Introduction

While the findings of Milgram’s famous Yale Obedience studies [1, 2] remain compelling after half a century, his explanation of these findings has become increasingly unconvincing. Milgram deceived his participants into participating as a Teacher in a learning experiment. This involved giving a Learner (a male confederate of Milgram’s) a list of word pairs, then going through the list, each time giving the first word of the pair and asking which of four options it had originally been paired with. Participants (Teachers) were instructed to inflict an escalating series of electric shocks on the Learner each time he made an error (though in fact the shocks were not real).

In the best-known variant of the study (the so-called baseline condition) the Learner was not physically present but could be heard giving various pre-scripted indications of pain and dissent at different shock levels, played from a tape recorder. In this case, some two-thirds of participants (65%) went all the way to the maximum shock level—an apparently lethal 450v. These studies on obedience to authority ‘shocked the world’ [3].

In his early publications, Milgram advanced various potential explanations for his findings. However, in his summary text Obedience to Authority, written roughly a decade later in 1974, he settled on the so-called ‘agentic state’ account [4]. This drew heavily on his interpretation of the notion of ‘the banality of evil’ [5]. It proposed that people have an inherent tendency to become so focussed on doing what is demanded of them by an authority figure as effectively and efficiently as possible that they lose sight of the consequences of their actions. In short, toxic obedience is seen to stem from the fact that people are unaware of the harm they are doing.

There are many reasons to doubt this account [6], including its inability to explain patterns of obedience and disobedience. Indeed, Milgram conducted around 30 versions of his basic paradigm (though not all were published, see [3]) in which levels of obedience ranged from 0% to 100%. Such a large variation in findings cannot be accounted for by the agentic state explanation [7].

However, potentially the most compelling evidence against the ‘agentic state’ explanation—and the core assumption that participants are inattentive to the Learner—has more to do with what happens while people obey rather than with whether they obey. Milgram’s own film of his studies, Obedience (1965), suggests that participants are deeply concerned with the plight of their victim and deeply ambivalent about what they are doing. Milgram [4] himself notes how his procedures ‘created extreme levels of nervous tension in some Ss’, [2]: nervous laughter, sweating, trembling, stuttering, even groaning. He notes how participants would sometimes make the shocks as brief as possible and would also demonstrate discomfort by hesitating to press the shock button, especially later in the studies when the shock levels were higher and when the Learner protested more vociferously [3]. There also appear to have been instances where participants tried to cue the Learner in to the correct response through the use of emphasis when they read out the word choices in each trial [8].

The first aim of this paper is to provide systematic data to back up these anecdotal observations. Do participants demonstrate concern for the Learner even as they administer shocks in the Milgram paradigm? Do they try to help him answer the memory test correctly and so avoid being shocked? More concretely, do they place greater verbal emphasis on the one correct answer when reading it out than on the three incorrect answers? If this can be demonstrated, it would provide strong and direct evidence against the idea that people shock because they are unaware of and unconcerned with the plight of the Learner.

But if participants are concerned for the Learner why do they still obey experimental instructions? What leads them to privilege the demands of the Experimenter to continue administering shocks over the demands of the Learner to stop? One answer, which Milgram himself alludes to in his unpublished experimental notebooks, is that this is a function of identification. In general terms, the more one sees oneself having a positive relationship with the source of an authority, the more likely one is to obey that authority. Specifically, a broad range of evidence supports the contention that the more participants identify with science, the more they obey the Experimenter’s instructions [9–11]. From this perspective, those who obey do so not through unawareness but because, whatever other concerns they have, they ultimately consider it the right thing to do. That is, obedient participants are engaged followers [12, 13].

In light of this debate, a second aim of this paper is to explore the relationship between identification with science (IS) and concern for the Learner (CL) while shocks are being delivered. This is measured by (a) the emphasis laid on the correct response so as to help participants avoid incorrect answers, (b) the hesitation in applying shocks should an incorrect answer be given (especially at higher shock levels), and (c) reported levels of stress when administering shocks.

Here we do not have a single prediction, but rather explore two alternative possibilities. The most straightforward is that there will be a negative relationship between IS and CL. That is, the more a participant identifies with science and is concerned with the successful completion of the study, the less they care about the experience and fate of the Learner—more concretely, the less they would try to help him avoid shocks, the less they would hesitate in applying such shocks, and the less stressed they would be by doing so.

However, perhaps more intriguingly, it is also possible that there will be a positive relationship between IS and CL. That is, science demands concern for all those involved in the investigatory process, including research participants. Hence increased salience of scientific identity would not only increase commitment to the Experimenter and to the study but also increase concern for the Learner. This would increase the extent to which participants would help the Learner avoid shocks, increase the hesitation before applying shocks, and increase the stress of doing so (even if, ultimately, their commitment to completing the study would prevail).

Method

Analysis

A Bayesian analysis was carried out (see S1 Dataset and S1 Statsmodel). This was motivated by the fact that there are several response variables and using frequentist statistical inference would provide no satisfactory level of control of the overall significance level for multiple tests apart from heuristics. Instead, with the Bayesian approach we can treat all statistical relationships in one set of simultaneous stochastic equations and make probability statements about any group of parameters of interest derived from the joint probability distribution of all parameters. Table 1 shows the overall model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Likelihoods for the response variables.

https://doi.org/10.1371/journal.pone.0209704.t001

Likelihoods.

We use the notation y ~N(μ, σ²) to denote that y has a normal distribution with mean μ and variance σ². The likelihoods and statistical models for the response variables are shown in Table 1.

• Eq. 1 RSPL: This simple linear model relates RSPL to the variation in the two conditions (Non-Science vs Science).
• Eq. 2 dAPQ = APQpost–APQpre. This relates dAPQ to the conditions with RSPL as the covariate.
• Eq. 3 Time2ShocksLast2 (time last 2 shocks): The average time to administer the final two shocks by condition and the average time to administer all other shocks (time2shocks).
• Eq. 4 ysc: The extent to which participants identified with science as a function of condition.
• Eq. 5 ynsc: The extent to which participants identified with being a student as a function of condition.
• Eq. 6 RSPL': This relates RSPL to the variation across the three conditions (Control vs Non-Science vs Science). Note that RSPL' is the same as RSPL for the values corresponding to Non-Science and Science.

Prior distributions.

Prior distributions for all parameters on the right-hand side of the equations are Normal distributions (mean 0, standard deviation 10). Hence the 95% prior credible intervals for these parameters are between ±20. All the standard deviations σ have half Cauchy distribution on the positive line with parameters median 0 and scale 5, so that the prior 95% credible intervals are approximately between 20 and 125. The model was fitted using Stan library in Matlab (http://mc-stan.org/users/interfaces/matlab-stan).

Convergence.

Four chains each with the number of simulations as 4000 were used and convergence was obtained, with all Rhat values equal to 1.

Results

There were no systematic differences between the Science and Non-Science groups with respect to age, their experience with computer games, their knowledge of computer programming, their prior experience of Virtual Reality (See S6 Text; S2 Fig), or their prior knowledge of the Milgram Obedience paradigm (S3 Text; S3 Fig).

The overall results are in Table 2 which shows summary statistics from the posterior distributions of the parameters of Table 1. Note that these are derived from the overall joint distribution of all the parameters. In what follows, P(A) means the posterior probability of A.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Summaries of the posterior distributions of the model.

https://doi.org/10.1371/journal.pone.0209704.t002

The factor Priming = 0 for Non-Science, and 1 for Science (i.e., X in Table 2). The intercept term corresponds to Non-Science when all covariates are 0. For each parameter the mean, standard error, and 2.5, 50, and 97.5 percentiles are given, and the posterior probability of the parameter being positive P(>0).

Manipulation check.

S4 and S5 Figs show the effect of the manipulation on participants’ identification with either students or scientists. Although participants in the Science condition appeared to have higher identification with science, there was no such effect in the Non-Science condition on identification with students. Importantly, participants gave nearly identical responses on the ‘identification with students’ questionnaire, regardless of the manipulation. However, participants in the Science condition had greater median answers on all but one question. This can be seen also in S5 Fig, which shows nearly identical mean PCA scores ynsc (non-identification with science) and ysci (identification with science) on the identification with student score. However, on the identification with science items, the ysci score is greater than the ynsc score. Hence, there is a very high probability that the ysci scores are greater for the Science group than the Non-Science group (P(β_ysc,1 > 0) = 0.992; see S4 Text). However, for ynsc, the equivalent probability is 0.5.

Discussion

This study addressed two questions. The first was whether participants, in a Virtual Reality reprise of one instance of the Milgram Obedience paradigm, display concern for the (male) Learner as reflected in the help they provide him by emphasizing the correct answers. The second was how this concern for the Learner—as reflected not only in helping but also in the hesitation to apply shocks and the level of stress provoked by doing so—is affected by identification with science. Would participants show more or less concern when such identification was accentuated?

The results relating to the first question are clear. Compared to a control condition in which no Learner was present, and participants simply read out the right answer on each of the 30 trials, those who were required to administer shocks to a virtual Learner (as per the Milgram paradigm) were more likely to emphasize the correct answer by reading it out more loudly than the incorrect answers (supporting H1). That is, even as they continue to impose electric shocks, participants do attend to the Learner and seek to help him avoid being shocked. It seems, then, that people do not apply electric shocks because they cease caring about the Learner, but rather despite the fact that they care.

The results relating to the second question are, at first glance, more mixed. On the one hand, the levels of help and the hesitation in applying shocks is greater where identification with science is made salient, supporting our H2b. Conversely, the level of stress is greater when identification with science is not made salient, supporting our H2a. As we suggest, however, one way of making sense of this apparent contradiction is to examine the interactions between our various ‘concern’ measures and the way that this depends upon levels of identification with science.

Thus, it is plausible to expect that if people do something to help the Learner, such as cueing them in to the right answer, and if the Learner still gets it wrong, then they (not you) become accountable for their plight. This is consonant with what we know about victim blaming in general [25, 26]. It is also consonant with findings from other research into the Milgram paradigm which suggest (a) that participants become increasingly keen for Learners to get the answer correct as shock levels increase and (b) that if the Learner continues to make errors, participants can become exasperated with them [8]. What is more, one might also plausibly expect such an impact of helping on stress precisely where one feels more concern for the Learner and hence questions of accountability for his plight are more acute. Assuming, then, that such concern is higher when science identity is salient, this would account for our finding of a (negative) relationship between helping and stress in the Science condition and of no such relationship in the Non-Science condition. It would explain why, in the Science condition, participants help more, hesitate more and, as a consequence, feel less rather than more stressed.

In sum, considering the complexity of relations between our concern measures, all the findings can be interpreted as consonant with our H2b: that identification with Science increases concern for the Learner. As we have argued, this can be explained by the fact that the Learner and the participant are part of the same scientific study and, especially when science identity is made salient, are more likely to be seen as members of a common ingroup and more likely to benefit from norms of concern for all those who take part in an experiment.

But what might be more surprising is that those high science identifiers who, across a range of studies, have been shown to be most likely to obey the Experimenter in shocking the Learner [9–11, 13] are also those who are most concerned for the Learner’s plight. This implies that obeying the Experimenter and caring for the Learner are not necessarily counterposed. Not only can one be—and participants are—concerned for both, but an increased tie to the Experimenter can be associated with an increased concern for the Learner.

Of course, such an interpretation of our findings has to be tentative, given that it is based on a post-hoc interpretation of our findings. What is not in doubt, however, is that obeying the Experimenter and showing concern for the Learner clearly co-exist and also that identification with science is an important moderator of both helping and stress in a Milgram-like paradigm. We are therefore in a position to draw two important lessons from our findings. One concerns Milgram’s agentic state account, premised on the lack of attention of the participant to the plight of the Learner. Even within a VR simulation (where participants know that no real harm is actually being done), participants are aware of the Learner’s suffering, reported being stressed by it, and help to try to prevent it. Previous evidence relating to this point was either indirect (showing that ‘inattentiveness’ as an explanation had problems in explaining various facets of Milgram’s findings) or else anecdotal (for a review see 11). We now have experimental evidence that Milgram’s ‘agentic state’ explanation of his own findings is incorrect.

The other lesson concerns our own alternative ‘engaged followership’ account of Milgram’s findings. As would be expected from this perspective, the extent to which one identifies with the science of the study plays an important part in determining levels of stress and helping. To date, this account has been mainly used to explain commitment to the study and hence levels of obedience to experimental instructions. Here we add an extra dimension by addressing the participant’s relationship to the Learner as well as the Experimenter. We certainly do not claim that our findings constitute support for ‘engaged followership’—not least because we did not have specific a priori hypotheses as to how identification with science would impact concern for the Learner. Our aim was more exploratory—to see whether identification would have effects and, if so, what they are. While the patterns proved to be more complex than we had anticipated, we certainly consider that we now have grounds for exploring further the way in which identification with science affects the various relationships in Milgram’s paradigm—those between participant and Learner as well as between participant and Experimenter (as suggested by [11, 12]).

There is also a third lesson which we wish to draw from this study. One of the great methodological challenges for psychology is to find ways of studying dramatic and impactful phenomena (such as toxic obedience) in ways that are both rigorous and ethical. Virtual Reality has been proposed as one potential solution [27–30] and indeed previous research shows that people experience and respond to a VR simulation of the Milgram paradigm in ways that are akin to the original [18]. This research goes further in showing that VR methodology is particularly valuable as a means of investigating relations between variables and hence elucidating psychological processes. Indeed, it is the first study in the context of the Milgram obedience paradigm in which VR has been used in this way, exploring the respective merit of ‘agentic state’ and ‘engaged followership’ models of obedience to authority. As in any research, there is a need to triangulate the present findings with others using different methods. Nonetheless, we would suggest that the strength of these findings points to the value of VR in providing rigorous data concerning hard-to-study phenomena. Moreover, the use of VR has made it possible to study the impact of individual differences amongst participants in order to understand the psychological components of the tendency to follow the Experimenter’s instructions (see [31] for a discussion).

To conclude, then, this study challenges the idea that people follow the instructions from an authority figure because they are unaware of or uninterested in the impact upon their victims. It shows that they are attuned to both the authority and the victim and that they seek to do their best by both. Participants strive to reconcile the contradictory demands placed upon them by the Experimenter and the Learner, inflicting shocks as demanded by the former while trying to help the latter avoid such shocks. It may be of little consolation to those on the receiving end, but our findings suggest that, in some circumstances at least, perpetrators can care for their victims even as they harm them.

Supporting information

Acknowledgments

The Milgram Virtual Reality program was based on an earlier version that had been developed by Dr Paulo Rodrigues, under a grant from The Character Project at Wake Forest University and the John Templeton Foundation, led by Dr David Gallardo at the University of Barcelona. We also thank Dr Angus Antley and Aitor Rovira of UCL who helped with finding an actor and the animations.

The research reported in this paper was supported by ESRC grant ES/L003104/1 ‘Beyond the banality of evil: A new understanding of conformity and atrocity’.