Asymmetric Attitude Malleability 12/8/2004 Page -1- Easier Done than Undone: Asymmetry in the Malleability of Implicit Preferences

Dual-process models imply that automatic attitudes should be less flexible than their selfreported counterparts; the relevant empirical record, however, is mixed. To advance the debate, we conducted four experiments investigating how readily automatic preferences for one imagined social group over another could be induced and/or reversed. Experiments 1 and 2 revealed that automatic preferences, like self-reported ones, could be easily induced by both abstract supposition and concrete learning. In contrast, Experiments 3 and 4 revealed that newly-formed automatic preferences, unlike self-reported one, could not be easily reversed by either abstract supposition or concrete learning. Thus, the relative inflexibility of implicit attitudes appears to entail, not immunity to sophisticated cognition, nor resistance to swift formation, but insensitivity to modification once formed. Asymmetric Attitude Malleability 12/8/2004 Page -3Easier Done Than Undone: The Asymmetric Malleability of Automatic Preferences Social psychologists have long noted that self-report measures of attitude, though useful and convenient tools, are vulnerable to a variety of validity-impairing biases including self-presentation (Schlenker, 1975; Baumeister, 1982), self-deception (Greenwald, 1988; Paulhus, 1993) and self-ignorance (Converse, 1970; Nisbett & Wilson, 1977; Wilson, Dunn, Kraft, & Lisle, 1989). As a consequence, alternative avenues of attitudinal assessment have been explored. One resourceful but unwieldy strategy has been to observe behavior covertly with a view to inferring the attitudes that underlie it (e.g., Crosby, Bromley, & Saxe, 1980). The more modern successor to this strategy has been to assess nondeclarative responses to attitude objects with the aid of specialized techniques (Bassili, 2001). Such techniques, collectively denominated implicit measures, have in common the property that they provide information about respondents’ automatic associations towards some object. Implicit measures come in a variety of shapes and sizes (Brauer, Wasel, & Niedenthal, 2000; De Houwer & Eelen, 1998; Greenwald & Banaji, 1995; Koole, Dijksterhuis, van Knippenberg, 2001; Vanman, Paul, Ito, & Miller, 1997; Hetts, Sakuma, & Pelham, 1999; Von Hippel, Sekaquaptewa, & Vargas, 1997) but most commonly take the form of compatibility tasks where targets, distractors, or responses vary with respect to one another in their degree of semantic or evaluative congruency (Bargh, Chaiken, Govender, & Pratto, 1992; De Houwer, 2003; Fazio, Sanbomatsu, Powell, & Kardes, 1986; Glaser & Banaji, 1999; Greenwald, Draine, & Adams, 1996; Greenwald, Klinger, & Liu, 1989; Greenwald, McGhee, & Schwartz, 1998; though see Koole & Pelham, 2003). The structure of such tasks is deliberately designed to make controlled responding difficult (Neely, 1977). In addition, their intent is often disguised to guard against the possibility of respondent reactivity (Fazio, Jackson, Dunton, & Williams, 1995). Implicit measures of attitude yield interpretable and replicable results (Banaji, 2001; Dovidio & Kawakami, 2001; Gemar, Segal, Sagrati, & Kennedy, 2001; Teachman, Gregg, & Woody, 2001). For example, the vast majority of people show an automatic preference for stimuli that connote or denote the self (Greenwald & Farnham, 2000), a finding that fits with converging evidence for the potency of the self-enhancement motive (Sedikides & Gregg, 2003). Such findings constitute prima facie evidence that implicit measures reflect something meaningful. However, exactly such measures capture remains Asymmetric Attitude Malleability 12/8/2004 Page -4elusive. On the one hand, the habitually low correlation between explicit and implicit measures of attitude (e.g., Blair, 2001) suggests that implicit measures possess substantial discriminant validity, perhaps stemming from their immunity to self-presentation (Kim, 2003; Steffens, 2004). On the other hand, the same low correlation can be read as implying that implicit measures are psychometrically weak, a claim reinforced by the fact that they sometimes fail to predict theoretically expected outcomes, or even to correlate with one another (Brauer et al., 2000; Bosson, Swann, & Pennebaker, 2000). Reassuringly, however, more recent research has found that implicit measures of attitude do cohere when measurement error is controlled for, and load on a latent factor that is distinct from that on which explicit measures of attitude load (although the two latent factors correlate positively; Cunningham, Nezlek, & Banaji, in press; Cunningham, Preacher, & Banaji, 2001; Gawronski, 2002). Moreover, implicit measures of attitude, and related constructs, have proven capable of predicting a range of phenomena (Egloff & Schmuckle, 2002; Greenwald, Nosek, & Banaji, 2003; Jellison, McConnell, & Gabriel, 2004; Koole & Pelham, 2001; Maison, Greenwald, & Bruin, 2001; Marsh, Johnson, & Scott-Sheldon, 2001; Nosek, Banaji, & Greenwald, 2002a; Teachman & Woody, 2003) including spontaneous behavior not predicted by explicit measures (Asendorpf, Banse, & Mucke, 2002; Perugini, 2004; Dovidio, Kawakami, Johnson, Johnson, Howard, 1999; Dovidio, Kawakami, & Gaertner, 2002; Fazio et al., 1995; McConnell & Leibold, 2001; Neumann, Hülsenbeck & Seibt, 2004; Spalding & Hardin, 1999). Indeed, meta-analysis reveals that the overall predictive power of implicit measures rivals that of explicit measures, and actually outstrips them in the domain of stereotyping and prejudice (Poehlman,Uhlmann, Greenwald, & Banaji, 2004). But although such demonstrations provide reassurance about the predictive validity of implicit attitudes, they do not directly address the question of what they assess. To answer that question fully, we must first determine what leads to the creation of automatic attitudes in the first place and what thereafter causes them to change—that is, we must gain an understanding of their antecedents. Such an enterprise is highly characteristic of science in general and of research on self-reported attitudes in particular (cf. Petty & Krosnick, 1995). Antecedents of Automatic Attitudes According to the classic dissociation model (Devine, 1989), automatic attitudes towards social groups form inexorably over time. In virtue of belonging to a culture, people Asymmetric Attitude Malleability 12/8/2004 Page -5cannot help being exposed to information, in the media and elsewhere, that links different social groups to positive and negative attributes. Because these links are repeatedly and chronically activated, their activation eventually becomes automatic for all members of that culture, requiring only that the social groups in question, or some representation of them, be perceived. If people later come to disagree, on principled grounds, with how a given social group is portrayed in their culture, they must inhibit the automatic associations that they have passively acquired during socialization and activate in their place the enlightened insights they have latterly arrived at. What distinguishes egalitarians from bigots, on this view, is not their automatic social attitudes, which are collectively shared and deeply ingrained, but rather their conscious social attitudes, which are individually entertained and readily malleable. Subsequent empirical research has called into question some of the more provocative predictions of the original dissociation model (Blair, 2002). Automatic attitudes are not activated in everyone invariably after all: variables such as attention (Castelli, Macrae, Zogmaister, & Arcuri, 2004; Gilbert & Hixon, 1991; Macrae, Bodenhausen, Milne, Thorn, & Castelli, 1997) and motivation (Devine, Plant, Amodio, Harmon-Jones, & Vance, 2002; Spencer, Fein,Wolfe, Fong, & Dunn, 1997; Lepore & Brown, 1997; Sinclair & Kunda, 1999; Wittenbrink, Judd, & Park, 1997) have also been found to play moderating and mediating roles. Nevertheless, one key prediction of the dissociation model remains potentially viable, namely, that automatic attitudes will generally show a greater resistance to change than their conscious counterparts. In particular, the dissociation model proposes that people will become egalitarian in their professed ideals before they become egalitarian in their spontaneous sympathies, a proposition consistent with much cross-sectional data on racial attitudes that have been assessed both explicitly and implicitly (Dovidio & Gaertner, 1998; Gaertner & Dovidio, 1986; Fazio, 2001). Moreover, the original dissociation model is joined by several cognate dual-process theories from which similar predictions about the relative malleability of automatic and self-reported attitudes could be derived (Chaiken and Trope, 2000; Sloman, 2002). Smith & DeCoster (1999) have postulated the existence of two complementary representational systems: a rule-based one, in which sudden transformations of serial representations (or symbols) occur, and an associative one, in which gradual transformations of connectionist Asymmetric Attitude Malleability 12/8/2004 Page -6representations (or weights) occur. In addition, Epstein & Pacini (1999) contend, as part of their cognitive-experiential self-theory, that the mind contains both a rational system, characterized by relative flexibility, and an experiential system, characterized by relative inertia. Finally, Wilson, Lindsey, & Schooler (2000) have drawn a distinction between attitudes that are consciously constructed online and attitudes that take the form of more enduring dispositions. Thus, there is ample theoretical precedent to postulate that automatic attitudes might be more stable than their self-reported counterparts. What is more, there is an array of supportive empirical evidence. Empirical Evidence for the Stability of Automatic Attitudes To begin with, automatic attitudes, unlike their self-reported counterparts, resist attempts at deliberate manipulation. Kim (2003) found that White participants blatantly instructed not to show automatic preferences while performing an Implicit Association Test (IAT; Greenwald et al., 1998) persisted in showing them. Moreover, naïve respondents tend not to spontaneously discover effective “faking” strategies (Banse, Seise, & Zerbes, 2001; Egloff & Schmuckle, 2002; Foroni & Mayr, in press; though see Lowery, Hardin, & Sinclair, 2001) even if they could in principle deploy them with the benefit of instruction or hindsight (Blair & Banaji, 1996; Steffens, 2004). There is also direct evidence that automatic attitudes, as originally hypothesized, reflect the “introspectively unidentified...traces of past experience” (Greenwald & Banaji, 1995, p. 8). In an elegantly designed study, Petty and Jarvis (1998) began by classically conditioning a preference in participants for one photographed face over another. Then, to induce a further preference based on perceived similarity, they led participants to believe that they shared more opinions in common with one of the people allegedly depicted in the photographs (e.g., Eddie) than with the other (e.g., Phil). Several different groups of participants were created. In some, both attitude manipulations were designed to induce the same preference (e.g., both manipulations portrayed Eddie as preferable to Phil); in others, both manipulations were designed to produce contrary preferences (e.g., one manipulation portrayed Eddie as preferable to Phil, and the other, Phil as preferable to Eddie). At the end of the experiment, suitable comparisons revealed that explicit ratings of Eddie and Phil were influenced only by perceived similarity. Crucially, however, performance on an evaluative priming task (cf. Fazio et al., 1995) was influenced by the face preferences that have been classically conditioned. This shows that, despite superficial appearances, and additional explicit Asymmetric Attitude Malleability 12/8/2004 Page -7manipulations, earlier learning subtly persisted. Similar findings have been obtained using introspection as a manipulation and rapid responding as an implicit measure (Wilson, Lindsey, & Schooler, 2000). Further evidence for the robustness of automatic attitudes comes from recent work on cognitive dissonance (Festinger, 1957; Harmon-Jones & Mills, 1999). Standard counterattitudinal manipulations that reliably cause changes in selfreported attitudes have been found to leave implicit attitudes unaffected, consistent with the latter being relatively more stable than the former (Gawronski & Strack, 2003; though see McDell, Banaji, & Cooper, 2004). Field data broadly consistent with such laboratory findings have also emerged. Hetts et al. (1999) documented a stepwise rise in self-esteem among different generations of East-Asians immigrants to the United States. Crucially, this rise was manifested on explicit measures (self-report questionnaires) before being manifested on implicit measures (evaluative priming tasks). Finally, data from vast web surveys show the attitudes toward age (young versus old) and academic disciplines (math versus arts) are comparable across the lifespan (Nosek, Banaji, & Greenwald, 2002b). This strongly suggests that at least some automatic attitudes are generationally stable. Empirical Evidence for the Malleability of Automatic Attitudes In stark contrast to the above findings, there is also a wealth of empirical evidence that existing automatic attitudes, initially theorized to be virtually invariant, are in fact exquisitely malleable (for a review, see Blair, 2002). For example, Dasgupta and Greenwald (2001) found that White participants exposed to favorable exemplars of Black Americans and unfavorable exemplars of White Americans showed weaker automatic preferences for their own race than did control participants; overtly expressed racial preferences, however, did not vary. This effect, sustained over several days, and replicated in a different domain, indicates that automatic attitudes may be amenable to change even when their conscious counterparts are not. In addition, Lowery et al. (2001) found levels of automatic race prejudice can be moderated by even casual social encounters, depending on racial backgrounds of those involved, an effect that is itself moderated by respondents’ relative status (Richeson & Ambady, 2003), and that generalizes to other domains, dramatically on occasion (Banaji, 2002). Other investigators (Foroni & Mayr, in press) have found that normative automatic preferences for flowers over insects can be significantly curtailed simply by having participants read a story presenting a fictional rationale for entertaining contrary preferences (i.e., flowers become radioactive, and insects Asymmetric Attitude Malleability 12/8/2004 Page -8nutritious, on a post-apocalyptic earth). Blair, Ma, & Lenton (2001) established that automatic associations could be reduced even without exposure to any external sources of information. Participants who summoned up imagery at odds with gender stereotypes showed attenuated automatic biases relative to participants who summoned up neutral, stereotypical, or no mental imagery. This result held for several different implicit measures, thereby infirming alternative explanations invoking response suppression or shifting response criteria. Several additional studies have documented substantial shifts in automatic attitudes based on recent experiences, variations in context, or variations in physiological state (Gawronski, Walther, & Blank, 2004; Karpinski & Hilton, 2001; Lane, Mitchell, & Banaji, 2004; Mitchell, Nosek, & Banaji, 2003; Rothermund & Wentura, 2001; Seibt, Hafner, & Neumann, 2004). Finally, investigations that have teased apart internal consistency from temporal stability have found that underlying automatic attitudes showed marked variability from one measurement occasion to another (Gregg & Sedikides, 2004; Steffens & Buchner, 2003). The Present Research What, then, are we to make of this paradoxical picture? Some research suggests that, either in absolute terms or relative to self-reported attitudes, automatic attitudes are comparatively inflexible. However, other research, equally telling, suggests that they are comparatively supple. For the moment, therefore, it seems advisable to adopt a pragmatic perspectivist approach (Banaji, 2002; McGuire, 1973). That is to say, the substantial positive assertion that automatic attitudes are supple, or the substantial negative assertion that they are inflexible, are both likely to be true in some contexts but false in others, each in important and interesting ways. The ultimate resolution of the matter must await the development of an integrative theory that specifies the key moderators marking out the chief boundary conditions under which automatic and self-reported attitudes exhibit one characteristic or the other (cf. Petty, 1997). However, given that such an integrative theory is a distant prospect, there are good grounds for adopting in the interim a research orientation designed to generate a broad body of findings that can representatively inform any subsequent integrative theorizing. That is to say, a bottom-up inductive approach, furnishing the raw materials for future theoretical edifices, may prove as useful as a topdown deductive approach, checking the rudimentary theoretical edifices so far assembled. As Sherlock Holmes once remarked to his faithful companion, Dr. Watson, “It is a capital Asymmetric Attitude Malleability 12/8/2004 Page -9mistake to theorize before one has data. Insensibly one begins to twist facts to suit theories, instead of theories to suit facts (Doyle, 1981, p. 165).” Bearing this in mind, we embarked on a coordinated series of laboratory studies— theory-testing in intent, but also reflecting an openness to alternative formulations implied by the data—to investigate the ease with which automatic and self-reported preferences for one imagined social group over another might be induced or undone. We derived our provisional theoretical starting-point from the classic dual-process models reviewed above, which are consensually regarded as implying that automatic and self-reported attitudes are affected to different degrees by cognitive processes of greater or lesser sophistication. In our view, dual-process models can also be more narrowly interpreted as implying that, whereas both self-reported and automatic attitudes should be responsive to what might be termed concrete learning, only self-reported attitudes should be responsive to what might be termed abstract supposition. Let us elaborate. Concrete learning we define as the act of cognitively assimilating multiple pieces of information about the characteristics of an object, or alternatively, of assimilating the same piece of information multiple times. Thus, reading a detailed account of some object, or undergoing an intensive session of associative conditioning (De Houwer, Thomas, & Baeyens, 2001), both qualify as instances of concrete learning. Abstract supposition, in contrast, we define as the act of hypothetically assuming that an object possesses particular characteristics. Thus, entertaining the idea, out of the blue, that a novel object is X or ~X, or that an existing object known to be X is in fact ~X (or vice versa), both qualify as instances of abstract supposition. The critical difference between the two is that, in the case of concrete learning, the characteristics of an object are implied by, or inferred from, an elaborate set or protracted series of prior instances, whereas in the case of abstract supposition, no corresponding set or series is available. In other words, the act of abstractly supposing that some state of affairs is the case involves entertaining cognitions that are purely formal and symbolic. As such, abstract supposition should activate explicit representations only (i.e., those that are “rule-based”, “rational”, and “constructed”) and leave implicit representations entirely unaffected (i.e., those that are “association-based”, “experiential”, and “dispositional”). Thus, if classic dual-process models are correct, then, by implication, the effects of explicit representations ought to register on explicit measures of attitude alone, but not on implicit measures of attitude. Asymmetric Attitude Malleability 12/8/2004 Page -10One merit of drawing this distinction between concrete learning and abstract supposition is that it permits us to investigate whether Devine’s (1989) dissociation model—in particular its prediction that automatic attitudes will generally show a greater resistance to change than their conscious counterparts—still holds in a qualified sense, specifically, under conditions where people are engaged in abstract supposition (capable of occasioning changes on explicit measures only), but not under conditions where people are engaged in concrete learning (capable of occasioning changes on explicit and implicit measures alike). One reason for suspecting that this might be the case is that egalitarian principles typically take the form of prescriptions about how one should think or feel. The “should” implies that one does not yet conform to an ideal that one consciously endorses. But, logically speaking, to consciously endorse an ideal, one must first be able to envisage it as a hypothetical state of affairs—with the aid of some suitably symbolic thought. If dualprocess models are correct, then it plausibly follows that cognitive acts of this ilk are powerless to shift more deeply ingrained biases in more primitive parts of the mind. Of course, the distinction between abstract cognition and concrete learning may not be hard and fast. In supposing that some hypothetical state of affairs obtains, people may naturally drift into imagination, which involves recruiting multiple instances from memory, together with all their affect-laden connotations. However, many distinctions retain their utility without being hard and fast: the existence of twilight does not mean that the distinction between night and day is bankrupt. Abstract supposition, insofar it is unaccompanied by vivid imagery, has the virtue of possessing the property of explicitness in particularly pure form, and as such, it is the construct of choice to deploy in determining the true sensitivity of automatic attitudes to mental manipulation. Across four experiments, then, we attempted to induce (Experiments 1 and 2), and thence to undo (Experiment 3 and 4), preferences for one imagined social group over another. In each study, participants were randomly assigned to conditions in which they abstractly supposed, or concretely learned, in a variety of ways, that two social groups, and their respective members, had attributes of contrasting valence. The relative impact of these manipulations on self-reported and automatic attitudes was subsequently assessed using both rating scales and the IAT so that inferences about their relative stability or malleability could be duly drawn. Asymmetric Attitude Malleability 12/8/2004 Page -11In using imagined social groups rather than real ones, our research can be regarded as exemplifying the recent trend towards flexibly exploring the dynamics of attitude formation and change experimentally using novel attitude objects (De Houwer et al., 2001; Fazio, Eiser, & Shook, 2004; Glaser, 1999; Greenwald, Pickerell, & Farnham, 2002; Mitchell, 2004; Mitchell, Anderson, & Lovibond, 2003; Olsen & Fazio, 2002). With regard to the current research question, our laboratory-based methodology has, despite its spartan appearance, a number of distinct advantages. First, by using previously unfamiliar stimuli as attitude objects, the scope for new attitudes to take root is maximized (Cacioppo, Marshall-Goodell, Tassinary, & Petty, 1992). Second, by retaining control of the attitudeinduction process, greater confidence is afforded that participants will construe attitude objects as intended, and that those objects will correspond for self-reported and automatic attitudes (Fishbein & Ajzen, 1975). Third, by not focusing on real-life social groups, selfpresentational biases that might otherwise complicate interpretation can be averted (Plant & Devine, 1998). We were aware, of course, that by using imagined social groups we ran the risk of compromising the realism of our investigation. To address this risk, we therefore included (in three of our four experiments) self-reported measures of attitude meaningfulness, and as a precaution redid all pertinent analyses using only those participants scoring high thereon. Experiment 1 Our goal in Experiment 1 was to test the hypothesis that we derived from dualprocess models, namely that whereas concrete learning would induce both automatic and self-reported preferences, abstract supposition would induce self-reported preferences only. Overview The experiment featured three conditions. In one condition (Suppose) participants engaged in abstract supposition: They hypothetically assumed that one social group possessed positive traits and another one negative traits. In the two remaining conditions participants engaged in concrete learning. They either (a) read a graphic story in which one group exemplified positive traits while another exemplified negative traits (Narrative) or (b) repeatedly rehearsed positive trait associations towards one group and negative trait associations towards another (Rehearsal). We deliberately operationalized concrete learning in different ways in order to achieve wider conceptual coverage of the construct. Extracting pieces of information from a text (Narrative) arguably qualifies as a more Asymmetric Attitude Malleability 12/8/2004 Page -12sophisticated form of learning than does rote rehearsal of paired associates (Rehearsal). We were therefore open to the possibility that the former might be relatively superior at inducing self-reported preferences and the latter relatively superior at inducing automatic preferences. However, our main hypothesis was that abstract supposition, although sufficing to induce self-reported preferences, would fail to induce automatic preferences, unlike either form of concrete learning, which would induce both types of preference. Participants Forty-six students from St. Anne’s Convent School in Southampton, UK, participated in this study in return for funds allocated to the refurbishment of their recreational quarters. All participants were female, studying for A-levels, and aged between