He Did What?: The role of diagnosticity in revising implicit evaluations

Abstract

Research suggests that implicit evaluations are relatively insensitive to single instances of new, countervailing information that contradicts prior learning. In six experiments, however, we identify the critical role of the perceived diagnosticity of that new information: counter-attitudinal information that is deemed highly diagnostic of the target's true nature leads to a complete reversal of the previous implicit evaluation. Experiments 1a and 1b establish this effect by showing that newly-formed implicit evaluations are reversed minutes later with exposure to a single piece of highly diagnostic information. Experiment 2 demonstrates a valence asymmetry in participants’ likelihood of exhibiting rapid reversals of newly formed positive versus negative implicit evaluations. Experiment 3 provides evidence that a target must be personally responsible for the counter-attitudinal behavior and not merely incidentally associated with a negative act. Experiment 4 shows that participants exhibit revision only when they judge the target's counter-attitudinal behavior as offensive and thus diagnostic of his character. Experiment 5 demonstrates the behavioral implications of newly-revised implicit evaluations. These studies show that newly-formed implicit evaluations can be completely overturned through deliberative considerations about a single piece of counter-attitudinal information.

Throughout much of their over 30 years of marriage, Barbara Kuklinski would have described her life with her husband, Richard, in mostly idyllic terms. “We had what seemed to be the perfect life,” she said. “We were the All-American family.” It was thus quite a shock to Barbara and her children, on the day that Richard was arrested, when she abruptly discovered that he had been, over his 30-year tenure as a hitman for a number of mafia families in New Jersey, one of the most prolific contract killers in American history. The man whom she had known since she was 19 years old, whom she described as a romantic that would often leave flowers for her, and with whom she had raised 3 children, was rather suddenly revealed to be a cold-blooded killer. He would later boast in interviews that he had been responsible for over 100 murders. Outside of what one detective described as a “normal, classic family existence,” Kuklinski had earned himself the nickname “Iceman”—a reference to his technique of freezing his victims in an industrial-sized freezer to obscure the time of death. After his conviction, Kuklinski would spend the rest of his life in a maximum security prison, until his death in 2006 (Ginsburg, 2001).

When faced, like Barbara Kuklinski, with a revelation that is strikingly inconsistent with our well-learned and long-held beliefs about someone, to what extent can this single new piece of information change our evaluations of someone? Contemporary research on evaluations suggests that minimal amounts of counter-attitudinal information of this sort can quickly influence our evaluations in some ways but not others. Although explicit, self-reported evaluations can rapidly and readily incorporate revelations that overturn previous learning, some theory and evidence suggest that implicit (i.e., unintentional, spontaneous) evaluations are relatively less sensitive to such information (e.g., Gregg, Seibt & Banaji, 2006; Rydell & McConnell, 2006; Rydell et al., 2006, 2007).

The claim that implicit evaluations are resistant to rapid revision is important because, in addition to suggesting that the processes underlying implicit versus explicit evaluations may differ (as we discuss in more detail below), it also has potential implications for our behavior. Indeed, implicit evaluations have been shown to uniquely predict our behavior in ways that explicit evaluations do not (Cameron, Brown-Iannuzzi, & Payne, 2012; Galdi, Arcuri, & Gawronski, 2008; Ferguson, 2007; Greenwald et al., 2009; McNulty et al., 2013; Perugini, Richetin, & Zogmaister, 2010; Towles-Schwen & Fazio, 2006; cf. Oswald et al., 2013). If implicit evaluations cannot be readily updated to accommodate a newly learned truth, they may ultimately guide us in ways that are inconsistent with our beliefs, resulting in potentially maladaptive behavioral outcomes.

In the present research, however, we show that implicit evaluations can, in fact, be highly responsive to single pieces of countervailing information, particularly when this information is extremely negative. We argue and provide evidence to suggest that this occurs because extreme negative information is judged as highly diagnostic—that is, especially revealing of a person's true nature or character. Such information, we propose, can rapidly and fully reverse an otherwise well-learned and evaluatively uniform implicit evaluation towards an individual.

In what follows, we first review evidence that implicit evaluations are relatively insensitive to single instances of countervailing behavior and discuss how contemporary theories explain such insensitivity. Next, we consider the possibility that implicit evaluations may be more responsive to minimal amounts of countervailing information than current findings suggest, identifying the extremity and diagnosticity of new information as a factor that predicts strong revision.

Evidence of insensitivity to single pieces of countervailing evidence

Several lines of work show that implicit evaluations become relatively insensitive to new information once established. First, some research suggests that implicit evaluations are relatively unresponsive to the negation of previously learned associations. In one study often cited as evidence of this assertion (see, e.g., Gawronski & Sritharan, 2010, p. 227), Gregg et al. (2006) had participants learn about two fictitious groups that strongly differed in their character—one group's actions were highly positive while the other's were highly negative. Following this, participants’ implicit evaluations of the groups were consistent with the induction. Next, Gregg and colleagues sought to “undo” this learning in various ways by telling participants reasons that the groups had changed character. These attempts ranged from assertions that the groups had changed to long and detailed narratives about how the groups had changed. Yet, none of these strategies successfully altered implicit evaluations (despite the fact that they readily shifted explicit evaluations).

Second, implicit evaluations seem to respond slowly and gradually to new information, such that any single piece of information does not appear to exert a potent influence. For example, Rydell and colleagues (2007) gave participants extensive exposure to evaluatively-consistent behavioral statements (100 trials) about a target individual named Bob (e.g., “Bob donates his time at the soup kitchen”). Afterwards, participants were given varying amounts of counter-attitudinal information (0, 20, 40, 60, 80, or 100 trials), followed by measures of their implicit and explicit evaluations of Bob. Whereas explicit evaluations were found to be highly responsive to the counter-attitudinal information—exhibiting revision after just 20 trials—implicit evaluations were more resistant, responding incrementally and linearly to the amount of counter-attitudinal information to which participants were exposed. It was not until participants had been exposed to 100 trials of counter-conditioning that there was unequivocal evidence that implicit evaluations reflected the counter-attitudinal information rather than the initial learning.

Some readers may be surprised by these results, given that implicit evaluations seem to shift rather easily in response to subtle contextual manipulations, such as being reminded of positive exemplars of an otherwise stigmatized group (e.g., Dasgupta & Greenwald, 2001; Ferguson & Bargh, 2004; Wittenbrink, Judd & Park, 2001). However, there is an important distinction between effects of contextualization, and changes of the sort we consider here. Contextualization reflects the activation of potentially well-established mental representations that were learned previously in different contexts. Because it is possible that these previously learned evaluations required a great deal of evaluative learning within a given context to develop, context dependence does not necessarily reveal anything about whether people can update their evaluation of a target based on learning new information (see also Fazio, 2007; Ferguson & Fukukura, 2012; Gregg et al., 2006, p. 16).

Why might implicit evaluations be resistant to minimal counter-evidence?

These lines of work indicate that implicit evaluations are less responsive to new information than explicit evaluations. This means that a revelation such as learning that one's husband is actually a contract killer for the mob could be expected to immediately alter one's self-reported evaluation, but potentially have little effect on one's implicit evaluation. Why might this be? Whereas some theories have argued that such insensitivity is in line with assumptions about the operation of implicit and explicit evaluations, others argue that implicit evaluations may be more alterable than current evidence suggests.

Dual-process models assert that such findings emerge because implicit and explicit evaluations are generated by two qualitatively distinct mental processes or systems (e.g., Bassili & Brown, 2005; Conrey & Smith, 2007; Rydell & McConnell, 2006; Rydell et al., 2006, 2007; Smith & DeCoster, 2000; Strack & Deutsch, 2004; Wilson, Lindsey, & Schooler, 2000; Gawronski & Bodenhausen, 2006, 2011). Implicit evaluations, on the one hand, are thought to emerge from associative mental processes that operate through the spreading activation of associations in memory. These mental associations typically develop and strengthen over time through repeated co-occurrence between an object and positively or negatively valenced experiences that occur in close spatial/temporal contiguity or that possess features that cue similarity between an object and other representations in memory. Importantly, associative processing is thought to operate independently of the perceived truth of the experiences one has with the object (Gawronski & Bodenhausen, 2006, 2011, 2014; Rydell & McConnell, 2006), meaning that one can hold an implicit evaluation completely at odds with one's explicit evaluation.

Explicit evaluations, on the other hand, are assumed to be governed by propositional processes that operate on the basis of rule-based logic, inference, and reasoning (Gawronski & Bodenhausen, 2006, 2011; Rydell & McConnell, 2006; Sloman, 1996; Strack & Deutsch, 2004). Propositional processes can operate abstractly and symbolically, giving rise to deliberative evaluations that may be either consistent or inconsistent with the spontaneous affective reactions that are the products of associative learning. In contrast to the associative mental processes underlying implicit evaluation, rule-based processes are capable of taking account of the validity of information and can thus negate past experiences if the consideration of current propositions warrants it. Thus, in contrast to implicit evaluation, explicit evaluation can more strongly and immediately reflect new information, even if this information is not the product of direct experience (“Did you hear what Jason did at the party Saturday?”) and even if it is inconsistent with one's prior beliefs (“Oh, I had thought that Elaine was a nice person; I can’t believe she did that!”).

Some dual process models, relying on such assumptions, assert that implicit evaluations are insensitive to counterevidence once they are established (Rydell & McConnell, 2006; Rydell et al., 2006, 2007). For example, Rydell and McConnell's Systems of Evaluation Model (SEM) posits that implicit evaluations rely on a slow-learning associative system that is largely (but perhaps not completely; see McConnell & Rydell, 2014) incapable of rapidly incorporating new information about a target, particularly after a lot of prior information has already been learned.

Other models allow for greater interaction between the two processes. For example, according to Gawronski and Bodenhausen's (2006, 2011, 2014) Associative—Propositional Evaluation (APE) model, the key distinction between implicit and explicit evaluation is whether implicit evaluations can incorporate validity information, particularly information that negates prior learning. Although new associations can be created that affirm the opposite of a previously-held evaluation, once an association has been created, it cannot be negated by new information that challenges its validity (e.g., “That information was false.”). Thus, counterevidence may be effective in changing implicit evaluations only when it affirms the opposite rather than negates prior learning (cf. Peters & Gawronski, 2011).

Empirical evidence and theory consistent with rapid revision

Although there is evidence suggesting that implicit evaluations are relatively unresponsive to single pieces of countervailing information, some research suggests that information can be rapidly incorporated into implicit evaluations. First, implicit evaluations can form towards novel targets extremely rapidly (Ashburn-Nardo, Voils & Monteith, 2001; Castelli, Zogmaister, Smith, & Arcuri, 2004; De Houwer, 2006; Gregg, Seibt, & Banaji, 2006; Otten & Wentura, 1999; Whitfield & Jordan, 2009)— sometimes in response to only a single piece of information about a target. For example, Gregg et al. (2006) found that when participants were first asked to merely suppose that novel, fictional groups possessed good or bad traits, they immediately exhibited implicit positivity toward the good versus bad group, suggesting that implicit evaluations were rather easy to create towards novel targets. Similarly, De Houwer (2006) found that when he informed participants that nonsense syllables would, later in the experiment, signal the immediate presentation of a positive or negative photo, they exhibited implicit evaluations in line with this information even before the pairings had occurred. If implicit evaluations can form so rapidly towards novel stimuli, as these studies demonstrate, it may also be that they can be revised rather quickly under certain circumstances.

Second, there is evidence that implicit evaluations can be at least somewhat modified by small amounts of behavioral information. For example, in one study, Whitfield & Jordan (2009; Study 3) had participants undergo an evaluative conditioning procedure, and then exposed them to 13 behavioral statements that were in the opposing direction from the conditioning. They found that participants’ relative implicit preference for one target over the other was responsive to these statements (though it was unclear from their analysis whether these 13 statements were enough to lead to full reversals of their previously-held preferences). Third, recent evidence also suggests that implicit evaluations may be responsive to negations under some circumstances. When Peters & Gawronski (2011; see also Boucher & Rydell, 2012) exposed participants to information about a novel target (e.g., “Mike lent money to a friend in financial trouble”) that was sometimes immediately revealed to be false, participants’ implicit evaluations did, in fact, reflect the validity of the learned information—a finding that was inconsistent with Peters and Gawronski's expectations and with a number of current dual process models. Interestingly, however, when participants were notified of the truth or falsity of the behavioral information after a short delay, their implicit evaluations were less sensitive to this information (see also Foerde & Shohamy, 2011). Though it is unclear why implicit evaluations were responsive to the truth of the behavioral statements in one case but not the other, these results nonetheless raise the possibility that there are circumstances under which countervailing information can influence implicit evaluation formation and change.

Some contemporary models of implicit and explicit evaluation are consistent with this growing body of evidence demonstrating that implicit evaluations can sometimes exhibit responsiveness to minimal counter-attitudinal information. For example, De Houwer and colleagues’ single-process model rejects the existence of a (slow-learning) associative system altogether, positing instead that implicit and explicit evaluations both rely on the formation of propositions about a stimulus. On this view, once a proposition has been formed in memory, it can then be retrieved automatically (e.g., Bar-Anan, De Houwer, & Nosek, 2010; Hughes, Barnes-Holmes, & De Houwer, 2011; see also De Houwer, in press), and because propositional processes are thought to operate rather quickly, implicit evaluations may thus also change rather quickly in response to new propositions.

As another example, the memory systems model (Amodio, 2014; Amodio & Ratner, 2011; see also Amodio & Devine, 2006) posits that, rather than implicit and explicit evaluations relying on a single process, implicit social cognition more generally is instead the product of multiple interacting systems that each have their own properties of acquisition and extinction. For instance, whereas (implicit) semantic learning of the sort that occurs when one learns to pair doctor with nurse may proceed slowly and extinguish only gradually, Pavlovian fear conditioning in animal models develops rapidly (e.g., Hermer-Vazquez et al., 2005; Yin & Knowlton, 2006), including during single trial episodes (e.g., see Cahill & McGaugh, 1990; Hilliard, Nguyen, & Domjan, 1997). The developmental and learning trajectories of implicit evaluations should be influenced by the type(s) of memory system(s) involved in the acquisition of new information (see also Ferguson, Mann, & Wojnowicz, 2014). If information is acquired through a (slower-learning) semantic memory process, it may develop and change rather slowly; however, if some types of implicit learning proceed quite rapidly, they may be updated in response to minimal amounts of information.

There is also a burgeoning literature over the last decade suggesting that implicit evaluation measures such as the Implicit Association Test (IAT; Greenwald, McGhee, & Schwartz, 1998; Greenwald et al., 2009) or Affect Misattribution Procedure (AMP; Payne et al., 2005) are unlikely to reflect just one type of process, whether associative or propositional. Sherman, Payne, and colleagues (e.g., Bishara & Payne, 2009; Conrey, Sherman, Gawronski, Hugenberg, & Groom, 2005; Sherman, 2006) have provided evidence that multiple processes underlie implicit evaluation measures—including processes that have been characterized as controlled in nature (Sherman, Klauer, & Allen, 2010). This work implies that implicit evaluations may not be under the sole purview of associative processes (see also Mitchell, De Houwer, & Lovibond, 2009), and it is up to researchers to identify and catalogue the features of the conditions under which implicit evaluations form and change in an attempt to better understand these processes. The present work is one attempt to this end.

The role of extremity and diagnosticity

Our goal in the current studies was to identify one type of minimal, countervailing information that could lead to rapid revision of implicit evaluations—specifically, the extent to which new information is judged as especially diagnostic of the person's true nature. In line with prior research, in order to signal such diagnosticity, we presented behaviors that were extreme in valence (e.g., Fiske, 1980). We examine the effects of single instances of both extreme positive and extreme negative counter-attitudinal verbal information. However, we predict and find evidence that extreme negative information is especially potent—a proposal in line with classic work on person perception suggesting that extreme negativity is particularly diagnostic of the person's disposition and likely future behavior (e.g., Knobe, 2003; 2006; Nadelhoffer, 2006; Malle & Knobe, 1997; Mende-Siedlecki, Baron, & Todorov, 2013; Reeder & Brewer, 1979; Reeder & Coovert, 1986; Reeder, Pryor, & Wojciszke, 1992; Trafimow & Schneider, 1994; Trafimow & Trafimow, 1999). In other words, we argue that extremely negative behavior is likely to overturn previously positive implicit evaluations because such behavior is often perceived as especially diagnostic of the person's true character.

Consider, for example, the revelation that someone has been convicted of a serious crime, not unlike the situation faced by Barbara Kuklinski. To the degree that one sees this revelation as revealing of the kind of person someone truly is, one may see a great deal of informational value in this single revelation, thus imbuing it with greater significance and ascribing it greater weight in one's evaluation of the target.

If the processes that govern implicit evaluation are entirely passive and responsive primarily to the amount of counter-attitudinal information such that each piece of information is given approximately equal weighting, we should expect that a single piece of information—no matter how relevant—ought to influence a well-established implicit response only mildly, if at all. However, if one's assessment of the diagnosticity of each instance of behavior matters, then even a single piece of highly diagnostic information ought to rapidly “undo” one's prior implicit tendencies.

The current studies

Across six studies, we tested whether participants’ implicit evaluations of a novel target were responsive to single, extreme, highly diagnostic verbal descriptions of behavior. To test this hypothesis, all of the experiments we report have a similar format. We first had participants undergo an extensive learning paradigm that consisted of 100 repeated co-occurrences of a target (referred to as “Bob” throughout the experiment) in conjunction with evaluatively consistent behavioral information—either positivity or negativity (e.g., Gawronski et al., 2010; Rydell & McConnell, 2006, Rydell et al., 2006, 2007; Rydell & Gawronski, 2009). We then exposed participants to their 101^st instance of Bob's behavior; however, unlike the previous 100 behaviors, this one was pre-selected to be: (a) inconsistent with the previous 100 instances of behavior, and (b) particularly extreme in valence, and thus especially high in perceived diagnosticity. We measured participants’ evaluations towards the target before and after they were exposed to this single new piece of information, thus assessing whether their evaluations incorporated the new information. Importantly, throughout our experiments, we use novel objects, and thus have control over participants’ entire evaluative history with the object, reducing the likelihood that contextualization can explain the changes in task performance we observe across our studies (see Fazio, 2007; Ferguson & Fukukura, 2012; Gregg et al., 2006).

In Experiments 1a and 1b, we demonstrate that participants’ implicit evaluations of Bob are sensitive to (two different) extreme negative behaviors. Experiment 2 extends our analysis by examining whether there is a valence asymmetry in participants’ readiness to update their impressions in light of extremely positive versus extremely negative information. In Experiment 3, we show, however, that not all extremely negative behaviors are rapidly incorporated into participants’ implicit responses. Rather, they must be clearly attributable to the target. To this end, we exposed participants to two similarly extreme behaviors but implicate either the target, or, alternatively, an individual who is merely incidentally associated with the target. We show revision in the former but not the latter case. In Experiment 4, we test our proposed mechanism of diagnosticity, providing evidence consistent with a mediational pathway in which those participants who judge an act to be extremely offensive are more inclined to view it as highly diagnostic of a person's true character, and to then show the most evaluative updating. Finally, in Experiment 5, we test the behavioral implications of revised implicit evaluations.

Following Simmons, Nelson, and Simonsohn's (2012) recommendation, we report how we determined our sample size, all data exclusions (if any), all manipulations, and all measures in each of the six studies reported. With respect to sample size, in our first study we attempted to recruit approximately 30 subjects per cell. We then used the effect size in this Experiment (1a) to determine that a sample size of n=90 per cell would give us sufficient statistical power to detect a difference of equivalent size 90% of the time. We thus recruited n=90 per cell for all remaining studies except for Experiment 4 (in which we test a mediational pathway and thus increased our intended sample size to the pre-determined N=240) and Experiment 5 (in which we test correlations with our implicit measures and thus increased our intended sample size to N=300).

Experiment 1a: Who is Bob, really?

Method

Participants

Fifty-seven undergraduates participated in exchange for credit or $5. The responses from two participants were excluded from analyses due to a lack of variance (that is, failing to follow instructions by pressing the same response key on all trials; see Payne et al., 2005, for similar exclusion criteria) on one or both implicit measures (described below). One participant pressed invalid response keys on an excessive number of trials (>10%) on both implicit measures, and was thus also excluded from analyses. This left a final sample of 54 participants.

Procedure

Participants completed a learning paradigm in which they read a total of 100 behavioral statements about a target individual named Bob (e.g., “Bob gave a hitchhiker a ride to a shelter”), identified by a picture presented at the top of the screen on each trial (see Rydell et al., 2007; Rydell & McConnell, 2006)¹. Participants’ task was to assess whether they thought each target behavior was characteristic or uncharacteristic of Bob by pressing the c key (characteristic) or u key (uncharacteristic). They received immediate feedback on their response, consisting of either the word “correct” printed in blue text, or “incorrect” printed in red text, followed below it by a summary of the meaning of the feedback (e.g., “Giving a hitchhiker a ride to a shelter is characteristic of Bob”). To ensure that participants formed an initially positive and well-rehearsed evaluation towards Bob, all 50 positive behavioral statements were said to be characteristic of Bob, while all 50 negative statements were said to uncharacteristic of him. The order of the behavioral statements was randomly determined for each participant. These statements have been used previously in attitude formation research, and are moderate in valence.

Assessing implicit and explicit evaluations

After the 100 learning trials, participants’ implicit evaluations of Bob were assessed using an Affect Misattribution Procedure (AMP; see Payne et al., 2005). Each of the 60 trials in the measure consisted of: (a) a prime image (75 ms) of either Bob (30 trials) or one of 6 unfamiliar college-aged white males (5 trials each), (b) a Chinese pictograph (100 ms), and, (c) a backward mask (random black and white noise), which remained on the screen until participants provided a response. Participants’ task was to indicate whether the Chinese pictograph was more or less pleasant than average. Following Payne et al.'s (2005) procedure, participants were warned that having just seen a positive or negative image immediately before the pictograph can sometimes bias their evaluation, and that we were interested in their ability to ignore such influences and provide an objective assessment of each pictograph. Previous research (e.g., Payne et al., 2005) has shown that people misattribute their automatic evaluation of the prime to their feelings about the pictograph, thus providing a measure of people's spontaneous and unintentional evaluations of the primes. Although some recent evidence by Bar-Anan and Nosek (2012) raised the possibility that evaluations measured by the AMP may be the result of participants intentionally evaluating the primes (rather than unintentionally rating the Chinese pictographs), more recent work by Payne et al. (2013) has shown that implicit evaluations as measured by the AMP do seem to reflect unintentional responses to the prime stimuli and that participants’ reported intentional evaluations of the primes appear to be post-hoc confabulations.

Changing impressions of Bob

To assess the effects of extreme countervailing information on participants’ implicit responses, participants were next told that they would learn a single new piece of information about Bob and that they should pay close attention because this new piece of information was more recent and could potentially violate the impression they had already formed. Participants in the experimental condition read the statement, “Bob was recently convicted of molesting children.” Participants in the control condition read the statement, “Bob recently bought a soda.” In both conditions, immediately after the statement, participants were reminded that this piece of information was characteristic of Bob and that they would be tested for their memory of it later in the experiment.

Participants then completed a Time 2 implicit evaluation measure, following the same protocol as the Time 1 measure except that a different set of 60 Chinese pictographs were presented during the AMP (to control for familiarity and mere exposure). Next, participants completed an explicit evaluation measure in which they judged Bob's likeability from 1 (very unlikeable) to 9 (very likeable) as well as how bad—good, mean—pleasant, disagreeable—agreeable, uncaring—caring, and cruel—kind they considered him to be, all on 9-point Likert scales (see Rydell et al., 2006, 2007). Because these items were highly inter-correlated (α=.912), they were averaged into a single composite explicit evaluation measure.

Finally, participants completed a short demographics questionnaire and were debriefed, thanked, and dismissed.

Results

To create a measure of participants’ implicit evaluations of Bob, we calculated, for each AMP, the proportion of times participants indicated that the target pictograph was more pleasant than average separately for Bob trials and neutral trials. We submitted these proportion-pleasant judgments to a 2 (time: 1 or 2) × 2 (condition: experimental or control) × 2 (prime: Bob or neutral) mixed-model analysis of variance (ANOVA) with time and prime as within-subjects factors and condition as a between-subjects factor. This analysis yielded the predicted 3-way interaction, F(1, 53)=12.351, p<.01, ɳ_p²=.189 (see Figure 1).

Figure 1.

Participants’ proportion-pleasant judgments towards Bob before (Time 1) and after (Time 2) learning either that he was convicted of molesting children (experimental) or that he bought a soda (control) in Experiment 1a.

Decomposing this interaction, in the experimental condition, there was the predicted significant interaction between time and prime, F(1,27)=8.216, p<.01, such that participants exhibited greater implicit positivity towards Bob (M=.56, SD=.28) than the unfamiliar faces (M=.49, SD=.21) at Time 1 (though this difference failed to reach significance in this experiment, F(1,27)=1.327, ns, in all subsequent experiments we find reliable evidence of formation at Time 1). At Time 2, as predicted, there was a complete reversal of the pattern at Time 1, such that participants exhibited significantly more implicit negativity towards Bob (M=.39, SD=.22) than towards the neutral targets (M=.55, SD=.21), F(1,27)=8.384, p<.01.

In the control condition, there was a predicted main effect of prime,F(1,26)=14.267, p<.001, such that participants exhibited greater implicit positivity towards Bob at both Time 1 (Bob: M=.57, SD=.23; Neutral: M=.46, SD=.17; F(1,26)=4.590, p<.05) and Time 2 (Bob: M=.65, SD=.24; Neutral: M=.39, SD=.20; F(1,26)=14.786, p<.001)².

Explicit evaluations

As predicted, participants reported more favorable explicit evaluations of Bob after learning he had bought a soda (M=7.69, SD=2.12), versus had recently been convicted of molesting children (M=4.66, SD=1.95), t(51)=5.422, p<.001. Thus, participants’ explicit evaluations were highly sensitive to the new information learned at Time 2—a finding that is consistent with previous research (e.g., Gregg et al., 2006; Rydell et al., 2007).

Discussion

Following an extensive learning paradigm that has been shown in past research to result in relatively entrenched implicit evaluations (e.g., Rydell et al., 2007), participants exhibited implicit positivity towards a novel target. More importantly, whereas neutral information had little impact on participants’ task performance on the Time 2 implicit measure, a single piece of diagnostic propositional information that was inconsistent with prior learning fully reversed participants’ Time 1 implicit responses. Whereas at Time 1 individuals trended towards greater implicit positivity towards Bob than novel, neutral targets, they exhibited significantly less implicit positivity towards Bob than neutral targets at Time 2. This reversal resulted in a wide disparity between experimental and control participants’ implicit evaluations towards Bob at Time 2, even though these implicit responses were essentially equivalent a minute earlier.

Though these results are consistent with the notion that implicit evaluations can be rapidly revised in light of just a single piece of countervailing information, there are, of course, many peculiarities of child molestation that may have made it a rather potent influence on task performance. To ensure that our findings were not unique to child molestation, in the next experiment we sought to replicate this pattern of results with a different, extremely negative behavior.

Experiment 1b: Who is Bob? II

Method

Participants

One hundred and eighty Mechanical Turk workers (http://www.mturk.com) completed an online experiment on, “Learning about a new person.” Five participants submitted a survey code but failed to complete all of the components of the study and were thus excluded from analyses. Responses from an additional 12 participants were excluded from analyses due to: (a) a lack of variance in their responses on one or both of the implicit measures, or (b) excessive pressing (>10% of trials) of incorrect response keys on the implicit measures (see Payne et al., 2005). This left a final sample of 163 participants.

Procedure

The procedure for Experiment 1b was identical to Experiment 1a with the following exceptions. First, rather than completing the experiment in a cubicle in the lab for academic credit or $5, workers on Amazon's Mechanical Turk service (http://mturk.com) were directed to a survey link and completed the entire study from their browser in exchange for a small sum of money. This change in procedure also afforded us an opportunity to draw from a more diverse population in terms of ethnicity, age, and socio-economic status than the undergraduate sample we drew from in Experiment 1a (e.g., Buhrmester et al., 2011). Second, participants completed a Time 1 explicit evaluation measure immediately after completing the first AMP, following the same protocol as the measure in Experiment 1a (Time 1: α=.963; Time 2: α=.986). This allowed us to test implicit and explicit evaluation change simultaneously. Third, at Time 2, instead of learning that Bob had been convicted of molesting children, participants in the experimental condition were told that: “Bob recently mutilated a small, defenseless animal.” On the basis of a short pre-test completed by 60 undergraduates that included a variety of different positive and negative behavioral statements, this behavior was determined to be moderately high in diagnosticity (on a scale from 1=not at all diagnostic to 7=extremely diagnostic; M=5.51, SD=1.48; for comparison, child molester item: M=6.63, SD=1.03) and of high negative valence (on a scale from −3=extremely negative to +3=extremely positive; M=−2.42, SD=.72; child molester: M=−2.93, SD=.41). Finally, we included an exploratory measure at the end of the experiment in which we assessed participants’ explicit memory for behaviors they saw during the initial learning task by giving them a 10-item recognition test (e.g., “Did this behavior appear in the earlier task: “Bob helped a foreign student find a place to live”?). No effects emerged on this measure and we thus do not discuss it any further.

After this exploratory measure, participants answered several demographics questions and were then given a code to enter on mturk to receive their payment.

Results

A 2 (time: 1 or 2) × 2 (condition: experimental or control) × 2 (prime: Bob or neutral) mixed ANOVA on participants’ proportion-pleasant judgments again yielded the predicted 3-way interaction, F(1,161)=40.178, p<.001, ɳ_p²=.200 (see Figure 2).

Figure 2.

Participants’ proportion-pleasant judgments towards Bob before (Time 1) and after (Time 2) learning either that he mutilated a small animal (experimental) or that he bought a soda (control) in Experiment 1b.

Decomposing this interaction, in the experimental condition, there was a significant interaction between time and prime, F(1,80)=41.156, p<.001, such that participants exhibited significantly greater implicit positivity towards Bob (M=.65, SD=.27) than the unfamiliar faces (M=.47, SD=.23) at Time 1, F(1,80)=18.562, p<.001. At Time 2, however, there was a full reversal of the pattern at Time 1, such that participants now exhibited significantly more negativity towards Bob (M=.38, SD=.29) than towards the neutral targets (M=.60, SD=.26), F(1,80)=18.246, p<.001. Interestingly, this effect was driven not just by (predicted) decreases in implicit positivity towards Bob from Time 1(M=.65, SD=.27) to Time 2 (M=.38, SD=.29), F(1,80)=40.937, p<.001, but also partly by an unpredicted increase in implicit positivity towards unfamiliar targets from Time 1 (M=.47, SD=.23) to Time 2 (M=.60, SD=.26), F(1,80)=16.621, p<.001. This latter finding for unfamiliar targets, although unpredicted, emerged across the studies. It is intriguing evidence that in addition to implicit evaluations of Bob changing based on the newly learned behavior, participants also changed their implicit evaluations of the control faces based on a comparison with Bob. As Bob became much worse, these strangers become relatively more positive. However, this pattern might also have emerged simply as a response bias due to one of the two classes of prime stimuli in this task being rather highly affectively charged, thus serving as a kind of default that influenced responses to the other stimuli.

In the control condition, there was once again the predicted main effect of prime, F(1,81)=16.120, p<.001. Participants exhibited equivalent levels of implicit positivity towards Bob relative to neutral targets at both Time 1 (Bob: M=.65, SD=.25; Neutral: M=.51, SD=.25; F(1,81)=13.421, p<.001) and Time 2 (Bob: M=.64, SD=.25; Neutral: M=.47, SD=.24; F(1,81)=15.592, p<.001).

Explicit evaluations

Because Experiment 1b included a Time 1 explicit evaluation measure, we were able to test revision of explicit evaluations. As anticipated, an interaction between time and condition emerged, F(1,161)=276.828, p<.001, ɳ_p²=.635 (see Figure 3), such that participants in the control condition exhibited equivalent positivity towards Bob at Time 1 (M=6.59, SD=.51) and Time 2 (M=6.60, SD=.84), F<1, ns, whereas participants in the experimental condition expressed strong positivity towards Bob at Time 1 (M=6.72, SD=.58), but negativity at Time 2 (M=3.13, SD=1.82), F(1,80)=279.173, p<.001.

Figure 3.

Participants’ explicit evaluations of Bob before (Time 1) and after (Time 2) learning either that he mutilated a small animal (experimental) or that he bought a soda (control) in Experiment 1b.

Discussion

In Experiment 1b, we again found evidence that implicit evaluations were highly sensitive to a single piece of extremely negative information. Whereas participants’ implicit responses in the control condition were largely unchanged from Time 1 to Time 2, those in the experimental condition exhibited a reversal, such that their positivity toward Bob at Time 1 was completely reversed at Time 2. Participants’ explicit evaluations mirrored their implicit responses, exhibiting sensitivity to new, countervailing information. These results suggest that the pattern of rapid revision that we observed in Experiment1 was neither unique to the particular extreme behavior we chose nor to the undergraduate sample we utilized.

So far in our experiments, participants formed an evaluation on the basis of highly positive information and then learned a new piece of extremely negative information. We have not yet tested, however, whether a well-rehearsed negative evaluation would be sensitive to new pieces of highly positive information. Although it may be easy to undo one's good deeds with a single, particularly egregious act of wrongdoing, it may be that a consistent prickly and disagreeable disposition can be considerably less easy to shake with single, extreme acts of kindness.

There are strong theoretical reasons to suspect that a heavily-entrenched negative evaluation should be relatively less sensitive to countervailing information than a similarly entrenched positive one. First, research shows that bad events generally exert a larger impact on individuals than equivalent positive events (e.g., Baumeister et al., 2001; Rozin & Royzman, 2001). As a result of this negativity bias, individuals exhibit stronger brain activation and behavioral reactions to negative events (Cacciopo et al., 1997); have their attention drawn more quickly to negative versus positive words (Pratto & John, 1991); and learn more quickly from punishment versus reward (Vaish, Grossman, & Woodward, 2008). In their classic article on the negativity bias, Rozin and Royzman (2001) argued that a beverage can be utterly tainted by brief contact with a single cockroach, but no amount of goodness can be added to a plate of cockroaches to make it even slightly more palatable. Perhaps so, too, a single instance of particularly egregious behavior can tarnish one's otherwise positive impression of Bob. However, many more good deeds may be necessary to change one's (implicit) mind about someone we have learned over time is distasteful.

Moreover, research on person perception has shown a valence asymmetry in response to learning about personality behavior. Whereas participants tend to view a single instance of morally bankrupt behavior as highly indicative of one's “true” character and to thus draw broad conclusions about one's likely future interactions with that individual, they are less likely to forgive past transgressions on the basis of a single instance of particularly altruistic behavior. This can occur, in part, because such behaviors are more likely to be discounted as deviations from an otherwise negative disposition, rather than as especially diagnostic of one's true nature (e.g., Malle & Knobe, 1997; Mende-Siedlecki, Baron, & Todorov, 2013; Reeder & Coovert, 1986; Reeder, Pryor, & Wojciszke, 1992; Trafimow & Schneider, 1994; Trafimow & Trafimow, 1999). This work suggests that when an initially seemingly likeable Bob molests a child, he is not merely a good person who did a bad thing; rather, he is now (irreversibly) a bad person. However, when a consistently disagreeable and unsavory Bob decides to donate a kidney to a stranger, he does not simply become a good person; he is still ultimately a bad person who did a good thing. Such negative impressions may thus be especially difficult to shake precisely because good behaviors can be seen as less diagnostic of one's true nature.

To test for the possibility of this kind of valence asymmetry in implicit evaluative revision, in the next study, we included both positive and negative evaluation inductions at Time 1. At Time 2, we provided participants with either a neutral piece of information (i.e., “Bob bought a soda”) or an evaluatively inconsistent piece of information—that is, a highly negative piece of information in one case or a highly positive piece of information in the other. The two primary goals of the study were to discern whether: (a) an established negative implicit evaluation can be overturned by a single, extreme positive behavior, and, (b) the amount of change in the positive and negative directions reveal a valence asymmetry.

Experiment 2: The non-equivalence of molestation and kidney donation

Method

Participants

Three hundred and thirty Mechanical Turk workers completed the experiment online. Twenty-three participants submitted a survey code but failed to complete all of the components of the study and were thus excluded from analyses. The final sample was thus 307 participants.

Pre-test

To select a behavior that was of the same extremity but opposite in valence from a child molestation conviction, we conducted a short pre-test (N=30) in which Mechanical Turk workers assessed the diagnosticity—on a scale from 1 (not at all) to 7 (extremely) —and valence—on a scale from −3 (extremely negative) to +3 (extremely positive)—of a collection of 10 different positive (e.g., “Bob helped a stranger on the road with a flat tire”) and negative (e.g., “Bob lied to his boss about how much he contributed to a team project”) behaviors.

On the basis of this pre-test, we selected, for the positive counter-induction, the behavior, “Bob recently donated one of his kidneys to a child in need he had never met before,” which was equivalent but opposite in valence to child molestation (|M_kidney|=2.70, SD_kidney=1.21; |M_molester|=2.90, SD_molester=.548; t<1), as well as diagnosticity (M_kidney=6.47, SD_kidney=1.25; M_molester=6.53, SD_molester=1.25, t<1).

Procedure

The study employed a 2 (Time: 1 or 2) × 2 (Time 1 induction: positive or negative) × 2 (Time 2 information: control or counter-attitudinal) × 2 (Prime: Bob or neutral) mixed design. To manipulate the valence of participants’ initial implicit evaluations towards Bob, we simply reversed the characteristic/uncharacteristic feedback for positive and negative induction participants—that is, all participants encountered the same 100 behaviors; however, participants in the negative induction received feedback that 100% of the positive statements were uncharacteristic of Bob and 100% of the negative statements were characteristic of him, while those in the positive induction received the opposite feedback (i.e., a learning induction procedure identical to experiments 1a and 1b).

Following the induction, participants completed Time 1 implicit and explicit evaluation measures. At Time 2, participants either received counter-attitudinal information (that is, kidney donation following a negative induction or child molestation following a positive induction) or neutral information (i.e., “Bob bought a soda”). Across all conditions, this information was provided using the same protocol as the previous experiments. Next, participants completed the Time 2 implicit and explicit evaluation measures, followed by questions assessing participants’ perceptions of the valence of the statement they read at Time 2 on a −3 (extremely negative) to +3 (extremely positive) scale, and several demographics questions.

Results

We submitted participants’ proportion-pleasant judgments to a 2 (Time: 1 or 2) × 2 (Time 1 induction: positive or negative) × 2 (Time 2 information: control or counter-attitudinal) × 2 (prime: Bob or neutral) mixed ANOVA, which revealed the predicted 4-way interaction, F(1,303)=43.946, p<.001, ɳ_p²=.127 (see Figure 4)³.

Figure 4.

Participants’ implicit evaluations of Bob before (Time 1) and after (Time 2) learning neutral (control) or counter-attitudinal information in Experiment 2.

Breaking down this interaction, we explored the effects of time separately for the conditions in which neutral information was provided at Time 2 and the conditions in which counter-attitudinal information was provided. In the control conditions consisting of neutral information at Time 2, there was a Time 1 induction by prime interaction, F(1,153)=65.819, p<.001, that was unqualified by time, F<1—that is, participants’ implicit responses towards Bob were significantly affected by the type of initial learning they received, but there was no evidence of any changes over time after learning neutral information at Time 2. When participants underwent a positive induction, there was a main effect of prime, F(1,77)=14.919, p<.001, such that participants exhibited significantly more implicit positivity for Bob (Time 1: M=.67, SD=.28; Time 2: M=.70, SD=.26) than unfamiliar targets (Time 1: M=.52, SD=.28; Time 2: M=.51, SD=.29) at both time points. When they underwent a negative induction, there was also a main effect of prime, F(1,76)=57.146, p<.001. As expected, participants exhibited significantly less positivity for Bob (Time 1: M=.37, SD=.30; Time 2: M=.41, SD=.32) than unfamiliar targets (Time 1: M=.73, SD=.23; Time 2: M=.73, SD=.26) at both time points.

In contrast, in the experimental conditions, the Time 1 induction by prime interaction was significant, F(1,150)=8.516, p<.01, but was qualified by the predicted 3-way interaction with time, F(1,150)=54.276, p<.001. When participants underwent a positive induction, there was a significant interaction between time and prime, F(1,75)=38.397, p<.001, such that participants exhibited significantly greater positivity towards Bob relative to unfamiliar targets at Time 1 (M_Bob=.65, SD=.24; M_neutral=.49, SD=.23), F(1,75)=15.564, p<.001, but exhibited the reverse pattern at Time 2 (M_Bob=.37, SD=.28; M_neutral=.49, SD=.23), F(1,75)=35.079, p<.001. Thus, participants in the positive induction fully reversed their implicit responses towards Bob after encountering diagnostic and extreme negative information about him at Time 2. This pattern replicates the results of the first two studies. Mirroring the pattern of results in Experiment 1b, there was also evidence that this reversal at Time 2 was driven both by the predicted decreases in implicit responses towards Bob, F(1,75)=34.102, p<.001, as well as by unpredicted increases in responses towards neutral targets, F(1,75)=16.628, p<.001.

When participants underwent a negative induction, there was also a significant interaction between time and prime, F(1,75)=16.353, p<.001. However, unlike the positive induction, there was no evidence of a full reversal—that is, at Time 1, participants exhibited significantly less positivity towards Bob (M=.38, SD=.27), than unfamiliar targets (M=.65, SD=.24), F(1,75)=41.458, p<.001, and although this difference was attenuated at Time 2, there was still evidence that participants evaluated Bob (M=.52, SD=.28) significantly less favorably than controls (M=.63, SD=.24), F(1,75)=8.078, p<.01. Thus, we did find evidence of moderate revision when the initial learning was negative, but no evidence of a reversal.

Explicit evaluations

Participants’ responses were once again highly intercorrelated (Time 1: α=.996; Time 2: α=.986) and were thus combined into a single composite explicit evaluation measure at each time point. A 2 (time: 1 or 2) × 2 (Time 1 induction: positive or negative) × 2 (Time 2 information: control or counter-attitudinal) mixed ANOVA on this measure revealed the predicted 3-way interaction, F(1,303)=503.559, p<.001, ɳ_p²=.624 (see Figure 5).

Figure 5.

Participants’ explicit evaluationstowards Bob before (Time 1) and after (Time 2) learning neutral (control) or counter-attitudinal information in Experiment 2.

Decomposing this interaction, in the positive induction, participants who received counter-attitudinal information at Time 2 initially exhibited high levels of positivity towards Bob at Time 1 (M=6.56, SD=.94), but very strongly revised their initial impressions in light of the child molestation conviction (M=2.36, SD=1.51), F(1,75)=461.627, p<.001. When participants received neutral Time 2 information, they exhibited strong positivity towards Bob at both Time 1 (M=6.61, SD=.76) and Time 2 (M=6.62, SD=.71), F<1, ns. Taken together, these results fully replicate the findings of the previous studies.

In the negative induction, participants who received counter-attitudinal information at Time 2 initially exhibited high levels of negativity towards Bob (M=1.14, SD=.49). At Time 2, they did indeed revise their initial impressions, developing increased positivity towards him in light of his altruistic behavior (M=2.36, SD=1.51), F(1,75)=89.405, p<.05, although not to the extent that they did so in response to the counter-attitudinal information in the positive induction—that is, just like their implicit evaluations, participants exhibited evidence of a valence asymmetry on their explicit evaluations as well. Participants who received neutral information at Time 2 remained unchanged in their negative opinions of Bob (Time 1: M=1.17, SD=.49; Time 2: M=1.22, SD=.53), F(1,76)=1.646, ns.

Perceptions of Time 2 behavior

Although we had chosen the counter-attitudinal behaviors in this study on the basis of their equivalence in a pre-test, we could examine whether participants saw the behaviors through a different lens given what they learned about Bob during the induction. Whereas the two behaviors were similar in absolute valence in the pre-test, significant differences in absolute valence for the positive versus negative behavior did emerge during the full study such that the molester behavior was more extreme than the kidney donation behavior (|M_kidney|=2.63, SD_kidney=.73; |M_molester|=2.93, SD_molester=.30; t(150)=3.356, p<.001).

There are two possibilities for the differences in results between the pre-test and the full study. The first is that there are significant differences in participants’ perceptions of the two behaviors but that we lacked sufficient power to detect these differences in the pre-test. The second, however, is that participants in the pre-test and in the full study viewed these behaviors through different lenses. Indeed, our analysis posits that a particularly heinous or altruistic act is never evaluated in isolation, but rather viewed in light of the previous impressions that one has already formed of the individual engaging in the behavior. Critically, we may be less inclined to see the donation of a kidney to a stranger, for instance, in highly positive terms if we already know of a great deal of Bob's previous transgressions. This would suggest that a prior, well-learned negative evaluation might cast a suspicious light on even highly positive behavior, in a way that would not occur for a well-learned positive evaluation when one encounters a highly negative behavior (e.g., Reeder & Coovert, 1986; Trafimow & Trafimow, 1999). Indeed, this is the essence of the asymmetry between morally positive and negative reversals.

Discussion

The primary contributions of Experiment 2 are two-fold. First, we replicated the finding that a single piece of extremely negative behavior can overturn previously positive implicit evaluations. Second, in line with research and theory on negativity dominance (Baumeister et al., 2001; Rozin & Royzman, 2001) and trait asymmetry (e.g., Reeder & Coovert, 1986; Reeder et al., 1992; Trafimow & Trafimow, 1999), we found evidence for a valence asymmetry in participants’ receptiveness to new, countervailing information. Whereas a single egregious act gave rise to a full reversal of participants’ positive implicit responses, a single particularly altruistic act was not enough to fully “undo” many instances of moderately negative behaviors.

It is important to note that participants did show evidence of being responsive to information that Bob had donated one of his kidneys to a stranger. That is, implicit responses towards Bob did change—a finding that is noteworthy given previous research suggesting that a single piece of information cannot easily change prior evaluations (e.g., Rydell et al., 2007). However, he was nonetheless still implicitly evaluated negatively relative to the control faces; rather than becoming positive in participants’ eyes, he had instead merely become less negative.

One limitation of this study is that we established the equivalence of the extremity of kidney donation and child molestation using a self-report questionnaire. However, it is possible that people respond to these two behaviors differently at the implicit level. What we have shown here, then, is provisional evidence of a valence asymmetry (in line with other research in person perception). However, further research may reveal that the effects of positive versus negative behaviors on revision can sometimes be more symmetrical than what we observed here.

We now have evidence over multiple studies that learning about a single, counter-attitudinal, extreme behavior results in significant implicit evaluative revision, particularly when this behavior is negative. We have argued that the reason for such sensitivity to a single piece of extreme information is because extreme behaviors of this sort have important implications for an individual's true nature or character—that is, they communicate important, diagnostic information about a target's likely future behavior, thus leading participants to imbue it with greater significance. (e.g., Fiske, 1980). However, thus far our evidence for diagnosticity has only been indirect. In the next two studies we directly test our account. In Experiment 3, we test the hypothesis that the extremely negative behavior has to be attributed to the actor (Bob) rather than merely associated with him. In Experiment 4, we collected direct evidence that the perceived extreme negativity of the behavior informs the perceived diagnosticity of the behavior, which in turn predicts revision.

Experiment 3: The company Bob keeps

On the view of some associative accounts of implicit evaluation formation and change, implicit evaluation can often be susceptible to a kind of “guilt by mere association” (e.g., Walther, 2002). Because the associative processes that are thought (by some) to govern implicit evaluation formation and change operate primarily on the basis of similarity and temporal contiguity, situations in which either: (a) a liked or disliked individual is paired in close temporal proximity with a target individual, or (b) a liked or disliked individual shares superficial features of similarity with a target individual, can potentially give rise to a “spreading attitude effect” in which one's evaluation towards one individual can be associatively transferred to the other (see also (Skowronski, Carlston, Mae, & Crawford, 1998). This perspective would predict that an extreme negative behavior that is merely associated with Bob may be enough to lead to implicit evaluative revision.

However, if the perceived diagnosticity of the behavior is necessary for revision, then participants should show that they are able to discriminate between a very bad behavior that Bob performs versus a very bad behavior that someone superficially associated with Bob performs. We tested this question about “guilt by mere association” by adding a condition in which we exposed participants to the same negative behavior as in previous studies (i.e., child molesting), but implicated someone with whom Bob was merely incidentally associated—a high school coach at a high school in Bob's home town. If guilt by mere association can give rise to a spreading attitude effect, there ought to be strong revision in this condition, even though Bob was not responsible for the behavior. However, if participants consider the diagnosticity of the new information, they should be inclined to discount this piece of information in their implicit and explicit evaluations of him.

Method

Participants

Two-hundred and thirty mturk workers completed the experiment online. Fifteen workers submitted a survey code but failed to complete all components of the study and were thus excluded from analyses. The responses from an additional 14 participants were excluded from analyses due to either a lack of variance in their responses or excessive pressing of incorrect response keys on one or both of the implicit measure. This left a final sample of 201 participants.

Procedure

The procedure in Experiment 3 was essentially identical to the previous studies, but with the addition of a third condition in which the new piece of information provided at Time 2 implicated someone incidentally associated with Bob as a child molester rather than Bob himself. The Time 2 behavioral statement in this condition read: “The football coach at a high school in Bob's home town was recently convicted of child molestation.” The statements in the Bob-molester condition and in the control condition were identical to Experiment 1a. The study was thus a 2 (Time: 1 or 2) × 3 (Time 2 behavior: Bob molester, high school coach molester, or soda) × 2 (prime: Bob or neutral) mixed design.

Results

A 2 (Time: 1 or 2) × 3 (Time 2 information: Bob molester, coach molester, or soda) × 2 (prime: Bob or neutral) mixed ANOVA on participants’ proportion-pleasant judgments once again yielded the predicted 3-way interaction, F(2, 198)=13.998, p<.001, ɳ_p²=.124 (see Figure 6)⁴.

Figure 6.

Participants’ implicit evaluations of Bob before (Time 1) and after (Time 2) learning information about his behavior or the behavior or someone incidentally associated with him in Experiment 3.

Decomposing this interaction, in the Bob-molester condition, replicating the previous results, there was a significant interaction between Time and Prime, F(1,65)=25.262, p<.001, such that participants exhibited significantly greater implicit positivity towards Bob (M=.60, SD=.26) than the unfamiliar faces (M=.50, SD=.22) at Time 1, F(1,65)=4.267, p<.05. At Time 2, however, participants now exhibited significantly more negativity towards Bob (M=.36, SD=.29) than towards the neutral targets (M=.62, SD=.25), F(1,65)=21.529, p<.001. Like the previous studies, this interaction effect was driven, in part, by large decreases in implicit positivity towards Bob from Time 1 to Time 2, F(1,65)=24.661, p<.001, but also partly by an increase in implicit positivity towards unfamiliar targets from Time 1 to Time 2, F(1,65)=11.493, p<.001.

In the control condition, there was a main effect of prime, F(1,66)=24.427, p<.001, that was unqualified by an interaction with time, F<1, ns. Participants exhibited greater implicit positivity towards Bob at both Time 1 (Bob: M=.69, SD=.23; Neutral: M=.48, SD=.23; F(1,66)=24.315, p<.001) and Time 2 (Bob: M=.64, SD=.24; Neutral: M=.46, SD=.23; F(1,66)=15.870, p<.001).

More important for the current study, the pattern of results in the coach-molester condition closely mimicked those in the control condition and not those in the Bob-molester condition. There was a predicted main effect of prime, F(1,67)=13.451, p<.001, that was unqualified by an interaction with time, F<1, ns. Participants exhibited greater implicit positivity towards Bob at both Time 1 (Bob: M=.67, SD=.24; Neutral: M=.51, SD=.26; F(1,67)=12.652, p<.001) and Time 2 (Bob: M=.66, SD=.26; Neutral: M=.52, SD=.25; F(1,67)=9.734, p<.01).

Moreover, when testing the 2-way simple effect between the coach-molester and control conditions, only a main effect of target emerged, F(1,133)=36.741, p<.001, that was unqualified by condition, F(1,133)=.036, ns.

Explicit evaluations

On the composite measure of explicit evaluations (Time 1: α=.926; Time 2: α=.983), there was a predicted interaction between time and condition, F(2,198)=168.737, p<.001, ɳ_p²=.630 (see Figure 7). Participants in the Bob-molester condition initially expressed strong positivity towards Bob at Time 1 (M=6.65, SD=.55), but exhibited a reversal of their previous evaluations at Time 2 (M=2.65, SD=1.67), F(1,65)=362.710, p<.001. Participants in the control condition exhibited high levels of positivity towards Bob at both Time 1 (M=6.69, SD=.63) and Time 2 (M=6.69, SD=.63), F<1, ns. Taken together, these results replicate the findings of the previous studies.

Figure 7.

Participants’ explicit evaluations of Bob before (Time 1) and after (Time 2) learning information about his behavior or the behavior or someone incidentally associated with him in Experiment 3.

More interesting, however, were the effects of time on participants’ explicit evaluations in the coach-molester condition. Here, participants were equivalently positive (relative to the other two conditions) towards Bob at Time 1 (M=6.60, SD=.85), but nonetheless exhibited a(n) (unpredicted) decrease in their positivity towards Bob at Time 2 (M=5.98, SD=1.75), F(1,67)=10.093, p<.01. There were also significant differences in participants’ explicit evaluations at Time 2 among the 3 conditions, F(2,198)=26.793, p<.001. As predicted, post-hoc comparisons on Time 2 explicit evaluations (Tukey's HSD) showed that the Bob-molester condition significantly differed from the coach-molester and control conditions, but that the coach-molester and control conditions did not significantly differ from each other.

Discussion

As our account predicted, participants were mostly unmoved (both implicitly and explicitly) by the revelation that someone in Bob's home town was convicted of molesting children, despite whatever incidental association existed between them, and despite the fact that our explicit instructions presumably led participants to conclude that the information was important and relevant. Whereas those who learned that Bob had molested children exhibited similar updating of implicit and explicit evaluations as observed in our previous studies, participants in the coach-molester condition exhibited a pattern that was essentially equivalent to those that learned that Bob had bought a soda. Thus, Experiment 2 shows that diagnosticity can predict both large amounts of change on the basis of small amounts of information—as evidenced in the Bob-molester condition—as well as no change whatsoever despite encountering highly negative information—as evidenced by the coach-molester condition.

Participants’ explicit evaluations became more negative from Time 1 to Time 2 after they had learned about the coach. Thus, although participants’ implicit evaluations were largely unaffected by such “uninformative” information, their explicit evaluations were. It may be that as participants were self-reporting their evaluations, they considered the reasons why we were presenting them with this information (e.g., Gricean norms; see Grice, 1975), simply responding according to demand. That is, although the information itself was not informative about Bob, participants may have considered the presentational and demand pressures surrounding that information while reporting their evaluations, while such information did not affect their implicit evaluations. We note, however, that participants’ explicit evaluations at Time 2 in the coach condition were not significantly different from those in the control condition, and were also significantly more positive than those in the Bob-molester condition, as predicted.

The results from this study support the role of diagnosticity in implicit evaluative change. However, in the next study, we sought evidence that people think that extremely negative actions are important because they are diagnostic about the person's likely future actions.

Experiment 4: The mediating role of diagnosticity

In this experiment, we tested whether extremity predicts revision via perceived diagnosticity. We contend that extreme negative behavior signals diagnosticity because extreme negative acts are relatively rare and are thus likely to be interpreted as being due to the person being dispositionally “bad” (e.g., see Mende-Siedlecki, Baron, & Todorov, 2013). Research on person perception, for example, shows that whereas people do not think a target's single honest act necessarily says anything about the target's stable disposition, a single dishonest act indicates that the target is fundamentally dishonest, even if that person was otherwise well behaved (see Reeder & Coovert, 1986). We argue, then, that to the extent that people find animal mutilation as extremely offensive, they will see that act as saying something stable and “true” about the target. Such perceived diagnosticity should then lead them to weigh that important information over any number of earlier-learned positive behaviors.

After the initial learning about Bob, we measured participants’ beliefs about whether they considered animal mutilation as offensive and worthy of condemnation or punishment. We then measured the extent to which they viewed the action as diagnostic of who Bob truly is, and whether it would predict his future behavior.

Method

Participants

Two-hundred and forty Mechanical Turk workers completed an online experiment on, “Learning about a new person.”

Procedure

The procedure was identical to Experiment 1b in which participants learned at Time 2 that Bob had recently mutilated a small, defenseless animal. Following the Time 2 implicit and explicit evaluation measures, participants completed a short questionnaire assessing: (a) the offensiveness of the Time 2 behavior (4 items; α=.879; e.g., “To what extent do you feel it is morally wrong to harm animals?”; “How offensive do you consider Bob's later behavior to be?”; on 1=not at all to 7=extremely Likert scales), and (b) the perceived diagnosticity of the Time 2 behavior for Bob's future actions (8 items; α=.765; e.g., “To what extent did the later information change your overall impression of Bob?”; “To what extent do you feel that the later information you learned about Bob is a reflection of his true nature or character?”; also on 1=not at all to 7=extremely Likert scales).

After this short questionnaire, participants provided demographic information and received a survey completion code used to receive payment.

Results

We first ensured that participants exhibited rapid implicit evaluative change in light of Time 2 information. To assess this, we conducted a 2 (time: 1 or 2) × 2 (prime: Bob or neutral) within-subjects ANOVA, which revealed the predicted significant time by prime interaction, F(1,239)=53.848, p<.001, ɳ_p²=.184 (see Figure 8). At Time 1, participants initially exhibited greater positivity towards Bob (M=.68, SD=.26) than towards unfamiliar strangers (M=.51, SD=.26), F(1,239)=60.944,p<.001. At Time 2, participants exhibited a full reversal of their evaluations at Time 2, expressing lower levels of positivity towards Bob (M=.49, SD=.31) than neutral targets (M=.59, SD=.28), F(1,239)=12.439, p<.001. Moreover, a significant effect of time emerged for both Bob primes, F(1,239)=62.747, p<.001, and neutral primes, F(1,239)=14.722, p<.001, replicating the previous results.

Figure 8.

Participants’ implicit evaluations of Bob before (Time 1) and after (Time 2) learning that he had mutilated a small animal in Experiment 4.

On the composite measure of explicit evaluations (Time 1: α=.919; Time 2: α=.964), there was a large effect of time, F(1,239)=659.074, p<.001, such that participants expressed strong positivity towards Bob at Time 1 (M=6.63, SD=.74), but strongly revised their impressions at Time 2 (M=3.57, SD=1.78)⁵.

We anticipated that participants’ perceptions of the offensiveness of animal mutilation would predict implicit evaluative change, and that this effect would be mediated by participants’ perceived diagnosticity of animal mutilation as an indicator of Bob's true character and, thus, his likely future behavior. To assess this, we conducted a bootstrapping analysis using the procedures outlined by Preacher & Hayes (2008; 1000 re-samples) for estimating direct and indirect effects of a potential mediator. First, we created a measure of participants’ implicit preference for Bob at each time point by constructing a difference score between their proportion-pleasant judgments for Bob trials and proportion-pleasant judgments for neutral trials. Participants’ Time 2 implicit Bob preference served as the dependent measure, with Time 1 implicit Bob preference serving as a covariate in the analysis. Offensiveness (M=5.69, SD=1.35) was entered as the predictor variable, and diagnosticity (M=4.50, SD=1.57) was entered as the proposed mediator.

This analysis indicated that the total effect of offensiveness on Time 2 implicit Bob preference (controlling for Time 1 Bob preference; total effect: −.07, p<.001; partial effect of Time 1 Bob preference on Time 2 Bob preference: −.10, ns) became non-significant when diagnosticity was included in the model (direct effect of offensiveness: −.03, ns). The total indirect effect of offensiveness on Time 2 implicit Bob preference was significant, with a point estimate of −.04 and a 95% confidence interval of −.07 and −.01. These findings are consistent with an account suggesting that the effect of offensiveness on Time 2 implicit Bob preference was meditated by perceived diagnosticity.

Correlations among measures

As predicted, offensiveness and diagnosticity were each strongly negatively correlated with time 2 implicit and explicit evaluations—that is, the more participants saw Bob's Time 2 behavior as offensive and diagnostic of his character, the more negative their implicit and explicit evaluations at Time 2 (see Table 1 for correlations between the two composite measures of offensiveness and diagnosticity and the evaluation measures).

Table 1.

Correlations among measures in Experiment 4.

	Implicit Bob Preference		Explicit Composite
	Time 1	Time 2	Time 1	Time 2
Offensiveness	.00	−.22***	.13*	−.57***
Diagnosticity	.05	−.28***	−.04	−.57***

^*p<.05.

** p<.01

^***p<.001

Discussion

Experiment 4 provides evidence that the offensiveness of an action is a key ingredient in rapid implicit evaluative change in response to single pieces of extreme information. Participants’ perceptions of the offensiveness of Bob's Time 2 action were a strong predictor of their likelihood of exhibiting such rapid changes. More importantly, the findings were consistent with a meditational account for the role of perceived diagnosticity of that offensive behavior. The more participants saw Bob's animal mutilation as an especially immoral and offensive action, the more they saw that behavior as reflective of his true character, and updated their implicit evaluations of him accordingly. Although in this experiment we did not manipulate diagnosticity, and instead relied on correlational evidence, these findings, together with those from Experiment 3, strongly suggest that extreme negative behaviors lead to implicit revision through perceived diagnosticity.

Overall, the first 5 studies together provide strong and consistent evidence that a single, extreme behavior leads to rapid implicit evaluative change. One important outstanding question, however, is whether such newly-revised implicit responses are predictive of participants’ behavior. Rapid changes of the sort we have observed in the preceding studies would be of little consequence if they did not ultimately guide our responses towards individuals. In our final experiment, we thus examined the consequences of single pieces of diagnostic, countervailing information with respect to participants’ behavioral intentions. After undergoing a procedure similar to the previous studies, participants were asked about their behavioral intentions towards Bob. Critically, we expected that, if participants’ evaluations have truly been updated, participants’ Time 2 implicit evaluations of Bob (but not their Time 1 implicit evaluations) would predict their behavioral intentions.

Another important objective of this final experiment was to compare the predictive validity of revised implicit evaluations with that of revised explicit evaluations. Because both implicit and explicit evaluations have, across our studies, evidenced strong revision in response to a single, extreme behavior, one might wonder whether they are redundant with each other. In the next experiment, we examine this question by testing the predictive validity of participants’ Time 1 and Time 2 implicit and explicit evaluations of Bob in subsequent intentions.

Experiment 5: The effects of revised evaluations on behavioral intentions

Method

Participants

Three hundred Mechanical Turk workers completed an online experiment on, “Learning about a new person.” Four participants submitted a survey code but failed to complete all of the components of the study and were thus excluded from analysis. Finally, in line with previous work (Payne et al., 2005), 7 participants were excluded from analyses due to a lack of variance or one or both implicit measures. The final sample was 289 participants.

Procedure

The first part of the procedure in Experiment 5 was identical to the experimental condition in Experiment 1a. All participants underwent a positive induction and then completed Time 1 implicit and explicit evaluation measures. Next, all participants were told that Bob had recently been convicted of child molestation, and then completed Time 2 implicit and explicit evaluation measures.

Following these tasks, participants answered a single-item measure of their behavioral intentions towards Bob, specifically: “How much would you try to organize your neighbors to try to prevent Bob from moving into your area?” on a Likert scale from 1 (not at all) to 9 (a lot).

Following this question, participants completed a short demographics questionnaire and were then given a survey completion code.

Results

First, to ensure that we had replicated our earlier findings, we conducted a 2 (time: 1 or 2) × 2 (prime: Bob or neutral) within-subjects ANOVA, which revealed the predicted significant time by prime interaction, F(1,289)=114.707, p<.001, ɳ_p²=.284 (see Figure 9). At Time 1, participants initially exhibited greater positivity towards Bob (M=.66, SD=.24) than towards unfamiliar strangers (M=.48, SD=.22),F(1,289)=85.938, p<.001. However, at Time 2, participants exhibited a full reversal of their Time 1 preferences, now exhibiting less implicit positivity towards Bob (M=.41, SD=.28) than neutral targets (M=.60, SD=.25), F(1,289)=56.850, p<.001. As in the previous studies, there was evidence that this effect was driven both by shifts towards greater negativity towards Bob, F(1,289)=122.388, p<.001, as well as shifts towards greater positivity towards neutral targets, F(1,289)=39.212, p<.001.

Figure 9.

Participants’ implicit evaluations of Bob before (Time 1) and after (Time 2) learning that he had mutilated a small animal in Experiment 5.

On the composite measure of explicit evaluations (Time 1: α=.953; Time 2: α=.933), there was an effect of time, F(1,289)=1390.608, p<.001, such that participants expressed strong positivity towards Bob at Time 1 (M=6.67, SD=.73), but strongly revised their impressions at Time 2 (M=3.35, SD=1.36).

To assess the effects of participants’ newly-revised implicit evaluations on their behavioral intentions, we created a difference score between participants’ proportion-pleasant judgments towards Bob and towards neutral targets on the AMP at Time 1 and 2. This measure thus reflects participants’ relative implicit preference at each time point. We then conducted a linear regression that included each of these relative implicit preference measures for Bob as well as participants’ explicit evaluation composites at each time point as predictors of the extent to which they would try to organize their neighbors to prevent Bob from moving into their area. In this analysis, participants’ Time 2 relative implicit preference for Bob, β=−.11, t(289)=−2.03, p<.05, as well as participants’ Time 2 explicit evaluations, β=−.74, t(289)=−7.51, p<.001, emerged as the only two significant predictors of participants’ behavioral intentions, β=−.16, t(100)=−2.040, p<.05 (see Table 2).

Table 2.

Regression analysis in Experiment 5.

	B	SE B	β
(Intercept)	(8.128***)	(1.191)
Time 1 implicit Bob preference	.466	.384	.064
Time 2 implicit Bob preference	−.658*	.324	−.117*
Time 1 explicit evaluation	.128	.171	.039
Time 2 explicit evaluation	−.737***	.098	−.423***

Note. Implicit preference is a differences score between participants' proportion-pleasant judgments of Bob and their proportion-pleasant judgments of neutral targets.

^*p<.05.

^***p<.001.

Discussion

We found that participants’ newly-revised implicit responses predicted their behavioral intentions towards Bob such that the more they implicitly disliked Bob, the more they said they would try to organize their neighbors to prevent him from moving into their area. These assessments were uniquely predicted by their explicit and implicit Time 2 evaluations towards Bob.

We note, however, that the unique predictive validity of our implicit measure in predicting participants’ behavioral intentions need not be a result of differences in the underlying processes that govern implicit and explicit evaluation. Payne and colleagues (2008) have argued that differences among implicit and explicit measures may be driven not by different processes, but rather by differences in the structure of the measures used to assess these evaluations (e.g., kinds of stimuli, timing). It is noteworthy, however, that our measure of behavioral intentions shared considerably more structural properties with our explicit evaluation measure than with the implicit measure and, yet, implicit evaluations nonetheless emerged as a unique predictor.

General Discussion

The results of six experiments show that implicit impressions can be quickly reversed by a single piece of counterevidence. After learning 100 unique positive pieces of behavioral evidence about a target, participants exhibited a positive implicit evaluation of that target. Yet, after learning just a single new piece of negative information (their 101^st in the experiment), they immediately showed a complete reversal of their evaluations of the target.

We found evidence for this rapid revision in response to two extremely negative behaviors (child molestation: experiments 1a, 2, and 3; animal mutilation: experiments 1b, 4, and 5), and one extremely positive behavior (kidney donation to a stranger: Experiment 2). However, we also found support—as our analysis had predicted—for a valence asymmetry. Whereas participants readily incorporated negative diagnostic behaviors into their implicit evaluations of previously uniformly positive targets, they were less inclined to cast aside many initial instances of negative behavior in light of a single act of altruism (Experiment 2). Extreme negativity is especially effective as counterevidence because it is seen as especially diagnostic of a person's true character (e.g., someone who mutilates a small animal is not merely engaging in a bad behavior, but rather is a bad person; Reeder & Coovert, 1986). And, in support of this, we showed in Study 4 that the more participants saw the negative act as offensive, the more they saw it as revealing something stable and predictive about the target, which in turn predicted the degree to which they updated their implicit impression of him.

We also identified situations in which extreme behaviors were less influential on revision. Though a strong associationistic view might suppose, for example, that implicit evaluations should promiscuously incorporate information about targets even when they are only incidentally associated with the behavior, we found no evidence, in the coach-molester condition in Experiment 3, to suggest that participants succumbed to mere “guilt by association” (cf. Walther, 2002)

Finally, we showed in Study 5 that revised implicit evaluations have unique predictive validity for behavioral intentions. Participants’ revised implicit evaluations of Bob significantly predicted whether they would be inclined to mobilize with their neighbors to drive Bob out of the neighborhood should he decide to move in. Participants’ implicit evaluation of Bob measured before the revision were not a predictor of this outcome, indicating that the revised implicit evaluation had effectively “silenced” the initial 100 behaviors participants had learned about Bob.

Reconciling these findings with current theories

Although our results unequivocally show that implicit evaluations can be rapidly updated in light of minimal counterevidence, it is less clear what kind of underlying processing is influencing task performance on the implicit measures in our studies—that is, whether task performance was primarily influenced by associative processes, propositional processes, or some interaction between the two. As one of us has argued elsewhere (Ferguson et al., 2014; see also Moors, 2014), it is quite difficult, on the sole basis of behavioral evidence, to identify the representational format or algorithmic nature (Marr, 1982) of the underlying processes governing observed behaviors. Moreover, given past work by Sherman, Payne and others (e.g., Bishara & Payne, 2009; Conrey, Sherman, Gawronski, Hugenberg, & Groom, 2005; Payne, 2001; Sherman, 2006), task performance on the Affect Misattribution Procedure (or any other implicit measure) is likely influenced by multiple processes—some of which may be propositional in nature—and the operating characteristics of such processes still remain unclear (e.g., Sherman, 2006).

Nevertheless, there is a precedent in the literature to classify the types of information we used in our learning paradigm—that is, verbal information requiring sentence comprehension—as propositional in nature (e.g., Gawronski & Bodenhausen, 2006, 2011, 2014; Rydell & McConnell, 2006; Rydell et al., 2006). If participants' processing of the behavioral statements did occur at least partially propositionally, then the rapid changes in task performance that we observed in our studies would appear to be inconsistent with those classes of dual process models that predict that task performance on implicit measures should be influenced only by (slow-learning) associative processes (Rydell & McConnell, 2006; Rydell et al., 2006, 2007; though see McConnell & Rydell, 2014, pp. 213-216).

Other dual process models, such as Gawronski & Bodenhausen's (2006, 2011) APE model, do allow for some kinds of interactions between associative and propositional processes in the formation and change of implicit responses. On this view, although it may be difficult to negate existing associations after they have been established, one might reflect new learning by creating new associations that affirm the opposite of a previously-learned association (e.g., Gawronski, Deutsch, Mbirkou, Seibt, & Strack, 2008). If the counterevidence that we presented about Bob is an affirmation that he is a decidedly awful person—rather than a negation of the proposition that he is a relatively good person—then our findings may be consistent with such a model.

There are other theoretical approaches that allow for even stronger influences of propositional information on implicit evaluations, including those that do not assume that task performance on implicit measures like the AMP is solely under the purview of associative processes. For example, De Houwer's single-process model (De Houwer, in press; Hughes et al., 2011) posits that both implicit and explicit responses are governed by a single propositional process. On this view, participants may have rapidly incorporated the new verbal information propositionally, thus allowing for rapid influences on their task performance on the implicit measure.

But, of course, the information that we presented about Bob, though propositional in its format, was also highly affectively charged. It may be that learning that someone has recently been convicted of a serious crime has such strong affective connotations that it ultimately results in rapid changes in task performance through a relatively associative process. For example, Amodio and colleagues (Amodio, 2014; Amodio & Ratner, 2011; see also Amodio & Devine, 2006) argue that there are multiple memory systems involved in implicit processing that each possesses its own learning characteristics and trajectories. With this model in mind, perhaps participants' formed their implicit responses towards Bob on the basis of a slower-learning, implicit semantic system, but then rapidly revised their implicit responses in light of new highly affective information about Bob on the basis of a faster-learning classical conditioning system.

Implicit versus explicit evaluations

A number of current theories have suggested that the relative insensitivity of implicit evaluations to a new, single countervailing (and “propositional”) piece of evidence implies that implicit and explicit evaluations rely on different processes or systems (Gawronski & Bodenhausen, 2006, 2011, 2014; Rydell & McConnell, 2006; Rydell et al., 2006, 2007). After all, if implicit and explicit evaluations exhibit different learning trajectories, it could be due to different processes. The current results, then, might lead one to speculate that implicit and explicit evaluations share underlying processing more than current theories have assumed. For example, it could be that performance on both types of measures emerged entirely through associative processes, entirely through propositional processes, or some interaction between associative and propositional processes.

However, just as it is difficult to identify the type of process(es) (in algorithmic terms; Marr, 1982) underlying implicit evaluations on the basis of patterns of change versus stability, it is similarly difficult to infer that a dissociation between implicit and explicit performance is due to different underlying processes, or that equivalence in performance implies that the same underlying processing. Without the use of corroborative computational modeling, that can explain behavioral evidence, it is hard to pinpoint the type of processing underlying task performance (see Ferguson et al., 2014). Moreover, there are numerous and important structural differences in implicit versus explicit evaluation measures (Payne et al., 2008), such as timing, type of response, type of stimuli, abstractness of stimuli, and so on that could serve to inflate the dissociation between the two measures. Perhaps in some circumstances these structural differences are overwhelmed by some other variable that drives performance in similar ways across the measures. Research will continue to identify cases where implicit and explicit evaluations are similar versus dissimilar, and inform an understanding of the processes at play for each type of measure.

Reconciling our findings with previous research

How do our findings speak to earlier work showing that implicit evaluations are, in the conception of Gregg and colleagues (2006), “easier done than undone”? One possibility is that the strategies that were used to induce revision in earlier work did not lead participants to view new information as highly diagnostic, or, at least, no more diagnostic than the information they learned earlier in the experiment. For example, in Gregg and colleagues’ study 4, participants were provided with a narrative that described one fictitious group as decidedly negative and the other as highly positive. They were then given a narrative counter-induction in which the two groups were described as having switched characters over time. Gregg et al. found that this narrative counter-induction was not influential. From our perspective, one reason for this lack of revision may have been that the rationale provided for the role reversals in the counter-induction did not negate the group's past behavior in participants’ eyes; instead, they may have seen each group's earlier behaviors as better examples of the group's true natures than their later behaviors.

Similarly, in Rydell et al.'s (2007) research, in which participants’ implicit evaluations changed only slowly in response to many instances of counterevidence, we would contend that no single behavior was sufficiently negative (or extreme) to trigger perceived diagnostic priority. When the initial behaviors, and the new, countervailing behaviors, are all moderate in their valence, there simply may not have been a strong enough signal to trigger the kinds of diagnostic judgments that can alter people's implicit impressions as they did in our studies

Were participants’ Time 1 implicit evaluations “erased”?

An important consideration in the interpretation of our findings is whether the counter-attitudinal information we provided at Time 2 “erased” (that is, replaced) the information participants learned initially, or merely created new associations that somehow overpowered initial learning. Research has shown that evaluatively inconsistent exposure to an object can often lead to contextualization—a process in which different aspects of a mental representation may be activated in response to different contextual cues in the environment (e.g., Rydell & Gawronski, 2009; Gawronski et al., 2010; see also Gawronski & Sritharan, 2010). If such a process were operating in our experiments, it would have important implications for the durability of the effects of the Time 2 information because it would suggest that the older implicit evaluations could perhaps resurface in certain contexts. Ultimately, we cannot distinguish, with our current data, between each of these two possibilities. However, there are a number of reasons why we suspect that contextualization is a less likely explanation for our findings. First, we made every effort to ensure that the circumstances under which participants learned Time 1 and Time 2 information were essentially identical and that no obvious contextual cues were available to allow participants to make sense of the evaluative inconsistencies to which we exposed them (cf. Gawronski et al., 2010; Rydell & Gawronski, 2009). Thus, even if contextualization processes could, in principle, have operated in the current studies, the necessary antecedents for such a mechanism may have been largely absent (but see AAB renewal effects in Gawronski et al., 2010).

Second, in contrast to the well-established and multi-faceted objects that have been shown to give rise to contextual effects (in evaluation) in previous work, because participants were exposed to novel targets in our experiments, we know participants’ entire evaluative history with Bob. What this evaluative history suggests is that even if Time 2 information gave rise to a contextualized, evaluatively inconsistent mental representation, the fact that a single new piece of propositional information was enough to rapidly give rise to this contextualization is a new finding that must be accommodated in this work.

Factors affecting assessments of diagnosticity

Although we have documented one particular set of circumstances in which minimal counterevidence can result in implicit evaluative revision, there are likely to be a number of relevant moderators. First, it is unlikely that an assessment of the diagnosticity of a new piece of information occurs in complete isolation; a more relative assessment may occur in which individuals evaluate not just the diagnosticity of current information, but also how later behavior compares to other behaviors that one has already learned (that may themselves be seen as diagnostic).

In practice, however, this relative judgment of diagnosticity may be less clear-cut than it was in our studies. Such assessments may require, for example, some explicit memory for particular instances of an individual's behavior that one can then compare and contrast with the current behavior—an assumption that may not necessarily be warranted (e.g., Dijksterhuis, 2004; Olson & Fazio, 2001, 2002; Rydell et al., 2006; see also Klein & Loftus, 1993; Tulving, 1993), particularly for well-established implicit responses that are the product of years of experience. Moreover, it is likely more often the case that participants come to the assessment that both earlier and later information is relevant to participants’ understanding of a person's character, which could, in many cases, lead to feelings of implicit ambivalence (e.g., Petty et al., 2006) rather than the full reversals we document in our studies.

Second, though we found that participants exhibit rapid reversals when they construed new information as highly diagnostic, we have not specified the conditions under which participants are likely to arrive at this assessment. We made the assumption, in line with previous work, that extreme, negative behaviors would be the most likely to trigger diagnostic assessments (Fiske, 1980; Knobe, 2003; 2006; Nadelhoffer, 2006; Malle & Knobe, 1997; Mende-Siedlecki, Baron, & Todorov, 2013, Reeder & Brewer, 1979; Reeder & Coovert, 1986; Reeder, Pryor, & Wojciszke, 1992; Trafimow & Schneider, 1994; Trafimow & Trafimow, 1999). However, there are sure to be important factors that govern the extent to which participants see new behaviors as gleaning information about a person's true character or not.

Although we are suggesting that assessments of diagnosticity can influence the likelihood of implicit evaluation change, this is not to say that it must be an unbiased assessment. Consider, for example, a scenario in which you come home to discover your romantic partner in a compromising position with your best friend. Our analysis suggests that whether this single act induces change in your presumably heavily-entrenched implicit (positive) evaluation of your partner will be driven by whether you deem this behavior as revealing of your partner's true nature. However, note that whether you ultimately come to this assessment or not will be influenced by various psychological mechanisms that serve to ameliorate the impact of negative events and threats to the self (e.g., Baumeister & Newman, 1994; Ditto & Lopez, 1992; Ditto et al., 1998; Dunning, 2012; Kunda, 1987, 1990; Lord, Ross, & Lepper, 1979)—mechanisms that may prevent you from seeing the situation objectively or clearly (“It was a one-time thing; it’ll never happen again”; “I know that s/he actually really loves me”) and thus alter your judgment of the diagnosticity of the behavior, with implications for the likelihood of rapid changes in implicit evaluations.

In short, one's own assessment of the diagnosticity of a behavior may be quite different from a more objective assessment provided by unbiased others, and there may be important barriers—barriers that were not present in the current studies— that prevent individuals from reasoning that a behavior is more important than past experience. What is clear from our results, however, is that when participants arrive at this assessment, they can readily update their implicit evaluations.

Limitations of the current work

We have presented evidence across multiple experiments that a single behavior can completely undo previously learned implicit evaluations. However, throughout these experiments we measured implicit evaluations exclusively using the AMP (Payne et al., 2005). Although the most recent evidence suggests that the AMP measures unintentional evaluations of the prime stimuli (see Payne et al., 2013), it will be valuable to test in future work whether the same kind of rapid revision occurs in performance on other implicit measures, as we would expect. Additionally, we have used a specific learning format in these experiments where single behaviors are presented individually and sequentially. Whereas in the current work the relation among the behaviors—and especially the relationship between the old and new behavior—is never offered to participants, in future work, it will be critical to test whether different types of learning, perhaps via more ecological formats, would lead to a similar type of revision of implicit impressions.

Conclusion

According to some current evidence and theory, when we discover information that is highly inconsistent with our well-learned memories about someone, we face what would seem to be a rather troubling and highly maladaptive situation: despite the implications that revelations of this sort may have for the validity of our past experiences, we may nonetheless harbor a lingering implicit evaluation of the individual that can continue to exert an influence on our behavior for some time.

Yet, the experiments we report here suggest a more adaptive outcome: that we may be able—like the participants in these studies—to quickly incorporate minimal counter-attitudinal information into our implicit evaluations. We may sometimes be able to cast earlier impressions aside and come to an arguably more adaptive and accurate implicit impression. We need only, on the basis of our analysis, to see such revelations as especially diagnostic of who someone truly is. Far from leading us astray, implicit evaluations can uniquely shape and guide our behavioral choices in accord with what we learn to be true of the world.