What are the neuropsychological consequences of complying to the delivery of an order?

Research Project II:

What are the neuropsychological consequences of complying to the

delivery of an order?

Name: Irene Arnaldo

Student Number: 11665696

Supervisors: Kalliopi Ioumpa & Emilie Caspar

Submitted: 30/01/20

What are the neuropsychological consequences of complying to the

delivery of an order?

Abstract

Previous research has shown that following orders affects underlying psychological processes, such as sense of agency and vicarious pain perception. Nevertheless, the hierarchical structure of coercive organisations entails that individuals often find themselves in the position of the _{​coercer​}instead of the _{​coerced agent. ​}In the present study we wanted to see how complying to the delivery of an order, as opposed to voluntarily transmitting it, changes vicarious pain network activity assessed with fMRI. Furthermore, we wanted to see if commanding a robot instead of a human to give a shock would have different implications. The MRI results did not support our main hypothesis that coerced commanders would display a reduced activation in the pain network compared to when they were free to decide, nor the hypothesis that the empathetic response would be higher when the executor was a robot instead of a human. We discuss the implications of these findings and highlight potential pathways for future research.

1. Introduction

In the 60s, after the world had witnessed two terrible wars, Standley Milgram conducted a series of controversial studies on obedience (1963, 1974). His research evidenced that humans tend to comply with coercive orders even when they entail severe harm to another person. Due to the ethical issues these experiments posed (Herrera, 2001), no follow-up studies were conducted, leaving the cognitive mechanisms underlying obedience unexplored.

As humans we are endowed with the ability to feel what others feel and empathetically share states with them (Decety, & Lamm, 2006; Keysers, & Gazzola, 2014). Empathy can be conceived as the vicarious experience of other people’s states (Lockwood, 2016). There is an extensive amount of literature showing that observing someone else's pain triggers an empathic response in the observer (Singer, & Lamm, 2009; Keysers, & Gazzola, 2014). For example, empathy involves unpleasant sensations when observing someone else in pain (Singer et al., 2004; Lepron, Causse, & Farrer, 2015). Vicarious pain perception activates the so-called pain matrix, a brain network that is activated when we undergo painful stimulation (Singer et al., 2004, Keysers, & Gazzola, 2014). In this way, our own pain system allows us to understand other people’s experiences (Lamm & Majdandzic, 2015). More specifically, the anterior insula (AI) and the anterior cingulate cortex (ACC) display overlapping activity when experiencing pain and empathizing with the pain of others (Singer et al., 2004). A recent study demonstrated that following orders to deliver a shock compared to freely choosing to do so decreased the activity in the so-called pain-matrix (Caspar, Ioumpa, Keysers, & Gazzola, under revision). Interestingly, subjects rated the intensity of the

pain as higher in the Free-Choice compared to the Coercive conditions, even if they knew the current was kept constant throughout the experiment. Furthermore, participants reported a decreased sense of responsibility under coercive circumstances. Notably, a recent set of experiments demonstrated that increased sense of agency and responsibility enhances the empathic response to pain (Lepron, Causse, & Farrer, 2015).

Sense of agency refers to the subjective experience of being in control over one’s own actions, and consequently their outcomes (Caspar, Christensen, Cleeremans, & Haggard, 2016). Implicit measures of sense of agency are thought to control for self-report biases oriented to increase social desirability (Bandura, 2006 ). A typical implicit measure for sense of agency is temporal binding (Dewey, & Knoblich, 2014). It refers to subjective time perceived between an intentional action and its consequences (Haggard, Clark, & Kalogeras, 2002). In this way, the interval between a voluntary key press and a tone will be perceived as shorter than the delay between two tones generated externally (Poonian, & Cunnington, 2013). Temporal binding effects that emerged from voluntary actions have been associated with conscious representations of actions and their consequences (Haggard, Clark, & Kalogeras, 2002). Notably, following coercive orders has been shown to reduce sense of agency, reflected in longer interval estimates compared to voluntary actions (Caspar, Christensen, Cleeremans, & Haggard, 2016).

In sum, coercion affects different psychological processes such us of sense of agency, responsibility reports , and vicarious pain activity (Caspar et al., 2016; Caspar, Ioumpa, Keysers, & Gazzola, under revision). However, most coercive institutions such as the military, have a hierarchical structure resulting in a chain of commands that exempt a large part of its members from being direct agents.

In the present study we wanted to emulate this hierarchy by introducing a third person: ‘the commander’_​, who will be both the _{​coerced and the ​coercer​}. Hence, participants will be in charge of giving the order to deliver or not to deliver a shock to another executor, under coercive and free circumstances. In addition, we wanted to see whether giving orders to a human or a robot would influence sense of agency, responsibility, and empathy for others pain. Therefore, we conducted a reciprocal paradigm with three conditions: _{​Coercive​}, Free-Choice Human_​, _{​Free-Choice Robot​}, where participants _​who were commander first, were victims second, and vice versa. The commander was always the subject inside the MRI scanner.

Our main hypothesis was that commanders would display a lower empathetic response towards the victim’s pain in the Coercive condition compared to Free-Choice Human condition. Based on previous results we expected a reduced activation in the ACC, insula, amygdala, temporo parietal junction (TPJ), and striatum, when participants give the order of delivering a shock under coercive circumstances compared to when they choose to give the order freely (see Caspar, Ioumpa, Keysers, & Gazzola, under revision). In addition, we predicted that commander’s empathetic response would be higher when a robot instead of another human executes their commands.

Moreover, we anticipated that this pattern would be reflected in responsibility ratings, whereby commander’s would feel less responsible for the outcomes of their orders in the

Coercive condition compared to Free-Choice Human condition. And, the reported responsibility would be lower when they give orders to another human instead of a robot. Furthermore, we expected these differences to be accompanied by changes in the perceived intensity of the pain. In this way, commanders would rate the victim’s pain as less intense in Coercive condition compared to Free-Choice Human, and higher in the Free-Choice Robot compared to Free-Choice Human executor.

Since our paradigm was reciprocal, we anticipated that subjects that were victim first would act vindicatevely when taking the role of commander by administering more shocks. Furthermore, following previous results we expected vindictive tendencies to correlate negatively with activation in areas like the amygdala and insula (see Caspar, Ioumpa, Keysers, & Gazzola, under revision).

On a different experimental paradigm with EEG recordings we acquired implicit measures of sense of agency. The present paper will only explore the behavioral measures acquired with this different experimental sample. Based on previous results, we did not expect any significant differences between conditions in implicit measures of sense of agency for commanders (see Caspar, Cleeremans, & Haggard, 2018), since when commanders coerce another executor instead of conducting the action themselves, they seem to psychologically distance themselves from the outcomes of their orders.

2. Methods and Materials

The present study collected two different sets of data due to the experimental constraints that different neuroimaging methods entail. The focus of this paper will be on fMRI data set. However, I will briefly introduce the EEG task as well and _​the behavioral measures of this experiment will be included in the results.

2.1 Procedure

A total of 88 right-handed participants were recruited. Their ages ranged from 18 to 36 years. A total sample of 42 participants took part in the fMRI study (n=12 males; n=30 females), while 48 subjects participated in the EEG study (n=24 female; 24=males). Prior to the experiment they provided their informed consent, and completed all seven questionnaires online (_{​https://www.limesurvey.org/​). The subjects were paired according to} their gender, and we ensured they did not know each other beforehand.

Participants arrived 15 minutes before the experiment so a second fMRI screening test could be performed. Subsequently, they were taken into a room where they received the instructions of the experiment, followed by an introduction to fMRI (or EEG).

Before the task, participants indicated their detection and pain thresholds using a DS7A current stimulator, and they were informed that the current would be held constant throughout the task. Two electrodes were placed on the participants’ left hand on the abductor pollicis muscle in order to produce a clear and visible muscle twitch and the threshold was increased by steps of 1mA until a mildly painful stimulation and muscle twitch were achieved.

Subjects were told that they would receive a bonus of 0.05 _​€ for each shock sent.​Afterwards, they were randomly assigned to the role of ‘ _{​commander​}’ or ‘ _​victim​’. The paradigm was

however reciprocal, so those that were commander first would fulfill the role of victim later, and vice versa.

Before the experiment they could practice the task for a couple of trials. Then, the commander was placed in a MRI scanner (or EEG room) to perform the task, while the victim sat in a different room watching nature documentaries. When the task was finished the roles were reversed. Finally, all subjects completed few questions regarding ‘How bad how sorry’ they felt (see supplementary materials), and they were compensated for their participation with 50_{​€ and an additional bonus that was based on the number of shocks they} administered during the task.

2.1.1 The task

In fMRI data acquisition is necessary to establish long-delays between events to secure that brain activity can be dissociated. Therefore, to ensure the separation between motor actions and outcomes (i.e victim’s pain), the fMRI paradigm had a very long action-outcome delay that precluded the use of implicit measures of sense of agency, which were acquired in the EEG paradigm.

The task consisted of three different conditions (Coercive, Free-Choice Human, Free-Choice Robot), that were randomized across participants. Due to the temporal restrictions of fMRI data acquisition trials last for a fairly long amount of time, therefore each condition was divided in two runs of 30 trials so participants could rest. In total commanders spent around 1.5h completing the task, and each run lasted approximately 15 minutes.

Each trial of the fMRI task begun with a fixation cross that lasted for 8-12 seconds. Afterwards, commanders heard a verbal instruction from the experimenter. In the Free-Choice conditions subjects were told ‘_{​You can decide​}’, while in the Coercive condition the instruction was either ‘_{​Give a shock ​}’ or ‘ _{​Don’t give a shock ​}’. Participants were told that the instructions were online. However, to ensure fine control over stimulus presentation during the experiment the instructions were pre-recorded. Nevertheless, to increase authenticity the instructions were recorded 4 different times with small voice variations and background noise.

In order to increase the psychological effect of receiving orders, during the Coercive condition the experimenter was present in the MRI room. The experimenter made it clear she was in the room during Coercive condition by speaking directly with the commander, and she used a fake microphone.

The verbal order was followed by a picture of two rectangles, a red one labelled ‘SHOCK’ and a green one labelled ‘NO SHOCK’, were displayed on the right and left sides of the screen. The position of the labels, and the rectangles paired with them, was randomized across trials but the action-outcome mappings were always congruent. Commanders had to press one of the two buttons to give the order to the executor of either delivering or not delivering a shock. There was no time limit for button press since we wanted to safeguard commander’s voluntary actions.

Then the mapping disappeared, and if the commander had selected ‘SHOCK’, then a red light turned on on the executor’s button box, while a green light was on when the order was ‘NO SHOCK’.

In the Coercive and Free-Choice Human conditions the executor was always a human and her hand pressing the button was visible. The motor action of button press was kept consistent across trials. In _​the Free-Choice Robot condition the executor was a robot, therefore the executor’s button press was automatized for this condition (see Fig 1).

Following executor’s button press another arrow pointed to the victim’s hand screen (top-right) to ensure they were attending to the trial outcomes (i.e Muscle twitch in ‘Shock’ trials). The arrow appeared consistently for 750ms before the display of the shock (also for NoShock trials) and disappeared 750ms after, and the shock lasted for 250ms.

Additionally, on each run a pain rating scale was randomly presented on 6 different trials to guarantee participants were attending the victim’s hand. They were requested to rate the perceived intensity of the shocks from ‘Not painful at all’ (‘0’) to ‘Very painful’ (‘1000’), even if they knew that the electric current was kept constant throughout the experiment. They were asked to move the red bar along a scale by using four buttons. The outer keys of the button box allowed them to move the bar in steps of +-100ms, whereas the inner keys moved the bar in steps of +-1ms. When it was a ‘NoShock’ trial they were instructed to rate the trial as ‘Not painful at all’. They had 6 seconds to provide an answer. In the same fashion, at the end of each run participants had to rate how responsible they felt for the outcomes of their actions.

In the EEG paradigm each run consisted of 60 trials, lasting around 20 minutes per condition. Overall, commanders spent 1hour performing the task. As mentioned above, EEG allows the acquisition of implicit measures of sense of agency. Therefore, the EEG paradigm introduced the variation of a tone occurring following the delivery (or not) of the shock. The delay between the executors’s key press and outcomes (‘Shock’, ‘NoShock’) varied randomly between 200, 500, 800 ms. Participants were requested to provide an estimate of the action-outcome interval. To do so, on a trial-to-trial basis, they had to move an arrow on a dimensional scale between 0 to 1000 ms. Like in the fMRI experiment the pain scales were randomized across trials, and the responsibility scale was displayed at the end of each run. Due to the high behavioral variability that we observed (i.e disobedience, not enough Shocks) an additional run was introduced at the end of the experiment so further analyses could be performed. This observation run consisted of 29 trials where participants did not have to do anything specific. To keep consistency with other conditions participants witnessed the same sequence of events (executor’s button press followed by shock/noshock delivery). However, there were some variations. First, the visual instruction ‘Tell the executor what to do’ was removed while the mapping of ‘SHOCK’ ‘NOSHOCK’ was maintained, and the auditory instruction was not played. Furthermore, the responsibility scale was not included in this condition.

Fig. 1. A) Display of the experimental set-up. The commander was inside the scanner while the ‘victim’ was outside the scanner, watching a documentary. In Free-Choice blocks, the experimenter was in the control room. In coercive blocks, the experimenter entered the scanner room, indicated her presence to the agent but remain hidden from her view. B) Display of one single MRI run.

2.2 Materials

2.2.1 Questionnaires

A total of seven different questionnaires were used to evaluate individual differences prior to the experiment. To assess individual differences in authoritarianism we employed the Right-wing Authoritarianism Scale (RWA; Altemeyer, 1981), and the Aggression-Submission-Conventionalism scale (ASC; Dunwoody, & Funke, 2016). To assess personality traits of narcissism, machiavellianism, and psychopathy, we used a short version of Dark Triad (SD3; Jones, & Paulhus, 2014), as well as the Levenson Self-Reported Psychopathy Scale to evaluate differences in psychopathy (LSRPS; Levenson et al., 1995). We used the Hypomania Check List to evaluate differences in mood and energy (HCL-32 R1; Angst et al., 2005). The Interpersonal Reactivity Index was employed to estimate

individual differences in empathy (IRI; Davis, 1980). In addition, differences in moral values were evaluated through the Moral Foundations Questionnaire (MFQ-30; Graham et al., 2008).

2.2.2 General Data Analyses

Behavioral results were analysed with frequentist and Bayesian statistics (Dienes, 2011). As opposed to frequentist analyses that assess the probability of the data under the null hypothesis, bayesian approaches provide not only evidence against the null hypothesis but also in favour of it (Ortega, & Navarrete, 2017). The interesting aspect of bayesian approaches is that they offer an estimation of how probable the data is under both the null, and the alternative hypothesis (Ortega, & Navarrete, 2017; Lee, & Wagenmakers, 2014). In this way, bayes factors quantify the evidence for one of the hypothesis. In other words, it estimates how probable the data is under one of the hypothesis relative to the other. For example BF_​10​, which corresponds to the p(data|H1)/p(data|H0), will inform about how likely the data is under the H1 compared to the probability of it occurring under H0. Therefore, a BF10=20 would indicate that the data is 20 times more likely under the H1 than H0. A BF _​10 below 1/3 or above 3 is interpreted as supporting H0 and H1, respectively.

Similarly, we will report inclusion BF when we are testing significance of effects. Inclusion BF compares two types of models, one model includes the main effect of the factor of interest while the other does not. Consequently, BF will estimate the probability of the data happening under the effect of interest relative to the model without that particular effect (Hinne, Gronau, van den Bergh, & Wagenmakers, 2019).

Bayes Factors and p-values were calculated using JASP (JASP Team, 2019), with the default priors implemented in JASP.

2.2.3 _{​Functional Magnetic Resonance Imaging (fMRI)}

MRI images were recorded using a 3-Tesla Philips Ingenia CX system and a 32-channel head coil. T1-weighted structural images were recorded with the following specifications: matrix = 240x222; 170 slices; voxel size = 1x1x1mm. Seven runs of functional images were recorded (matrix M x P: 80 x 78; 32 transversal slices in ascending order; TR = 1.7 seconds; TE = 27.6ms; flip angle: 72.90°; voxel size = 3x3x3mm, including a .349mm slice gap). MRI data processing was conducted in SPM12 (Ashburner et al., 2014). EPI images were slice time corrected and realigned to the mean image. _​High quality _​T​1 images were coregistered to the mean EPI image and segmented. The coregistered gray matter segment was normalized onto the MNI gray matter template and the resulting normalization parameters applied to all EPI images. Afterwards, images were smoothed with a 6mm kernel.

During the first level analyses we defined different regressors for the Shock and NoShock trials. Condition type was modelled in separate runs. A time window of 1.500ms was selected for these regressors. They started 750ms before the shock, which had a duration of 250ms, until 500ms after the moment of the shock, which coincided with arrow presentation.

The rest of the task was modelled with additional regressors that accounted for: auditory orders (between 8-12s after the beginning of the trial), the pain rating scale (which appeared 1s after the arrow disappeared) and the responsibility rating scale (which appeared at the end of each run 1s after the arrow disappeared) were included in the same regressor, and motion parameters. Overall, we had four regressors of interest plus six motion parameters. In addition, coercive trials were participants disobeyed were modelled separately as prosocial or antisocial disobedience regressors.

In order to test our main hypothesis we defined the contrast as Shock-NoShock instead of examining the shock condition alone. Therefore, we ended up defining two main contrasts of interest to see how giving the order of delivering or not delivering a shock affected vicarious prain processing. To test our hypotheses we conducted a whole brain analyses by applying the contrasts [(Free-Choice human Shocks - NoShocks) - (Coercive Shocks - NoShocks)] and, [(Free-Choice robot Shocks - NoShocks) - (Free-Choice human Shocks - NoShocks)]. At the second level, we localized where these contrasts were significant across participants in a random effect one-sample t-test. Results were thresholded at _{​p < .001 and family-wised} error (FWE) corrected at the cluster level.

3. Results

3.1 Behavioral Results

Number of shocks delivered_​: The following data includes data in the fMRI and EEG paradigm N=90. Participants were randomly instructed to order the delivery of a shock in 30/60 trials in the Coercive condition. In total 15 participants disobeyed prosocially, 11 of them disobeyed antisocially, and 17 disobeyed by contradiction. Nevertheless, only 22 reported having intentionally disobeyed (10 by contradiction, 12 prosocially, and 1 antisocially) (see Annex)

In the Free-Choice conditions participants were instructed to freely choose whether to command the delivery or not of a shock on each of the 60 trials per condition. On average, commanders ordered the administration of 24.39/60 (SD=6.34, minimum:0/60, maximum=60) shocks when the executor was a human, and 24.21 (SD=16.34 minimum: 0/60, maximum:60/60) shocks when they gave orders to a robot executor.

We performed a Pearson’s correlation between the total number of shocks participants received in the Free-Choice conditions when they were victim first, and the total number of shocks they sent in the Free-Choice conditions when they were commanders. The results showed a robust trend to act vindictively (r=.37, p=.015).

Fig. 2​: Pearson’s correlation between the total number of shocks received in the Free-Choice

conditions for those who were victim first, and the number of shocks they delivered in the Free-Choice conditions when they became commanders.

Exploratory frequentist and bayesian correlations were conducted to see if the total number of shocks sent in the Free-Choice conditions was correlated with any personality traits. First we conducted bayesian statistics and then perform frequentist correlation on the significant relationships, correcting for multiple comparisons with FDR. To control for the potential confound of vindictive tendencies we pooled participants according to their role order (Commander First, Victim First). Bayesian and frequentist correlations indicated that the total number of shocks that participants gave when they were commander first was positively correlated with DT_Machiavelisim (r=.49, p(corrected with FDR)=0.002, BF_​10​=18.97), and with levenson’s LSPRS_P1 scale (r=.59, p=.002, BF_​10​=1161.40). On the other hand, bayesian and frequentist correlations showed that the total number of shocks that participants gave when they were commander second (Victim first) was only correlated with the measure of MFQ_Purity (r=.46, p<.001, BF _​10​=24.68). Indicating that higher scores on this scales correlated positively with number of shocks sent.

Responsibility Ratings_​: After each experimental run participants were requested to rate how responsible they felt for the outcome of their orders, responsibility ratings were acquired for both fMRI and EEG samples. To test whether responsibility ratings were significantly different across conditions we conducted a two-way repeated-measures ANOVA _{​on the} total sample, N=89, with Condition (Free-Choice Human, Free-Choice Robot, Coercive) as within subjects factor, and Role Order (Victim First, Commander First) as between subjects factor. The analyses yielded to a valid n=85, since four participants did not take part in one of the Free-Choice conditions.

Mauchly’s test indicated that the assumption of sphericity had been violated(χ2(2) = 26.19, p < .001), therefore degrees of freedom were corrected using Greenhouse-Geisser estimates of sphericity (ε = 0.79). The analyses showed a main effect of Condition, F(1.58, 130.36)=53.19, p<.001,_​η​p​2​_{=.39, which was largely confirmed by BF=5.004e14. ​The effect of}

Role Order was not significant, F(1,83)=1.89, p=.17, η_​p​2​_{=.02, and evidence for the null} hypothesis was confirmed by bayesian statistics with a BF=.33.

Post-hoc comparisons using the Tukey HSD test with Bonferroni adjusted alpha levels confirmed our hypothesis that responsibility ratings would be significantly lower in the Coercive condition (M=55.23, SD=27.96) compared to the Free-Choice Human (M=77.55, SD=22.34), (p<.001), confirmed by BF_​10​=4.38, and compared to the Free-Choice Robot (M=84.56, SD=19.1), (p<.001), proved by BF_​10​=6.59. Furthermore, participants reported feeling significantly more responsible in the Free-Choice Robot condition compared to the Free-Choice Human (p=.003), which was confirmed by BF_​10​=23.35, confirming the hypothesis that subjects would report feelings significantly more responsible when the executor was a robot instead of a human.

To understand if executing the action directly or giving orders to an executor would influence responsibility ratings, we conducted a two-tailed independent sample t-test on the responsibility ratings (Free-Choice Human, Coercive) comparing the agent (Caspar, Ioumpa, Keysers, & Gazzola, under revision) and commander experiment.

There was no significant effect of the type of role (Agent or Commander) in the responsibility ratings for the Free-Choice Human condition, t (120)= -1.72, p>.05, d=-.35, which yielded to a BF_​10​=.79, despite agents (M=86.13, SD=14.17) reported higher scores than commanders (M=79.45, SD=20.56). This effect was also not present in the Coercive condition (t(120)=1.12, p>.05, d=.23), with a BF _​10​=.37, where agent’s ratings (M=51, SD=24.17) were lower than commander’s scores of responsibility (M=57.25, SD=27.94). Therefore, bayesian statistics do not confirm that there is evidence for H0, suggesting a lack of sensitivity.

Reaction Times_​: We performed two-way repeated-measures ANOVA on the time elapsed between the presentation of the instructions and button press in the Commander, with Condition (Free-Choice Human, Free-Choice Robot, Coercive), and Trial (Shock, NoShock) as within-subjects factors. We only recorded this data for the fMRI sample, N=43, because these data was not recorded for the EEG experiment, and due to the behavioral variability our final sample (valid n=40). The main effect of Condition was significant, F(2, 78)=29.11, p<.001, _​

η

_​

_=.43

_​

significant, F(1,39)=.61, p=.44, _​

η

_​

=.02, which generated a BF=.13 that provided

evidence for the null hypothesis.

Post-hoc comparisons using the Tukey HSD test using Bonferroni alpha adjusted confidence intervals, showed that participants were significantly faster in the Coercive condition (M=1.13, SD=0.56) compared to the Free-Choice Human (M=1.73, SD=0.76), (p<.001), which yielded to a BF _​10​=5.33. Participants were also significantly faster in the Coercive compared to the Free-Choice Robot (M=1.79, SD=0.79), (p<.001), confirmed with a BF_​10​=2.87. However, there were no significant differences in reaction times between the Free-Choice conditions, (p=.56) which was proved by a BF_​10​=.09.

Due to some technical difficulties data for the Agent’s reaction times could not be accessed, making it impossible to compare the two studies.

F_​ig. 3_{​: ​Behavioural results Commander​. Red columns represent Coercive condition , green} columns Free-Choice Human condition, and blue columns represent Free-Choice Robot conditions. A) Participants reported feeling significantly more responsible in Free-Choice Human, and Free-Choice Robot, compared to the Coercive condition. B) Participants gave orders significantly faster when they were in the Coercive condition compared to the Free-Choice Human, and Free-Choice Robot conditions. C) There were no significant differences between conditions in How Bad participants reported feeling. D) Participants felt significantly more sorrier for the victim in the Free-Choice Human condition compared to the Coercive condition. All tests were two-tailed. Error bars represent standard errors, p-values less than 0.05 are flagged (*), p-values smaller than 0.01 are flagged, and p-values less than 0.001 are flagged(***).

How bad how sorry_​: At the end of the experiment participants had to rate on a 7-point Likert scale from -3 (not very bad/sorry at all) to +3 (very bad/sorry) how they felt on each condition as a commander. Two separate two-way repeated-measures ANOVA with Condition (Coercion, Free-Choice Human, Free-Choice Robot) as within subjects factor, and Role-Order (Commander First, Victim First) on the total sample were conducted on the ‘bad’ and ‘sorry’ scores. Since not all participants underwent all conditions, the analyses yielded to a valid N=85.

On the ‘sorry’ scores, Mauchly’s test indicated that the assumption of sphericity had been violated(χ2(2) = 32.08, p<.001), therefore degrees of freedom were corrected using Greenhouse-Geisser estimates of sphericity (ε = 0.76).

The analyses yielded a main effect of Condition, F(1.51, 127.22)=6.31, p<.006, _​

η

_​

supported by BF=7.59. Post-hoc tests revealed that participants reported feeling significantly more sorry in the Free-Choice Human (M=.69 ,SD=1.67) compared to the Coercive condition

(M=.08 , SD=1.87), (p=.005) that yielded to BF _​10​14.08, which confirmed the evidence for the alternative hypothesis. There were no significant differences between Free-Choice Robot (M=0.48, SD=1.74) and the Coercive sorry scores, (p=.17, which resulted in BF _​10​=.69, and when they were compared to the Free-Choice Human scores, (p=.182), which yielded to BF_​10​=.66. Therefore, the null hypothesis for Free-Choice Robot conditions was not confirmed by Bayesian analyses.

In addition, Role-Order had no effect in how sorry participants felt (F(1, 84) =.29, p>.05,

η

_​

_=.003),

_​

On the ‘bad’ scores, Mauchly’s test indicated that the assumption of sphericity had been violated, χ2(2) = 38.09, p<.001, therefore degrees of freedom were corrected using Greenhouse-Geisser estimates of sphericity (ε = 0.73).

The effect of condition was not significant when participants reported how bad they felt, F(1.46, 121.02)=2.78, p=.08, _​

η

_​

η

_​

both factors respectively: BF=.57 and, BF=.44.

We performed two independent sample two-tailed t-test to compare ‘how sorry’ and ‘how bad scores’ between the Agent (Caspar, Ioumpa, Keysers, & Gazzola, under revision), and the Commander experiment, for each Condition (Coercive, Free-Choice Human).

How bad scores did not differ significantly between the groups neither in the Free-Choice Human, t (121)=-.31, p>.05, d=-.06, nor in the Coercive, t(122)=.79. p >.05, d=.15, supported by a BF _​10 = .22, and BF_​10​= .27, respectively. Although not significant, agent’s bad scores were higher in the Free-Choice condition (M=.35, SD=1.69) than commanders (M=.24, SD=1.82), while in the Coercive condition agents attained lower scores (M=-.43, SD=1.85) than commanders (M=-.14, SD=1.92).

There were also no significant differences in how sorry participants felt depending on their role in administration of shocks in either the Free-Choice, t(122)=-1.03, p>.05, d=-.201, nor the Coercive, t(122)=-.02, p>.05, d=.004, conditions. The results were confirmed by a BF_​10​=.33, and BF_​10​=.21. Nevertheless, agents reported feeling more sorry in the Free-Choice Human (M=1, SD=1.73), and Coercive (M=.05, SD=1.91) conditions, than commanders in the Free-Choice (M=.65, SD=1.7), and Coercive (M=.05, SD=1.89) conditions.

Pain Scale_​: The pain scale was randomly presented in a total of 6 trials per run, where participants had to rate on a scale from 0 (‘not painful at all’) to 1000 (‘very painful’) the perceived intensity of the victim’s pain. We observed that some participants used the pain scale incorrectly by rating NoShock trials as more painful than Shock trials. Therefore, participants who on average did not rate Shock trials at least 100 points higher as NoShock trials in one of the main Conditions (Coercive, Free-Human, Free-Robot) were excluded from the analyses.

Due to some technical difficulties during the experiment pain scores during the observation run were not recorded for all participants, n=30, compromising the effect size. Therefore, we decided to perform two separate two-way repeated measures ANOVA.

A two-way ANOVA was conducted on the pain scores for shocks trials with condition (Coercive, Free-Human, Free-Robot), as within-subjects factors, and role order (Victim First, Commander First) as between subjects factors, valid n=46. The main effect of condition was not significant, F(2,66)=.87, p>.05, _​

η

_​

hypothesis that commander’s estimates of the victim’s pain would be significantly different across conditions. The main effect of role order was also not significant, F(1,33)=1,61, p>.05, _​

η

_​

The same analyses was repeated including the four conditions (Coercive, Free-Choice Human, Free-Choice Robot, Observation). This ANOVA, valid n=18, confirmed that neither condition, F(3,51)=.79, p>.05, _​

η

_​

p>.05, _​

η

_​

victim’s pain intensity.

To see if the perceived intensity of the victim’s pain would be modulated by the type of role participants were assigned to (Agent, Commander), an two independent sample t-test were conducted. Pain ratings did not differ significantly between groups for either the Free-Choice Human, t(72)=-.99, p>.05, d=-.23, with a BF _​10​= .37. This same pattern was found for the Coercive conditions, t(80)=1.21, p>.05, d=.27, which resulted in a BF _​10​=.44. Therefore, bayesian statistics did not provide additional support in favour of the null hypothesis.

Agents evaluated victim’s pain slightly more intense (M=499.61, SD=204.06) than commanders (M=457.14, SD=166.27), in the Free-Choice Human condition. Although, under coercive circumstances agents gave on average lower ratings (M=431.79, SD=223.18), than commanders (M=488.07, SD=195.74).

SOA_​: In the EEG experiment, a total of 48 subjects estimated the time interval between button press and tone after each trial. However, due to the high variability participants exhibited in their behavior 10 participants had missing data-points. Therefore, we run a three-way repeated-measures ANOVA on the interval estimates of the EEG sample (valid N=38) , with Condition (Coercive, Free-Choice Human, Free-Choice Robot) and Trial Outcome (Shock, NoShock) as within-subjects factors, and Role Order (Commander first, Victim first) as between subjects factor.

The main effect of Condition was not significant, F(2, 72)=.08, p_​

= .92

_​

η

_​

p​

= .002

​

η

_​

p​

= .02

​. ​

supported by bayesian statistics with BF=.02, and BF= .15 respectively. Role-Order did not have an effect either in implicit measures of sense of agency, F(1,36)=.119, p>.73,_​

η

_​

p​

= .003

​

with a BF=.08.

In this way, bayesian and frequentist approaches confirmed the hypothesis that there would be no observable differences in implicit measures of sense of agency between conditions when participants are commanders instead executors.

Table 1: Descriptive Statistics of Responsibility Ratings

Condition Commander First Victim First

N M SD N M SD

Coercive 43 57.28 29.13 42 54.18 27.50

Free-Robot 43 89.15 14.09 42 80.23 22.35

Free-Human 43 79.64 21.43 42 75.94 23.33

Table 2: Descriptive Statistics _{​How Bad/Sorry Scores for the Different Conditions}

Condition BAD SCORES SORRY SCORES

N M SD N M SD

Coercive 87 -.14 1.92 87 .05 1.89

Free-Robot 85 .08 1.80 86 .47 1.74

Free-Human 86 .24 1.82 87 .65 1.70

Table 3: Descriptive Statistics ​Pain Scores for

each Condition and Trial Outcome Condition_Trial Shock N M SD Coercive 49 488.07 195.74 Free-Robot 41 466 166.71 Free-Human 43 457.14 166.27 Observation 30 499.43 202.53

Table 4: Descriptive Statistics ​for Implicit Measures of Sense of Agency for each Condition and Trial Outcome

Condition_Trial Shock No Shock

N M SD N M SD Coercive 45 -.03 .19 47 .02 .15 Free-Robot 42 -3.86 3.33 44 1.37 2.31 Free-Human 42 -3.82 3.722 46 6.40 1.98

Note: Sample sizes differ due to the high variability in behavior which affects the number of data points per variable.

fMRI Results: _​We collected fMRI data from a total of 42 participants but the final sample only included 23 of them. A minimum of 7 shocks per Free-Choice condition were required to include these participants in the analyses. Therefore, fourteen participants were not included because they did not give enough shocks during the Free-Choice conditions to establish a reliable Shock - NoShock contrast. For this same reason, one participant was excluded from the neuroimaging analyses because he disobeyed antisocially in all coercive trials. Additionally, two participants were not included in the analyses because they disobeyed by contradiction in 27/60 and 30/60 trials. Two participants had to be excluded because of movement artifacts.

Based on previous results we expected to observe enhanced activation in regions associated with vicarious pain perception when participants were freely giving the order of giving a shock to a human executor compared to when they did so under coercive circumstances We also hypothesized that activity in these regions will be higher when participants are freely choosing to give the order to delivery a shock to a robot compared to when they give it to a human executor.

To test these hypotheses we took a whole-brain analyses approach, by using the contrasts: a) [(Free-Human Shocks - NoShocks) - (Coercive Shocks - NoShocks)], and _​b)​[(Free-Robot Shocks-NoShocks) - (Free-Human Shocks- NoShocks)].

These contrasts revealed there were no significant differences in the activity of the vicarious pain network between conditions. In order to control for potential confounds in the analyses we conducted a quality-check contrast that took into account all conditions: [ All Shock- All NoShock]. This contrast confirmed that regions involved in vicarious pain perception were actively involved when subjects observed the victim’s hand receiving a shock (FWE corrected at cluster level, t=3.51, p<.001, minimum cluster size=160).

Fig. 4. _{​MRI results​. (ALL Shocks - ALL Noshocks)} ​contrast. Peak coordinates can be seen in Table 1. _​ FWE ​at cluster level (160 voxels), t=3.51, ​p ​< .001.

4. Discussion

Broadly the aim of this study was to examine the neuro psychological effects of introducing an additional layer in a coercive set-up, where participants were not directly coerced to deliver a shock but to order someone else to administer a shock. This was achieved by collecting various measures that included fMRI neuroimaging data, self-reports, as well as implicit behavioral measures. To this end, mainly we wanted to test whether being free order the delivery of a shock would lead to higher vicarious pain brain activity compared to the delivery of coerced instructions. Additionally, we also expected that giving orders to a robot instead of a human would lead to higher activity in the apin network.

fMRI Results

The main hypothesis of the fMRI Commander experiment was not confirmed in this study. We expected higher brain activation for pain perception when commanders were freely choosing to delivery an order compared to when they were coerced to do so. Furthermore, we also hypothesized this activation would higher when the orders were given to a robot instead of human executor. These differences were expected in areas associated to empathy, which include the ACC, striatum, amygdala, TPJ, and insula/IFG. Our MRI results did not provide any support for the alternative hypothesis, suggesting that the type of condition participants were in did not modulate the empathetic response towards the victim’s pain.

The lack of significant differences between conditions in the Commander experiment appeals to the hypothesis that giving orders instead of directly executing them distances individuals from the outcomes of their actions, blurring any potential differences between voluntary and coerced orders. Nevertheless, our results need to be interpreted cautiously. The fMRI sample exhibited highly variable behavioral responses, such as a large number of disobedience trials, or the delivery of very few shocks in free-choice conditions. This high behavioral heterogeneity requires larger sample sizes, and differences between conditions might be dissolved by the extreme group variability not suitable to conduct group statistics. This is especially relevant in our paradigm since we modelled obedience and disobedience trials with different regressors. Therefore, the psychological consequences of being coerced might not be fully captured in obedience trials if participants exhibited such variable behavior (obedience vs. disobedience) during the coercive runs.

Implicit Behavioral Measures

The results gathered from the EEG sample confirmed our hypothesis that when participants are given the role of the Commander differences in sense of agency between conditions disappear as opposed to when they are Agents. This means that sense of agency is not enhanced when individuals are free to choose the outcome for the victim compared to when they are coerced to command those outcomes. Therefore, as Caspar and colleagues (2018) argued, relying on someone else to execute the action reduces implicit measures of sense of agency over the outcomes of the orders, suggesting that motor actions are necessary for a full sense of agency to manifest.

There is evidence showing that participants tend to act vindicatevely when they fulfill the role of victim first (Caspar et al., 2016; Caspar, Ioumpa, Keysers, & Gazzola, under revision). Our hypothesis that commanders would tend to act vindictively was supported. In this way, participants that had received many shocks as a victim sent more shocks when they were commanders.

Our formalization of vindicativeness did not correlate significantly with any individual personality traits. Nonetheless, we saw that the number of shocks sent by participants that were victim first, thereby the one that would show vindictive tendencies, correlated positively with a subscale of the Moral Foundations Questionnaire (MFQ_Purity). In this subscale participants need to evaluate behaviors that pertain virtues of chastity, wholesomeness and control of desires. Hence, the more severe attitudes participants have towards deviations from moral values the more shocks they would send when they have the chance to give

back the shocks. Thus, this might be an important trait that modulates the desire for retaliation, which should be replicated by further studies.

In addition, we found that psychopathy traits were correlated with the number of shocks sent when participants were commanders first (see Annex). Previous studies (Caspar et al., 2018) had also identified it as a potential factor for the diffusion of responsibility when people coerce others, which might be a variable especially relevant for future studies on disobedience.

Self-Reported Measures

Based on the Agent study (Caspar, Ioumpa, Keysers, & Gazzola, under revision) we had hypothesized that participants would perceive the victim’s pain as less intense when they were coerced to give the order compared to when they were freely choosing to do so. Nevertheless, our results showed that differences in perceived pain disappear across conditions when participants are assigned the role of Commanders. This is in line with the results presented above, suggesting that giving orders instead of directly executing them distances subjects from the outcome of their actions as their perceptual reports of the victim’s pain are also not modulated by obeying orders.

We found significant differences between conditions in participant’s self-reports of responsibility. The analyses confirmed our hypothesis that participants would feel significantly more responsible when they were giving orders to a robot executor than when they did to a human. Moreover, they also reported feeling significantly less responsible under coercive circumstances compared to when they giving voluntary orders.

These findings, map nicely onto the differences found in reaction times between coercive and free-choice circumstances, which suggest that participants took more time to reflect in the free-choice conditions. Despite the fact that differences between Free-Choice Robot and Free-Choice Human in responsibility ratings were not accompanied by changes in reaction times.

These results point out the need for a cautious interpretation of self-reported measures since they are not necessarily accompanied by changes in implicit measures, like reaction times or sense of agency. It’s therefore important to address that responsibility scores might be driven by social desirability factors, whereby subjects think they should feel more responsible when they give orders to a robot than to a human.

Commander’s reports of how bad they felt about delivering the order of a shock in exchange of money were not influenced by the type of condition participants were in. However, Bayesian statistics did not provide additional evidence for the lack of differences between conditions. On the other hand, they reported significantly more sorry when they freely gave orders to a human executor than when they were coerced to do so. According to frequentist statistics the type of executor however did not play in how participants felt about the pain of the other, but bayesian statistics did not provide further support for these results. Taking all together although frequentist statistics indicate lack of significant differences, it is difficult to draw conclusions regarding self-reported measures due to the lack of sensitivity that bayesian statistics point out.

Agents vs. Commanders

Exploratory analyses comparing the behavioral measures acquired in the Agent and Commander studies did not identify any significant differences between the two studies. Nevertheless, bayesian statistics indicated that the null hypothesis could not be confirmed for either responsibility ratings, nor for the pain scale. These results suggest a lack of sensitivity to compare the two experiments. In order to increase statistical power future studies could explore the possibility of a within-subjects design that allows the comparison of Agent and Commander measures within subjects. Comparing differences in vicarious pain activity within-subjects would not only increase the statistical power but would also allow to overcome the limitations that heterogeneous behavior poses. In addition, it would be an interesting way to see the effects that adding an additional coercive layer could have on behavior, in terms of readiness to disobey.

Future Research

This study also opened the door for studying the underlying dynamics of disobedience. The heterogeneity of behavior does not allow drawing conclusions with approaches that rely on group averages. Therefore, modeling individual dynamics can be a good alternative to study obedience. Agent based modeling allows to demonstrate how complex dynamical interactions work, whereby situational and psychological predispositions can be taken into account. As Hesp and colleagues (2019) argue traditional statistical approaches rely on the assumption of ergodicity, which assumes that the data for every individual will follow the same pattern at any point in time. Following their work, I propose the use of agent-based modeling to model socio-emotional dynamics that emerge in coercive dyads. This type of would allow us to disentangle which individual prior-preferences, and situational factors lead to different types of disobedience, that can be prosocial, antisocial, or disobedience by contradiction.

5. Conclusions

Recent research has shown a renewed interest in Milgram’s famous studies (1963) to understand the mechanisms underlying obedience. It has been shown that when individuals comply to the execution of an order their sense of agency is reduced (Caspar et al., 2018), shocks appear as less painful, and the neural response associated to vicarious pain perception is reduced (Caspar, Ioumpa, Keysers, Gazzola, under revision). In this study we showed that introducing a second layer in the coercive hierarchy makes differences between conditions in implicit measures of sense of agency, and vicarious pain activity disappear. Furthermore, our hypothesis that giving orders to a robot instead of a human would increase the vicarious pain activity was also not proven. The results of our study contribute to the understanding of the effects that coercive structures can have on individuals. The interpretation that giving orders to another executor distances subjects from the outcomes of the orders is, however, unlikely to be a complete explanation. Ultimately, future experiments should increase the power of their studies to facilitate group comparisons.

ANNEX DISOBEDIENCE Behavioral Results

In total 15 participants disobeyed prosocially, 11 of them disobeyed antisocially, and 17 disobeyed by contradiction. Nevertheless, only 22 reported having intentionally disobeyed (10 by contradiction, 12 prosocially, and 1 antisocially).

We only report participants that acknowledged having intentionally disobeyed for the fMRI sample (N=16, M=10.94 SD=7.24 ). Due to issues with the data for the EEG sample participant’s reports on intentionality and disobedience could not be accessed, therefore we excluded all participants that only disobeyed in one trial, as it was probably a mistake (N=11, M= 8.45, SD= 6.92 )

Fig.5_{​: Histogram showing group frequencies on the number of trials participants intentionally} disobeyed from the EEG (red) and fMRI (blue) samples.

Exploratory correlations with frequentist and bayesian approaches were conducted that included: responsibility ratings, number of shocks, how bad scores, how sorry scores. To correct for multiple comparisons the false discovery rate (FDR) was used. We decided to report the correlations that were significant in both approaches.

Responsibility ratings in the coercive condition were negatively correlated with number of shocks given (r=-.386, p (FDR) =.003, BF _​10​= 110.49). Indicating that the more responsible participants felt in the coercive condition the more inclined they were to disobey. Consequently, prosocial disobedience was positively correlated with responsibility ratings under coercive circumstances (r=.4, p= .0001).

Fig. 6_{​: Pearson's correlation between responsibility ratings and number of shocks sent in the coercive}

condition.

We conducted a repeated measures ANOVA with Trial-behavior (Obedience, Disobedience trials) and Trial-outcome (Shock, NoShock) as within-subjects, and Role-order (Victim First, Commander First) as between subjects factors, n=8. Reaction times did not differ significantly between obedience or disobedience trial, F(1,6) = 0.627, p >.05, _​

η

_​

_=.02

_​

did Trial-outcome have an effect on reaction times, F(1,6)= 0.75, p>.05, _​

η

_​

_=.02

_​

effect or Role-order, F(1,6)=.83, p>.39, _​

η

_​

_=.04

_​

approach, which yielded to a confirmation that our data is more probable under the null hypothesis, Trial-behavior (BF= .59), Trial-outcome (BF=.49), and Role-order (BF=.59).

Table 6​: Table displaying the reasons participants gave for having intentionally disobeyed.

Type of Disobedience Nº Prosocial Trials Nº Antisocial Trials Reason

Prosocial 5 0 ‘Because I had the liberty to do. I love when people can make their own choice.’

5 0 ‘I thought that more than 3 shocks was too much for the victim. But as the time went by I stopped following my rule. ‘

30 0 ‘This is because it was my choice and I did not want to.’

5 0 ‘I had no incentive for obeying: the financial incentive was very weak. there was no consequence for me when i disobeyed and I had already explored the relationship with the executor in the free-choice human executor scenario.’

Contradiction 26 4 ‘When I tried to give shock to the victim her hands jumped a lot. It was unnecessary pain!’

14 16 ‘Because I was choosing to inflict the shock. I wanted to make the decision on my own.’

10 5 ‘It made it more interesting.’ 12 15 ‘It helped me reduce the stress.’

Note_​:_{​ Not all participants that disobeyed are included in this table since not all of them gave} explicit reasons to why they disobeyed.

REFERENCES

● Altemeyer, B. (1981). Right-wing authoritarianism. University of Manitoba press.

● Angst, J., Gamma, A., & Meyer, T. D. (2009). S22-03 Update on recent research with the hypomania checklist HCL-32. European Psychiatry, 24, S118.

● Bandura, A. (2006). Toward a psychology of human agency. ​Perspectives on psychological science​,

1​(2), 164-180.

● Caspar, Ioumpa, Keysers, & Gazzola (under revision) Obeying orders reduces vicarious brain activation towards victim’s pain

● Caspar, E. A., 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.

● 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.

● Davis, M. H. (1980). A multidimensional approach to individual differences in empathy. JSAS Catalog of Selected Documents in Psychology, 10, 8

● Decety, J., & Lamm, C. (2006). Human empathy through the lens of social neuroscience. The scientific World journal, 6, 1146-1163.

● Dewey, J. A., & Knoblich, G. (2014). Do implicit and explicit measures of the sense of agency measure the same thing?. ​PloS one​, ​9​(10), e110118.

● Dunwoody, P. T., & Funke, F. (2016). The Aggression-Submission-Conventionalism Scale: Testing a new three factor measure of authoritarianism.

● Engbert, K., Wohlschläger, A., & Haggard, P. (2008). Who is causing what? The sense of agency is relational and efferent-triggered. ​Cognition​, ​107​(2), 693-704.

● Graham, J., Haidt, J., & Nosek, B. (2008). Moral Foundations Questionnaire, MFQ 30 revised in July 2008. Extracted, 8, 2008.

● Haggard, P., Clark, S., & Kalogeras, J. (2002). Voluntary action and conscious awareness. Nature neuroscience, 5(4), 382.

● Herrera, C. D. (2001). Ethics, deception, and 'those Milgram experiments'. Journal of applied philosophy, 245-256.

● Hesp, C., Steenbeek, H. W., & Van Geert, P. L. (2019). Socio-emotional concern dynamics in a model of real-time dyadic interaction: parent-child play in autism. ​Frontiers in psychology​, ​10​, 1635.

● Hinne, M., Gronau, Q. F., van den Bergh, D., & Wagenmakers, E. J. (2019). A conceptual introduction to Bayesian Model Averaging.

● Jones, D. N., & Paulhus, D. L. (2014). Introducing the short dark triad (SD3) a brief measure of dark personality traits. Assessment, 21(1), 28-41.

● Keysers, C., & Gazzola, V. (2014). Hebbian learning and predictive mirror neurons for actions, sensations and emotions. ​Philosophical Transactions of the Royal Society B: Biological Sciences​,

369​(1644), 20130175.

● Lamm, C., & Majdand žić, J. (2015). The role of shared neural activations, mirror neurons, and morality in empathy–A critical comment. Neuroscience Research, 90, 15-24.

● Lee, M. D., & Wagenmakers, E. J. (2014). ​Bayesian cognitive modeling: A practical course​. Cambridge

● Lepron, E., Causse, M., & Farrer, C. (2015). Responsibility and the sense of agency enhance empathy for pain. Proceedings of the Royal Society B: Biological Sciences, 282(1799), 20142288.

● Levenson, M. R., Kiehl, K. A., & Fitzpatrick, C. M. (1995). Assessing psychopathic attributes in a noninstitutionalized population. ​Journal of personality and social psychology​, ​68​(1), 151.

● Lockwood, P. L. (2016). The anatomy of empathy: Vicarious experience and disorders of social cognition. Behavioural brain research, 311, 255-266.

● Milgram, S. (1963). Behavioral study of obedience. The Journal of abnormal and social psychology, 67(4), 371.

● Milgram, S. (1974). Obedience to authority: An experimental view. New York: Harper & Row.

● Ortega, A., & Navarrete, G. (2017). Bayesian Hypothesis Testing: An Alternative to Null Hypothesis Significance Testing (NHST) in Psychology and Social Sciences. In ​Bayesian inference​. IntechOpen. ● Poonian, S. K., & Cunnington, R. (2013). Intentional binding in self-made and observed actions.

Experimental brain research​, ​229​(3), 419-427.

● Singer, T., & Lamm, C. (2009). The social neuroscience of empathy. ​Annals of the New York Academy

of Sciences​, ​1156​(1), 81-96.

● Singer, T., Seymour, B., O'doherty, J., Kaube, H., Dolan, R. J., & Frith, C. D. (2004). Empathy for pain involves the affective but not sensory components of pain. Science, 303(5661), 1157-1162.

SUPPLEMENTARY MATERIAL S1

1._​ ​How bad did you feel when your delivered a shock in exchange of money?

In the free-choice human condition

(-3) --- (-2) --- (-1) --- (0) --- (+1) --- (+2) --- (+3)

Not bad at all Very Bad

In the free-choice robot condition

(-3) --- (-2) --- (-1) --- (0) --- (+1) --- (+2) --- (+3)

Not bad at all Very Bad

In the coercive condition

(-3) --- (-2) --- (-1) --- (0) --- (+1) --- (+2) --- (+3)

Not bad at all Very Bad

2.​ ​How sorry did you feel when your delivered shocks to the ‘vcitim’?

In the free-choice human condition

(-3) --- (-2) --- (-1) --- (0) --- (+1) --- (+2) --- (+3)

Not sorry at all Very sorry

In the free-choice robot condition

(-3) --- (-2) --- (-1) --- (0) --- (+1) --- (+2) --- (+3)

Not sorry at all Very sorry

In the coercive condition

(-3) --- (-2) --- (-1) --- (0) --- (+1) --- (+2) --- (+3)

Not sorry at all Very sorry

1.​ ​Please describe in a few word how did you feel during the experiment

––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––– –––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––

2.​ ​Which role (commander or victim) did you prefer and why ?

––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––– ––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––

3.​ ​As commander which condition did you prefer (free-choice human, free-choice robot, or coercive) and why?

––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––– ––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––