NOTE-827: Haggard — Sense of agency in the human brain

Contribution

The review fixes a narrow sense of “sense of agency”: the experience of controlling a specific motor act and, through it, an external outcome. It sets this apart from the felt capacity to act (self-efficacy) and from the social attribution of an action to an agent. It then argues against the prevailing view that the experience is computed only retrospectively. It cites evidence that action selection and outcome prediction contribute before the outcome arrives, and that this prospective part is what psychosis loses.

Key insight

The feeling of having made something happen is not a readout of a single prediction-error comparator. It is assembled from signals available at different times: premotor signals of selecting and initiating an action, predictions of its outcome, and feedback once the outcome arrives. The early signals can raise the feeling even when they carry no information about actual control, and can sustain it on trials where the expected outcome never comes.

Assumptions

• The construct is the instrumental, motor one. “I use the term sense of agency to refer to an experience that accompanies the performance of a specific motor act” (p. 2). The social-science and computational uses, “a felt capacity to act (self-efficacy), without reference to any specific motor act”, are set aside. Bandura 1982 is cited for the former (ref. 1).
• Implicit measures stand in for the experience. A difference in intentional binding between two well-chosen conditions “can potentially be interpreted as a difference in sense of agency”. Shifts in time perception alone “are not diagnostic”, since attention, causality, drugs and adaptation also move them (p. 6).
• Voluntariness is taken as given though contested. “The precise definition of a voluntary action remains controversial” (p. 4).
• A review. Every finding below is reported from cited studies, often the author’s own group’s. Sample sizes and effect sizes are not given in the text.

Key results

• Definition (pp. 3–5). Agency involves ownership of the body movement (“my body moved”) and “the cognitive experience of voluntary control over that body movement (‘I voluntarily made it move’)”. Involuntary movements (reflexes, brain stimulation) produce the peripheral experiences and ownership “but not central experiences”, and “are never accompanied by a sense of agency”. “The core of sense of agency, therefore, is the association between a voluntary action and an outcome.”
• Explicit judgements are biased (p. 5). There is “a tendency to overestimate one’s own agency”, stronger “when the outcome of an action is positive”. Implicit and explicit measures “tend to be only weakly correlated across individuals”, though both respond to reward, action–outcome interval and the actual degree of control.
• Intentional binding (p. 6, Fig. 2). Voluntary actions are perceived as shifted towards their outcomes, and outcomes towards the actions. The effect is measured against baselines where action or tone occurs alone; in the standard task the tone follows the keypress after 250 ms. A TMS-evoked twitch “has the opposite effect”. In the interval-estimate variant, intervals are always under one second, and shorter estimates mean more agency.
• Instrumental agency and social attribution are distinct (p. 6). “It remains unclear whether social attribution of actions to agents depends on an antecedent ability to compute and perceive instrumental agency, or vice versa.”
• Action preparation (p. 7). A stronger early readiness potential went with stronger binding. Temporal coincidence of a TMS twitch with ongoing preparation did not produce binding: “The cognitive preparation of an action plan needs to match the muscular movement precisely”.
• Action selection (pp. 7–9).
  • A meta-analysis of seven studies found that anodal tDCS over left DLPFC enhanced binding “only in those studies in which participants themselves selected between alternative actions”.
  • Subliminal primes that eased or hindered selection moved explicit agency judgements over an outcome they did not predict. “The contribution of action selection fluency to sense of agency is, in a sense, illusory”, but “may be a rule of thumb that normally works”.
• The comparator model and its limits (pp. 9–11, Fig. 3). A forward model predicts the sensory consequences of the motor command from an efference copy. A zero prediction error means “the event is caused by one’s own action”. Three objections are made:
  • The model “may successfully explain the phenomena of ‘non-agency’” but not the positive “buzz”, since it “does not generate any neural signal that could cause this experience”.
  • It attenuates the perception of the very outcomes that are the goals of action.
  • People can feel agency over arbitrary outcomes “long after the action occurred”.
• The prospective component (pp. 11–12, Fig. 4). When a tone followed the action on 50% of trials, action binding was stronger on trials where the tone came, a retrospective effect. On trials with no tone, binding was stronger in blocks where the tone was likely (75%) than where it was less likely (50%). So “an additional, prospective component of the sense of agency is present even if the outcome does not in fact occur”.
• Brain mechanisms (pp. 12–14).
  • A meta-analysis of 15 neuroimaging studies found the temporoparietal junction, including the angular gyrus, as the correlate of “non-agency”. The only area “consistently associated with positive self-agency was the anterior insula”.
  • The angular gyrus was more active for low agency ratings only after incompatible primes, and TMS over it abolished the effect of the primes. “The parietal cortex may not simply match outcomes to actions retrospectively.”
  • tDCS and theta-burst TMS over pre-SMA reduced binding.
  • The comparator-in-parietal-cortex view is doubted, because parietal activity depends on the result of the comparison more than on its process. “The key neural correlate of sense of agency might lie in the connectivity between frontal and prefrontal motor areas that initiate action, and parietal areas” (p. 14).
• Pathology (pp. 14–16). Conditions can be hyperagentic or hypoagentic. In schizophrenia, patients’ action binding “depended entirely on the actual occurrence of the tone”, whereas in healthy volunteers it “depended on the probability that the tone might occur”. “Psychosis thus seemed to abolish the normal prospective element of agency.” An earlier small study found binding stronger, not weaker, in patients with positive symptoms.
• Coercion (pp. 16–17, Fig. 5). Participants took turns giving and receiving shocks. Under an experimenter’s instruction, the perceived interval between keypress and tone (synchronous with the shock) was longer than under free choice. The auditory N1 to the tone was reduced “both when the action lead to a painful shock, and when it did not”. “Obeying instructions reduces the sense of agency, producing an experience closer to passive movement than to voluntary action.” This “does not, of course, legitimate the defense of ‘only obeying orders’”.
• Conclusion (p. 18). The circuit “operates both prospectively … as well as retrospectively”. The post hoc inference view “has perhaps been overstated”. “The distinctive experience of initiating a voluntary motor action appears to be necessary, though not sufficient, for sense of agency in normal circumstances.”
• Box 1 (pp. 23–24). In law, someone “who lacks a sense of agency regarding an action cannot be held criminally responsible”, though courts are “rightly skeptical”. The English “loss of control” defence is discussed: rats with prior control over a stressor acquired resilience to later uncontrollable stress, so prolonged abuse might remove that resilience.
• Box 2 (pp. 24–25). Infants learn to become agents. Paralysed users of neuroprostheses report, anecdotally, a feeling of agency over the prosthesis. “The cognitive mechanisms underlying instrumental agency may be almost unlimited in scope.”

Claims

id	claim	strength	support
C1	The sense of agency requires a voluntary motor command; involuntary movement gives ownership without agency	moderate	binding reversed by TMS twitches; review of the cited studies, pp. 3–4, 7
C2	Explicit agency judgements are biased towards self and towards good outcomes, and correlate only weakly with implicit measures	moderate	cited studies, p. 5
C3	Action-selection fluency raises the sense of agency independently of actual control	moderate	subliminal-priming studies and a seven-study tDCS meta-analysis, pp. 7–9
C4	The sense of agency has a prospective component present on trials where the outcome does not occur	moderate	one binding experiment (75% against 50% tone probability), pp. 11–12
C5	In schizophrenia, binding depends on outcome occurrence, not outcome probability	moderate	one larger patient study, p. 16, Fig. 4
C6	The neural correlate is frontal–parietal connectivity, not a single comparator site	weak	an argument from where activations fall, p. 14
C7	Coercive instruction reduces the implicit sense of agency and the neural response to the outcome	moderate	one experiment, pp. 16–17, Fig. 5
C8	The comparator model explains non-agency but not the positive experience of agency	weak	conceptual argument, pp. 10–11

Concepts

• Sense of agency. “The experience of controlling one’s own motor acts and, through them, the course of external events” (p. 2). Not self-efficacy, and not the social attribution of authorship.
• Sense of ownership. “The feeling of ‘mineness’” of a body part or mental state. Present in involuntary movement, where agency is absent.
• Instrumental control. “The capacity to initiate an action, and thus bring about an intended change in the environment” (glossary).
• Intentional binding. The mutual compression in perceived time of a voluntary action and its outcome, used as an implicit measure of agency.
• Prospective and retrospective agency. Contributions available before the outcome (selection, initiation, prediction) and after it (feedback, postdiction).
• Hyperagentic and hypoagentic. Excessive or reduced experience of one’s own causation and control, by analogy with hyperkinetic and hypokinetic movement disorders.
• Volition. “The process of preparing, initiating and executing an action under one’s own control”. Its traditional hallmark is that the agent “could have done otherwise” (glossary).

Connections

The review builds on the comparator account of motor control (Wolpert, Blakemore and Frith) and on Frith’s account of delusions of control. It argues against Wegner’s The Illusion of Conscious Will (2002) on how far agency is post hoc inference, and takes from Synofzik et al. (2008) the point that the comparator cannot produce the positive feeling. It cites Bandura twice. The 1982 paper is cited for the self-efficacy sense of “agency” the review sets aside. The 2001 Annual Review paper (LIT-1005) is cited for the claim that agency is strongest with “a strong motivation to act, a clear action goal and a specific cortical motor command” (p. 4). That last element is not a psychological claim Bandura’s paper is known for, and it could not be checked here.

Bearing on the record

• No existing THEORY holds the review’s findings. A grep of theory.d, claims.d and questions.d for “sense of agency”, “intentional binding”, “comparator”, “efference” and “forward model” found nothing on the experience of agency. Two accounts are filed from this reading:
  • THEORY-322: the sense of agency draws on prospective signals, not only retrospective comparison (C3, C4, C5).
  • THEORY-300: coercive instruction reduces the implicit sense of agency (C7).
• QUESTION-048 asks what report, no-report, deficit and architectural evidence each establish about experience. This review is a worked case for a narrow experience. Explicit reports are biased and only weakly correlated with implicit measures (C2). The implicit measures, though, are not diagnostic alone, only as differences between conditions. The schizophrenia data are deficit evidence that dissociates two components.
• CLAIM-062 (Superseded) kept agency apart from phenomenal experience. Haggard’s “sense of agency” is the phenomenal side of that split: an experience that accompanies agency, can be absent when movement occurs and present when control is lacking. The review supports keeping the two apart, and shows they can dissociate in both directions (C3, C1).
• QUESTION-070 and QUESTION-071. The coercion result (C7) measures an agent’s experience of what she does under another’s instruction. It bears on QUESTION-070’s “whose agency does the wrong run through”: the instructed agent experiences her own act as closer to passive movement. It says nothing about manipulation, which is covert, whereas the instruction here was overt. It does not touch THEORY-040 or THEORY-235, whose subject is the agent’s structure or her access to an influence, not her felt control.
• Will-organization line. Box 1’s resilience point (prior control over a stressor protects against later uncontrollable stress) is reported from rat studies in one sentence and is not filed.
• What the LIT’s summary got wrong. It said the review “sets prediction-based (comparator) accounts against retrospective, inferential ones”. On Haggard’s account the comparator model is itself retrospective: it computes agency after delayed feedback arrives (p. 11). The contrast the review draws is prospective against retrospective signals, and comparator against an associative, connectivity account. The LIT’s summary is corrected.
• No instruction for machine-learning practice; nothing for the anthology. Box 2’s remark on brain-machine interfaces is anecdotal and gives no practice.

Limitations

• A review by an author whose group ran many of the studies reviewed. No effect sizes, samples or replications are given in the text, and none of the primary studies was read here.
• The prospective component rests on one binding design (Moore and Haggard 2008) and its patient extension (Voss et al. 2010).
• Implicit measures are themselves under-validated as measures of an experience. The review grants that shifts in time perception “are not diagnostic of a sense of agency” in themselves.
• The coercion result is one experiment in one paradigm, with financial reward attached to the shock key.
• The legal discussion is the author’s framing, not evidence. He calls the relevance of neuroscientific evidence to the law “correctly, a topic of intense debate”.

Open questions

• Does social attribution of agency depend on instrumental agency, or the reverse? The review says this is unclear (p. 6).
• Is the reduced agency under coercion a reduced experience, or a reduced report of it? Interval estimates are reports of time. The N1 result is the part that does not depend on the report.
• Does the prospective component survive outside the binding paradigm, for example in explicit judgements made without primes?
• What signal would carry the positive “buzz” of agency if the comparator outputs nothing on success? The connectivity view names a place but not a signal.