In recent years, cognitive scientists have been trying to explain a curious psychological quirk called the ‘Knobe effect’. A new paper by Chandra Sekhar Sripada and Sara Konrath in Mind & Language offers an answer.
(As this is a long post, I’ve formatted it as a PDF so you can print it out and read it over a coffee if you prefer)
A little under a decade ago, a young cognitive scientist called Joshua Knobe, now at Yale University, ran a series of simple experiments with people in a Manhattan Park. He asked half of them to read the following story:
The vice-president of a company went to the chairman of the board and said, ‘We are thinking of starting a new program. It will help us increase profits, but it will also harm the environment.’ The chairman of the board answered, ‘I don’t care at all about harming the environment. I just want to make as much profit as I can. Let’s start the new program.’ They started the new program. Sure enough, the environment was harmed.
He then asked his participants, “Did the Chairman intentionally harm the environment?”. The vast majority — around 82% — said yes, he did.
Meanwhile, Knobe asked the other half of his interviewees to read exactly the same passage, but with the word ‘harm’ replaced with ‘help’, and then had them answer the question, “Did the Chairman intentionally help the environment?” Now most people (77%) said no, he did not1,2.
These results — and they’ve been replicated many times since — are a bit surprising. The mechanics of both situations are identical: the chairman signs off a project irrespective of its environmental impact, although in one scenario the outcome is bad (the environment gets harmed), and in the other, it’s good (the environment is helped). So shouldn’t the outcomes, whether good or bad, be seen as equally intentional or unintentional? Well, these findings suggest they’re not: it looks like the ‘moral flavour’ of the outcomes of actions — that is, whether they’re good or bad — shapes whether or not we construe those actions as having been performed intentionally.
Yet it’s natural to think that our moral judgments about the outcomes of people’s actions are themselves sensitive to whether or not those outcomes were intended or mere side effects. There’s a world of difference between me walking up to someone in the street and punching them square in the face, and accidentally hitting someone in the face as I’m trying to provide directions to another passerby (a feat I once achieved on London’s Tower Bridge). In the first case, my intentions would be bad, and so the outcome is judged as a moral offence; the latter was an accident, a side-effect of trying to help someone else, and therefore not morally blameworthy. It seems that judgments about intentionality should come first, with the moral judgment deriving from that.
In the scenarios Knobe explored, however, the influence sometimes goes the other way: when outcomes are bad, the agent whose actions lead to those outcomes is judged to have acted intentionally, even if the outcomes are technically side effects. But this isn’t true when positive outcomes, such as helping the environment, emerge as side effects of actions taken for other reasons, like pursuing profit. This asymmetry in judgments about intentionality is now well-known as the Knobe effect.
How to explain it? Perhaps the most obvious solution turns on the moral quality of the outcomes in question. When they’re good, and arise as side effects in the pursuit of other ends, they’re seen as just that: unintentional side effects, not the sort of thing to praise. But when the same actions lead to bad outcomes we apply the concept of intentionality differently, and treat analogous side effects as being intentionally produced. This helps us make sense of blaming people for the bad outcomes they create, whether on purpose or not.
Alternatively, ascriptions of intentionality may turn not so much on outcomes per se, but on judgments about the moral status of the person behind the relevant actions. If they violate a moral norm, by harming the environment, then we see them as immoral. This causes us to switch to a ‘moral blame’ mode, which makes us more likely to say that they acted intentionally — just as most most people do in the Chairman’s harm scenario.
These slightly different explanations are both, however, rooted in judgments about good and bad, right and wrong (whether they be directed at outcomes or agents). In the philosophical lexicon, they’re based on ‘normative factors’. And normative explanations of the Knobe effect are currently the most popular.
What might link normative judgments with intentionality judgments? One possibility is the negative emotions that are typically aroused when we hear about people doing things that lead to bad or immoral outcomes. Perhaps these feelings feed into how we deploy the concept of intentional action, so that perceived violations of moral norms — like harming the environment — are treated as intentional outcomes, even though they’d be seen as unintentional side effects if the outcomes were positive.
Unfortunately, there are reasons to believe this isn’t the case. The evidence comes from patients with damage to brain regions associated with processing negative emotions. These patients, who have sustained damage to the ventromedial prefrontal cortex (VMPFC), are not impaired intellectually, and have participated in a number of studies on moral cognition, in which they show a markedly different pattern of moral judgments compared with the general population.
Take the well-known Footbridge dilemma: an out of control train is on course to hit and kill a group of five walkers stuck on the tracks ahead, and the only way to stop the train is to push a very large and heavy man off a footbridge over the tracks into the train’s pain, stopping the train but killing the man. Is it morally permissible to push the man to his death to save five others? Most people say no, and their response in this and similar cases is associated with greater activity in the VMPFC, suggesting that the emotional aversion to harming another person inhibits this choice. The brain-damaged patients, by contrast, typically favour more a more utilitarian — some might say colder — logic, and the majority say it is OK to push the man. Yet despite this deficit in processing negative emotions, they still show the Knobe effect, indicating that it does not depend on these emotions3.
These findings do not mean that the normative theories of the Knobe effect are wrong; it just means they have to come up with another mechanism apart from the negative emotions elicited by violations of normative standards to explain the Knobe effect. In any case, normative theories, while very popular, are not the only game in town. Another contender has been developed by Chandra Sekhar Sripada4 (co-author with Sara Konrath of the new paper5 to be discussed shortly). Sripada’s account has the imposing title of the deep-self concordance model (DSCM), and it takes a second to get your head around. But it’s a neat model, so it’s worth it.
The basic idea is that when we encounter cases like Knobe’s Chairman scenario, we make two different judgments: one about his attitudes and values — is he anti- or pro-environmental? — and another about whether the outcomes of his actions are consistent with those attitudes. Neither of these judgments is normative; they are descriptive judgments about the facts of the case. In both scenarios we’ve been looking at, the Chairman reveals an anti-environmental attitude: he says he “doesn’t care about the environment”. (Though he doesn’t explicitly say that he’s against the environment and wants to cause environmental damage, his indifferent attitude is the sort leads to anti-environmental activities.) In the case where the project he approves harms the environment, the outcome is concordant with his anti-environmental stance, and so the outcome is seen as intentional. However, when the project has pro-environmental consequences, these are seen as discordant with his attitude, and are therefore not considered to be intentional outcomes.
In short, the normative explanations claim that ascriptions of intentionality are driven by judgments about the rightness and wrongness of actions and outcomes – that is, on evaluative judgments about the moral status of agents and their deeds. The DSCM, by contrast, focuses on descriptive judgments about the evaluative attitudes that agents hold.
A real-world example of the difference between forming evaluative judgments of an agent, and attributing evaluative attitudes to an agent, will bring out these distinctions a little more clearly. Let’s say George W. Bush gives a speech about abortion. Listening to this, we’re likely to make two distinct kinds of judgment: what we think the content of his speech is (the descriptive aspect), and how we feel about that content (the evaluative, normative aspect). Specifically, we’ll come to a view as to whether Bush advocates pro- or anti-abortion sentiments: we will ascribe certain evaluative attitudes (pro- or anti-abortion) to Bush based on descriptive facts about what he said in his speech. At the same time, we’ll decide whether his moral views on abortion are themselves moral or immoral — that is, we’ll make a normative judgment about the evaluative attitudes that we ascribe to Bush.
Notice that there’s likely to be much greater agreement about the descriptive side of this equation. Most people who hear Bush talk about abortion will recognise that he’s against it; that’s the evaluative (normative) attitude most people will ascribe to Bush. Yet there’s much greater disagreement about whether this makes Bush moral or not, as this judgment depends on the separate normative evaluation of Bush (based on the normative attitudes we ascribe to him). As descriptive judgments about the content of other people’s evaluative attitudes often go hand-in-hand with evaluative judgments of those very same people, they can seem like the same kind of judgment, but it’s clear that they are not.
To reiterate: normative theories suggest that whether we say someone did something intentionally or not depends on evaluative (normative) moral judgments about the agent, his behaviour or the outcomes it produces. The DSCM, on the other hand, proposes that judgments of intentionality depend of descriptive judgments about the evaluative (moral) attitudes of other people, and — again descriptively — whether their behaviour and the outcomes produced are concordant with those attitudes: if they are, the action is judged to be intentional; if not, then unintentional. And so this is what Sripada and Konrath set out to test.
To do so, they ran the original Chairman case by 240 study participants: half read the ‘harm’ version, and half the ‘help’ version. Participants then answered a series of question, each designed to tap into different normative and descriptive aspects of the Chairman’s behaviour. The questions were as follows, and each was answered using a seven-point sliding scale:
1. How much do you agree with the statement “The Chairman intentionally harmed [helped] the environment”? (Scale ranged from ‘strongly agree’ to ‘strongly disagree’.)
2. In your view, how good or bad is the outcome that the environment is harmed [helped]? (‘Very bad’ to ‘very good’.)
3. In your view, what is the Chairman’s moral status? (‘Very moral’ to ‘very immoral’.)
4. What are the Chairman’s values and attitudes towards the environment? (‘Very pro-environment’ to ‘very anti-environment’.)
5. In the vignette above, the Chairman’s action brings about an outcome in which the environment is harmed [helped]. In your view, to what extent is the Chairman the kind of person who will, in other contexts and situations, bring about outcomes similar to this one? (‘Very likely’ to ‘very unlikely’.)
6. What are your own values and attitudes towards the environment? (‘Very pro-environment’ to ‘very anti-environment’.)
Question 1 of this series simply gauges whether or not people are more likely to ascribe intentionality to the Chairman in the harm condition than the help condition. They did, providing another confirmation of the Knobe effect. The other questions bear on the possible factors that might underlie this effect. Questions 2 and 3 probe normative evaluations of the outcome of the Chairman’s actions, and his moral status, respectively. Question 4 asks a descriptive question about the values and attitudes of the Chairman. Question 5 is a similarly descriptive judgment about the likelihood that the Chairman’s behaviour will generalise to other situations— that is, whether the behaviour in this case is perceived as an aspect of the Chairman’s ‘deep self’, his enduring attitudes and values; or whether it’s seen as a one-off with no relevance to how we view the Chairman’s deep self. (So Questions 4 and 5 constitute the ‘deep self’ variables.) Question 6, which probes study subjects’ own normative attitudes towards the environment, was included to look at the role of these considerations in judging the Chairman. (The order in which these questions were presented was systematically varied between subjects to eliminate potential order effects.)
Sripada and Konrath didn’t simply looking at the subjects’ responses, and try to work out which factors are driving the Knobe effect. Instead, they employed a technical statistical tool called structural path analysis to achieve the same end. I’m not going to try to explain this in detail, but here is the basic idea.
The starting point is that the Chairman case people read about — either the harming or helping scenario — has an effect on intentionality judgments (the Knobe effect). There are a number of mediating factors that could potentially explain this effect: some normative, others descriptive. So the psychological process involved here can be represented as a series of boxes, starting on the left with Case (harm or help), moving right through the various mediating factors, and ending up with Intentionality Judgments on the right. These boxes can then be connected with arrows or ‘paths’ that represent potential causal connections between each box. The crucial point is that structural path analysis enables a pathway coefficient to be calculated for each causal path, which denotes how strong the effect is on the next box (moving rightward). This all makes much more sense when you actually look at a structural path analysis; Figure 1 shows the first example Sripada and Konrath came up with.
The way that Sripada and Konrath devised Figure 1 is as follows. Firstly, the whole point of this study is to see how reading about the Chairman harming or helping the environment (the two cases that can sit in the ‘case’ box) influences intentionality judgments. And Sripada and Konrath wanted to see which factors mediate this effect. The questions they had subjects answer tapped into four different factors: Goodness/Badness Judgments (of the outcome); Moral Status (of the Chairman); Chairman Values/Attitudes; and Generalizability (whether the Chairman would behave similarly in other contexts). These four potential meditating factors therefore sit between Case on the left, and Intentionality Judgments on the right. And because Case is presumed to affect judgments of these four factors, there are causal arrows leading from Case to the meditating boxes. Likewise, because these four factors are hypothesised to affect intentionality judgments, there are arrows leading from the former to the latter.
What do the numbers on each path tell us? These are the pathway coefficients, which reflect the strength of the effect of each leftward box on the rightward box to which it points. The details of how these are calculated need not delay us here, but in essence the logic is this. Different people read one case or the other, and then they answer a question about how good or bad the outcomes were, say. On the whole, people who read about the Chairman harming the environment will tend to say ‘bad’ or ‘very bad’, and the converse in the other, helping, case. So there will be a correlation between Case (harm or help) and Goodness/Badness Judgments about the outcome. Similarly, we’d expect a correlation for Case and Moral Status, Chairman Attitudes and Generalizability. Each of these four factors may in turn be correlated with Intentionality Judgment. (In addition, there is a path going straight from Case to Intentionality Judgments, which captures the effects of case that are not mediated through one of the other variables.)
Structural path analysis take the correlations in the raw data and, on the basis of the proposed set of interconnections, calculates the pathway coefficients, which for technical reasons are not simply the correlations between each box. Rather, they reflect the causal contribution of one to the next. The triple asterisks on some of the paths denote highly statistically significant effects (p < 0.001); the dotted lines represent insignificant causal connections.
So what does Figure 1 tell us? Let’s move from left to right. The pathway coefficients indicate that Case has a statistically significant effect on all four proposed meditating variables, being particularly high for Goodness/Badness Judgments, Chairman Attitudes, and Generalizability, and lower for Moral Status. At this point, we might well think, “So all factors play an important role in intentionality judgments”.
But let’s move rightward once again, from the four variables to Intentionality Judgment. Now the pathway coefficients tell a different story. The coefficients leading from the normative factors — Goodness/Badness and Moral Status — are in fact very low and statistically insignificant, suggesting that they are not playing a strong causal role in Intentionality Judgments.
However, the pathway from the Chairman Attitudes and Generalizability boxes to Intentionality Judgment, as well as the direct pathway from Case to Intentionality, make a much more profound contribution. Score 1 for the DSCM.
The model shown in Figure 1 — that is, the specific causal connections leading from Case to Intentionality Judgments — was devised on the basis that all four variables (normative and deep-self) could all play a role. However, structural path analyses make it possible to ask, of the model as a whole, how well it ‘fits’ the data it has to account for. For the model in Figure 1, the fit was ‘good’.
However, good isn’t great, and there might be room for improvement. Thankfully, structural path analysis also suggests which paths may be modified in order to get a better fit with the data. Revising Figure 1 along these lines gives the model in Figure 2 (see above).Figure 2. Revised structural path analysis of intentionality judgments
Notice that in this revised model, the arrows connecting the two normative factors of Moral Status and Goodness/Badness with Intentionality Judgment have been done away with. Instead, a new arrow has been introduced in which Goodness/Badness Judgments have an effect, albeit a small one, on judgments about the Chairman’s Attitudes, which in turn has an effect on Intentionality Judgments.
Sripada and Konrath carried out further analysis of the network of cause and effect shown in Figure 2. Specifically, they calculated mediation effects for the two normative factors in the model, and the two deep-self variables. Case manipulations — whether you read the harm or help scenario — affect Intentionality Judgments, and mediation effects capture how much of this influence is mediated through the various variables. In the model shown in Figure 2, Sripada and Konrath found that the two deep-self variables mediate 55% of the influence of Case manipulations on Intentionality Judgments; just 4% is mediated through Goodness/Badness Judgments, which in this model act indirectly by affecting judgments about Chairman attitudes. (Moral Status, as it was not connected to any downstream boxes, does not mediate any effect on Intentionality Judgments in this model.)
Statistically, this model has an almost perfect fit with the data, further bolstering the deep-self concordance model, and relegating normative factors to at best a minor causal role in generating the Knobe effect. If the DSCM model is correct, or at least on the right track, then it undermines claims for the pervasive effect of moral considerations on what is construed as intentional action. When it comes to judging whether someone produced an outcome intentionally, it isn’t morality that matters, but consistency: specifically, the concordance between a person’s values and attitudes, and the outcomes of their actions.
Sripada and Konrath didn’t stop here. They also looked at how people make sense of the Knobe effect, by getting a further batch of participants to read both the harm and help scenarios, followed by a note explaining that researchers have demonstrated an asymmetry in intentionality judgments in these cases (in slightly more user-friendly language!). They were then asked what might explain this effect. The vast majority cited normative factors — precisely the explanatory factors that the structural path analyses undermine.
Why are people so introspectively inaccurate in this case? The fact is that our introspective powers frequently prove to be limited, and we often have little access to the unconscious and automatic processes that shape our decisions and judgments. Instead, we typically latch on to the most salient and accessible features of a problem, and cite those as playing the key role in our thought processes – whether or not they really are playing the causal role we assign to them. Sripada and Konrath propose that in the case of the Chairman scenarios, the normative factors — the fact that he did something that led to bad outcomes, and could be an immoral person — are highly salient, and easy to bring to mind. The ideas underlying the deep-self concordance model, by contrast, are much less salient and introspectively accessible. If this is right, it would not only explain why regular folk can be misled as to what’s important in the Chairman case, but also why so philosophers have focused on normative factors in explaining the Knobe effect – they are human, after all!
1. Knobe, J. Intentional action and side effects in ordinary language. Analysis 63, 190–94 (July 2003).
2. Jones, D. The good, the bad and the intentional. The Psychologist 22, 666–669 (2009).
3. Young, L., Cushman, F., Adolphs, R. et al. Does emotion mediate the relationship between an action’s moral status and its intentional status? Journal of Cognition and Culture 6, 265–278 (2006).
4. Sripada, C. S. The Deep Self Model and asymmetries in folk judgments about intentional action. Philosophical Studies 151, 19–76 (2010).
5. Sripada, C. S. & Konrath, S. Telling more than we can know about intentional action. Mind & Language 26, 353–80 (June 2011).