Tuesday, March 6, 2018

Learning Moral Rules

While evolutionary psychology has led to a proliferation of (often outlandish and essentializing) claims about innate human traits and tendencies, the view that human morality is innate has a long and reputable history. Indeed, broadly evolutionary accounts of morality go back to Darwin and his contemporaries. Views that posit innate cognitive mechanisms specific to the domain of morality (viz., moral nativism) are of more recent vintage. The most prominent contemporary defenders of moral nativism adopt a perspective called the “linguistic analogy” (LA), which uses concepts from the Chomskian program of generative linguistics to frame issues in the study of moral cognition.[1] Here, I present one of LA’s key data points, and propose an alternative, non-nativist explanation of it in terms of learning. 

The data point on which I’ll focus concerns the proposed explanation for certain observed patterns in people’s moral judgments, including in response to trolley cases. In the sidetrack case (fig. 1), most people judge that it would be permissible for a bystander to save five people by pulling a switch that would divert the trolley onto a sidetrack where one person would be struck and killed. However, in the footbridge case (fig. 2), most people judge that it would not be permissible for a bystander to save five people on the track by pushing someone bigger than himself off the bridge into the path of the train to stop it.

The results from cross-cultural studies of the trolley problems and similar dilemmas suggest that subjects’ judgments are sensitive to principled distinctions like the doctrine of double effect, where harms caused as a means to a good outcome are judged morally worse than equivalent harms that are mere side effects of an action aimed at bringing about a good outcome.[2]

To explain the acquisition of these implicit rules, LA invokes an argument from the poverty of moral stimulus. For example, Mikhail argues that to judge in accordance with the doctrine of double effect involves tracking complex properties like ends, means, side effects, and prima facie wrongs such as battery. It’s implausible that subjects’ sensitivity to these abstract properties is gained through instruction or learning. Rather, a more plausible explanation is that humans are endowed with an innate moral faculty that enables the acquisition of a moral grammar (which includes the set of these rules).[3]

I believe that other research from language acquisition and the cognitive sciences more broadly points to the availability of a different explanation of how these implicit rules could be acquired, via learning mechanisms not specific to the moral domain. Evidence suggests that children employ powerful probabilistic learning mechanisms early in their development.[4] With these mechanisms, children are able form generalizations efficiently, on the basis of what might otherwise appear to be sparse data.

Consider the following example of a study on word learning: 3-4 year old subjects who heard a novel label applied, for example to one Dalmatian extend the label to dogs in general.[5] When applied to three Dalmatians, subjects extend the label to Dalmatians only. In the latter case, though the data is consistent with both candidate word meanings (dog, Dalmatian), the probability of observing three Dalmatians is higher on the narrower hypothesis.

I propose that a similar process of inference could account for the acquisition of implicit moral rules. There may be sufficient information contained in the stimuli to which individuals typically are exposed in the course of their early development – including the reasoning and response patterns of adults and peers in their environment – to account for their ability to make such moral distinctions. Consider the act/omission distinction. Cushman et al. found that subjects judge in accordance with what they call the ‘action principle’, according to which harm caused by action is judged morally worse than equivalent harm caused by omission.[6] Children observe this distinction in action. A child may be chided more harshly for upsetting a peer by taking a cookie away from her than for upsetting a peer by failing to share his own cookies, for example. With the probabilistic learning under consideration, it may take surprisingly few such observations for children to generalize to a more abstract form of this distinction. Observing the distinction at play in a few different types of scenarios may be sufficient for a learner to generalize, and go beyond tracking the distinction just in the particular cases observed to infer a general model that could have given rise to data they have encountered.

Of course, further investigation is needed to comparatively assess these two proposals. I’ll end by noting that the debate over moral nativism has both theoretical and practical implications. If the non-nativist account is right, this points to a view of our capacity for moral judgment as more malleable and amenable to intervention and improvement than the nativist account suggests. On the other hand, some (though not all) take the nativist account, if correct, to invite a skeptical view about morality.

Theresa Lopez
Department of Philosophy
University of Maryland

[1]Dwyer, S., Huebner, B., and Hauser, M. 2010: The linguistic analogy: motivations, results and speculations. Topics in Cognitive Science, 2, 486–510.
[2] Hauser, M., Young, L., and Cushman, F. 2008: Reviving Rawls’ linguistic analogy. In W. Sinnott-Armstrong (Ed.), Moral psychology, Vol. 2, The Cognitive Science of Morality: Intuition and Diversity. Cambridge, MA:
MIT Press, 107-144.
[3] Mikhail, J. 2011: Elements of Moral Cognition. Cambridge: Cambridge University Press.
[4] Xu, F., and Griffiths, T.L. 2011: Probabilistic models of cognitive development: Towards a rational constructivist approach to the study of learning and development. Cognition, 120, 299-301; Perfors, A., Tenenbaum, J. and Regier, T. 2011: The learnability of abstract syntactic principles. Cognition, 118, 306-338.
[5] Xu, F., and Tenenbaum, J. B. 2007: Word learning as Bayesian inference. Psychological Review, 114, 245–272.
[6] Cushman, F., Young, L. and Hauser, M. D. 2006: The role of conscious reasoning and intuition in moral judgment: testing three principles of harm. Psychological Science, 17, 1082-1089.


  1. Hi Theresa, thanks for this interesting post. A couple of quick questions.

    1) Would you mind unpacking the explanation for why children tend to restrict the application of "dalmation" to dalmations when the training set involves three as opposed to one? I'm not quite sure what you mean by saying that the probability of observing 3 dalmations is higher on the narrower hypothesis. I would have guessed something like: we infer a higher conditional probability of (spots/dalmation) with the larger group.

    2) Alison Gopnik defends the theory theory against Chomskian nativism by taking issue with the assumption that innate structures are immutable. She seems to think that infants are born with some moral theories, but that they are tested and revised more or less as you describe. Are you in her camp, or are you making the stronger claim that infants aren't born with any distinctly moral theory at all and that Bayesian mechanisms together, perhaps together with some sort of innate mechanism for guessing, are all that needs to be supposed?

    1. Hi Randy,

      Thank you for the comments. Re: 1) Yes, exactly – I am referring to the conditional probability of the set of observations (e.g. X: three spotted dogs) given the hypothesis (e.g. H1: novel label = Dalmatian, H2: novel label = dog). This probability (the likelihood in Bayes’ rule) reflects the learner’s expectations about what will be observed, such as the expectation that the examples observed will be random and representative. If the label refers to dogs generally, learners would expect some variation when presented with multiple examples. When the label is applied to three Dalmatians, it appears to be a suspicious coincidence that leads the learner to infer (intuitively) a higher p(X/H1) than p(X/H2), and consequently (in this case) a higher p(H1/X) than p(H2/X).

      Re: 2) As a general matter, I’m sympathetic to and quite in line with Gopnik’s thinking on development (e.g. that infants have powerful learning capacities). I believe that children possess early-emerging (pro-)social and normative capacities that are plausibly innate and are not restricted to the moral domain, but that are important parts of moral development. And because of these capacities and even more general capacities for probabilistic learning, it may be the case that children’s early-emerging moral capacities can be explained without positing an innate moral faculty (e.g. innate moral theories).

      Do you have her views in The Philosophical Baby in mind, or a different source? I’d like to look into it further. I’ve heard Gopnik suggest that the nativist-empiricist divide is often drawn too sharply. I take it that I’m largely in her camp, though not if she believes there are innate moral theories.

    2. Hi Theresa, thanks for that, very clarifying. Mainly I am thinking about a paper Gopnik wrote called "The theory theory as an alternative to the innateness hypothesis." She has it posted on her website I believe. I think you are right that she does believe that there is a fuzzy boundary between these views.

  2. Hi, Theresa (quick intro: I'm not a philosopher; I just like philosophy, and post more or less often in philosophy blogs when I find something interesting and have time to comment).

    That's an interesting hypothesis; regarding the implications, in my view if the nativist account is false, that supports the thesis that either a substantive moral error theory is true, or else an epistemic one is (barring culture relativism).

    I'll try to briefly explain my take on this:

    Different societies have different local rules; let's call them "legal" rules. Some of those legal rules are written, though (broadly) historically, most weren't. Either way, members a society generally learn the basic rules - most people are no legal experts, but in general we know what's allowed and what's not in our daily lives. I don't know whether there is a specific rule-acquisition mechanism for those rules, but either way, that does not invite skepticism about legal rules, because the content of those rules is given by each society (in different ways, such as traditions, parliaments, courts, etc.).

    Yet, when it comes to moral rules, the content is not given by each society (even if what's morally acceptable for a person to do depends on the specifics of a situation, which vary from society to society).
    For instance, the following are examples from Deuteronomy (there are many others in the OT):

    22:23 If there be a young lady who is a virgin pledged to be married to a husband, and a man find her in the city, and lie with her; 22:24 then you shall bring them both out to the gate of that city, and you shall stone them to death with stones; the lady, because she didn't cry, being in the city; and the man, because he has humbled his neighbor's wife: so you shall put away the evil from the midst of you.

    In addition to the legal rule, there is a clear moral claim or implication that the punishment is deserved. The person who wrote that was badly mistaken about that and about many other moral matters (*). However, beliefs like that - and many others - are often not contradictory (barring an analytical reduction of moral terms to non-moral ones, but even then, there would be no way of proving a contradiction, since the reduction is not known).

    Without a species-wide moral sense, it seems to me moral rules would be on the same boat as legal rules, in the sense that people would be learning culture-specific rules, but not culture-independent ones. However - and on this, unlike legal rules - that would result in false moral beliefs people can't get out of. For example, if an ancient Israelite (at some point in history, etc.) believed that in their society, the law was to stone to death people who meet the criteria described above, she was correct - she properly learned the culture-specific rule. But if she also believed that the rule was just and the punishment deserved, she was in error. However, if her moral sense was culturally acquired, then she had no way to realize otherwise. Maybe she should have concluded that the rules were not given by a powerful creator, but still, that does not resolve the issue of whether they're just.

    The system you propose seems to be capable of handling acquisition of legal rules without a specific rule-acquisition system - though the latter would also work; I do not know whether there is such specific system -, but I don't see how it might help acquire moral rules. Maybe you have a more specific account that can find a way around this?

    (*) If you are a Christian reader and you think OT rules were just, the punishments deserved, etc., please feel free to pick another example, say some Soviet rules, or Aztec rules, or any other of your choosing. The basic point I'm trying to make here does not depend on that.