by someone else?
Experimental paradigms often isolate one level for good reason. Problems arise when results are generalized upward without evidence.
A motor-action task can reveal something about action-effect attribution. It does not by itself determine whether a person has meaningful authorship over a complex decision.
Disorders and unusual experiences of agency
Altered agency has been studied in schizophrenia, functional neurological symptoms and other conditions, as well as after neurological injury.
These literatures are important because they show that agency experience can diverge markedly from ordinary expectations.
They also require caution.
Group-level differences on an agency task are not diagnostic markers for an individual. Different disorders can produce superficially similar task effects through different mechanisms, and medication, attention, symptom state and motor performance can all influence results.
A systematic review of intentional binding in neuropsychiatric disorders has found heterogeneous effects, which is exactly what one would expect from a measure that may mix temporal prediction with agency.
The appropriate conclusion is that altered agency is scientifically relevant across conditions, not that one task identifies a disease.
Agency and error correction
Agency is easier to study when something goes wrong.
If a cursor deviates from the intended path, a person may compensate without consciously noticing small perturbations. Larger deviations become explicit and can be attributed externally.
This suggests graded monitoring rather than a single binary judgment.
Error correction can occur below conscious awareness, while explicit agency judgments are made afterward from a broader set of cues.
This is another reason the sense of agency should not be treated as a direct readout of motor computation. Multiple levels of monitoring contribute at different times.
Agency, habit and automaticity
Skilled action often becomes automatic.
A pianist does not consciously choose each finger movement. A practiced driver does not verbally plan every steering correction. Fluent action can feel highly agentic even when conscious deliberation is minimal.
This complicates simplistic equations between agency and conscious control.
More deliberation is not always more agency. In skilled action, excessive conscious control can even disrupt performance.
The relevant question is whether the person can appropriately initiate, monitor, interrupt and revise action when needed.
That is a richer capacity than continuous conscious supervision.
Agency and delegation
Delegation is another useful boundary case.
A manager can intentionally delegate a task and remain accountable for the decision without controlling every action performed by the delegate. A user can ask an AI system to draft text and retain some authorship through goals, selection and revision while not having generated every sentence.
This shows why agency should be represented as structured contribution rather than a simple yes/no variable.
The more complex the system, the more useful it becomes to distinguish intention, initiation, execution, monitoring, correction, approval and responsibility.
Contestability as an agency condition
For technology-mediated action, the ability to contest an outcome may be as important as the feeling of control.
If a system makes a recommendation but the user can inspect evidence, reject the result and choose another path, meaningful agency may remain high even if the machine contributed substantially.
If the user feels fully in control but cannot understand, challenge or reverse the system's action, felt agency may exceed effective agency.
This distinction has direct relevance to human oversight in AI governance. The science of agency does not supply the policy standard by itself, but it explains why subjective confidence is an insufficient test of control.
What the evidence does not show
The evidence does not show that the sense of agency is generated by one brain region or one comparator mechanism.
It does not show that intentional binding is a universally valid implicit measure of agency.
It does not show that feeling in control means a person actually controlled the relevant causal process.
It does not show that disrupted agency experience automatically indicates psychiatric disorder.
It does not show that agency and body ownership are the same construct.
It does not show that more agency is always better. Skilled collaboration, delegation and trust can involve intentionally shared control.
And it does not show that an automated system preserves human agency merely because a human clicked the final button.
Boundary conditions and open questions
One open question is how different cues are weighted. Motor prediction, action selection, sensory consequences, prior beliefs and social context may matter differently across tasks.
A second concerns measurement. The field needs converging methods that distinguish agency from temporal prediction and generic causal inference.
A third concerns development and culture. Expectations about authorship, control and collective action vary across social settings, yet much experimental work relies on narrow laboratory samples.
A fourth concerns impairment and adaptation. Neurological and psychiatric conditions can alter agency experience, but heterogeneity is substantial and task effects do not map cleanly onto diagnosis.
A fifth concerns machine mediation. As AI increasingly operates between intention and outcome, classic agency models will need to accommodate causal chains in which human goals, automated inference and environmental feedback are tightly interleaved.
Institute synthesis: agency needs an audit trail
The Institute distinguishes experienced agency from accountable agency.
This is an Institute synthesis rather than an established scientific taxonomy.
Experienced agency asks whether an action or outcome feels self-caused. Accountable agency asks whether a person can meaningfully trace, understand, contest and take responsibility for the process that produced an outcome.
The distinction becomes useful when technology increases fluency.
A system can make action feel easier while making causal structure harder to see.
That means subjective ease should not be the only measure of human agency in a technological environment.
What this changes for Inner Technology
Agency is a capacity question because action is where intention meets consequence.
A person can have values, insight and plans while repeatedly discovering that outcomes are being shaped elsewhere: by habit, social pressure, interface design or automated systems.
Inner Technology therefore needs more than a feeling of empowerment.
It needs practices of attribution: What did I choose? What did the system choose? What evidence changed the result? Where could I have intervened? Could I explain or contest what happened?
The goal is not maximum control over everything.
It is accurate participation in the things for which we are prepared to say, I did this.
Editorial limitation note
This article is a research reference, not a diagnostic or treatment page. It does not infer an individual capacity, disorder or cause from a single subjective experience, questionnaire score, laboratory task or neural finding. Institute interpretations are labelled as synthesis and should be read as disciplined editorial inference, not as settled scientific constructs.
Research references
- David N, Newen A, Vogeley K. The "sense of agency" and its underlying cognitive and neural mechanisms. Consciousness and Cognition. 2008;17(2):523-534. DOI: 10.1016/j.concog.2008.03.004. https://pubmed.ncbi.nlm.nih.gov/18424080/
- Chambon V, Haggard P. Sense of control depends on fluency of action selection, not motor performance. Cognition. 2012;125(3):441-451. DOI: 10.1016/j.cognition.2012.07.011. https://pubmed.ncbi.nlm.nih.gov/22902284/
- Tsakiris M, Longo MR, Haggard P. Having a body versus moving your body: neural signatures of agency and body-ownership. Neuropsychologia. 2010;48(9):2740-2749. DOI: 10.1016/j.neuropsychologia.2010.05.021. https://pubmed.ncbi.nlm.nih.gov/20510255/
- No evidence in favor of the existence of "intentional" binding. Psychological Science. 2024. PubMed PMID 38635224. https://pubmed.ncbi.nlm.nih.gov/38635224/
- Botvinick M, Cohen J. Rubber hands 'feel' touch that eyes see. Nature. 1998;391:756. DOI: 10.1038/35784. https://doi.org/10.1038/35784