In Regard for Reason in the Moral Mind, Joshua May (Reference May2018) argues that our “moral judgments are often governed by rule-based inference” (p. 55). Here, inference is understood as a mode of reasoning in which “beliefs or similar propositional attitudes are formed on the basis of pre-existing ones” (p. 8). To support this view, May cites a body of evidence suggesting that participants take consequences, personal harm, and other factors into consideration when making moral judgments. On the basis of this evidence, May concludes that, in making moral decisions, “we often rapidly infer the moral status of an action in part by relying on general principles” (p. 70).
The evidence May cites admits of at least two interpretations, however. On May's interpretation, the transition from consequences, intentions, and other factors to moral judgments is underwritten by an inference from general principles. On an alternative interpretation, moral cognition is instead underwritten by non-inferential, subpersonal, domain general valuation mechanisms. This commentary defends the latter interpretation. Moreover, it suggests that, insofar as May fails to consider this latter possibility, the book's argumentative strategy is weakened. May's argument proceeds by elimination: He aims to show that the evidence regarding the emotions is not as compelling as generally thought, and hence that “what's left is inference” (p. 71). But if emotion and reason are not the only explanatory options to choose from – if valuation mechanisms are a viable, interpretive possibility – then we should not be so quick to conclude that “moral judgments are generated by fundamentally cognitive and rational processes” (p. 18).
Roughly, domain general valuation mechanisms refer to a body of computational, neural, and behavioral mechanisms thought to underwrite decision making and economic choice (for an excellent overview, see Glimcher et al. Reference Glimcher, Camerer, Fehr and Poldrack2009). On one way of describing these mechanisms more precisely (in what is called “reinforcement learning”), decision making is underwritten by multiple distinct decision systems, each relying on a distinct computational approach to predicting future reward and value (Sutton and Barto Reference Sutton and Barto1998). Here, reward refers to the intrinsic desirability of a given state, whereas value refers to the total, expected, future reward of a given state. The distinction allows for a state that is not intrinsically rewarding to nonetheless be assigned value, contingent on its causal relations to future rewarding states.
Human decision making is thought to depend on at least three distinct, domain general valuation mechanisms (for reviews, see Dayan & Abbott Reference Dayan and Abbott2001; Dayan & Niv Reference Dayan and Niv2008; Rangel et al. Reference Rangel, Camerer and Montague2008). A Pavlovian valuation mechanism governs automatic approach and withdrawal responses to appetitive and aversive stimuli, respectively (Mackintosh Reference Mackintosh1983). A second, model-free valuation mechanism caches positive and negative state-action pairs, and assigns values to actions based on their previous outcomes. A third, model-based mechanism explicitly represents possible choices and determines the sequence of actions that maximizes value. This procedure is typically represented by a decision tree: Each node in the tree representing a possible choice, where total value is the sum of the rewards minus the sum of the punishments along a given branch. Notably, however, none of these mechanisms involve rule-based inferences or an appeal to general rules. Rather, as their name suggests, domain-general valuation mechanisms rely on a basic notion of value and the calculation of value to underwrite our everyday decision making.
These (and possibly other) domain general valuation mechanisms are increasingly thought to play a role in moral cognition. Taking a computational approach, for example, Crockett (Reference Crockett2013; Reference Crockett, Forgas, Jussim and Van Lange2016a; Reference Crockett2016b) argues that these mechanisms - and their interactions - can account for how the features of a moral situation can be transformed into a moral decision and, by extension, a moral action. Analogously, but adopting a more explicitly neuroscientific approach, Shenhav and Greene (Reference Shenhav and Greene2010 p. 671, Table 1) show that participants’ ratings of moral acceptability are correlated with degrees of activation in their posterior cingulate cortex and ventromedial prefrontal and medial orbitofrontal cortices, that is, with brain activations relatively similar to those seen in instances of valuing physical goods and actions. These mechanisms have further been implicated in the processing of normative statements (Berns et al. Reference Berns, Bell, Capra, Prietula, Moore, Anderson, Ginges and Atran2012), experiences of trust (Fehr et al. Reference Fehr, Fischbacher and Kosfeld2005), the feeling of compassion (Montgomery et al. Reference Montgomery, Kappes, Crockett, Sinnott-Armstrong and Miller2017), and the phenomenon of implicit bias (Huebner Reference Huebner, Brownstein and Saul2016).
So how might these domain general valuation mechanisms explain the specific findings May cites in favor of reasoning in moral cognition? I focus on the components of consequences, personal harm, and the means versus by-products distinction, though analogous arguments can be made for May's discussion of intentions and actions versus omissions (see especially Crockett Reference Crockett2013; Cushman Reference Cushman2013; though, see also Ayars Reference Ayars2016)
First, May (p. 58) argues that ordinary moral judgments are “sensitive to the quantity of harmful consequences that follow from an agent's options,” including relatively complex considerations, and takes these to be evidence of the role of rule-based inference as characterized previously. However, domain general valuation mechanisms can equally account for such assessments of consequences, and notably can do so without appealing to general principles or rules. Specifically, both the model-free and model-based mechanisms characterized above predict instrumental state-action pairs; that is, they predict the consequences of various action alternatives as selected in various future states. Similarly, though again using a more neuroscientific emphasis, in the study cited above, Shenhav and Greene (Reference Shenhav and Greene2010) demonstrate that participants make surprisingly fine-grained assessments of consequences - even taking uncertainties and risks in account - using precisely those domain general valuation systems that underwrite other types of decisions, including economic decisions.
Second, May argues that principles relating to harm, notably principles relating to force, contact, and battery, play an important role in moral cognition. As above, the suggestion seems to be that such assessment requires rule-based inference. May suggests that although there is not precision in this literature, a quip from Paul Bloom can characterize the problem, suggesting, “Here is a good candidate for a moral rule that transcends space and time: If you punch someone in the face, you'd better have a damn good reason for it” (Bloom Reference Bloom2013, p. 10). However, as in the case of consequences, assessments of personal harm, including such factors as force and contact, can be accounted for by using domain general valuation mechanisms.
Cushman (Reference Cushman2015) makes just such a case, suggesting that the model-free system can explain why participants provide inconsistent responses to the trolley problem. On Cushman's view, the model-free system elicits aversion toward direct, physical harm, where this aversion “can be understood as the consequence of negative value assigned intrinsically to an action,” namely, the physical harm (Cushman Reference Cushman2015, p. 59). This learned aversion in turn prevents participants from endorsing the pushing of the bystander in the “Footbridge” case, but still allows them to endorse the pulling of the switch in the “Switch” case. Notably, again, however, such an assessment depends on the cached attribution of value rather than on an appeal to general principles.
Third, May argues that moral cognition involves making a sophisticated distinction between an outcome that is a direct means to one's end and one that is merely a by-product of one's endeavors. Although the empirical results are, by May's own account, rather mixed on this front, the basic implication of May's argument remains the same: moral cognition depends on inference and, in particular, on some set of morally relevant general principles. Again, however, domain general valuation mechanisms may do the trick without appealing to principles at all.
Crockett (Reference Crockett2013) cites a special kind of interaction between the Pavlovian and model-based systems to defend this latter possibility. Specifically, Crockett appeals to the phenomenon of “pruning,” in which the Pavlovian system “cuts” the branches of a model-based decision tree in the face of aversive alternatives represented early on in the decision tree. Crockett argues that in means cases, as when an individual is directly used to stop a trolley, the harm – killing the individual – is sufficiently high up on the tree that this alternative is “pruned,” causing participants to deem it an unacceptable alternative. By contrast, in by-product cases, as when a trolley avoids killing five individuals but then kills one individual, the benefits of saving the five occurs before any pruning can take place. Consequently, “the side-effect death is incidental to saving the five individuals, so it can be safely pruned away while preserving the contribution of saving five toward the overall action value” (Crockett Reference Crockett2013, p. 365). Domain general mechanisms thus provide a plausible explanation of the means versus by-product distinction in moral cognition.
To summarize: Based on his assessment of these and the other factors, May concludes that “a clear picture is emerging from the science of moral judgment. We often rapidly infer the moral status of an action in part by relying on general principles that identify as morally relevant various features of agents, actions, and outcomes” (p. 70). However, moral cognition may instead involve the assessment of consequences, personal harm, and means versus by-product effects (as well as the other factors) without a cognitive appeal to general principles. Moreover, as noted above, insofar as May aims to provide an argument by elimination, May's case for moral reasoning is not as strong as it might first appear.
In Regard for Reason in the Moral Mind, Joshua May (Reference May2018) argues that our “moral judgments are often governed by rule-based inference” (p. 55). Here, inference is understood as a mode of reasoning in which “beliefs or similar propositional attitudes are formed on the basis of pre-existing ones” (p. 8). To support this view, May cites a body of evidence suggesting that participants take consequences, personal harm, and other factors into consideration when making moral judgments. On the basis of this evidence, May concludes that, in making moral decisions, “we often rapidly infer the moral status of an action in part by relying on general principles” (p. 70).
The evidence May cites admits of at least two interpretations, however. On May's interpretation, the transition from consequences, intentions, and other factors to moral judgments is underwritten by an inference from general principles. On an alternative interpretation, moral cognition is instead underwritten by non-inferential, subpersonal, domain general valuation mechanisms. This commentary defends the latter interpretation. Moreover, it suggests that, insofar as May fails to consider this latter possibility, the book's argumentative strategy is weakened. May's argument proceeds by elimination: He aims to show that the evidence regarding the emotions is not as compelling as generally thought, and hence that “what's left is inference” (p. 71). But if emotion and reason are not the only explanatory options to choose from – if valuation mechanisms are a viable, interpretive possibility – then we should not be so quick to conclude that “moral judgments are generated by fundamentally cognitive and rational processes” (p. 18).
Roughly, domain general valuation mechanisms refer to a body of computational, neural, and behavioral mechanisms thought to underwrite decision making and economic choice (for an excellent overview, see Glimcher et al. Reference Glimcher, Camerer, Fehr and Poldrack2009). On one way of describing these mechanisms more precisely (in what is called “reinforcement learning”), decision making is underwritten by multiple distinct decision systems, each relying on a distinct computational approach to predicting future reward and value (Sutton and Barto Reference Sutton and Barto1998). Here, reward refers to the intrinsic desirability of a given state, whereas value refers to the total, expected, future reward of a given state. The distinction allows for a state that is not intrinsically rewarding to nonetheless be assigned value, contingent on its causal relations to future rewarding states.
Human decision making is thought to depend on at least three distinct, domain general valuation mechanisms (for reviews, see Dayan & Abbott Reference Dayan and Abbott2001; Dayan & Niv Reference Dayan and Niv2008; Rangel et al. Reference Rangel, Camerer and Montague2008). A Pavlovian valuation mechanism governs automatic approach and withdrawal responses to appetitive and aversive stimuli, respectively (Mackintosh Reference Mackintosh1983). A second, model-free valuation mechanism caches positive and negative state-action pairs, and assigns values to actions based on their previous outcomes. A third, model-based mechanism explicitly represents possible choices and determines the sequence of actions that maximizes value. This procedure is typically represented by a decision tree: Each node in the tree representing a possible choice, where total value is the sum of the rewards minus the sum of the punishments along a given branch. Notably, however, none of these mechanisms involve rule-based inferences or an appeal to general rules. Rather, as their name suggests, domain-general valuation mechanisms rely on a basic notion of value and the calculation of value to underwrite our everyday decision making.
These (and possibly other) domain general valuation mechanisms are increasingly thought to play a role in moral cognition. Taking a computational approach, for example, Crockett (Reference Crockett2013; Reference Crockett, Forgas, Jussim and Van Lange2016a; Reference Crockett2016b) argues that these mechanisms - and their interactions - can account for how the features of a moral situation can be transformed into a moral decision and, by extension, a moral action. Analogously, but adopting a more explicitly neuroscientific approach, Shenhav and Greene (Reference Shenhav and Greene2010 p. 671, Table 1) show that participants’ ratings of moral acceptability are correlated with degrees of activation in their posterior cingulate cortex and ventromedial prefrontal and medial orbitofrontal cortices, that is, with brain activations relatively similar to those seen in instances of valuing physical goods and actions. These mechanisms have further been implicated in the processing of normative statements (Berns et al. Reference Berns, Bell, Capra, Prietula, Moore, Anderson, Ginges and Atran2012), experiences of trust (Fehr et al. Reference Fehr, Fischbacher and Kosfeld2005), the feeling of compassion (Montgomery et al. Reference Montgomery, Kappes, Crockett, Sinnott-Armstrong and Miller2017), and the phenomenon of implicit bias (Huebner Reference Huebner, Brownstein and Saul2016).
So how might these domain general valuation mechanisms explain the specific findings May cites in favor of reasoning in moral cognition? I focus on the components of consequences, personal harm, and the means versus by-products distinction, though analogous arguments can be made for May's discussion of intentions and actions versus omissions (see especially Crockett Reference Crockett2013; Cushman Reference Cushman2013; though, see also Ayars Reference Ayars2016)
First, May (p. 58) argues that ordinary moral judgments are “sensitive to the quantity of harmful consequences that follow from an agent's options,” including relatively complex considerations, and takes these to be evidence of the role of rule-based inference as characterized previously. However, domain general valuation mechanisms can equally account for such assessments of consequences, and notably can do so without appealing to general principles or rules. Specifically, both the model-free and model-based mechanisms characterized above predict instrumental state-action pairs; that is, they predict the consequences of various action alternatives as selected in various future states. Similarly, though again using a more neuroscientific emphasis, in the study cited above, Shenhav and Greene (Reference Shenhav and Greene2010) demonstrate that participants make surprisingly fine-grained assessments of consequences - even taking uncertainties and risks in account - using precisely those domain general valuation systems that underwrite other types of decisions, including economic decisions.
Second, May argues that principles relating to harm, notably principles relating to force, contact, and battery, play an important role in moral cognition. As above, the suggestion seems to be that such assessment requires rule-based inference. May suggests that although there is not precision in this literature, a quip from Paul Bloom can characterize the problem, suggesting, “Here is a good candidate for a moral rule that transcends space and time: If you punch someone in the face, you'd better have a damn good reason for it” (Bloom Reference Bloom2013, p. 10). However, as in the case of consequences, assessments of personal harm, including such factors as force and contact, can be accounted for by using domain general valuation mechanisms.
Cushman (Reference Cushman2015) makes just such a case, suggesting that the model-free system can explain why participants provide inconsistent responses to the trolley problem. On Cushman's view, the model-free system elicits aversion toward direct, physical harm, where this aversion “can be understood as the consequence of negative value assigned intrinsically to an action,” namely, the physical harm (Cushman Reference Cushman2015, p. 59). This learned aversion in turn prevents participants from endorsing the pushing of the bystander in the “Footbridge” case, but still allows them to endorse the pulling of the switch in the “Switch” case. Notably, again, however, such an assessment depends on the cached attribution of value rather than on an appeal to general principles.
Third, May argues that moral cognition involves making a sophisticated distinction between an outcome that is a direct means to one's end and one that is merely a by-product of one's endeavors. Although the empirical results are, by May's own account, rather mixed on this front, the basic implication of May's argument remains the same: moral cognition depends on inference and, in particular, on some set of morally relevant general principles. Again, however, domain general valuation mechanisms may do the trick without appealing to principles at all.
Crockett (Reference Crockett2013) cites a special kind of interaction between the Pavlovian and model-based systems to defend this latter possibility. Specifically, Crockett appeals to the phenomenon of “pruning,” in which the Pavlovian system “cuts” the branches of a model-based decision tree in the face of aversive alternatives represented early on in the decision tree. Crockett argues that in means cases, as when an individual is directly used to stop a trolley, the harm – killing the individual – is sufficiently high up on the tree that this alternative is “pruned,” causing participants to deem it an unacceptable alternative. By contrast, in by-product cases, as when a trolley avoids killing five individuals but then kills one individual, the benefits of saving the five occurs before any pruning can take place. Consequently, “the side-effect death is incidental to saving the five individuals, so it can be safely pruned away while preserving the contribution of saving five toward the overall action value” (Crockett Reference Crockett2013, p. 365). Domain general mechanisms thus provide a plausible explanation of the means versus by-product distinction in moral cognition.
To summarize: Based on his assessment of these and the other factors, May concludes that “a clear picture is emerging from the science of moral judgment. We often rapidly infer the moral status of an action in part by relying on general principles that identify as morally relevant various features of agents, actions, and outcomes” (p. 70). However, moral cognition may instead involve the assessment of consequences, personal harm, and means versus by-product effects (as well as the other factors) without a cognitive appeal to general principles. Moreover, as noted above, insofar as May aims to provide an argument by elimination, May's case for moral reasoning is not as strong as it might first appear.