Asymmetry of Intentionality Judgments in Causal-Deviance Scenarios

When Joshua Knobe documented an asymmetry between people’s judgments of intentionality toward negative and positive side effects in his CEO scenario, many, including Knobe himself, hypothesized that this difference was due to subjects’ differing moral attitudes toward the CEO in the two cases. In particular, subjects seemed to be blaming the CEO for harming the environment in the former case but not praising or blaming him in the latter and, in one way or another, these attitudes toward the CEO were thought to influence subjects’ intentionality ratings (e.g., Knobe, 2006; Adams & Steadman, 2004; Alicke, 2008).

A number of studies have called these theories into question, though, arguing that Knobe-like effects can be observed even in morally neutral scenarios. These alternative theories fall roughly into two categories: those that posit some kind of inferential mechanism to account for the Knobe asymmetry and those that appeal to a particular feature of the CEO case, like trading-off (Machery, 2008), to account for this difference. The former theories often emphasize the importance of norm violation in Knobe’s “harm” version of the CEO scenario for ascribing intentionality (e.g., Sloman et al., in press; Uttich & Lomborozo, 2010).

However, some recent work I’ve done in collaboration with Fiery Cushman suggests that subjects’ moral attitudes may still be influencing their intentionality judgments at least to a small degree.  In our experiment, we presented about 650 subjects (on Mechanical Turk and the Brown University campus) with one of six scenario contexts crossed with one of three moral variants—virtuous, neutral, or reprehensible (making for 18 stimuli in total). All of these scenarios were causally deviant, meaning that an agent’s intention to perform some action caused the agent to perform some different, unintended action that nevertheless produced the agent’s desired outcome. For example, one of our scenarios involved a woman named Janet who wanted to redirect a train in order to save some children’s lives (virtuous), fulfill an employee’s request (neutral), or kill her boss (reprehensible), but who did so by inadvertently tripping on an emergency lever. After reading this prompt, subjects were asked to rate on a 1–7 scale: “To what extent did Janet intentionally redirect the train?”

Our data showed two interesting trends. First, we were surprised to find that subjects’ mean intentionality judgments for all 18 of our scenarios were quite high—often higher than the midpoint (4) between no intentionality (1) and complete intentionality (7). Pooling across all 18 scenarios, the average rating was 4.69; pooling across just the “neutral” scenarios, it was 4.23.  This consistent finding suggests that people are willing to ascribe at least some intentionality to an outcome even when its precipitating action was clearly unintentional, contrary to Searle’s (1983) assumption that these outcomes are wholly unintentional.

Second, and as relates to the Knobe Effect, we observed significant differences in intentionality judgments between morally reprehensible and neutral cases (p < .001), and also between morally virtuous and neutral cases (p < .01), when collapsing across all scenarios. When comparing the means of each of the six individual contexts, the reprehensible-vs.-neutral difference was present in all six cases (p < .01, treating scenario as the random effect), while the virtuous-neutral difference was present in 4 out of 6 cases (p = .15). The means for our virtuous, neutral, and reprehensible conditions were 4.84, 4.23, and 5.01, respectively.

We cannot find an explanation for these effects by appealing to the trade-off (Machery, 2008), reason against acting (Turner, 2004) or need for consideration (Scaife & Webber, forthcoming) hypotheses because they are specific to cases of side-effects—whereas our causal deviance cases achieve an “ambiguity of intentionality” without the use of a side effect. Consequently, though these features may still play a role in explaining the original Knobe Effect, our data suggest that they are not the only factors relevant to Knobe’s asymmetry in intentionality judgments.

The inferential accounts of Uttich & Lomborozo (2010), Sloman et al. (in press), and Holton (2010) do not seem to apply to our scenarios, either. In particular, our agents do not violate any norms in the virtuous cases, nor does norm violation appear to be particularly informative in the reprehensible cases given that, unlike in the CEO scenario, our agents all have clearly stated intentions. Inferential theories like Sloman and colleagues’, which appeal to the CEO’s desire for harming or helping the environment, may be able to account for our effects. However, we have conducted some preliminary tests of desire attribution in our scenarios, and so far the data suggests that these patterns of attribution cannot explain the observed intentionality ratings.

Instead, our data seem to fit nicely with the moral-appraisal theories. For example, according to Adams and Steadman’s pragmatics account, our subjects were, in part, using intentionality to blame the agents in the reprehensible cases and praise them in the virtuous cases. Likewise, Alicke’s model might predict that subjects would give higher ratings for intentionality in the reprehensible and virtuous cases in order to validate their visceral desires to blame or praise the agents.

Of course, applicability of our results to Knobe’s scenario is limited. Given the variety of theories that have been proposed and tested as contributors to the effect in the original CEO cases, it is likely that no single theory completely explains the phenomenon.  Indeed, we observed a much smaller effect for causal deviance cases than has been reported for the CEO cases, suggesting that we can explain at best a fraction of the CEO effect.

We're eager for any feedback on this data, our interpretation of it, or other implications that we have not considered. Thanks!

Comments

[Adam Limehouse]

Just a thought, is it possible that some of these effects could be represented as greater or lesser degrees of magical thinking in the subjects being tested? Magical thinking being described (vaguely) as the propensity to ascribe intention and/or consciousness inappropriately.

Another issue, do you have data to share re: demographics? What do you suppose could be some possible differences to be found if you took this survey into the African Bush or (barring malaria and such) into Carteret County, NC?

[Adam Bear]

Hi Adam,

Thanks for the feedback! If I understand you correctly, I agree that it's possible that subjects are making inappropriate intentionality judgments, for example, because of task demands. The question remains, though, whether this can explain the *asymmetry* we found between neutral and virtuous/reprehensible judgments. That's not to say it can't, but this would need to be explained.

Regarding demographics, our subject pool comprised mostly users of Mechanical Turk, who are all from the US (perhaps some from NC) and who tend to have more liberal political views. The rest of our subjects came from Brown University, which, as you probably know, isn't a very diverse population! It's an interesting question how people's intentionality judgments vary across cultures. Unfortunately, I'm not familiar with any studies that have looked at that in particular, but I'm sure they exist!

[Chandra Sripada]

Hi Adam,

Cool results! I think you are right that moral factors such blame reactions have a role here. But I also think at least some of this effect is explained by covert mental state attributions. Many people tend to focus on *desire* as the key mental state for intentionality judgments. My Deep Self Concordance Model says that in addition to ‘surface attitudes’ such as desires, intentionality judgments are also sensitive to people’s *deeper evaluative attitudes*, i.e., their cares, values, and commitments. In particular, the model says people are more likely to judge an agent intentionally brought about an outcome if the outcome concords with the agent’s underlying deep attitudes.

So here is a hypothesis about what is going on: In the morally-infused scenarios, people attribute deep attitudes to Janet (e.g., that she deeply hates her boss, is committed to ending her boss’s life, does not care about harming others, and so on), and these attributed deep attitudes are concordant with the outcome she brings about (redirecting the train and killing her boss). So intentionality ratings are higher. In the neutral scenario, they don’t attribute deep attitudes to Janet (what attitudes follow from the fact that she is following a request?), so intentionality ratings are lower.

To test this hypothesis, you could use questions like the following:

* It deeply matters to Janet that the train is redirected.
* Janet cares deeply that the train is redirected.
* It is critical to Janet that the train is redirected.
* Janet is committed to redirecting the train.

Other questions could probe the ‘cross-situational ‘traitiness’ of her motive (deep attitudes are typically more trait-like):

* In the story above, [a person is harmed] or [children are helped] or [insert neutral outcome]. How much do you agree with the following statement:
Janet is likely to, in other contexts and situations, bring about outcomes like this one.

I’ve used questions along these lines to probe attributions of deep attitudes in previous studies. See for example, http://sitemaker.umich.edu/sripada/files/sripada_-_mental_state_attributions_and_the_side-effect_effect.pdf

I would predict that attributions of deep attitudes would track with intentionality judgments across your vignettes, and this might be interesting to test.

[Adam Bear]

Hi Professor Sripada,

Thanks for the input! As a matter of fact, Fiery and I have discussed how your "deep self" view might explain our data, and I absolutely agree that the moral conditions might involve "deeper evaluative attitudes" than those in the neutral conditions. We actually have it on our agenda to test some of your hypotheses, and your suggestions for how to test the model are very helpful. If we keep working on this project, we will definitely want to look at how your theory fits in.