Is Morality Unified? Evidence that Distinct Neural Systems Underlie Moral Judgments of Harm, Dishonesty, and Disgust

Is it worse to cheat on an exam or to eat your dog? Consideration of these acts feels very different, yet we tend to classify a diverse variety of acts (e.g., theft, lying, assault, disrespect, incest) as “morally wrong.” This common linguistic label does not ensure that moral judgments of such diverse acts are similar enough to be studied as a single kind of judgment. Nonetheless, research investigating moral judgments has tended to treat moral judgments pertaining to vastly different contents as comparable or even equivalent.

Many theorists suggest that there is a universal, innate human moral faculty (Hauser, 2006; Dwyer, 1999; Mikhail, Sorrentino, & Spelke, 1998; Rawls, 1971). Such theories posit that our moral judgments are fueled neither by emotion nor by conscious deliberation but by an unconscious evaluative mechanism that makes judgments based on innate, often inarticulable principles analogous to Chomsky's universal grammar. Others have emphasized the causal role of emotions in moral judgments (Wheatley & Haidt, 2005; Haidt, 2001). Still others have suggested that emotional responses and cognitive computations work in tandem to give rise to moral judgments (Dupoux & Jacob, 2007; Greene, Nystrom, Engell, Darley, & Cohen, 2004). The existence of any such shared, uniform mechanism—be it rational, affect-driven, or both—would predict shared neural responses across moral domains.

Some philosophers (Prinz, 2008), psychologists (Blair, Marsh, Finger, Blair, & Luo, 2006; Moll, Zahn, de Oliveira-Souza, Krueger, & Grafman, 2005; Haidt, 2001), and anthropologists (Shweder, Much, Mahapatra, & Park, 1997) have questioned the assumption that morality is homogenous enough to be treated as a unified faculty. Behavioral data indicates that some moral judgments are rooted in conscious reasoning whereas others are not (Cushman, Young, & Hauser, 2006) and that psychopathy scores correlate with moral judgments in some areas more than others (Glenn, Iyer, Graham, Koleva, & Haidt, 2009). These findings have underscored the need to adopt a more fine-grained approach to studying morality.

Such an approach may be especially fruitful in identifying the neural underpinnings of moral judgment. There is evidence that the neural correlates of harm judgments differ depending on whether the harms are direct or indirect (Greene et al., 2004; Greene, Sommerville, Nystrom, Darley, & Cohen, 2001), as well as whether they are intended or unintended (Young & Saxe, 2008; Borg, Hynes, Van Horn, Grafton, & Sinnott-Armstrong, 2006). There is also some evidence that different brain areas are recruited by the judgment of scenarios involving bodily harm compared with those that do not (Heekeren et al., 2005). Independent studies of dishonesty (Greene & Paxton, 2009) and disgust (Borg, Lieberman, & Kiehl, 2008; Moll, de Oliveira-Souza, et al., 2005) report distinct patterns of brain activations associated with judgments of moral transgressions within these broad categories, suggesting that these judgments are distinct from one another and from harm judgments. In addition, different morally relevant emotions such as guilt, anger, and embarrassment are associated with dissociable networks of brain regions (Immorino-Yang, McColl, Damasio, & Damasio, 2009; Zahn et al., 2009; Kedia, Berthoz, Wessa, Hilton, & Martinot, 2008; Moll et al., 2007; Takahashi et al., 2004; Moll, de Oliveira-Souza, & Eslinger, 2003; Berthoz, Armony, Blair, & Dolan, 2002; Dougherty et al., 1999). Across these studies, activation of the dorsal medial pFC (DMPFC), an area implicated in self-referential processing (Northoff et al., 2006) and thinking about others (Amodio & Frith, 2006), especially in ambiguous circumstances (Jenkins & Mitchell, 2010), is most widely associated with morally relevant processing (Moll, Zahn, et al., 2005). Importantly, although independent studies have been conducted within multiple moral areas (e.g., harm, disgust, and dishonesty), methodological differences between studies render direct comparisons difficult.

Another obstacle to understanding the neural basis of moral judgment is the widespread tendency to contrast moral scenarios to each other and to neutral scenarios without considering whether participants judge acts in those scenarios to be morally wrong. Notably, one study (Greene et al., 2004) found significant activation differences between those who judged the same act to be morally permissible or impermissible within the category of physical harm. It remains to be seen whether these differences extend to other moral domains. Accordingly, we sought to determine whether any neural substrate is common to judgments of moral transgressions (as compared with neutral scenarios) within or across various domains, or with judgments that an act is morally wrong (as compared with not morally wrong) within or across domains. It was predicted that judgments of different categories of moral transgressions would be related to different patterns of activity in the brain.

More specifically, it was predicted that disgusting moral transgressions would evoke activity in areas associated with affective processing, such as the amygdala, OFC, and insula (Jabbi, Bastiaansen, & Keysers, 2008; Britton et al., 2005), whereas dishonest transgressions would elicit activity in areas activated when inferring others' mental states, such as the TPJ (Saxe & Kanwisher, 2003), the posterior cingulate cortex (PCC), and the DMPFC (Rilling, Sanfey, Aronson, Nystrom, & Cohen, 2004). In line with the bulk of fMRI studies of moral reasoning that have focused on harm (e.g., Greene et al., 2001), it was predicted that the dorsolateral pFC (DLPFC), ACC, and the aforementioned mentalizing areas, as well as areas implicated in action understanding (posterior STS, inferior parietal lobule [IPL], SMA; Grafton, Arbib, Fadiga, & Rizzolatti, 1996) and, more specifically, in processing socially informative sensory information (posterior STS; Allison, Puce, & McCarthy, 2000) would show activity in response to moral transgressions involving bodily harms.

The neural correlates of “wrong” and “not wrong” judgments within and across story types were compared. As activation differences between “wrong” and “not wrong” judgments have only been found within the category of harm (Greene et al., 2004) and other studies (Borg et al., 2006) have failed to find significant differences, no specific hypotheses were made about the neural correlates of these judgments.

Three rounds of pilot testing involving 75 participants at Princeton University and Dartmouth College yielded 14 acceptable scenarios in each of four categories: Disgusting, Harmful, and Dishonest moral scenarios, and Neutral (morally irrelevant) scenarios. At the start of pilot testing, participants were told that they would read multiple scenarios and make several ratings of each one. Participants were instructed to accept the details as specified in the scenario. For example, if a person is described as teasing another person, the participant should accept teasing as an accurate description of the behavior. Participants were also instructed that if they fill in details, those details should be what the participant takes to be most likely (e.g., if the story says that Jim asked Matt a question, the participant should assume that Jim was speaking loudly enough for Matt to hear him). Then, participants were given a sample scenario and the following descriptions of the ratings they were to make for each scenario:

• Is the act morally wrong? This question is not about whether the act is or should be illegal. Nor is it about whether you would do the act or advise friends to do it. The question is simply about whether that particular act is morally wrong in the specified circumstances. In the above example, the question is not whether it is wrong in all circumstances to tell somebody to shut up but whether it is wrong in the particular scenario for Mike to tell John to shut up.

• Is the act harmful? This question is not about whether harm would result if this kind of act became common or if other people found out about it. The question is only about whether this act causes any harm in the particular circumstances as specified.

• Is the act dishonest? This question is about whether the act depends on intentional deception, promise breaking, cheating, theft, or something like that. It is not about whether everyone in the scenario knows exactly what is going on.

• Is the act disgusting? This question is not about whether it is morally horrendous. It is, instead, about whether the act is disgusting in the way that most people find it disgusting to lick toilets or to have oral sex with their grandparents.

After reading the instructions, the pilot participants made the dichotomous (yes or no) judgments listed above for each scenario. Scenarios were considered acceptable for use in the fMRI study if they adhered to the following criteria: A Neutral scenario was deemed acceptable if 80% or more of the pilot participants judged the associated act to be morally “not wrong” and no more than 20% of pilot participants deemed the act to be disgusting, harmful or dishonest. The moral scenarios (Disgusting, Harmful and Dishonest scenarios) were deemed acceptable if at least 80% of the pilot participants judged the associated act as representative of its intended moral domain (e.g., Disgusting) and no more than 20% of pilot participants deemed the act to be representative of the other two moral domains (e.g., Harmful, Dishonest). These criteria helped ensure that the transgression scenarios were representative of one and only one moral domain. In addition, the moral scenarios were acceptable only if no less than 30% and no more than 70% of pilot participants considered them to be morally “wrong.” This inherent ambiguity afforded sufficient response variability to contrast “wrong” and “not wrong” responses within and across moral categories. The scenarios selected for the fMRI study did not significantly differ in length between categories (mean lengths for all four categories ranged from 44 to 46 words; F(3, 55) = 1.95, ns). For further analyses of the contents of scenarios in different categories, refer to Table A3, Table A4, and Figure A1 of the Appendix.

For the fMRI study, each scenario was converted into scrolling black text on a light blue background using LiveType software (Apple, Inc.). This ensured that participants read the same words at approximately the same time. Each scenario scrolled for 22 sec, and the full text remained on the screen for 2 sec after scrolling had ceased.

Thirty-eight naive, right-handed adults were scanned in a 3.0-T Phillips scanner. All participants were right-handed English speakers with normal or corrected-to-normal vision. Before scanning, the experimenter obtained informed consent and explained the task to the participant. Specifically, participants were told that they would read scrolling text scenarios describing the actions of a person. After each scenario, they were prompted to press a button indicating whether they considered the main character's actions to be “wrong” or “not wrong,” and then to press a second button indicating the confidence level of their judgment (“High” or “Low”). See Figure 1 for a schematic representation of a single trial.

Between reading and judging scenarios, participants viewed a fixation cross for varying amounts of time and pressed a button whenever the cross changed color. The order of the stimuli and the intervening fixation periods was optimized for hemodynamic deconvolution by the AFNI program RSFgen (Cox, 1996).

After leaving the scanner, participants were provided with a written copy of all of the scenarios they had viewed while being scanned, as well as definitions of the terms “disgusting,” “harmful,” and “dishonest” (the same definitions used in pilot testing). After reading these definitions, participants were asked if the actions described in each scenario could be characterized as disgusting, harmful, and/or dishonest. In addition, participants rated if each scenario evoked emotion using a Likert scale from 1 (not at all) to 4 (intensely).

The study was performed using a Philips Intera Achieva 3-T scanner (Philips Medical Systems, Bothell, WA) with a SENSE (SENSEitivity Encoding) head coil. All stimuli were presented to the participant via a Panasonic DT-4000U DLP projector. Anatomical images were collected using a high-resolution 3-D Magnetization-prepared rapid gradient-echo sequence (124 slices, 1.2 mm thick, FOV = 240 mm, acquisition matrix = 256 × 256). Functional images were collected in seven runs using echo-planar functional images sensitive to BOLD contrast (TR = 2000 msec, TE = 30 msec, flip angle = 90°, 3 mm in-plane resolution). During each of the seven functional runs, 180 sets of axial images (36 slices/volume, 3.5 mm slice thickness with 0.5 mm skip) were collected in an interleaved fashion across the whole brain. Axial images were not angulated. See Figures A2 and A3 in the Appendix for echo-planer magnetic resonance images of a single participant to illustrate brain coverage.

Functional and anatomical images were analyzed with AFNI (Cox, 1996). Transient spikes in the signal were suppressed with the AFNI program 3dDespike. As the slices of each volume were not acquired simultaneously, a timing correction procedure was used. All volumes were motion corrected, normalized to the standardized space of Talairach and Tournoux (1988), and smoothed with a 6-mm FWHM smoothing kernel before conversion to percent signal change from the mean.

A regression analysis using 3dREMLfit in AFNI was conducted on each participant's preprocessed EPI volumes to estimate the individual hemodynamic response to each of the five stimulus conditions. Stick functions indicating the onset times of each stimulus were convolved with a 30-sec block function. Second-order trends and the participant's head motion were included as regressors at this stage to account for signal changes due to scanner drift and motion artifact. The convolved hemodynamic response for each condition for each participant was submitted to a series of t tests conducted with 3dMEMA (Chen, Saad, & Cox, 2010), an AFNI program that adopts a mixed-effects meta-analysis approach and takes into account both within- and between-participants variability, making it relatively robust to outliers.

Unless otherwise noted, all reported clusters survived a statistical threshold of p < .05, corrected for multiple comparisons as stipulated by AlphaSim within AFNI (afni.nimh.nih.gov/pub/dist/doc/manual/AlphaSim.pdf). AlphaSim carries out Monte Carlo simulations by randomly generating uncorrelated three-dimensional images using the overall dimensions and voxel geometry of a data set, followed by convolution with a Gaussian function to account for correlations between nearby voxels. Using the parameters of our data set (FWHM = 6 mm and a brain-only mask), we ran 1000 iterations over a skull-stripped whole brain mask, which indicated that a voxel-wise threshold of p < .005 and a cluster size threshold of 675 mm³ (25 × 27 mm³ voxels) would provide a family-wise false positive rate of .05.

Consistent with the possibility that different categories of moral judgment are subserved by different cognitive systems, postscanner questionnaire results indicated that participants classified scenarios into only the appropriate categories. Disgusting scenarios were rated as disgusting but neither dishonest nor harmful, χ²(5, n = 1560) = 1113.07, p < .0001. Dishonest scenarios were rated as dishonest but neither disgusting nor harmful, χ²(5, n = 1560) = 1369.45, p < .0001, and Harmful scenarios were rated as harmful but neither dishonest nor disgusting χ²(5, n = 1560) = 1302.26, p < .0001. Neutral scenarios were rated as neither disgusting nor harmful nor dishonest χ²(5, n = 1560) = 1508.47, p < .0001.

Table A2 in the Appendix contains the number of “wrong” and “not wrong” decisions for all moral and neutral scenarios, as well as the variance of participants' responses for each scenario and for each category. A one-way ANOVA of the three moral categories (Harm, Dishonesty, Disgust) with response variance across participants for each scenario as the dependent variable was performed. As Levene's test of equality of variances indicated that the assumption of homogeneity of variance was violated (F(2, 39) = 3.56, p < .05), the Brown–Forsythe F ratio is reported. There were differences between moral categories with respect to response variability, F(2, 31.42) = 6.84, p < .01. The Games–Howell post hoc test was used to detect differences between categories because it does not assume equal variances (Games & Howell, 1976). Disgust scenarios elicited less variable responses than both Harm and Dishonesty scenarios (both p's < .05), whereas Dishonesty and Harm scenarios did not differ in response variability (p = .81).

Whereas syntactic complexity was kept consistent across categories (see Table A4), semantic content varied by design. In the interest of ecological validity, scenarios were constructed to resemble as closely as possible how they might be described colloquially. As such, scenarios in each category differed with respect to their characteristic semantic content. In addition, Disgust scenarios were naturally more emotionally evocative than Neutral, Dishonest, or Harmful scenarios. These between-category differences and within-category similarities ensured that scenarios were classified easily into their respective moral domains and only their respective moral domains. See Figure A1 and Table A3 in the Appendix for analyses of emotional evocativeness and semantic content, respectively. Also refer to the Appendix for a full list of scenarios used in the fMRI study.

To identify the neural correlates of processing different types of moral transgressions, we conducted two sets of analyses. In the first set of analyses, we compared brain activity associated with the judgment of each type of moral transgression to that of morally neutral scenarios, as well as to transgressions in the remaining two moral categories. A moral transgression was defined as an act in a moral scenario that was judged to be morally wrong. Previous behavioral results indicate that individuals differ in which acts they classify as morally wrong versus merely conventionally inappropriate (Haidt & Hersh, 2001). Further, there exist significant differences in brain activity between participants who judge the same physical harms to be either morally permissible or impermissible (Greene et al., 2004). Limiting the first set of analyses to moral scenarios deemed wrong by the participants, therefore, afforded cross-subject comparisons most relevant to participants' judgments of moral wrongness. Please refer to Table A1 for similar analyses including brain activity associated with all moral stories as compared with neutral, irrespective of the associated decision (“wrong”/“not wrong”). Importantly, because all conditions, both moral and neutral, involved moral reasoning, the contrasts outlined in Table A1 do not reveal the neural correlates of moral reasoning. Rather, these differences in the hemodynamic response reflect differences associated with the processing of moral scenarios relative to neutral scenarios, irrespective of the ensuing moral judgment. The following comparisons contrast the hemodynamic responses associated with moral transgressions compared with neutral scenarios. “Moral transgressions” entered into the analysis for each participant include only those scenarios that he or she deemed to be morally “wrong.” Our primary question was whether different moral transgressions would evoke hemodynamic responses consistent with a unified or a disunified (separate systems) account of moral judgment.

The judgment of transgressions involving physical harms compared with neutral scenarios was associated with increased activity in the left DLPFC, a region of DMPFC including the ACC, the SMA, the IPL, and the posterior STS (see Table 1). Similarly, compared with judgments of both Disgusting and Dishonest transgressions, judgment of Harmful moral transgressions was associated with increased activity in an area of the DLPFC that included the SMA as well as the IPL and the anterior aspect of the precuneus (see Table 2).

As predicted, the judgment of Dishonest transgressions compared with neutral scenarios was correlated bilaterally with increased activity in the DMPFC, the TPJ extending superiorly into the IPL and the PCC, as well as increased activity in the left DLPFC (Table 1). Dishonest scenarios elicited increased activity in the left DLPFC and bilaterally in the TPJ compared with both Disgusting and Harmful transgressions (Table 2).

The Disgusting versus Neutral scenario contrast revealed increased activity bilaterally in the DMPFC, amygdalae, ACC, and PCC, as well as the right temporal pole (TP), left DLPFC, and a region in the left inferior frontal cortex encompassing the frontal operculum (FO) and the anterior aspect of the insula (see Table 1). Compared with both Dishonest and Harm scenarios, Disgust scenarios elicited increased activity bilaterally in the amygdalae, TPs, and lingual gyri, in the left FO/anterior insula and fusiform gyri, as well as the right subgenual ACC (sgACC) and posterior portion of the precuneus (see Table 2).

To explore whether there was any activation common to all transgression types relative to neutral scenarios, we performed a conjunction analysis. One 28-voxel cluster in the DMPFC was independently activated in the comparison of each moral transgression type compared with neutral scenarios (surviving a threshold of p < .05, corrected, for all comparisons). To investigate whether this region was activated specifically by moral transgressions, this cluster was then applied as a mask to each individual's data to extract the average hemodynamic response for the “not wrong” scenarios for each category type. This secondary ROI analysis revealed that Dishonest and Harmful scenarios judged to be “not wrong” also activated this region more strongly than Neutral scenarios (p < .005; p < .05, respectively). See Figure 3.

No significant activation differences between “wrong” and “not wrong” judgments within moral categories emerged. Also, judgment confidence (high vs. low) was not a significant predictor of brain activity within any system, with the exception of Disgust. Within the network of areas involved in the judgment of disgusting transgressions, confidence level modulated activity in the PCC. Greater judgment confidence was associated with greater activity in this area (Talairach coordinate of center of mass: −1, −48, 26; 30 voxels).

The behavioral and imaging results reported here suggest that multiple distinct cognitive systems support moral judgment (see Figure 2). These systems are consistent with multiple intercategory differences in semantic content and the associated processing demands of that content. Judgments of transgressions involving physical harm activated areas commonly associated with action understanding. Judgments of dishonest transgressions evoked activity in areas associated with representing other people's beliefs. Judgments of disgusting transgressions evoked activity in areas associated with affective processing.

Transgressions involving physical harms recruited areas associated with understanding and imagining actions (Grafton et al., 1996): the IPL, the STS, and the SMA. Additionally, the DLPFC and ACC were more activated by the judgment of harms than for that of neutral scenarios, consistent with previous studies implicating these areas in moral reasoning about physical harms (Greene et al., 2001). Further, compared with Dishonest and Disgusting transgressions, the judgment of transgressions involving physical harms elicited greater activity in areas associated with action understanding, such as the SMA and IPL, as well as a region of the precuneus implicated in self-centered visuospatial imagery strategies (Cavanna & Trimble, 2005). These results suggest that in comparison with other types of moral transgressions, understanding moral transgressions involving physical harms relies more on processing and imagining the physical actions described in these scenarios.

The judgment of Dishonest transgressions was correlated with increased bilateral activity in a region encompassing the TPJ and IPL that is implicated in reasoning about mental (Saxe & Kanwisher, 2003; Fletcher et al., 1995) and nonmental (Apperly, Samson, Chiavarino, Bickerton, & Humphreys, 2007) representations, simultaneously holding on-line multiple possible perspectives (Decety & Jackson, 2004), and attentional reorienting more generally (Corbetta, Patel, & Shulman, 2008; Astafiev, Shulman, & Corbetta, 2006). Dishonest transgressions also preferentially activated the left PCC and DMPFC, areas robustly associated with inferring others' mental states (Rilling et al., 2004). Similarly, compared with judging other moral transgressions, judging Dishonest transgressions was associated with DLPFC and TPJ activity (Table 2). These findings suggest that understanding dishonest transgressions requires mentalizing, or, more generally, a shifting of attention between deceitful actors' conflicting words and actions. Interestingly, however, Dishonest and Harmful scenarios did not significantly differ from each other in the frequency of belief or communication verbs (see Table A3 in the Appendix). This suggests that the increased bilateral activity in the TPJ for Dishonest, as compared with Harmful, transgressions (see Table 2) does not reflect the processing of mental state words per se but how the words, in combination, evoke different cognitive operations such as mentalizing. Alternatively, the bilateral activation of TPJ for Dishonest scenarios may reflect other cognitive operations associated with this region such as the discrimination of self and other (Cheng, Chen, Lin, Chou, & Decety, 2010; Decety & Sommerville, 2003). For example, the Dishonest scenarios (e.g., academic dishonesty) may have hit particularly “close to home” with our college-aged participants, thereby evoking a need to distance oneself from the protagonist.

The PCC is also associated with social–emotional processing (Britton et al., 2005) and the evaluation of emotionally evocative stimuli (Maddock, 1999). Therefore, its activation in the Dishonest versus Neutral contrast could also reflect the greater emotional arousal and evaluative processing elicited by Dishonest transgressions relative to Neutral scenarios. The activations observed in this contrast could reflect a greater need for deliberation when judging Dishonest, as compared with Neutral, scenarios.

Disgusting transgressions were associated with increased bilateral activity in the amygdalae, DMPFC, and PCC, areas implicated in social–emotional processing (Britton et al., 2005) as well as in the right TP, implicated in binding visceral affective responses to complex stimuli (Olson, Plotzker, & Ezzyat, 2007), in understanding social norm violations (Mendez, Anderson, & Shapira, 2005), and in social conceptual understanding more generally (Zahn et al., 2006). This comparison also revealed increased activity in the anterior portion of the left insula extending into the FO, a region associated with the experience of disgust as well as disgusting mental imagery (Jabbi et al., 2008). Results of this contrast suggest that the judgment of disgusting moral transgressions relative to neutral scenarios involves affective processing as well as social evaluation. This result is consistent with the higher number of emotion words contained in Disgust scenarios (see Table A3), as well as the higher ratings of elicited emotion for these scenarios (see Figure A1).

Compared with Dishonest and Harmful moral transgressions, judgments of Disgusting moral transgressions were correlated with bilateral activity in areas also identified in the Disgust versus Neutral contrast that are involved in general emotional processing (amygdalae, TP), as well as emotional processing specific to disgust (left FO/insula). These contrasts also revealed activity in the sgACC, an area implicated in emotional arousal and interoception (Critchley, 2004). Both contrasts also revealed increased activity in a region of the precuneus that has been implicated in the attribution of emotions to the self and others (Ochsner et al., 2004), possibly indicating that emotional empathy was preferentially engaged by scenarios involving sexual disgust. The bilateral lingual and parahippocampal gyri activations observed in both of these contrasts have previously been associated with processing erotic stimuli (Hu et al., 2008). Finally, the activation of the fusiform gyrus by Disgust scenarios could reflect increased imageability for disgusting transgressions relative to other transgression types, as activity in the fusiform gyrus has been consistently associated with mental imagery (e.g., Ishai, Ungerleider, & Haxby, 2000). Together, these activation differences highlight the many inherent characteristics of content and processing demands that differentiate moral transgressions that involve sexual disgust from other types of moral transgressions.

The differences in semantic content between moral categories almost certainly yielded concomitant differences in cognitive operations engaged by those contents. Such semantic and processing differences are expected in scenarios designed to be as natural and emblematic of their respective categories as possible. Indeed, the moral categories used here most likely differ in a number of ways left unspecified (e.g., imageability). These many differences underline the fact that moral transgressions are intrinsically diverse and differ systematically for several reasons, only some of which can be quantified by analyses of scenario contents, behavioral ratings by participants, and associated neural activity. Taken together, the diversity of content and concomitant cognitive and emotional processes coupled with the differences in brain activity between categories of moral transgressions further illustrate that “moral” judgment is best investigated as a superordinate category of judgments that share “family resemblances” (Rosch & Mervis, 1975) rather than any defining feature or set of features.

In addition to differences between scenarios that were consistent across participants, in at least some cases, the cognitive processes elicited by the same moral scenarios likely differed somewhat between participants. For instance, interesting work by Kedia et al. (2008) and Moll et al. (2007) indicates that agency perspective affects the neural correlates of moral processing. Although all scenarios were written in the third person, it cannot be ascertained whether individual participants took the perspective of the victim or transgressor while reading. Similarly, although many studies (Immorino-Yang et al., 2009; Zahn et al., 2009; Kedia et al., 2008; Moll et al., 2003, 2007; Takahashi et al., 2004; Berthoz et al., 2002; Dougherty et al., 1999) indicate that experiencing different moral emotions is associated with neural activation differences, we cannot ascertain what specific emotion(s) participants were experiencing while reading each scenario, which likely differed between participants for many reasons, including the agency perspective that participants took while reading a scenario. For instance in Dishonesty scenarios, different participants could have experienced indignation or guilt, depending on the agency perspective they assumed while reading. Similarly, for Harm scenarios, different participants could have differentially experienced compassion for the victim and/or contempt for the aggressor to varying degrees. For these reasons, characterizing the moral emotions evoked by different transgression types is beyond the scope of the present study. This endeavor and, more generally, the detailed characterization of the many differences that exist between categories of moral transgressions remains an important direction for future research.

A region in the DMPFC was independently activated by the contrast of each moral transgression type compared with Neutral scenarios. This region was also activated by Dishonest and Harmful scenarios judged to be “not wrong” relative to Neutral scenarios. Together, these findings suggest that the activity observed in this region reflects evaluative processes engaged by the moral scenarios that are not peculiar to a particular moral decision (“wrong” or “not wrong”). This interpretation is consistent with at least two lines of research investigating the characteristics of this region.

First, activity in this region is known to be modulated by ambiguity both in social (Jenkins & Mitchell, 2010) and nonsocial (Volz, Schubotz, & von Cramon, 2003) contexts. Our moral scenarios were ambiguous by design to allow for comparisons of “wrong” and “not wrong” judgments of the same scenarios. This ambiguity was not present in the neutral scenarios and is reflected in the high variability of responses to moral scenarios as compared with neutral scenarios (see Table A2 in the Appendix). Thus, the processing of ambiguous information represents one explanation for the shared DMPFC activation observed in the moral versus neutral comparisons.

Second, this area is consistently activated when thinking about other people (Van Overwalle, 2009; Amodio & Frith, 2006; Gallagher & Frith, 2003; Goel, Grafman, Sadato, & Hallett, 1995) and in self-referential processing (Northoff et al., 2006). It is likely that the evaluation of moral transgressions, as compared with neutral scenarios, involves a deeper evaluation of individuals' motives and actions and how they relate to one's own values. Of particular relevance to the current study, a recent meta-analysis found this region to be involved not only in moral judgment tasks, but also in mentalizing and in understanding goal-directed actions (Van Overwalle & Baetens, 2009). This explanation is not mutually exclusive with that of ambiguity resolution. Indeed, the observed activity in this region is consistent with its known role in evaluating other people in ambiguous situations (Jenkins & Mitchell, 2010).

Because participants were asked to employ moral reasoning (i.e., to answer the question, “Was it morally wrong?”) on all trials, both moral and neutral, characterizing the neural underpinnings of moral reasoning, as compared with other types of reasoning, is beyond the scope of the current study. Rather, this study was primarily concerned with characterizing the processing of what participants deemed to be moral transgressions. Nevertheless, because moral reasoning was common to moral and neutral trials, the finding that the DMPFC was activated by the conjunction of all moral transgression types relative to neutral scenarios (see Figure 3) and by the conjunction of all moral story types relative to neutral stories (see Table A1 in the Appendix) suggests that some property of the moral scenarios relative to the neutral scenarios was driving this activation—possibly their inherent ambiguity.

Regardless of the exact mechanisms involved, the present results suggest that the overlapping activation in DMPFC reflects aspects of processing moral transgressions that are not peculiar to the decision of moral wrongness. This suggests that this region is not the seat of a unified moral faculty in the strongest sense—that is, a faculty that is dedicated to judgments of moral transgressions in particular. The present results are consistent with a more general role of this region in the evaluation of human behavior, of which moral action is a subcategory. Thus, lesions or disruptions to this region would likely result in deficits or differences in making moral judgments due to the disruption of evaluative processes (e.g., ambiguity resolution, mentalizing) that are not specific to the domain of morality.

Although we had hoped to compare “wrong” and “not wrong” judgments within categories, we lacked sufficient power to do so. Although our sample was relatively large (n = 38), subtle activation differences between “wrong” and “not wrong” judgments may have been diluted by the length of the written scenarios used, which limited the number of exemplars per condition and precluded isolating the specific time at which an individual reached a moral judgment for each scenario. Future studies employing designs with greater temporal constraints may find more significant differences between “wrong” and “not wrong” moral judgments. However, these differences appear to be subtle in comparison with differences between moral transgression types.

Differences between aspects of the content of the scenarios (such as whether they depict physical violence, deceit, or sexual acts that elicit disgust) were much more robust. This suggests that previously observed patterns of activity that characterize moral judgment may reflect, at least in part, processing of characteristics of these scenarios independent of whether acts are judged morally wrong.

These results provide empirical support for philosophical arguments against the existence of a functional or anatomical module common and peculiar to all moral judgments (Prinz, 2008; Sinott-Armstrong, 2008). Although consistent networks of brain regions have been reliably implicated in some dimensions of cognition using diverse methods (e.g., theory of mind; Saxe, 2009), no network seems to be common to the judgment of a diverse variety of moral transgressions. Separate systems were found to characterize different kinds of moral judgment. The only region of anatomical overlap between these systems (DMPFC) is robustly associated more generally with self-referential processing, thinking about other people and processing ambiguous information.

It is likely that moral judgment is even more multidimensional than what is suggested here, given that there remain other domains of morality that were not examined in the current study (e.g., disrespect, betrayal of an in-group, fairness). These results suggest that, just as disparate systems are now understood to subserve aspects of cognitive faculties once thought to be monolithic (e.g., memory, attention), distinct systems subserve different types of moral judgment. Future research may benefit from working toward a taxonomy of these systems as Haidt and Graham (2007) have suggested. Such a taxonomic approach should help reduce the risks of false negatives and overgeneralization. The risk of false negatives is present whenever diverse stimuli are combined within a single experimental condition. Lumping together diverse moral transgressions risks failing to detect the component processes activated by one type of transgression that are not also activated by other types of transgressions. Similarly, caution should be taken in generalizing neural and cognitive processes across moral domains. Conflation of different types of moral judgment for analysis likely contributes to inconsistency in the current literature regarding, for instance, whether emotion causes (Haidt, 2001) or emerges as a mere consequence of (Huebner, Dwyer, & Hauser, 2009) moral judgment and whether the principles that guide moral judgments are consistent across demographic groups (Hauser, Cushman, Young, Jin, & Mikhail, 2007) or vary as a function of socioeconomic status and political affiliation (Haidt & Graham, 2007).

The current results are consistent with the hypothesis that morality is not a natural kind in the mind but, rather, is comprised of multiple dissociable mental processes. That said, there could exist a cognitive process common and peculiar to moral judgments that is not detectable using the current methods. The hemodynamic differences observed here appear to be modulated primarily by differences in the semantic content of moral scenarios, and the general cognitive processes they engage, rather than their perceived moral wrongness. Irrespective of whether subsequent studies find evidence for a unified moral faculty, future research using fMRI and standard paradigms will likely benefit from taking an increasingly fine-grained approach—from isolating distinct categories of moral behavior for study, and from investigating the specific component processes that contribute to different types of moral judgments. In so doing, we will be most likely to generate consistent and precise findings across studies, and to gain a more accurate and thorough understanding of the manifold judgments, we term “moral.”

We thank Liane Young, Joshua Greene, Fiery Cushman, Bertram Malle, Rebecca Saxe, and the Moral Psychology Research Group for helpful discussion and feedback and Daniel Leopold, Christina Chick, Cameron Shirazi, Anne Krendl, and Cecilia Tilli for their contributions to this project. We also wish to thank Dartmouth College, Princeton University, the Australian National University and the Rotman Institute of Philosophy for institutional support.

Reprint requests should be sent to Thalia Wheatley, Psychological and Brain Science, Dartmouth College, Hanover, NH 03755, or via e-mail: [email protected].