Side-Effect Effect Without Side Effects: The Pervasive Impact of Moral Considerations on Judgments of Intentionality more

Side-Effect Effect Without Side Effects: The Pervasive Impact of Moral Considerations on Judgments of Intentionality Florian Covaa, Hichem Naarb a Florian Cova, Département d’Etudes Cognitives, Ecole Normale Supérieure ; Institut Jean Nicod, Centre National pour la Recherche Scientifique, florian.cova@gmail.com. bHichem Naar, School of Social Sciences, Philosophy, University of Manchester, hm.naar@gmail.com This research was supported in part by a Grant from the French Agence Nationale de la Recherche (ANR) (ANR Blanche: SoCoDev). We thank Edward Cokely, Adam Feltz, Ana Ganho, Joshua Knobe, and two anonymous reviewers for useful comments on previous versions of this paper. Abstract Studying the folk concept of intentional action, Knobe (2003a) discovered a puzzling asymmetry: most people consider some bad side effects as intentional while they consider some good side effects as unintentional. In this study, we extend these findings with new experiments. The first experiment shows that the very same effect can be found in ascriptions of intentionality in the case of means for action. The second and third experiments show that means are nevertheless generally judged more intentional than side effects, and that people do take into account the structure of the action when ascribing intentionality. We then discuss a number of hypotheses that can account for these data, using reactions times from our first experiment. 2 Introduction In a series of seminal papers, Joshua Knobe (2003a, 2004, 2006) gives us reasons not to believe that our judgments about whether an action is intentional depend only on psychological considerations. By running a series of experimental studies on people’s intuitions about intentional action, he obtained results that strongly suggest that people’s ascriptions of intentionality are massively influenced by considerations – including, according to Knobe, moral considerations – that are generally not considered as playing any significant role in our folk psychology. In his famous 2003 study, Knobe (2003a) investigated whether people judge foreseen side effects of an action as performed intentionally. In a series of vignette experiments, he found an interesting asymmetry: people considering a bad side effect (e.g., harming the environment) were much more likely to judge the agent as performing it intentionally than people considering a good side effect (e.g., helping the environment). This striking asymmetry has since been replicated many times (for example, Nichols & Ulatowski, 2007) and with many other scenarios (for example, Mallon, 2008), including in other languages and cultures (Knobe and Burra, 2006), in infants (Leslie et al. 2006, Pellizzoni et al. 20091), in people suffering from the Asperger Syndrome (Zalla & Machery, ms) and in patients with cerebral lesions to the prefrontal cortex (Young et al. 2006). It has also been shown to depend on individual differences such as extraversion (Cokely and Feltz, 2009) or philosophical training (Feltz and Cokely, forthcoming). So far, the growing body of studies about this asymmetry has focused exclusively on people’s ascription of intentionality in the case of side effects,2 hence the name currently given in the literature to the asymmetry: the “side-effect effect” (its other name being the “Knobe effect”). In this paper, our aim is to contribute to this area of research by investigating 3 ascriptions of intentionality in the case of means for action. According to the philosophical literature on intentional action, replicating such results in the case of means seems unexpected: since means are typically intended and part of the agent’s plan, there is a consensus that means are generally intentional, a consensus that is largely absent about the intentional status of side effects. By contrast, we show in our first study that an effect similar to the “side-effect effect” can be observed in the case of means. But our second study stresses that people also take into account the structure of the action in their ascriptions of intentionality and that, ceteris paribus, means are judged more intentional than side effects. These results suggest that the influence of moral considerations upon ascriptions of intentionality isn’t limited to side effects and is broader than we could originally have thought (Knobe, 2010). Thus, the findings are consistent with an increasing amount of research showing that moral considerations have a deeply pervasive impact on folk psychological concepts and affect attributions not only of intentionality but also of mental states, such as ‘knowing’ (Beebe & Buckwalter, 2010) or ‘deciding’ (Pettit & Knobe, 2009). Nevertheless, we argue that it is not yet clear which kind of moral evaluations impact on ascriptions of intentionality, and how this could be the case. 1. Experiment 1 Our first experiment was designed to investigate whether a phenomenon similar to the “sideeffect effect” could also be observed for other parts of actions. An action can be divided into three distinct types of components: its goal, the means necessary to achieve this goal, and its potential effects, including side effects. Following this analysis, we designed 12 pairs of scenarios describing actions. In four pairs (borrowed from Knobe, 2003a and Mallon, 2008), 4 only the side effect was changed within pairs: one scenario featured a morally good side effect while the other featured a morally bad side effect (morally good side effect vs. morally bad side effect). In 4 other pairs of (original) scenarios, only the means were changed: one scenario of each pair featured morally good means while the other scenario featured morally bad means (morally good means vs. morally bad means). Finally, for the four remaining pairs, only the moral valence of the action’s goal was modified (morally good goal vs. morally bad goal). The “Means” and “Side Effect” scenarios featured a man intending to achieve a goal but explicitly not caring about what must be done in order for the goal to be fulfilled. In the “Goal” condition, by contrast, the agent cared very much about the goal they were trying to achieve. At this stage, one might worry that the distinction between means and side effects is somewhat blurry. For example, I may knowingly burn gasoline in driving to the office. But this might be both conceptualized as a means (driving the car is a means to getting to the office and burning gasoline is a means to making the car run) or as a side effect (burning gasoline is an unfortunate side effect of car travel).3 So, how should we understand the difference between means and side effects? A way to draw the distinction is the following: an event E is a means to my goal G only if E is a necessary component of the causal explanation of my bringing about G. If it isn’t a component of the causal explanation, but is still an effect of my reaching G, then it is a side effect. Let’s take as an example the famous pair of scenarios that constitute the Trolley Problem. In the Bystander case, a runaway trolley is headed for five people who will be killed if it proceeds on its present course. The only way to save them is to hit a switch that will turn the trolley onto an alternate set of tracks where it will kill one person instead of five. In the Footbridge case, a trolley threatens to kill five people. You are standing next to a large stranger on a footbridge that spans the tracks, in between the oncoming trolley and the five 5 people. The only way to save the five people is to push the man off the bridge, onto the tracks. He will die but his body will stop the trolley from reaching the others. In the first case, considering that your goal is to save the five people on the tracks, killing the person on the alternate set of tracks is a side effect because I can perfectly explain how I succeeded in saving the five only in saying that I hit the switch and diverted the trolley. In the second case, my killing the person is a means because there’s no way I could explain how I saved the five without mentioning my throwing a person onto the tracks. Of course, this criterion doesn’t solve all difficulties. It can be argued that, even in the Footbridge case, the death of the stranger can be conceived as a side effect: after all, I do not have to mention that he died to explain how his body stopped the train. But, in most cases, this criterion will be sufficient. For instance, according to this criterion, there’s no way my killing the other person would be conceived as a means in the Bystander case. This criterion also has the advantage of being consistent with criticisms that have been addressed to Machery’s experimental studies about ascriptions of intentionality to side effects (Machery, 2008; for critics, see Mallon, 2008; Phelan and Sarkissian, 2009; Wright and Bengson, 2009). Machery has compared two scenarios, the Free Cup case, in which the agent, because he feels dehydrated, orders a smoothie and receives it in a commemorative cup, and the Extra Dollar case, in which the agent orders a smoothie and has to pay one dollar more than usual to get it, and found that participants were far more likely to say that the agent intentionally paid one dollar more (95%) than he intentionally obtained the commemorative cup (45%). Some have argued that these results did not tell us something interesting about ascriptions of intentionality in the case of side effects because paying one dollar more was a means rather than a side effect. This criticism is consistent with our criterion: it is impossible to give a causal explanation of the fact the agent got a smoothie without mention of his giving the extra dollar. 6 Using this criterion, we gave “Means” and “Side Effect” scenarios two very distinct structures. In “Means” scenarios, an agent who wants to reach a goal G is told that he needs first to perform an action M that will cause G to happen, and M is what we considered as a means. In “Side Effect” scenarios, an agent who wants to reach a goal G is told that he needs first to perform an action M that will cause G to happen, and that performing M will also cause another event S to occur. In these cases, S is not part of the causal chain that links M to G and we considered such an event S to be a side effect. Below is a pair of ‘Side Effect’ scenarios (drawn from Mallon, 2008): Bad Side Effect: A terrorist tells his leader: “We’re planning to bomb a nightclub. It will allow us to kill many Americans and it will also kill some Australians who happen to be there.” The leader answers: “Well, it would be good to kill Australians, but I don’t really care. All I care about is killing as many Americans as possible.” The terrorist blows up the nightclub and Australians are killed. Good Side Effect: A terrorist tells his leader: “We’re planning to bomb a nightclub. It will allow us to kill many Americans and it will also drive down property costs, helping a nearby orphanage to expand.” The leader answers: “Well, it would be good to help the orphanage, but I don’t really care. All I care about is killing as many Americans as possible.” The terrorist blows up the nightclub and the orphanage expands. Next, a pair of “Means” scenarios (all “Means” scenarios are detailed in the Appendix): Bad Means: The vice-president of a company tells the chairman of the board: “We have found a new way to increase productivity. If we shorten our workers’ coffee break, they will waste less time at the cafeteria and work more.” Good Means: The vice-president of a company tells the chairman of the board: “We have found a new way to increase productivity. If we give our workers a one-hour break for taking a nap, they will be in better health and work more.” 7 The chairman answers: “I don’t care about them having a shorter break. All I want is to increase productivity. Let’s do this.” The workers’ break is shortened. Did the chairman intentionally shorten the workers’ break? The chairman answers: “I don’t care about them taking a nap. All I want is to increase productivity. Let’s do this.” The workers are given a one-hour naptime. Did the chairman intentionally give the workers a one-hour naptime? 1.1. Participants and Procedure For the experiment, we used French versions of the scenarios since participants were French. Participants were student recruited at the Ecole Normale Supérieure (ENS Ulm) of Paris. There were 30 participants (9 were females and 21 were males) and the age mean was 24.2 (standard deviation: 5.7). Before the experiment started, all participants were asked if they knew about the “side-effect effect” (or “Knobe effect”). None of them did. All scenarios were displayed on a computer screen and split into three parts. At the end of each scenario, subjects had to answer by “YES” or “NO” to a question about the scenario. This could be a morality question: “Can we praise/blame X for doing Y?” or an intentionality question: “Did X intentionally Y?”. The two questions always bore on the element of the scenario that changed within the pair to which it belonged (goal, means, or side-effect). The experiment was designed so that, for each pair of scenarios, participants had to answer one morality question and one intentionality question. It resulted that each participant answered the same number of morality and intentionality questions. 1.2. Results For the “bad” scenarios, the morality question asked participants if the agent could be blamed 8 for having brought about the targeted goal / means / side effect. In the “good” scenarios, they were asked if the agent could be praised for it. Table 1 indicates for each condition the proportion of participants who gave a positive answer to the morality question: Bad Good Goal Means Side Effect 87% 88% 78% 80% 28% 17% Table 1. Participants’ positive answers to the morality question (Experiment 1) For the intentionality question, Table 2 shows how many participants answered that the agent brought about the goal / means / side effect intentionally: Goal Means Side Effect Bad 95% 82% 62% Good 90% 42%*** 10%*** Table 2. Participants’ positive answers to the intentionality question (Experiment 1) Data were grouped by participant. We ran a two-factor ANOVA using as factors the part of the action that was modified (goal / means / side-effect) and the moral valence of these parts (bad / good). We found a significant effect of the part of action that was modified (F(2)=(66.3, p<.001,η η ρ 2 ρ 2 =0.4), a significant effect of moral valence (F(1)=64.1, p<.001, ρ 2 =0.25) and a significant interaction effect (F(2)=12.1, p<.001, η =0.11).4 Two-tailed wilcoxon tests allowed us to see that the difference between bad and good items were significant in the case of means and side-effects but not in the case of goals. 5 Thus, we were successful in replicating the “side-effect effect” at the level of means (and – as a side effect – in showing that it’s also present in French). Indeed, participants were more likely to say that the agent brought about a means intentionally when it was a morally bad means than when it was a morally good means. This suggests that the so-called “side-effect effect” is not limited 9 to side effects but also present at the level of means. 2. Experiment 2 Experiment 1 allowed us to observe that moral valence has an impact on how intentional a means is considered to be. Now, we would like to ask whether being a means in itself is something that can be taken into consideration when people ascribe intentionality judgments. Do people make a sharp distinction between means and side effects when it comes to ascribing intentionality? It seems plausible that means are typically considered more intentional than side effects, since means, as opposed to side effects, are genuine components of the action itself and thus are planned and embedded in the intention of the agent. Moreover, experiments in moral psychology have recently stressed the importance of the means / side effects distinction, particularly in the empirical study of moral dilemmas similar to the famous Trolley Problem. Overall, it seems that people are more willing to condemn the sacrifice of one person in order to save many when this sacrifice comes as a means (as in the Footbridge case) rather than a side effect (as in the Bystander case) (Cushman et al., 2006; Hauser et al., 2007). If intentionality judgments are inputs to morality judgments, it might be that this difference is due to the fact that means are perceived as more intentional than side effects. In Experiment 1, intentionality ratings were higher for means than for side effects (62% versus 36%). But this difference cannot be interpreted directly, since there was no minimal pair composed of one “Means” and one “Side Effects” scenarios: the effect could be due to differences in other features of the scenarios. So, to test precisely the impact of the Means / Side Effect distinction on intentionality judgments, we designed the following minimal pair of scenarios: 10 Mice (Means): A terrorist sets up a bomb in the basements of an embassy in order to blow it up. It is about to explode. A security agent learns about the situation. He enters the basement with a colleague who suddenly tells him: “Look! There are mice all over the bomb. To access it, we’ll have to disturb them.” The agent replies: “I’m perfectly aware of this, but I don’t care, all that matters is defusing the bomb”. The security agent repeatedly claps his hands to make the mice go away then defuses the bomb. Mice (Side Effect): A terrorist sets up a bomb in the basements of an embassy in order to blow it up. It is about to explode. A security agent is aware of the situation. He’s about to enter the basement when his colleague tells him: “If you enter the basement, it will disturb the mice that live there.” The agent answers: “I’m perfectly aware of this, but I don’t care, all that matters is defusing the bomb”. The security agent enters the basement, defuses the bomb and disturbs the mice. After each scenario, subjects had to answer the following question: “Did the agent intentionally disturb the mice?” Participants had to answer on an eleven-point scale ranging from -5 (“NO”) to 5 (“YES”), 0 being the midpoint. Our hypothesis was that people would judge disturbing the mouse (a neutral nonmoral event) more intentional in the case of a means than in the case of a side effect. 2.1. Participants and procedure Participants were 60 students recruited at Université Pierre et Marie Curie (Paris VI) in Paris. Twenty-four were females and two didn’t report whether they were male or female. The age mean was 19.7 (standard deviation: 4.2). Thirty-one participants received the Means condition and twenty-nine the Side Effect condition. 11 2.2. Results For the Means condition, the average intentionality rating was 2.39 (19% of participants gave an answer inferior to 0, 10% an answer equal to 0, and 71% an answer superior to 0). For the Side Effect condition, the average intentionality rating was -0.03 (45% of participants gave an answer inferior to 0; 7% an answer equal to 0; and 48% an answer superior to 0).6 A twotailed t-test revealed a statistically significant difference, N=60, t=2.6, df=55.2, p=.01, d=1.0. Thus, our hypothesis was confirmed: a neutral item is more thought to be the product of intentions when it is a means than when it is a side effect. 3. Experiment 3 In order to strengthen and replicate the results we obtained in Experiment 2, we ran a third study in which we compared again the intentionality judgments in means and side effects using (i) a new pair of scenarios and (ii) a within-subject design. The new pair of scenarios7 was the following: Burglar (Means): During the night, a burglar introduces himself into a house whose owners are abroad. If he switches on the light, there are great chances it will cast his shadow on the window. But, as the neighbors are also abroad, it won’t alert the police of his presence. The burglar has an accomplice waiting outside in a car. He must signal to him that the way is Burglar (Side-Effect): During the night, a burglar introduces himself into a house whose owners are abroad. If he switches on the light, there are great chances it will cast his shadow on the window. But, as the neighbors are also abroad, it won’t alert the police of his presence. The burglar has forgotten his pocket torch. In absence of light, it is impossible for him to 12 clear by casting his shadow on the window. To warn his accomplice, the burglar switches on the light, which casts his shadow on the window. thoroughly search the drawers. To have a better view, the burglar switches on the light, which casts his shadow on the window. After each scenario, subjects had to answer the following question: “Did the burglar intentionally cast his shadow on the window?” Participants had to answer on an eleven-point scale ranging from -5 (“NO”) to 5 (“YES”), 0 being the midpoint. 3.1. Participants and procedure Participants were 40 students recruited at Université Pierre et Marie Curie (Paris VI) in Paris. Twenty-eight were females. The age mean was 22.4 (standard deviation: 4.0). Participants received both scenarios. 20 received the MEANS condition first and 20 the SIDE EFFECT condition first. 3.2. Results For the MEANS condition, the average intentionality rating was 4.8 (all participants gave an answer superior to 0). For the SIDE EFFECT condition, the average intentionality rating was -0.63 (50% of participants gave an answer inferior to 0; 5% an answer equal to 0; and 45% an answer superior to 0).8 A two-tailed paired t-test revealed a statistically significant difference, N=40, t=8.4, df=39, p<.001, d=1.1. As in our previous experiment, participants tended to judge means as more intentional than side effects. There was no order effect. 4. Possible accounts 13 In the precedent sections, we have presented evidence that a phenomenon similar to the “sideeffect effect” also affected ascriptions of intentionality for means of action, though means were on average considered more intentional than side effects. How are we to account for these results? Many accounts have been proposed to account for the original “side-effect effect” (see Feltz, 2007 for a review). In this section, we select some of them and investigate whether they can be applied to the phenomenon we uncovered. 4.1. Pettit and Knobe’s account According to Pettit and Knobe’s account (Pettit and Knobe, 2009; Knobe, 2010), the semantic properties of “intentionally” must be understood by comparison with the semantics of gradable terms: an agent is said to have done x intentionally only when his pro-attitude towards x exceeds a certain amount y that constitutes the “default point”. The key point of Pettit and Knobe’s account is that y is determined by the moral valence of x, that is, whether x is good or bad. For morally good actions, the default point is to have some sort of “proattitude” while, for morally bad actions, the default is to have some kind of “con-attitude”. Thus, in cases of morally bad actions, an indifferent agent (neither pro nor con) is above the default point, and his action is judged intentional while, in cases of morally good actions, an indifferent agent is far below this same threshold, making his action unintentional. For Knobe, intentionality judgments, though they are influenced by moral evaluations of the consequences of one’s action, are nevertheless inputs to our judgments of praise and blame (Knobe, 2006). As we saw in our first experiment, judgments of praise and blame show the same asymmetry as intentionality judgments. For Knobe, it is because praise and blame judgments are driven by intentionality judgments that are in turn influenced by moral 14 evaluations of the consequences of one’s action. Since moral evaluations of the consequences themselves (in terms of good or bad) are distinct from praise-blame judgments (in terms of praiseworthy or blameworthy) and do not depend on them, this doesn’t create infinite regress (Knobe and Mendlow, 2004). When applied to the phenomenon we uncovered, this account can explain the results we obtained in our first experiment: given that we expect people to desire good events and to be reluctant to bring about bad events, the “default point” is set below indifference on the “pro-attitude” scale in the case of bad events and beyond this very same point in he case of good events. As a result, when the agent expresses indifference (“I don’t care”), as in means and side-effects scenarios, people will tend to judge bad events as intentional (pro-attitude is beyond the default point) and good events as not intentional (pro-attitude is below the default point). However, when agents express strong pro-attitudes, such as in goals scenarios, both bad and good events will be judged intentional (pro-attitude being far beyond the default point). It can also explain why means are ceteris paribus considered more intentional than side effects: when an event constitutes a means (rather than a mere side effect), we must add to the agent’s attitude towards this event an instrumental desire that enhances the pro-attitudes of the agent and increases the probability that they will rise and pass over the “default point”. 4.2. Nadelhoffer’s “blame bias” hypothesis According to Nadelhoffer (2004a, 2004b, 2006), the original “side-effect effect” is due to affective biasing. When we are not driven by emotions and properly apply our concept of intentional action, we usually do not consider side effects to be intentional. Nonetheless the perceived blameworthiness or praiseworthiness of the agent can induce emotions that bias us and lead us to attribute intentionality to this agent’s action. Thus, the asymmetry between 15 morally good and morally bad cases is explained by the fact that we are prone to blame indifferent agent in morally bad cases but not to praise them in morally good cases. How can this hypothesis account for our data? We can advance the hypothesis that people’s “default view” about means is that they are intentional, which would explain the results of our second experiment. Then, we can postulate that we are emotionally driven to withhold intentionality from morally good actions performed by indifferent agents because people that are insensitive and indifferent to goodness elicit strong negative feelings and attributions of blame. 4.3. Pragmatic account According to Adams and Steadman (2004a, 2004b, 2007), the “side-effect effect” is the product of conversational implicatures. Alhough our concept of “intentional action” doesn’t apply to side effects, people use the word “intentionally” in morally bad cases to conversationally imply that the agent deserves blame because they think that saying that the agent’s action was not intentional would on the contrary imply that they do not blame him for what he did. This hypothesis can be expanded to ascriptions of intentionality in the case of means in the following way: means are by default judged intentional but, on a pragmatic level, people prefer to say that morally good means brought about by indifferent agents are not intentional because they think that acknowledging that their actions were intentional would imply that they deserve praise for what they did. Such an account would explain both the results of our first and second experiment. 4.4. Multiple concepts accounts and “mixed” account 16 Nichols and Ulatowski (2007) have proposed that the original “side-effect effect” is due to the fact that people interpret “intentionally” in two different ways: a knowledge-based interpretation, according to which an agent did x intentionally if he believed that x would happen as a result of his action; and a desire-based interpretation, according to which an agent does x intentionally if he did x because he desired x to happen). According to Nichols and Ulatowski, people are “flexible” in their use of the term, using the first sense in some cases (e.g. the morally bad cases) and the second in other cases (e.g. the morally good cases). In a similar way, Cushman and Mele (2008) proposed that there are in fact three distinct concepts of intentional action. According to them, almost all people think that having the desire to do x is sufficient to do x intentionally when one x-es. However, while (i) some people consider that believing that x would happen as a result of their action is also a sufficient condition for doing x intentionally, (ii) some others consider that desire is a necessary condition and that belief cannot be sufficient and (iii) a third group treat belief as a sufficient condition in the case of morally bad actions but desire as a necessary condition in other cases. These accounts cannot accommodate all our data. To explain why means are more intentional than side effects (as we found in Experiments 2 and 3), we must consider that the “desire” condition takes into account instrumental desires that are oriented towards means. However, this would imply that means, whether they are morally good or bad, are intentional according to Nichols and Ulatowski’s two meanings and Cushman and Mele’s three concepts, and that we shouldn’t observe the effects we found in our first experiment. On the contrary, if we use these accounts to explain the results of our first experiments by making the hypothesis that, according to the “desire” condition, instrumental desires are not considered as desires, so that agent do not bring about intentionally means that are not intrinsically desired, then we are 17 unable to explain why means were more intentional than side effects in Experiments 2 and 3. Nevertheless, multiple concepts accounts can be useful to us if we combine them with Nadelhoffer’s emotional bias account (date?) to obtain a certain “mixed” account. In a series of experiments, Cokely and Feltz (2009) have showed (i) that some people used a “belief-issufficient” concept of intentional action, while others used a “belief-is-insufficient” concept of intentional action; and (ii) that people who are high in extraversion, and thus with looser regulation of affective reactions, are more prone to treat morally bad and morally good cases differently. These results are coherent with the following hypothesis: a set of people uses a “belief-is sufficient” concept while another set uses a “desire-is-necessary” concept and negative emotions associated to blame attribution bias some people and lead them to attribute intentionality in morally bad cases or withhold it in morally good cases. This account can explain (i) why we observe asymmetries in ascriptions of intentionality, whether for means or side effects, and (ii) why means are more intentional than side effects. In the absence of emotional bias, side effects are intentional only according to the “belief-is-sufficient” concept, whereas means are intentional according to both the “belief-is-sufficient” and the “desire-isnecessary” concepts. This account appears to fit our data better than the other available accounts. Nadelhoffer and Adams and Steadman’s accounts suppose that, in absence of bias, side effects are judged unintentional. However, the results we obtained in our second experiment about the Side Effect cases in Experiments 2 and 3 rather suggest that, in the absence of moral considerations, side effects are judged intentional by half of the participants and unintentional by the other half (a result coherent with Machery’s Free Cup case). These results are more consistent with the idea according to which half of the participants used a “belief-issufficient” concept while the other half used a “desire-is-necessary” concept. Note that Pettit and Knobe’s account can also accommodate these data if we add the auxiliary hypothesis that, 18 in absence of moral considerations, half of the participants place the “default point” above indifference and the other half below. So these two accounts seem to be the best we actually have for the asymmetry we observed in the case of means. 4.5. Non-moral accounts All the accounts we have proposed so far consider that the asymmetries we observed in judgments of intentionality depend, in one way or another, on moral evaluations. Nevertheless, some have proposed accounts of the “side-effect effect” according to which these asymmetries are not driven by moral considerations. According to the “reasons against” accounts, we judge a side effect intentional if it constitutes a reason for refraining to act. “Reasons against” accounts come in three ways: a side effect is intentional if it counts as a reason not to act (i) either from the agent’s perspective (Turner, 2004; Hindriks, 2009), (ii) or from “our” (i.e., the observer’s) perspective, (iii) or from any of these two perspectives (Nanay, 2010). In a similar way, Machery (2008) puts forward the “trade-off” hypothesis, according to which we consider a side effect intentional when we conceptualize it as a cost incurred by the agent. These theories do not refer directly to moral considerations and thus avoid the paradoxical aspect of theories according to which intentionality ascriptions depend on moral evaluations. Nevertheless, none of these theories seem fit to account for our data. If we consider that means are by default intentional, then they cannot account for the effect observed in our first experiment, insofar as all means, whether they are a reason against (or a cost), will be intentional. But if we now make the hypothesis that people consider that means are by default unintentional, unless they count as a reason not to act (in which case they are intentional), then we cannot account for the difference between means and side effects in 19 Experiments 2 and 3, in which we had “neutral” means (that do not count as a reason not to act) judged as fully intentional. As a result, although these theories might have a value in accounting for the original “side-effect effect”, they do not seem to be the best candidate for our purposes. 4.6. Moral considerations and the folk concept(s) of intentional action We have thus surveyed a number of existing theories and adapted them to the case of means. We also argued that only those theories that gave a role to moral considerations were able to account for the data we collected. Are there reasons to think that some of these theories are better than the others? It is important to determine which one of these accounts comes closer to the truth, for each one of them draws a different picture of the nature of our folk concept of intentional action. If Pettit and Knobe’s account is true, then, this concept integrates moral and, more broadly, evaluative components. This is at odd with most of the philosophical analyses that have been proposed so far. If Nadelhoffer or Adams and Steadman’s account is correct, then the impact of moral evaluations does not tell us anything interesting about our core concept of intentional action, which can be analyzed as follows: to be intentional is to be intended. Finally, if our “mixed” account is correct, asymmetries do not reveal something about our concept of intentional action but this concept is not one, rather two concepts. The main question can be summarized in the following way. Do the asymmetries we observed teach us something about the content of our concept of intentional action? Is Pettit and Knobe’s account right (in which case the answer is “yes”) or should we prefer any of the available alternative accounts? In the following section, we use reaction times collected in our first experiment to compare these different theories. 20 5. Reaction times from Experiment 1 As the scenarios of our first experiment were divided in three parts, the third part containing only the final sentence of the scenario and the question, we were able to measure participants’ reaction times, starting from the moment the third part of the scenario was displayed on the screen. Given the fact that some of the accounts we surveyed make precise predictions about what the reaction times would look like, we wanted to verify whether our data confirmed or contradicted these predictions. We can schematically oppose two kinds of predictions about the temporal order of intentionality and morality questions. As already mentioned, Pettit and Knobe claim that intentionality judgments, although informed by moral considerations about the positive or negative value of actions, are nevertheless inputs for the blame / praise judgments. Since our morality question bore on the blameworthiness / praiseworthiness of the agent, we can make the assumption that Pettit and Knobe’s hypothesis makes the prediction that intentionality judgments will in general come (i.e., faster) before morality judgments. Thus, their account is an “intentionality-first” account. In contrast, according to Nadelhoffer’s blame-bias hypothesis, Adams and Steadman’s pragmatic account and the “mixed” account we proposed, in the case of bad side effects and good means, morality judgments (bearing on the blameworthiness of the agent) should come first and cause subjects to consider bad side effects as intentional. These are “morality-first” accounts.9 Having discarded one subject, whose mean reaction time was more than two standard deviations away from the mean, we grouped data by subjects and compared reaction times for intentionality and morality questions using a paired t-test. We found a marginally significant effect (p=0.07), with answers to the intentionality questions taking less time than answers to 21 the morality questions (6755ms versus 7357ms).10 This implies that, overall, intentionality judgments tend to occur faster than morality judgments, which is consistent with Pettit and Knobe’s “intentionality-first” hypothesis. Nevertheless, this result might also be consistent with “morality-first” hypotheses. The “morality-first” hypotheses we presented above apply only to bad side effects and good means. Therefore, we should take a closer look at the data. Table 3 details the mean reaction times along three factors (type of question, part of the action, and moral valence). Goal Means Side Effect Bad Good Bad Good Bad Good Intentionality question 6576 (3599) 5990 (4897) 6033 (3140) 6817 (5460) 8641 (5525) 6476 (3543) Morality question 7113 (4172) 6517 (3787) 6979 (4137) 7555 (5599) 8389 (5611) 7591 (3681) Table 3. Mean reaction times (ms) and standard deviations along type of questions, part of the action and moral valence. As shown by Table 3, intentionality judgments are always formed faster than morality judgments, except for bad side effects, in which case morality judgments come first. Although this reversal is nowhere near statistical significance,11 “Morality-first” hypotheses still might be true for side effects: the “side-effect effect” might be due to the influence of morality 22 judgments on intentionality judgments. Still, we can note two difficulties for “morality-first” hypotheses. The first difficulty is that, as we have just shown, an effect similar to the side-effect effect can be found at the level of means, and that, as far as means (and even good means) are concerned, intentionality judgments are way faster than morality judgments. Saving “morality-first” hypotheses would require dissociating accounts for the two effects, which seems like a costly solution. The second difficulty is that reaction times for bad side effects are longer than for any other condition. At first sight, these results don’t fit with Nadelhoffer’s account or our “mixed” hypothesis, according to which the influence of morality judgments on intentionality judgments is an emotional bias and relies on fast non-reflexive responses. Nevertheless, it could be argued that longer reaction times are due to “unbiased’ intentionality answers that need to overcome the emotional bias and that doing so takes more time. This is a sensible answer since, for bad side effects, negative intentionality answers show slower times than positive intentionality answers (10,366 ms versus 7507ms).12 Note also that the pragmatic account should predict that positive intentionality answers, being the product of pragmatic enrichment rather than the use of “intentionally” in its literal and direct meaning, would be slower than negative answers, which is clearly not the case. In the same way, pragmatic accounts should predict that, in the case of good means, positive answers (literal answers) will take less time than negative answers (that are due to pragmatic enrichment). This, once again, is not the case (8276ms versus 5787ms).13 Thus, within “morality-first” accounts, our data seem to fit better “emotional bias” accounts than the pragmatic account. To conclude, the data regarding reaction times in our experiments do not allow us to decide between the two kinds of hypotheses (“intentionality-first” versus “morality-first”), 23 although they might lead us to prefer the emotional bias hypotheses to the pragmatic account within “morality-first” accounts. Further work is thus required, using more precise methods to record reaction times, for instance by using auditory rather than visual stimuli so as to eliminate variations due to reading time. 6. Conclusion In this paper, our aim was to investigate attributions of intentionality for means of actions using experimental methods. We found that such ascriptions were shaped by two kinds of considerations. First, in Experiment 1, we expanded Joshua Knobe’s findings to means by producing evidence that strongly suggests that ascriptions of intentionality are influenced by moral considerations. Second, in Experiments 2 and 3, we were able to confirm that the difference between means and side effect (that is, in the structure of the action) was taken into account and that means were judged more intentional than side effects. These two factors (moral valence and structure of the action) also seem to interact. As shown in our first experiment, goals are not sensitive to moral valence, as opposed to means and side effects. As a result, the influence of moral considerations seems to be modulated by the part of the action that is judged. A more careful examination of the relation between these two kinds of consideration will be needed in order to understand the mechanisms that lead us to attribute intentionality. Finally, our results did not allow us to identify the kind of relationship that leads moral evaluations to impact on ascriptions of intentionality. Is this impact a mere bias or does it reveal something about our underlying core concept (or concepts) of intentional action? This question is crucial to understand what makes part of an action intentional, whether this part is a means or a side effect. 24 Notes 1. Note that, in these two experiments, the word used in the question was not ‘intentionally’ but ‘on purpose’ or the Italian word ‘apposta’. 2. Nevertheless, studies about another puzzling phenomenon, the “Skill Effect”, can be considered as bearing on ascriptions of intentionality in the case of means (Knobe, 2003b). 3. We thank an anonymous referee for raising this problem and for this example. 4. We also analyzed our results using non-parametric tests. Using a paired wilcoxon test, we found a significant effect of valence: N=30, V=401, p<.001. Then we ran three paired wilcoxon tests using part of the action as a factor. We found significant differences between Goals and Side Effects conditions (N=30, V=3.5, p<.001), Goals and Means conditions (N=30, V=29.5, p<.001) and Means and Side Effects conditions (N=30, V=277.5, p<.001). 5. Here are the detailed results of the three wilcoxon tests. For Goals: N=30, V=12, p=.23. Means: N=30, V=265.5, p<.001. Side Effects: N=30, V=309, p<.001. 6. In the SIDE EFFECT condition, the average answer (-0.03) is very close to 0, the middle of the scale. How should we understand this “intermediary” status? Is it that people have no idea 25 what to answer and thus chose a random answer, or is it that half the participants answered a straight YES and the other half a straight NO? To answer this question, we examined the distribution of the participants’ answers. We noticed that the distribution was not centered on the middle of the scale but bipolar (centered on -4 and 4). It looked like participants were divided between two main groups: those who answered “YES” and those who answered “NO” with few people giving intermediary answers. 7. These scenarios are variations on an original case designed by Pierre Jacob. We thank him for his imagination. 8. As for Experiment 2, we found a bipolar distribution for the SIDE EFFECT condition (centered on -5 and 4). 9. Additionally, both kinds of accounts predict (although not for the same reasons) a positive correlation between reaction times for intentionality questions and reaction times for morality questions – a correlation that we did observe (r=0.42, t=6.1, df=172, p<0.001). To check for correlation, we grouped reaction times by subject, type of question, part of the action and valence of the action for a total of 348 items (two groups of 174 items). 10. N=30, t=-1.9, df=28, p=0.07, d=0.3. 11. We ran a three-factor ANOVA on reaction times with the type of question, the valence of the action and the part of the action as factors. We found no significant interaction between these three factors. 12. We used a t-test on the available data, without grouping answers by subjects and found a marginally significant effect (N=58, t=-1.9, df=37.3, p=0.07, d=0.29). 13. We used a t-test on the available data, without grouping answers by subjects but didn’t find a significant effect (N=58, t=-1.6, df=33.7, p=0.12, d=0.44). 26 References Adams, F. and Steadman, A. (2004a). Intentional action in ordinary language: core concept or pragmatic understanding. Analysis, 64, 173-181. Adams, F. and Steadman, A. (2004b). Intentional action and moral considerations: still pragmatic. Analysis, 64, 264-267. Adams, F. and Steadman, A. (2007). Folk concepts, surveys and intentional action. In C. Lumer and S. Nannini (eds) Intentionality, Deliberation and Autonomy: The ActionTheoretic Basis of Practical Philosophy. Aldershot: Ashgate Publishers. Beebe, J. and Buckwalter, W. (2010). The epistemic side-effect effect. Mind & Language, 25, 474-498. Cokely, E.T. and Feltz, A. (2009). Individual differences, judgment biases, and Theory-ofMind: Deconstructing the intentional side effect asymmetry. Journal of Research in Personality, 43, 18-24. Cushman, F.A., Young, L. and Hauser, M.D. (2006). The role of reasoning and intuition in moral judgments: Testing three principles of harm. Psychological Science, 17, 10821089. Cushman, F. and Mele, A. (2008). Intentional action: two and half folk concepts. In J. Knobe and S. Nichols (eds) Experimental Philosophy. New York: Oxford University Press, NY. Feltz, A. (2007). The Knobe effect: A brief overview. The Journal of Mind and Behavior, 28, 265-277. Feltz, A. and Cokely, E.T. (forthcoming). Individual differences in theory-of-mind judgments: Order effects and side effects. Philosophical Psychology. Hauser, M.D., Cushman, F.A., Young, L., Kang-Xing, R. and Mikhail, J. (2007). A dissociation between moral judgments and justifications. Mind & Language, 22, 1-21. 27 Hindriks, F. (2008). Intentional action and the praise-blame asymmetry. The Philosophical Quarterly, 58, 630-641. Knobe, J. (2003a). Intentional action and side-effects in ordinary language. Analysis, 63, 190193. Knobe, J. (2003b). Intentional action in folk psychology: an experimental investigation. Philosophical Psychology, 16, 309-324. Knobe, J. (2006). The concept of intentional action: a case study in the uses of folk psychology. Philosophical Studies, 130, 2, 203-231. Knobe, J. (2007). Reason explanation in folk psychology. Midwest Studies in Philosophy, 31, 90-107. Knobe, J. (2010). Person as scientist, person as moralist. Behavioral and Brain Sciences, 33, 315-329. Knobe, J. and Burra, A. (2006). Intention and intentional action: a cross-cultural study. Journal of Culture and Cognition, 1-2, 113-132. Knobe, J. and Mendlow, G. (2004). The good, the bad and the blameworthy: understanding the role of evaluative reasoning in folk psychology. Journal of Theoretical and Philosophical Psychology, 24, 252-258. Leslie, A., Knobe, J. and Cohen, A. (2006). Acting intentionally and the side-effect effect: ‘Theory of mind’ and moral judgment. Psychological Science, 17, 421-427. Machery, E. (2008). The folk concept of intentional action: philosophical and experimental issues. Mind & Language, 23, 165-189. Mallon, R. (2008). Knobe versus Machery: testing the trade-off hypothesis. Mind & Language, 23, 247-255. Nadelhoffer, T. (2004a). Praise, side effects and intentional action. Journal of Theoretical and Philosophical Psychology, 24, 196-213. 28 Nadelhoffer, T. (2004b). Blame, badness and intentional action: a reply to Knobe and Mendlow. Journal of Theoretical and Philosophical Psychology, 24, 259-269. Nanay, B. (2010). Morality of modality? What does the attribution of intentionality depend on? Canadian Journal of Philosophy, 40, 28-40. Nichols, S. and Ulatowski, J. (2007). Intuitions and individual differences: the Knobe effect revisited. Mind & Language, 22, 346-365. Pellizzoni, S., Siegal, M. and Surian, L. (2009). Foreknowledge, caring and the side-effect effect in young children. Developmental Psychology, 45, 1, 289-295. Pettit, D. and Knobe, J. (2009). The pervasive impact of moral judgment. Mind & Language, 24, 586-604. Phelan, M. and Sarkissian, H. (2009). Is the trade-off hypothesis worth trading for? Mind & Language, 24, 164-180. Turner, J. (2004). Folk intuitions, asymmetry and intentional side effects. Journal of Theoretical and Philosophical Psychology, 24, 214-219. Wright, J. & Bengson, J. (2009). Asymmetries in folk judgments of responsibility and intentional action. Mind & Language, 24, 237-251. Young, L., Cushman, F., Adolphs, R., Tranel, D. and Hauser, M. (2006). Does emotion mediate the effect of an action’s moral status on its intentional status? Journal of Cognition and Culture, 1-2, 291-304. Zalla, T. and Machery, E. (manuscript). The concept of intentional action in Asperger Syndrome. 29 Appendix: Detailed Results for “Means” Scenarios (Experiment 1) Below, the “Means” Scenarios with their average intentionality rating are provided. Scenarios were translated from French. All * indicate statistical significance calculated by means of Fisher exact tests (° = p>0.1, * = p<0.05, ** = p<0.01, *** = p <0.001). Bad A man named Andre tells his wife: “My father decided to leave his immense fortune to only one of his children. To be his heir, I must find a way to become his favorite child. But I can’t figure out how.” His wife replies: “Your father always hated his neighbors and has declared war to them. You could do something that would really annoy them, even if you don’t really care about whether that would work. Andre decides to set fire to the neighbors’ car. Did he intentionally harm the neighbors? The vice-president of a company tells the chairman of the board: “We have found a new way to increase productivity. If we shorten our workers’ coffee break, they will lose less time at the cafeteria and work more.” The chairman answers: “I don’t care about them having a shorter break. All I want is to increase productivity. Let’s do this.” The workers’ break is shortened. Did the chairman intentionally shorten the workers’ break? % 100 Good A man named Andre tells his wife: “My father decided to leave his immense fortune to only one of his children. To be his heir, I must find a way to become his favorite child. But I can’t figure how.” His wife replies: “Your father always admired people who gave money to humanitarian-aid organizations. You could give money to such an organization, even if you don’t care about it.” Andre decides to give money to the humanitarian-aid organization. Did he intentionally help a humanitarian-aid organization? The vice-president of a company tells the chairman of the board: “We have found a new way to increase productivity. If we give our workers a one-hour break for taking a nap, they will be in better health and work more.” The chairman answers: “I don’t care about them taking a nap. All I want is to increase productivity. Let’s do this.” The workers are given a onehour naptime. Did the chairman intentionally give the worker a one-hour naptime? % 20*** 93 40** 30 A soldier tells his commander: “We have captured an enemy combatant. He could tell us where our enemies are hidden. But he’s stubborn and won’t talk to us. We need to torture him so he will speak.” The commander replies: “I don’t care about what happens to him. All I want is to know where our enemies are hidden. Torture him!” The prisoner is tortured by the soldier and talks. Did the commander intentionally have the prisoner tortured? A member of a gang tells his leader: “I’ve just been told the police is after us. We need to find a discrete way of getting rid of our money, so they’ll have nothing to charge us with.” The leader replies: “The most discrete way is to give all our money to weapons smugglers. I don’t care about helping them but this is the best way.” The money is given to the weapons smugglers. Did the leader intentionally help the weapons smuggler? 100 A soldier tells his commander: “We have captured an enemy. He could tell us where our enemies are hidden. But he’s badly injured and could die soon. We need treat him and save his life so he will speak.” The commander replies: “I don’t care about what happens to him. All I want is to know where our enemies are hidden. Take care of him!” The military physician treats him and his life is saved. Did the commander intentionally have the prisoner healed? A member of a gang tells his leader: “I’ve just been told the police was after us. We need to find a discrete way of getting rid of our money, so they will have nothing to charge us with.” The leader replies: “The most discrete way is to give all our money to an humanitarian-aid organization. I don’t care about helping them but this is the best way.” The money is given to the humanitarian-aid organization. Did the leader intentionally help the organization? 73° 33 33 31