|Home | About | Journals | Submit | Contact Us | Français|
Classic social psychology studies demonstrate that people can behave in ways that contradict their intentions—especially within the moral domain. We measured brain activity while subjects decided between financial self-benefit (earning money) and preventing physical harm (applying an electric shock) to a confederate under both real and hypothetical conditions. We found a shared neural network associated with empathic concern for both types of decisions. However, hypothetical and real moral decisions also recruited distinct neural circuitry: hypothetical moral decisions mapped closely onto the imagination network, while real moral decisions elicited activity in the bilateral amygdala and anterior cingulate—areas essential for social and affective processes. Moreover, during real moral decision-making, distinct regions of the prefrontal cortex (PFC) determined whether subjects make selfish or pro-social moral choices. Together, these results reveal not only differential neural mechanisms for real and hypothetical moral decisions but also that the nature of real moral decisions can be predicted by dissociable networks within the PFC.
Psychology has a long tradition demonstrating a fundamental difference between how people believe they will act and how they actually act in the real world (Milgram, 1963; Higgins, 1987). Recent research (Ajzen et al., 2004; Kang et al., 2011; Teper et al., 2011) has confirmed this intention–behavior discrepancy, revealing that people inaccurately predict their future actions because hypothetical decision-making requires mental simulations that are abbreviated, unrepresentative and decontextualized (Gilbert and Wilson, 2007). This ‘hypothetical bias’ effect (Kang et al., 2011) has routinely demonstrated that the influence of socio-emotional factors and tangible risk (Wilson et al., 2000) is relatively diluted in hypothetical decisions: not only do hypothetical moral probes lack the tension engendered by competing, real-world emotional choices but also they fail to elicit expectations of consequences—both of which are endemic to real moral reasoning (Krebs et al., 1997). In fact, research has shown that when real contextual pressures and their associated consequences come into play, people can behave in characteristically immoral ways (Baumgartner et al., 2009; Greene and Paxton, 2009). Although there is also important work examining the neural basis of the opposite behavioral finding—altruistic decision-making (Moll et al., 2006)—the neural networks underlying the conflicting motivation of maximizing self-gain at the expense of another are still poorly understood.
Studying the neural architecture of this form of moral tension is particularly compelling because monetary incentives to behave immorally are pervasive throughout society—people frequently cheat on their loved ones, steal from their employers or harm others for monetary gain. Moreover, we reasoned that any behavioral and neural disparities between real and hypothetical moral reasoning will likely have the sharpest focus when two fundamental proscriptions—do not harm others and do not over-benefit the self at the expense of others (Haidt, 2007)—are directly pitted against one another. In other words, we speculated that this prototypical moral conflict would provide an ideal test-bed to examine the behavioral and neural differences between intentions and actions.
Accordingly, we used a ‘your pain, my gain’ (PvG) laboratory task (Feldmanhall et al., 2012) to operationalize this core choice between personal advantage and another’s welfare: subjects were probed about their willingness to receive money (up to £200) by physically harming (via electric stimulations) another subject (Figure 1A). The juxtaposition of these two conflicting motivations requires balancing selfish needs against the notion of ‘doing the right thing’ (Blair, 2007). We carried out a functional magnetic resonance imaging (fMRI) experiment using the PvG task to first explore if real moral behavior mirrors hypothetical intention, and second, to examine if these two classes of behavior are subserved by the same neural architecture. We hypothesized that people would imagine doing one thing, but when faced with real monetary incentive, do another—and that this behavioral difference would be reflected at the neurobiological level with differential patterns of activity.
Fourteen healthy subjects took part in this study: six males; mean age and s.d. 25.9±4.6, completed a Real PvG, Imagine PvG and a Non-Moral control task in a within-subject design while undergoing fMRI. Four additional subjects were excluded from analyzes due to expressing doubts about the veracity of the Real PvG task on a post-scan questionnaire and during debriefing. Two additional subjects were not included because of errors in acquiring scanning images. Subjects were compensated for their time and travel and allowed to keep any earnings accumulated during the task. All subjects were right-handed, had normal or corrected vision and were screened to ensure no history of psychiatric or neurological problems. All subjects gave informed consent, and the study was approved by the University of Cambridge, Department of Psychology Research Ethics Committee.
In the Real PvG subjects (Deciders) were given £20 and asked how much of their money they were willing to give up to prevent a series of painful electric stimulations from reaching the wrist of the second subject (the Receiver—a confederate). The more money the Decider chose to relinquish, the lower the painful stimulations inflicted on the Receiver, the key behavioral variable being how much money Deciders kept (with larger amounts indicating that personal gain was prioritized over Receiver’s pain). The task comprised a series of eight screens per trial across 20 trials. Each trial began with a screen displaying the running amount of the subject’s bank total (£20 on Trial 1) and current trial number. Subjects then had up to 11s to decide upon and use a visual analogue scale (VAS) to select the amount of money they wanted to spend on that trial (up to £1) and thus the corresponding painful stimulation to be administered to the Receiver. This 11-s phase was partitioned into the ‘Decide’ and ‘Select’ periods. The Decide screen was presented for a fixed 3s during which subjects were asked to think about their decision, so that when the select screen appeared, subjects could move the cursor to make their selection any time within the next 8s. This design was used in order to introduce a variable jitter within the trial sequence. After making a selection, subjects saw a 3-s display of their choice before experiencing an 8-s anticipation phase—during which subjects were told their choice was being transmitted over the internal network to the other testing laboratory where the Receiver was connected to the electric stimulation generator. Following this anticipation period, subjects viewed a 4-s video of the stimulation being administered (Video event) to the Receiver, or no stimulation if they had opted to spend the full £1 permitted on a given trial. Subjects viewed a video feed of the Receiver’s hand during stimulation administration. Finally, subjects used a 13-point VAS to rate their distress levels on viewing the consequences of their decision, before viewing a 4-s inter-trial-interval. At the conclusion of the 20 trials, subjects were able to press a button to randomly multiply any remaining money between 1 and 10 times, thus giving a maximum possible financial gain of £200. (See Supplementary Materials for descriptions of the Imagine PvG and Non-Moral tasks.)
MRI scanning was conducted at the Medical Research Council Cognition and Brain Sciences Unit on a 3-Tesla Trio Tim MRI scanner by using a head coil gradient set. Whole-brain data were acquired with echoplanar T2*-weighted imaging (EPI), sensitive to BOLD signal contrast (48 sagittal slices, 3mm thickness; Repetition Time (TR)=2400ms; Time to Echo (TE)=30ms; flip angle=78°; Field of View (FOV)=192mm). To provide for equilibration effects, the first seven volumes were discarded. T1-weighted structural images were acquired at a resolution of 1×1×1mm. Statistical parametric mapping software was used to analyze all data. Pre-processing of fMRI data included spatial realignment, co-registration, normalization and smoothing. To control for motion, all functional volumes were realigned to the mean volume. Images were spatially normalized to standard space using the Montreal Neurological Institute (MNI) template with a voxel size of 3×3×3mm and smoothed using a Gaussian kernel with an isotropic full width at half maximum of 8mm. In addition, high-pass temporal filtering with a cutoff of 128s was applied to remove low-frequency drifts in signal.
After pre-processing, statistical analysis was performed using the general linear model (GLM). Analysis was carried out to establish each participant’s voxel-wise activation during the following events: making the decision of how much money to keep/which stimulations to administer (Decide event; Figure 1B) and watching the stimulation be administered (Video event; Figure 1B). Activated voxels were identified using an event-related statistical model representing each of the experimental events, convolved with a canonical hemodynamic response function and mean-corrected. Six head-motion parameters defined by the realignment were added to the model as regressors of no interest. For each fMRI experiment, contrast images for the Decide and Video events were calculated using GLMs and separately entered into full factorial analyses of variances (ANOVAs).
For group statistics, ANOVAs were used. For all three tasks (Real PvG, Imagine PvG and Non-Moral), the Decide event and the Video event were used in the following contrasts: (i) Real PvG>Imagine PvG, (ii) Imagine PvG>Real PvG and (iii) Real PvG>Non-Moral. A parametric regression analysis was used to explore which brain regions showed a correlation with Money Kept across the Real PvG task. We used a 1–6 parametric regressor weighted to the money chosen per trial—corresponding to the VAS scale used during the Decide event (Figure 1C). No significant activity was found for a parametric regression analysis for the Imagine PvG task. We report activity at P<0.001 uncorrected for multiple spatial comparisons across the whole brain and P<0.05 Family Wise Error (FWE) corrected for the following a priori regions of interest (ROIs; attained by independent coordinates): anterior insula, posterior cingulate cortex (PCC), medial and dorso-medial PFC (mPFC; dmPFC), hippocampus, temporoparietal junction (TPJ), amygdala and dorsolateral PFC (dlPFC). Coordinates were taken from previous related studies1.
Our study was motivated by the observation that moral action does not always reflect moral principle. Based on this, we anticipated that when the opportunity for making real money was salient, participants would favor financial self-interest (at the expense of the Receiver’s pain) more during the real condition when compared with the hypothetical condition. This prediction was confirmed with subjects keeping significantly more money in the Real (£15.77, s.d. ±3.56) vs Imagine PvG task (£14.45, s.d. ±2.94; t=2.52; P=0.025; paired samples t-test, two-tailed; Figure 1D). Importantly, subjects showed no obvious strategy acquisition effects for keeping money over time (see Supplementary Analysis for details). There was no significant correlation between their ratings of the believability of the task and their behavioral performance (Money Kept), r=–0.22, P>0.1. Furthermore, amount of Money Kept could not be explained by subjects modifying their decisions in response to reputation management or feelings of being watched (Landsberger, 1958; r=0.284; P=0.325, see Supplementary Methods for details). Self-reported distress ratings following the viewing of the Video event revealed that the Real PvG was no more distressing than imagining the painful stimulations in the Imagine PvG task (t=0.13; P=0.89; paired samples t-test, two-tailed; Figure 1E). This suggests that the emotional manipulation of watching an aversive video of the moral decision (when compared with viewing a blue screen and simulating the feedback of the decision) had no differential effect on participants’ distress. There was, however, a significant difference between the distress levels reported in the Real PvG compared with the Non-Moral task (t=–2.29; P=0.039; paired samples t-test, 2 tailed; Figure 1E).
In line with the traditional research (Greene et al., 2001), we first compared moral decisions in the Real PvG to decisions in the Non-Moral task, which revealed bilateral amygdala and anterior cingulate cortex (ACC; the Decide event in the Real PvG contrasted with the Decide event in the Non-Moral Task [Table 1])—two regions that are known to process emotionally aversive stimuli (Bechara et al., 2003), especially during emotional conflict (Etkin et al., 2011). That decisions made during the Real PvG reveal patterns of activation within emotion processing areas likely reflects the fact that moral decisions are more emotionally arousing than decisions made within a non-moral context.
To specifically elucidate the differences between real and hypothetical moral decisions, we compared the Decide event (Figure 1B) for the Imagine and Real PvG tasks, highlighting the brain regions distinct to each condition. Significant activation in the PCC, bilateral hippocampus and posterior parietal lobe—all regions essential in imagination and prospection (Schacter et al., 2007)—were greater for hypothetical moral decisions (Figure 2A). Applying a priori ROIs derived from research on the brain’s construction system (Hassabis and Maguire, 2009) revealed a remarkably shared neural system with hypothetical moral decisions (Table 2). Additional a priori ROIs drawn from the moral literature—mPFC and dlPFC (Greene et al., 2001)—also showed greater activation for imagined moral choices. Parameter estimates of the beta values for these ROIs confirmed that these regions were more sensitive to hypothetical moral decisions, relative to real moral decisions (Figure 2A). In contrast, activation in the bilateral ventral TPJ [BA 37], bilateral amygdala, putamen and ACC were more active for real moral decisions (Figure 2B; Table 3). As with the previous contrast, we first applied a priori ROIs and then examined the parameter estimates to ensure that the amygdala and TPJ were significantly more active during real moral decisions. These regions are well documented within the social neuroscience literature and have been closely associated with processing stimuli with emotional and social significance (Phelps, 2006).
We ran a conjunction analysis of all moral decisions to determine if there is a common neural circuitry between real and hypothetical moral decisions (real moral decisions compared with non-moral decisions, along with imagined moral decisions compared with non-moral decisions (Real PvG Decide+Imagine PvG Decide). The results revealed that moral decisions, regardless of condition, shared common activation patterns in the bilateral insula (extending posterior to anterior), middle cingulate (MCC), bilateral dlPFC and bilateral TPJ extending into the posterior superior temporal sulcus (BA 40—which differs from the peak coordinates found for real moral decisions; Figure 2C; Table 4).
Although subjects’ distress ratings across moral tasks were not significantly different [F(1)<1, P=0.99 (Figure 1E)], we wanted to first ensure that the video feedback event in the Real PvG was not driving activation during the Decision event and then examine the Deciders’ socio-affective engagement with the task. Accordingly, we compared neural activation during the Video feed event (Figure 1B) and predicted greater anterior insula activation indexing empathy for pain (Singer et al., 2004) in the Real PvG when compared with the Imagine PvG. As we expected, a comparison analysis of the Video event for the Real vs Imagine PvG (Table 5) revealed greater bilateral anterior insula activation when participants watched real administrations of painful stimulations. This further supports the well-documented proposal that the anterior insula codes for the empathic experience of another’s pain (Singer et al., 2004) and suggests that participants experienced greater socio-affective engagement in the Real version of the task.
One strength of the use of multiple trials in the PvG task is that it gives Deciders the option to either maintain a purely black (keep £20; maximize shocks) or white (keep £0; remove shocks) moral stance, or to position themselves somewhere within the moral ‘gray area’ (£0<keep<£20). This not only has ecological validity, reflecting people’s tendency to qualify moral decisions but also allows us to investigate brain regions associated with different shades of ‘moral gray.’ We therefore conducted a parametric regression analysis and found that increasingly self-interested behavior on the Real PvG task (when Deciders kept more money, parametrically weighted on a scale from 1 to 6) was associated with increased activity in the dorsal ACC (dACC), bilateral dlPFC and orbital frontal cortex (OFC; Figure 3A; Table 6), regions sensitive to cognitive control (Ochsner and Gross, 2005) reward (Kringelbach, 2005) and, in particular, monetary gain (O’Doherty et al., 2001). Activations associated with decreasing self-interest (pro-social behavior) were predominately localized to the mPFC/rostral ACC (rACC), temporal pole and anterior insula (Figure 3B; Table 7; parametric regressions run on the Imagine PvG did not reveal any similar regions). Critically, these parametric analyses also clarified that the activations found during real moral decisions (regardless of the nature of the decision) were not simply an effect of reward classification for monetary gain or loss.
To test for potential motivating factors driving the selfish and pro-social behavior found in the Real PvG, we explored individual differences. Using post-scan questionnaires scores as covariates in a correlated regression for the Real PvG Decision event revealed differential activations for empathic concern and perspective taking (Davis, 1983) and self-reported similarity ratings. Subgenual ACC correlated with increased empathic concern (Table 8) while decreasing empathic concern and perspective-taking activated left putamen, dACC, bilateral dlPFC and bilateral OFC (Table 9). Finally, similarity ratings negatively correlated with the amount of Money Kept and elicited activation in the right anterior insula, while increasing similarity ratings correlated with activation in the ACC mPFC, dmPFC and left OFC (Tables 10 and and11,11, respectively).
This study examined the moral dynamic of self-gain vs other-welfare during real and hypothetical conditions. Our behavioral results show that moral decisions with real consequences diverge from hypothetical moral choices, verifying the ‘hypothetical bias’ effect (Kang et al., 2011). Compared with imagining their moral actions, people who make moral decisions under real conditions keep more money and inflict more pain on another subject. Although the research exploring real moral action is limited (Moll et al., 2006; Baumgartner et al., 2009; Greene and Paxton, 2009), our results stand in stark contrast to findings demonstrating that people act more morally than they think they will (Teper et al., 2011). Our results also contradict the accumulated research illustrating a basic aversion to harming others (Greene et al., 2001; Cushman et al., 2012). We contend that this is likely due to the fact that many of the moral scenarios used within the moral literature do not pit the fundamental motivation of not harming others (physically or psychological) against that of maximizing self-gain (Haidt, 2007). Accordingly, our findings reveal that engaging the complex motivations of self-benefit—a force endemic to many moral decisions—can critically influence moral action.
Our fMRI results identify a common neural network for real and hypothetical moral cognition, as well as distinct circuitry specific to real and imagined moral choices. Moral decisions—regardless of condition—activated the insula, MCC and dorsal TPJ, areas essential in higher order social processes, such as empathy (Singer et al., 2004). This neural circuitry is well instantiated in the social neuroscience literature and fits with the findings that moral choices are influenced by neural systems whose primary role is to facilitate cooperation (Rilling and Sanfey, 2011). The TPJ has been specifically implicated in decoding social cues, such as agency, intentionality and the mental states of others (Young and Saxe, 2008). For example, TPJ activation correlates with the extent to which another’s intentions are taken into account (Young and Saxe, 2009) and transiently disrupting TPJ activity leads to interference with using mental state information to make moral judgments (Young et al., 2010). Although there is a large amount of research indicating that the TPJ codes for our ability to mentalize, there is also evidence that the TPJ activates during attentional switching (Mitchell, 2008). In addition, one study revealed that patients with lesions to the TPJ do not show domain-specific deficits for false belief tasks (Apperly et al., 2007). Although these differential findings suggest that the specific functionality of the TPJ remains unclear, we propose that TPJ engagement during real and imagined moral decisions suggests a similar mentalizing process is at play in both real and hypothetical moral decision-making: when deciding how much harm to apply to another, subjects may conscript a mental state representation of the Receiver, allowing them to weigh up the potential consequences of their decision. This neural finding reinforces the role of the TPJ—and thus the likely role of mental state reasoning and inference—in moral reasoning.
However, we also found distinct neural signatures for both real and imagined moral decisions. In line with the literature, hypothetical moral decisions were specifically subserved by activations in the PCC and mPFC—regions also implicated in prospection, by which abridged simulations of reality are generated (Gilbert and Wilson, 2007). Although the overall pattern of brain activation during these hypothetical moral decisions replicates the moral network identified in previous research (Greene et al., 2001), the fact that the PCC and mPFC are activated both during prospection and during hypothetical moral decision-making implies that this region is recruited for a wide spectrum of imagination-based cognition (Hassabis and Maguire, 2009). Thus, either hypothetical moral decisions and imagination share a similar network or hypothetical moral decisions significantly rely on the imperfect systems of prospection and imagination. Further research exploring whether the PCC and mPFC are specific to hypothetical moral decisions, or recruited more generally for imagining future events, would help clarify their roles within the moral network.
In contrast, real moral decisions differentially recruited the amygdala. These results are consistent with the vast literature implicating the amygdala in processing social evaluations (Phelps, 2006), emotionally relevant information (Sander et al., 2003) and salient stimuli (Ewbank et al., 2009). Research on moral cognition further implicates amygdala activation in response to aversive moral phenomena (Berthoz et al., 2006; Kedia et al., 2008; Glenn et al., 2009); however, this finding is not systematically observed in moral paradigms (Raine and Yang, 2006). In line with the literature, it is possible that in the Real PvG task the amygdala is coding the aversive nature of the moral decision; however, distress ratings indicated that both conditions were perceived as equally aversive. Accordingly, an alternative interpretation is that the amygdala is monitoring the salience, relevance and motivational significance (Mitchell et al., 2002) of the real moral choice space. Decisions, which produce real aversive consequences (i.e. lose money or harm another), are far more salient and meaningful than decisions that do not incur behaviorally relevant outcomes. The amygdala is also commonly recruited for decisions which rely on social signals to emotionally learn positive and negative associations (Hooker et al., 2006). It is possible that the amygdala activation found for real moral decisions is signaling reinforcement expectancy information of both the positively (self-benefit) and negatively (harm to another) valenced stimuli (Blair, 2007), which then subsequently guides behavior (Prevost et al., 2011). This theory not only accounts for the differential behavioral findings between the real and hypothetical conditions but also it is consistent with the more general theoretical consensus regarding human moral cognition (Moll et al., 2005), which emphasizes how lower order regions like the amygdala modulate higher order rational processes (Dalgleish, 2004).
Our fMRI results further indicate that there are dissociable neural mechanisms underlying selfish and pro-social decisions. In the Real PvG, decisions that maximized financial benefit (selfish decisions) correlated with activity in the OFC, dlPFC and dACC—regions that support the integration of reward and value representations (Schoenbaum and Roesch, 2005), specifically monetary gain (Holroyd et al., 2004) and loss (Bush et al., 2002). Furthermore, the dACC was found to negatively correlate with empathic concern scores and positively correlate with self-reported similarity ratings in the Real PvG task. Together, this suggests that the dACC may be monitoring conflicting motive states (Etkin et al., 2011). However, the dACC has been further implicated in a variety of other functions, including emotion regulation (Etkin et al., 2011), and weighing up different competing choices (Mansouri et al., 2009). Thus, it is equally plausible that the dACC is processing the conflicting negative emotions involved with choosing to harm another for self-gain (Amodio and Frith, 2006).
In the PvG task, the morally guided choice is to give up the money to prevent harm to another. Unlike selfish decisions, such pro-social decisions showed significantly greater activation in the rACC/mPFC and right temporal pole, demonstrating that the nature of real moral decisions can be predicted by dissociable networks within the PFC. The rACC/mPFC is a structure engaged in generating empathic feelings for in-group members (Mathur et al., 2010) and for coding feelings of altruistic guilt and distress during theory of mind tasks (Fletcher et al., 1995). Clinical data have also shown that lesions to this area stunt moral emotions, such as compassion, shame and guilt, and contribute to overall deficits in emotional processing (Mendez and Shapira, 2009). In fact, research has demonstrated the rACC/mPFC as a region that responds specifically to the aversion of not harming others (Young and Dungan, 2011). Based on this, we propose that the rACC/mPFC activation found for pro-social decisions could be attributed to the empathic response generated by the emotional aversion (distress) of harming another—a key motivational influence and proximate mechanism of altruistic behavior.
Theorists have pointed to the importance of studying moral cognition in ecological valid and consequence-driven environments (Casebeer, 2003; Moll et al., 2005). Our results illustrate that specific regions of the moral network subserve moral choices—regardless of whether they are real or imagined. However, we also found a divergence between real moral behavior and hypothetical moral intentions—which was reflected in the recruitment of differential neurobiological systems. Thus, if morality is a domain where situational influences and the impact of imminent, real consequences can sway our decisions, then it is crucial that cognitive neuroscience investigate moral decision-making under real conditions. This seems especially relevant in light of this new neurobiological evidence, supporting what the philosopher Hume presciently noted—‘the most lively thought is still inferior to the dullest sensation’ (Hume, 1977).
Supplementary data are available at SCAN online.
This research was supported by the Medical Research Council Cognition and Brain Sciences Unit.
1We used a priori coordinates to define ROI in our analysis. All ROIs were selected on the basis of independent coordinates using a sphere of 6-10 mm (sphere size was defined by the corresponding structure) and corrected at P<0.05 FWE and were attained through MarsBaRs. Peak voxels are presented in the tables at P<0.001 uncorrected and images are shown at P<0.005 uncorrected.