Only giving orders? An experimental study of the sense of agency when giving or receiving commands

Emilie A. Caspar; Axel Cleeremans; Patrick Haggard

doi:10.1371/journal.pone.0204027

Abstract

In human societies, agents are assumed to experience being the author of their own actions. These basic motoric experiences of action are influenced by social hierarchies, leading to surprising and morally significant results. Here we ask whether, under coercion, the sense of agency and responsibility pass from the person who receives orders to the person who gives them. Volunteers took turns to play the roles of ‘commander’, ‘agent’ or ‘victim’ in a task where the commander coerced the agent to deliver painful shocks to the ‘victim’. We used ‘intentional binding’ as an implicit measure of sense of agency in both commanders and agents, in conditions of coercion and free-choice. We observed a reduced sense of agency when agents received coercive instructions, relative to when they freely chose which action to execute. We also found that sense of agency in the commanders was reduced when they coerced agents to administer the shock on their behalf, relative to when they acted by themselves. This last effect was associated with the commander’s self-reported level on a psychopathy scale. Thus, coercion resulted in neither commander nor agent feeling agency for the effect of the action, as measured through implicit methods. Our results could have profound implications for social decision-making and social regulation of moral behaviour.

Citation: Caspar EA, Cleeremans A, Haggard P (2018) Only giving orders? An experimental study of the sense of agency when giving or receiving commands. PLoS ONE 13(9): e0204027. https://doi.org/10.1371/journal.pone.0204027

Editor: Marcello Costantini, University G d'Annunzio, ITALY

Received: March 2, 2018; Accepted: August 31, 2018; Published: September 26, 2018

Copyright: © 2018 Caspar et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: Source Data. The behavioural data reported in this paper have been published on Open Science Framework (OSF – DOI 10.17605/OSF.IO/478TJ).

Funding: EAC was supported by the FRS-F.N.R.S (Belgium). PH was supported by EU FP7 grant VERE WP1, AHRC grant L015145/1 to PH, and by ERC Advanced Grant HUMVOL. PH was additionally supported by an ESRC Professorial Research Fellowship. AC is a Research Director with the F.R.S.-FNRS (Belgium). This work was partly supported by BELSPO IAP grant P7/33 and by ERC Advanced Grant RADICAL to AC.

Competing interests: The authors declared that they had no conflicts of interest with respect to their authorship or the publication of this article.

Introduction

All human societies require and endorse individual responsibility: people are aware of their actions and their consequences. While all, or almost all, human actions have some external consequences, actions made in a social context are particularly important because their effects impact not only the agent herself but also other people. Thus, the feeling of responsibility is a central aspect of the experience of agency in social contexts, and could be considered as a precursor to morality [1]. The experience of responsibility is a core aspect of the mechanism that constrains people to behave responsibly. At the same time, social factors can also constrain responsibility: sometimes, individuals are expected to obey coercive orders. For example, people under military command are expected to obey orders unconditionally, and their responsibility is not assessed in the usual way. The continuous balance between taking full responsibility for individual actions, and passing responsibility to commanding authorities, goes to the heart of how most societies work.

Nonetheless, history shows that the effects of coercion on the experience of responsibility can profoundly affect moral behaviour [2]. Stanley Milgram famously conducted a series of experiments that profoundly modified views of responsibility [3–4]. Milgram found that most people could be coerced into administering what appeared to be painful punishments. However, those experiments could not reveal how coercion influences behaviour. Today, the mechanisms of coercion remain poorly investigated.

We [5] recently measured the subjective experience of acting under coercion, and the effect of compliance on neural processing. In one condition, an experimenter ordered one participant to make a manual action that caused both an auditory tone, and either a simultaneous financial penalty or a calibrated painful electric shock to a co-participant. These harmful outcomes produced a financial gain for the participant. In another condition, participants chose for themselves how often to inflict the harm, and thus how much financial gain to obtain. Results showed that the social context of coercion reduced people’s individual sense of agency over their actions, as measured by the intentional binding effect, an implicit measure showing compression of perceived temporal intervals between an action and its outcome [6,7], relative to a free-choice condition. Explicit ratings of responsibility also showed a reduced feeling of responsibility under coercion. In addition, we found that event-related potentials to the tone exhibited reduced amplitude in the coercive condition, compared to a condition where participants freely chose which action to perform. These results confirmed the potential for social situations to modulate one’s own sense of agency, and thus moral conduct.

Acting under coercion therefore reduces the feelings of both agency and responsibility for one’s own actions. But are the sense of agency and responsibility then passed on to the person who gives the orders, as social and legal concepts of a chain of command might imply (i.e., “passing the buck”), or do they simply diffuse and disappear [8]? For example, military officers are held responsible for the actions of men they command, and government ministers are (in principle, if not always in practice) held responsible for the actions of civil servants working in their departments. In the present experiment, we assessed what happens to the feelings of agency and responsibility when one person commands another what to do.

To distinguish between these possibilities, we developed a paradigm in which three volunteer participants were tested together, each playing in turn the roles of commander, agent and “victim”. To assess the experience of responsibility and agency of each participant, we used both implicit and explicit ratings. Feeling of responsibility were evaluated through explicit ratings [1]. To assess the implicit sense of agency, we used an established behavioural proxy based on estimated action-outcome intervals (‘intentional binding effect’, [6,9]), rather than explicit self-reports–since the latter are strongly influenced by social desirability biases [10] and do involve different psychological processes [11]. In the intentional binding paradigm, participants have to estimate the delay between their action (i.e. a keypress) and an outcome (i.e. a tone). If the movement is voluntary, the perceived time is shorter than in a condition in which the movement is involuntary (for instance, triggered by Transcranial Magnetic Stimulation (TMS) pulse over the motor cortex, [6]). This suggests that identifying oneself as the author of an action modifies time perception, by reducing it (see [12] for a review). Thus, in our study, we focus on whether coercion increases the perceived interval between an action and its outcome, i.e., reduces the sense of agency.

According to a “socially optimistic” hypothesis, when commanders give coercive orders that are followed by the agents, the commanders should feel a strong sense of agency and responsibility, even if agents do not. Commanders should thus judge that action-outcome intervals are short relative to a condition where they simply observe the agent choosing and making the action and report high responsibility ratings. According to an alternative “socially pessimistic” hypothesis, no difference between these two situations is expected–when one person gets another to do something for her/him, neither person actually feels agency and responsibility for the outcome. In this scenario, the person who issues the orders would not feel responsible because even though she decides how to act, she does not actually carry out the act herself. The person who executes the orders would likewise not feel agency and responsibility for the outcomes because even though she carries out the action, she is not the one who chose which action was to be carried out.

Such issues—agency in social chains—, was explored in a previous study. Pfister et al. [13] investigated whether performing an action (= leader) that leads a second agent (= follower) to perform another action could influence implicit SoA. There was no evidence of subjective compression of time, neither for leaders nor for followers, over the final outcome of the chain (see their Experiment 2), which is consistent with our ‘socially pessimistic hypothesis’. However, in their paradigm, the leaders had no strong decisional power, since they could only press a single key when they wanted to do, a key that had no ecological consequences in the external world. This may have thus reduced their SoA over the final consequence of the chain. In contrast, the commander in our experiment could freely decide what order to give and when to give it [14], thus controlling the action of the follower, and hence potentially influencing sense of agency. Given that the research questions examined in the present study have been poorly investigated in previous literature, we also included some more exploratory analyses, in an effort to relate our sense of agency measures to other relevant psychological variables, such as personality traits as measured through self-report questionnaires.

Experiment 1

Experiment 1: Method

Materials and procedure.

Participants completed several questionnaires online at least three days before participating: The Interpersonal Reactivity Index (IRI, [15]), the Levenson self-report psychopathy scale (LRSP, [16]) and the Social Dominance Orientation scale (SDO, [17]). The order of the questionnaires was counterbalanced across participants.

On arrival at the experimental laboratory, participants read an information sheet about the experimental procedure and the aim of the experiment. The three co-participants signed their individual consent forms simultaneously, ensuring that they were all aware of the others’ consent. Roles were assigned based on where participants chose to sit when they arrived in the room. One participant was the commander, another one was the agent and the third one was the “victim”. These roles were switched during the experiment, making the procedure fully reciprocal. The agent and the “victim” were sat at a table, face to face. A keyboard was placed between them, oriented towards the agent but visible by all participants. The commander sat at another, taller, table, placed near to the agent. The experimental task was controlled by a computer located on the agent’s right side, with the screen visible only to the agent and to the commander. The agent was instructed to press one out of two keys on the keyboard at a time she chose after the start of the trial, using either the right index finger or the right middle finger. This caused a tone to occur. The delay between key press and tone was set to vary randomly between 200, 500, and 800 ms. If a shock was delivered, the shock occurred at the same time as the tone. The commander and agent then both made independent estimates of the delay between the agent’s key press and the tone. The ‘victim’ also completed the task of interval estimates in order to keep them engaged during the task, and to reinforce the symmetry of the situation. However, our research questions focussed on the experience of agents and commanders, so the data from the ‘victim’ were not further analysed. They were informed that the delay would vary randomly on a trial-by-trial basis, between 0 and 1,000 ms (they were reminded that 1,000 ms makes 1 s). Participants were also told to make use of all possible numbers between 1 and 1,000, as appropriate, using the full scale as appropriate, and avoiding rounding. Each participant received a paper sheet with 60 empty boxes for their time estimates in each condition of the task. Participants’ answers were hidden from view of the other participants by a barrier, so as to avoid one participant influencing another.

A short training session allowed participants to practice the interval estimation procedure. Two experimental conditions followed. All participants started with a specific endowment (€20). In the free-choice condition, agents were instructed that they could freely choose to increase their remuneration for the experiment by delivering a painful electric shock to the “victim”, using the appropriate key on the keyboard F to deliver a shock and H to not deliver the shock). Whether the agent delivered a shock or not, a similar tone occurred after the key press. They were told that they were totally free to choose how to act. The agents earned €0.05 each time they decided to deliver a painful electric shock to the “victim”. They earned no money if they decided not to deliver a shock. In this condition, the commander was simply a passive observer. In the coercive condition, the commander stood next to the agent and ordered the agent, on each trial, to deliver or not a painful electric shock to the “victim” (see Fig 1). The commander was completely free to decide which order to give, and could thus control the agent’s actions. If the commander ordered the agent to deliver a shock, both commander and agent earned €0.05. The experimenter sat in a corner of the experimental room, and never interfered with participants’ decisions.

Download:

Fig 1. Experimental set-up of Experiment 1.

In the free-choice condition, the agent was free to choose to deliver a shock or not (upper row). The commander sat at another table, passively watching. In the coercive condition, the commander was standing rather than seated, and ordered the agent to deliver a shock or not to the “victim” (lower row). Each participant took each role in turn (columns).

https://doi.org/10.1371/journal.pone.0204027.g001

There were 60 trials per condition (20 trials at each action-tone delay, in randomized order), giving a total of 360 trials (120 per roles). The order of the free-choice and coercive was counterbalanced across triads, but fixed within each triad.

Painful stimuli were delivered using a constant current stimulator (Digitimer DS7A) connected to two electrodes placed on the back of “victims”‘ left hand, visible to the agent and the commander. Individual pain thresholds were determined for all three participants after they had signed the consent form, and before starting the experiment. This threshold was determined by increasing stimulation in steps of 1 mA, following the procedure described in Caspar et al. [5] until participants reported a distinct pain sensation. The mean stimulation level selected by this procedure was 23.84 mA (SD = 7.14, pulse duration: 200 μs). All participants viewed the setting of each other’s thresholds.

In a post-session questionnaire, we asked participants to rate how responsible they felt during the experiment when they were the agent, and when they were the commander, as a percentage score. We also asked participants to describe other feelings during the experiment. The first item assessed how bad they had felt about giving shocks to the other participant, the second how sorry they were (answers were provided on 7-point Likert scales; “-3”; not very; “3”; very bad), the third how much they had looked at the other person’s face during the experiment (“-3”; not much; “3”; a lot) and how much pain they thought the other person had felt (“-3”; none at all; “3”; a lot). Participants had to answer both describing their experience as agent, and also as commander. In a first open question, participants were invited to explain if they administered more, less or a similar number of shocks to the “victim” when they were commander in comparison with when they were a free agent. They were invited to give open commentary about their decisions during the experiment. On a scale ranging from -3 (not very bad) to +3 (very bad), participants reported a mean score of .51 (SD = 1.73) when in the role of the agent and a mean score of .77 (SD = 1.81) when in the role of the commanders. This difference was not significant (p > .5). In a second open question, participants were invited to describe briefly their feelings and thoughts during the experiment. These descriptions confirmed that the participants were aware that their actions caused physical pain to their co-participant, and that they generally “felt bad” about doing so.

At the end of the experiment, participants were paid separately based on the original endowment, plus any additional financial gain during the experiment.

Participants.

Thirty-six right-handed and naïve participants were recruited by triads. All participants were female to avoid possible effects of gender on relations of command and coercion. Our recruitment procedure included checks to ensure that participants were not close friends or relatives of others in their triad. To estimate the sample size, we used the effect size of Experiment 2 in Caspar et al. [5]. That study had an effect size d_z of 0.630 (based on the means and SDs of the within-subjects free-choice (mean: 367, SD: 119) and the coercive (mean: 426, SD: 131) conditions). To achieve a power of .90 for this effect size, the estimated sample size was 29 (GPower: [18]). Participants received between €20 and €29 for their participation. The following exclusion criteria were decided in advance of the experiment: failure to produce temporal estimates covarying monotonically with actual action-tone interval, or failure to follow instructions. To identify participants whom the action-tone intervals did not gradually increase with action-tone intervals, we performed linear trend analysis with contrast coefficients -1, 0, 1 for the three delays. One participant was excluded due to a non-significant linear trend analysis. The mean age was 21.57 (SD = 1.98). All participants provided written informed consent prior to the experiment. The study was approved by the local ethical committee of the Université libre de Bruxelles (008/2016), and was carried out in accordance with the approved guidelines. Individuals pictured on Fig 1 were not real participants. They gave written informed consent (as outlined in PLOS consent form) to publish these case details.

Experiment 1: Results

No participants withdrew from the experiment, and none reported any distress either after testing or at follow-up. During follow-ups, participants were invited to answer different questions about how they had lived this experiment in the different roles. Answers could be given by email or by phone (see Supporting Information S1 Text).

Number of shocks.

When participants acted as agents, they freely chose to administer painful electric shocks to the “victim” in 29.23/60 trials (CI₉₅ = 24.58–33.87, min 1, max 60). When participants acted as commanders, they freely chose to administer painful electric shock to the “victim” in 31.57/60 trials (CI₉₅ = 26.89–36.25, min 3, max 60). Because each participant undertook both roles, we compared these choices directly with a dependent-samples t-test. The difference was significant (t(34) = -2.090, p = .044, Cohen’s d = .353, see Fig 2A), thus showing that participants delivered more shocks to the ‘victim’ when they were commanders than when they were agents.

Download:

Fig 2.

A) Number of shocks. Number of shocks freely administered by participants when they were in the roles of agent and of commander. B) Responsibility ratings (%) in post-experiment questionnaires. All tests were two-tailed. Errors bars represent standard errors. *** indicates a p value < = .001. * represents a p value between .01 and .05.

https://doi.org/10.1371/journal.pone.0204027.g002

Responsibility ratings.

Participants gave high responsibility ratings when they were free agents (Mean 89%, CI₉₅ = 83–94) and also when they were free commanders (Mean 81.5%, CI₉₅ = 72–91). The difference between these roles was not significant (p > .1). In contrast, participants reported lower responsibility when they received orders (mean 41%, CI₉₅ = 30–52) compare to when they were free to decide which key to press (t(33) = 7.992, p < .001, Cohen’s d = 1.37) or free commander roles (t(33) = 5.296, p < .001, Cohen’s d = .908, see Fig 2B).

Interval estimates.

We analysed both commanders’ and agents’ interval estimates using a Type III unbalanced repeated-measures ANOVA, with Role (Agent, Commander), Condition (Free-choice, Coercive) and Outcome (Shock, No shock) as within-subject factors and Order of the role as a between-subjects factor (see Fig 3A). This design was unbalanced, with 3 participants (8.5% of the total sample) had no data in at least one condition, because they chose either to deliver no shocks at all, or to deliver only shocks. The main effect of Role was not significant (p > .4), suggesting that interval estimates did not differ when participants were agents or commanders. The main effect of Condition was significant (F(1,29) = 8.254, p = .008, η²_partial = .222), with coercion leading to longer interval estimates (implying less sense of agency) than free choice (482 ms, CI₉₅ = 442–522 and 457 ms, CI₉₅ = 420–493), suggesting that coercion reduces the implicit feeling of agency, as measured through this method. The main effect of Outcome was also significant (F(1,29) = 4.723, p = .038, η²_partial = .140), with longer interval estimates when a painful outcome was given (476 ms, CI₉₅ = 439–513) than when no painful outcome was delivered (463 ms, CI₉₅ = 424–502). Most importantly for our hypothesis, the interaction Role x Condition was significant (F(1,29) = 7.640, p = .01, η²_partial = .209). Paired comparisons indicated that interval estimates were shorter in the free-choice (437 ms, CI₉₅ = 397–477) than in the coercive condition (487 ms, CI₉₅ = 445–528) for agents (t(34) = 4.343, p < .001, Cohen’s d = .734), but that this difference (free-choice: 475 ms, CI₉₅ = 432–518 –coercive: 477 ms, CI₉₅ = 433–520) was not significant for commanders (p > .3). To ensure that the absence of statistical difference was not due to lack of sensitivity in our data, we computed Bayes Factors (BFs, [19]) to refine these results. A BF between 1/3 and 3 indicates a lack of sensitivity. A BF below 1/3 or above 3 is typically interpreted as supporting for the null hypothesis, or for the alternative hypothesis, respectively. To compute a BF, we used Experiment 2 of Caspar et al. [5]. We selected this experiment because it is the first study published which compared a free-choice and a coercive condition in a paradigm involving electric shocks, and an implicit measure of sense of agency. No previous studies investigated SoA for commanders in a similar paradigm, so we assumed that this previous study of the agent role would provide a reliable basis to compute BFs for individuals from a similar population playing a commander role. To calculate the BF, we computed the expected mean and SD of the difference between the free-choice condition and the coercive condition [mean of plausibility(population value/theory) = 56 ms] and SD of plausibility(population value/theory) = 61 ms). The BF was 0.23, thus lending support to H0. This suggests that coercing another individual did not differ from passively observing the situation in terms of our sense of agency measure. The same interaction can also be considered by comparing the two conditions in which agents/commanders had full decisional power (the free-choice condition for agents and the coercive condition for commanders): in this case, interval estimates were shorter for agents than for commanders (t(34) = 2.331, p = .026, Cohen’s d = .394). When agents and commanders had no decisional power (i.e., the coercive condition for agents and the free-choice for commanders), their interval estimates did not differ (p > .4). The other interactions were not significant (all ps > .1). Neither the main effect of Order (p > .8), nor its interaction with other factors (all ps > .3) were significant. A series of additional analyses were then conducted. Given that participants were free to decide the number of harmful actions to deliver when they were agents and commanders, we also reanalysed the data using this information (number of shocks freely delivered when commanders–number of shock freely deliver when agents) as a covariate. This covariate was not significant (p > .9) and the general pattern of results was unchanged. We also reanalysed the data by correcting each participant’s interval estimates by subtracting the real action-tone delay from the judged action-tone delay for each trial in order to control the unbalanced distribution of delays (200, 500, 800 ms) across within-subject factors (see S1 Fig). The general pattern of results was unchanged (see Section a in S2 Text), but the Outcome factor was no longer significant (p > .5). Finally, we additionally conducted further analyses including Delay between the agent’s keypress and the tone as a within-subject factor in order to investigate the effects of duration of the action-outcome intervals. Section b in S2 Text shows the analyses, which mainly showed that the coercion effect for agents was stronger for long action-effect intervals than for short action-effect intervals.

Download:

Fig 3.

A) Interval estimates. Estimated action-tone intervals (ms) in free-choice and the coercive conditions for agents and commanders, pooling across the three action-outcome delays. The conditions have been named according to the agent’s perspective: in the free-choice condition, the agent was free to choose what to do and the commander was simply a passive observer. In the coercive condition, the agent followed the commander’s coercive instruction, while the commander was free to decide which order to give. B). Correlation. Correlation between LRSP score and the effect of indirect agency, defined as the difference between interval estimates of commanders vs agents in free choice conditions). All tests were two-tailed. Errors bars represent standard errors. *** indicates a p value < = .001. * represents a p value between .01 and .05.

https://doi.org/10.1371/journal.pone.0204027.g003

We conducted further exploratory analyses. To assess whether or not some self-reported personality traits could influence the “coercion effect” (i.e., the difference between the free-choice and the coercive condition) a forced-entry regression method was used to fit a multiple linear regression model. We centred the overall score of all predictor variables (IRI, LRSP and SDO) before building the model. Two participants did not answer all the questions of the online questionnaire, so their data were not taken into account. We observed that none of the predictor variables modified the coercion effect, neither for agents (all ps > .1), nor for commanders (all ps > .1). S1 and S2 Tables show supplemental results for similar correlations performed with the individual subscales of these instruments, where available.

We also investigated the effect of personality on an estimate of the “direct agency effect”, defined as difference in interval estimates when freely chose as a commander (and coerced another person to enact their choice), relative to when they chose and acted directly as an agent. A negative score on this difference measure indicates that interval estimates were shorter, and thus sense of agency stronger, when commanding others, compared to acting directly. Given that multiple correlations were performed, we applied a Bonferroni correction (α/3 = 0.016). We found a strong relation between LSRP score and the direct agency effect (t(32) = 2.921, p = .007, Beta = .497). The higher participants scored on LSRP (i.e., more psychopathic traits), the less they tended to feel a sense of agency when making free choices in the commander role, relative to the direct agent role—producing a positive score on our indirect agency measure (see Fig 3B). Importantly, the same analyses conducted on the corrected interval estimates (real action-tone delay—judged action-tone delay for each trial) also lead to similar results, with a strong correlation only with the LSRP (t(32) = 3.106, p = .004, Beta = .531). Supplementary analyses revealed that the secondary psychopathy subscale (impulsivity, poor behavioural control) of the LSRP was a better predictor (t(32) = 3.263, p = .003, Beta = .491) than the primary psychopathy subscale (also named ‘callousness’, p > .3). Other predictors (IRI, SDO) were not significant (all ps > .05). S3 Table shows results for similar regressions for specific subscales. We additionally investigated whether or not our direct agency measure could predict the difference between the number of shocks administered by agents and commanders, but we found no significant relation (p > .8).

Experiment 1: Discussion

Hierarchies frequently involves a form a diffusion of responsibility between those who make decisions, and those who translate them into action. To our knowledge, the present study is the first attempting to systematically explore how individuals’ sense of agency varies between the role of decider and the role of executor in an ecological paradigm.

First, we observed that agents reported longer interval estimates in the coercive condition than in the free-choice condition, replicating previous results [5]. Importantly, in this experiment, the orders were not given by an experimenter as previously, but rather by another participant who was free to decide which order to give. This suggests that the reduction of the implicit experience of agency is a common feature of coercive situations. The basic effects of coercion on sense of agency do not require a social structure or institution of “authority” [4]. Rather, coercive modulation of agency may arise whenever individuals find themselves in low-power relations [20], even when these are transient and reversing, as here. This finding adds generality and ecological validity to previous reports of coercion effects on sense of agency [5]. We additionally found a modest but significant main effect of outcome, with longer interval estimates, implying lower sense of agency, when actions produced shocks compared to when they did not. Reduced sense of agency for negative outcomes has been reported previously, using event timing estimates [21–22], although this effect was not statistically significant in our previous study with painful outcomes [5]. We speculate that the size, and thus significance, of this effect may depend on differences between experiments in the emotional valence associated with the outcomes (negative vocalisations versus painful electric shocks associated with a monetary reward in [22] and [5] respectively).

Second, we observed a striking increase in interval estimates, corresponding to a reduced sense of agency, when participants’ instructed another person to carry out their intentions, compared to when they themselves made the action. When commanders coerced another participant to implement their free choices, rather than doing it themselves, they appeared to distance themselves from the outcome of their intentions, and felt less agency. Our implicit measures showed lower sense of agency for such indirect, mediated coercive actions than for direct freely-chosen actions. Interestingly, exploratory analyses revealed that this effect was significantly related to individuals’ psychopathy scores. The higher participants scored on primary psychopathy, the greater the difference between interval estimates when coercing another agent, vs when acting as free agent oneself was. Thus, we may have identified an important trait variable relevant to the diffusion of responsibility that occurs when people coerce others, which should be confirmed in future studies.

Interestingly, we found a dissociation between implicit feeling of agency and explicit report of responsibility for commanders, but not for agents, in Experiment 1. Commanders reported high responsibility scores when they gave orders to the agent, but showed greater interval estimates, implying a reduced sense of agency, relative to performing an action themselves. This dissociation might be explained by the fact that explicit measures are strongly biased by social desirability factors: commanders know that they should feel responsibility and thus report that they do feel it, even though they may not actually feel it. Implicit measures of agency are less influenced by such demand factors. Importantly, the commanders’ high explicit responsibility ratings confirm that they understood the psychological issues associated with the role of commanding.

Our first experiment involved an important confound between low-level and high-level features of social roles. In particular, our participants made manual actions to trigger the tones and shocks when they acted as agents, but merely gave spoken instructions, which ultimately and indirectly lead to tones and shocks, when they acted as commanders. Thus, the difference between agent and commander roles is confounded with the direct causation of the outcome. The vast majority of sense of agency studies, particularly those using implicit measures, have focused on direct causal control over outcomes via manual actions [6,23–24]. Therefore, the lower sense of agency for commanders rather than agents might reflect an absence of manual action and direct causation, rather than a specific effect of social role. In situations of coercion, commanders indeed do not act by themselves. Rather, they impose their decisions onto a third party, who actually executes them. In neurocognitive terms, coercion splits the action control system across two individuals, so that the higher cognitive processes of action selection take place in the commander’s brain, while the lower-level motoric processes of action execution take place in the executor’s brain. Previous studies have highlighted the importance of sensorimotor signals for the sense of agency [25–29]. We suggest that commanders’ reduced sense of agency when they coerce others in Experiment 1 may reflect the lack of the normal sensorimotor information that underpins experiences of direct agency.

In a second experiment, we tested this hypothesis directly, by introducing a situation where the commander gave their coercive instructions not by verbal instruction as in Experiment 1, but by a manual action that more directly lead to the shock. In order to control for the presence of manual actions in all the experimental conditions, commanders had to press a key that delivered an instruction also in the free-choice condition. To preserve a feeling of freedom for agents, we thus told them that they could freely decide to follow or not this command. Indirectly, we thus offered the possibility to agents to obey or disobey the commander’s instructions. Previous research showed that an increased effort [30] and an increase in the number of alternative actions [31] both increase the feeling of agency. Choosing to go against coercive instructions should therefore lead to a greater sense of agency than being coerced. This modification also allowed us to control for causal events preceding the agent’s keypress [32]. Indeed, one could assume that the reduction of agency in the coercive condition for agents is due to the occurrence of some event (i.e. the commander’s instruction) preceding the keypress, rather than due to the coercive content of that event per se. This additional antecedent event is not present in free-choice conditions, thus potentially explaining the difference between conditions. However, experiment 2 balances this potential antecedent event across conditions. Specifically, in the free-choice condition of Experiment 2, commanders instructed the agents when to press a key (just as in the coercion condition), but the agents were free to decide whether to follow the instruction or not. Thus, in experiment 2, coercion and choice were different contexts, but the events were identical across conditions.

Finally, we also included a ‘predictability’ factor in the coercive condition in Experiment 2: the instruction that the commander sent by pressing one of two keys was either fully predictable from the key pressed (coercive predictable condition) or only predictable on 50% of trials (coercive unpredictable condition). Previous studies have showed that unpredicted outcomes reduce SoA [33] and should thus constitute a ceiling effect in the reduction of agency.

Again, the socially optimistic hypothesis predicts shorter interval estimates when the commander’s orders are applied by the agent in comparison with a situation in which the agent does not follow those orders. We further predicted that the commander’s more direct motoric involvement in causing the shock would increase their feeling of agency when they gave orders that are executed by the agent. Importantly, in Experiment 2 we could not directly compare agents and commanders’ interval estimates since participants in each role judged the delay between their own motor action and the resulting tone. Thus, the action judged was not physically identical for agents and for commanders, and the action-outcome intervals differed. Some of the analyses performed in Experiment 1, namely the direct agency effect, could thus not be investigated in Experiment 2. However, whenever possible, we performed statistical analyses comparing Experiments 1 and 2.

Experiment 2