The most obvious source of preferences in evaluative judgments is the valence of the target itself. When encountering valenced targets, evaluations are automatically activated from memory (Fazio et al., 1986). Likewise, judgments of liking can be influenced by stimulus features previously associated with positive or negative events by reactivating this positivity or negativity (Fazio and Olson, 2003). If multiple features of a target are available, their respective evaluations are integrated into an overall evaluation of the target (e.g., Massaro and Friedman, 1990). An approach to explain this integrative process are additive accounts of impression formation (e.g., Anderson, 1962, 1971; Kahneman et al., 1993) stating that the values of each single attribute of a stimulus are averaged to reach a net evaluation. The outcome of this process is positive or negative affect (Russell, 2003, 2009), which is then used as the basis for the evaluative judgment (Schwarz, 1990, 2002, 2012).

A more indirect source of preferences is stimulus consistency. Maintaining consistency, or coherence, between various aspects of the world is a fundamental aim of the mind (Singer, 1999; Gawronski and Strack, 2012), and humans have developed a strong sensitivity for the presence or absence of coherence (Antonovsky and Sagy, 1986). Consequently, inconsistency between different pieces of information, or between new information and existing knowledge is aversive. It potentially interferes with effective and unconflicted action (Harmon-Jones et al., 2012), or it disrupts the process of establishing meaning from coherent relations in the external world (see the meaning maintenance model, Heine et al., 2006; Proulx and Heine, 2008; Proulx et al., 2010; Randles et al., 2011).

Further supporting evidence of the affective impact of stimulus consistency on preferences stems from research on processing fluency (Reber et al., 2004), which is the content-independent efficiency with which a stimulus is processed. Consistency, or coherence, increases processing fluency (Topolinski, 2011). Various experiments have shown that high processing fluency is experienced as hedonically positive (Winkielman et al., 2003; Reber et al., 2004). The sort of fluency that is most closely related to stereotype consistency is semantic, or conceptual, fluency (Whittlesea, 1993), which has been shown to elicit positive affect (Topolinski et al., 2009). For example, Topolinski and Strack (2009b) presented participants with word triads that either shared a common associate (e.g., RABBIT CLOUD CREAM; implying WHITE; coherent) or were only random words (e.g., DREAM BALL BOOK; incoherent) and asked for participants’ spontaneous preference for these word triads. Although participants were not aware of the hidden semantic coherence, they preferred coherent over incoherent triads (Topolinski and Strack, 2009a,b,c), and even showed subtle smiling while reading coherent (compared to incoherent) triads (Topolinski et al., 2009). Moreover, valence has been shown to influence coherence judgments (Topolinski and Strack, 2009b). Participants were presented with semantically coherent and incoherent word triads and were asked to intuitively judge the triads’ coherence. In addition, independent from actual coherence, the triads’ valence was manipulated by constructing coherent and incoherent word triads out of relatively positive (HOLY LIQUID FRESH; implying WATER) or negative words (SALT RAIN DROWN; also implying WATER), respectively. It was found that not only coherent triads were judged to be coherent more often than incoherent triads, but that positive triads, too, were judged to be coherent more frequently than negative triads. Although this finding pertains to rather neutral non-social word triads, and the dependent measure was coherence intuitions, this evidence shows that valence and consistency, if pit against each other, can show additive effects.

In sum, from both inconsistency-threat (e.g., Heine et al., 2006; Harmon-Jones et al., 2012) and fluency approaches (e.g., Topolinski and Strack, 2009a,b) the prediction can be derived that consistency compared to inconsistency induces positive affect. Importantly, the positive affect resulting from consistency should be independent from actual stimulus content (e.g., its valence, see Topolinski and Strack, 2009b).

In social psychology, a classical example of striving for consistency is holding and maintaining knowledge about stereotypes (Weber and Crocker, 1983; Devine, 1989; Devine and Sharp, 2009). The present approach examines stereotype disconfirmation (e.g., Macrae et al., 1999; Sherman et al., 2012), the exhibition of behaviors or features of a stereotypical target that are inconsistent with the current stereotype.

Generally, stereotype-inconsistent compared to consistent information has been found to receive more attention (Brannon et al., 2017), to be processed more deeply (Fiske, 1980; Bargh and Thein, 1985; Pezzo, 2003), and, as a consequence, to be remembered better (Hastie and Kumar, 1979; Bargh and Thein, 1985; Schützwohl, 1998; for a review, see Stangor and McMillan, 1992; but see, for opposite effects, Fyock and Stangor, 1994; Johnston, 1996; Trope and Thompson, 1997; for an integration, see Sherman et al., 2012). This memory bias is even stronger, when the (counter-)stereotypic information is of special relevance to the perceiver (Süssenbach et al., 2016).

Thus, there is abundant research on the influence of stereotype disconfirmation on memory. However, the effects of stereotype disconfirmation on evaluative judgments has been less extensively examined. Importantly, up to now, there seems to be no research on the influence of consistency on preference for the target person. Previous research found that individuals who described others in stereotype-consistent ways were evaluated more positively than those who described others in stereotype-inconsistent ways (Castelli et al., 2005). The focus of the present research, however, is the question how an (in)consistent target person is evaluated him- or herself. Some experiments on memory biases for stereotype-inconsistent information assessed evaluative judgments for inconsistent targets but not in a way to enable testing differences in preference between consistent and inconstant targets (due to a within-target manipulation of consistency; Förster et al., 2000, 2004; or due to the fact that only inconsistent targets were shown, Förster et al., 2004). Another experiment that assessed evaluations of stereotype-inconsistent target persons was conducted by Bless et al. (2001). German participants read a description of a member of the ethnic group of Sinti and Roma (negative stereotype in Germany). The target person was described rather favorably and therefore inconsistent with that stereotype. However, because this experiments’ design focused on factors other than consistency vs. inconsistency, participants only received inconsistent target persons, so no comparison between consistent and inconsistent target persons was possible.

Furthermore, there is also work on the evaluation of stereotype-inconsistent behavior of target persons. A study by Lambert and Wyer (1990), for instance, had participants read a description of a target person who was labeled as either being a priest (highly moral) or a businessman (moderately moral) and who exhibited an immoral behavior (stealing). Even though the general evaluation of priests was positive, the evaluation of the priest showing immoral behavior was just as negative as the evaluation of a businessman showing immoral behavior (Lambert and Wyer, 1990). Participants’ evaluative judgment of the target person was not influenced by his group membership, but by his individual behavior. However, a comparison between stereotype-consistent and -inconsistent behavior was not possible because participants only evaluated target persons displaying stereotype-inconsistent (immoral) behavior.

One series of experiments actually compared evaluative judgments of stereotype-consistent and -inconsistent target persons (Mendes et al., 2007). Reactions toward typical and atypical members of a social category were compared. Specifically, participants interacted with a confederate who was either a typical or an atypical member of his or her social category (e.g., white vs. black with high vs. low socio-economic status; or Asian with or without southern accent). The result was that stereotype-consistent interaction partners were preferred over inconsistent ones and elicited less physiological threat responses (Mendes et al., 2007). This evidence is in line with the predictions from inconsistency-threat and fluency that inconsistency should elicit more negative evaluations.

To our knowledge, up to now there has been no attempt to manipulate stereotype valence and stimulus consistency orthogonally (but see the mixed evidence on stimulus valence and fluency due to repetition, Grush, 1976; Dijksterhuis and Smith, 2002; Schellenberg et al., 2008).

The present research goes beyond earlier approaches by systematically mapping stereotype valence against consistency. Furthermore, consistency itself was manipulated more rigorously. While earlier work (e.g., Förster et al., 2000, 2004; Mendes et al., 2007) used only a single target person and chose the single stereotype-(in)consistent behavior or feature arbitrarily (which allows material effects), the present experiments used multiple targets, each with multiple features, that were participant-generated (see Pilot Studies 1 and 2). Additionally, stereotype-consistent and -inconsistent features were sampled from the same stimulus pool, thus preventing material effects. However, the main additional contribution of the present work is the independent manipulation of valence and consistency, which is outlined in the following.

Concluding from the above reviewed literature on stereotype disconfirmation (e.g., Mendes et al., 2007) and processing fluency (Topolinski and Strack, 2009a,b,c) we predicted that stereotype valence and stereotype consistency would independently influence preferences, with more positive evaluations for positive compared to negative stereotypes, and also for consistent compared to inconsistent targets, with no interactional effects between valence and consistency. The previously reviewed experiment by Topolinski and Strack (2009b), showing that coherence and valence showed additive effects for consistency judgments, can be seen as supporting evidence. The present experiments examined whether this independent contribution also holds for preferences for stereotyped target persons. This research agenda has the theoretical significance of testing whether psychological processes found in basic cognitive research, in this case, the additive principle of valence and semantic coherence, also holds for social perception and person impression formation. Thereby, we explore the underlying information processing steps that drive social behavior and test which role well-established cognitive effects play in a social context.

In Experiment 1, we compared consistent target persons to valence inconsistent as well as semantically inconsistent targets. This experiment addressed the aforementioned question how a sweet hooligan is evaluated compared to a brutal hooligan. In Experiments 2–4, we compared consistent targets to semantically inconsistent targets, but kept valence within a target person constant. For instance, we expected an open-minded midwife to be evaluated more negatively than a caring midwife. In Experiments 5 and 6, we examined neutral stereotypes and again compared semantically consistent with inconsistent targets. We expected a literate butcher to be more negatively evaluated than a strong butcher. Thus, in all three cases we expected that consistency compared to inconsistency would increase fluency-induced positive affect, which should, in turn, lead to more favorable evaluations.

We determined the required sample sizes with a power analysis (G^∗Power 3, Faul et al., 2007), attempting a power of 0.80, an alpha-level of 0.05, and estimating the effect size based on previously published effects of coherence on liking for word triads outside a social context (Topolinski and Strack, 2009a, Experiment 3, p. 1484, N = 61, d_z = 0.35; Topolinski and Strack, 2009c, Experiment 1, p. 611, N = 22, d_z = 0.84). The pooled effect size was d_z = 0.48, which equals f = 0.24. Therefore, we used f = 0.24 as an estimate for the expected size of the consistency effect. A power analysis for a repeated-measures ANOVA with these parameters (as well as a correlation of 0.5, and non-sphericity correction of 1) yields N_req = 37. Because all experiments were combined with other studies, some of which required larger sample sizes, most Experiments exceed the required sample size; only Experiment 4 fell below this criterion. As the observed effect sizes in Experiments 1–4 were much smaller than expected, we drastically increased the sample sizes for the last two experiments to about 200 participants each. Additionally, Experiments 5¹ and 6² were preregistered.

We report all manipulations and measures. No subjects were excluded from the analyses and all preparatory steps prior to the main analyses are described in the text. All experiments were lab-based and took place at German universities; participants were predominantly students majoring in various disciplines. We used Medialab/Direct RT for data collection.

In two pilot studies, a new pool of stimuli was developed based on semantic and affective associations of participant samples comparable to the participants in the later experiments.

Evaluations of targets in the same category were averaged and then entered into a 2 (valence: positive, negative) × 2 (consistency: consistent, inconsistent) repeated measures ANOVA. A main effect of valence emerged, F(1,49) = 525.48, p < 0.001, ηp2 = 0.92, 90% CI = 0.87–0.93, indicating that positive stereotypes (M = 5.28, SD = 0.73) were evaluated more positively than negative stereotypes (M = 2.16, SD = 0.61). Additionally, a main effect of consistency emerged, F(1,49) = 14.11, p < 0.001, ηp2 = 0.22, 90% CI = 0.07–0.37, indicating that consistent targets (M = 3.94, SD = 0.62), were evaluated more positively compared to inconsistent targets (M = 3.50, SD = 0.63). However, this main effect was qualified by an interaction between consistency and valence, F(1,49) = 35.34, p < 0.001, ηp2 = 0.42, 90% CI = 0.24–0.55. Simple comparisons found that for positive stereotypes, consistent targets (M = 5.85, SD = 0.83) were evaluated more positively than inconsistent targets (M = 4.71, SD = 1.06), t(49) = 6.63, p < 0.001, d_z = 0.94, 95% CI = 0.60–1.27. However, for negative stereotypes, consistent targets (M = 2.02, SD = 0.90) tended to be rated as less positive than inconsistent targets (M = 2.29, SD = 0.74), t(49) = -1.70, p = 0.10, d_z = 0.24, 95% CI = -0.04 to 0.52. Results are illustrated in Figure 1.

Experiment 1 found a main effect of stereotype valence, which is a rather trivial finding. While there also was a main effect of consistency, this effect is due to the larger difference in ratings of positive compared to negative targets, as can be seen in Figure 1. This particular impact of valence-inconsistent information on positive stereotypes (e.g., the brutal midwife) can be interpreted as a classic negativity bias (Rozin and Royzman, 2001) or a contrast effect due to shifting standards (Biernat and Manis, 1994; Biernat et al., 1998).

The significant interaction effect rather suggests that the net evaluation of the target persons simply depended on the total amount of positive or negative information provided, which is in line with additive accounts of information integration (e.g., Anderson, 1962, 1971): the higher the number of positive attributes ascribed to a target person, the more positive the evaluation of the target person.

The fact that consistency did not influence judgments positively is intriguing because an extensive body of research has shown effects of consistency on preference (e.g., Topolinski and Strack, 2009b; Winkielman et al., 2012). This can be due to at least two factors. First, the affect elicited by the strong stereotype valence of the present stimulus pool (cf., Fazio and Olson, 2003) might be much stronger than the affect elicited by consistency-driven fluency (cf., Russell, 2003), since fluency-driven affective variations are very subtle in nature (Winkielman and Cacioppo, 2001; Topolinski et al., 2009). The affect feeding into the evaluative judgment would then be driven by the stronger affect elicited by stimulus valence, overriding the weaker affect elicited by consistency-driven fluency. This issue will eventually be addressed in Experiments 5 and 6.

Secondly, in the present set-up, the consistency manipulation itself was confounded with valence, because the inconsistent attribute was selected from opposite-valence stereotypes. Thus, the inconsistent attribute was not only semantically inconsistent with the given stereotype, but also evaluatively. In the critical case of a negative stereotype with an inconsistent (positive) attribute, these two forms of inconsistency (semantic and affective) work against each other. The negative affect that should be elicited by the inconsistency-driven lack of processing fluency is countered by the positive affect elicited by positive valence of the attribute itself.

To address this latter issue, Experiment 2 used a semantic consistency manipulation free from confounding affective consistency, by using inconsistent attributes of the same valence. For example, lazy is not part of the hooligan stereotype, but still negative. Thus, we kept the net target valence constant (both the stereotype hooligan and the inconsistent attribute lazy are negative). This more subtle affective variations induced by mere semantic consistency might be more likely to feed into preference judgments, thus making an effect of consistency-driven fluency more likely to emerge.

The 2 (valence: positive, negative) × 2 (consistency: consistent, inconsistent) repeated measures ANOVA on the preference judgments yielded a main effect of valence, with more positive evaluations of target persons from positive (M = 5.91, SD = 0.83) compared to negative stereotypes (M = 2.23, SD = 0.71), F(1,43) = 344.10, p < 0.001, ηp2 = 0.89, 90% CI = 0.83–0.92. Consistency had no effect, F(1,43) = 0.59, p = 0.45, ηp2 = 0.01. There was a tendency toward an interaction between valence and consistency, F(1,43) = 2.74, p = 0.11, ηp2 = 0.06; however, simple comparisons showed no significant differences [positive valence: t(43) = 1.64, p = 0.11; negative valence, t(43) = 0.47, p = 0.64]. Findings are illustrated in Figure 2.

Consistency did not influence preference judgments for valenced stereotypical target persons. This was the case even though the present purely semantic consistency manipulation was designed completely independent from affective matching (in contrast to Experiment 1), and involved a larger number of inconsistent attributes in the inconsistent condition (three of five instead of one of five attributes in Experiment 1). Moreover, the manipulation check (using an independent sample of participants) confirmed that the present attribute sets effectively manipulated coherence.

Apart from the possibility already mentioned in Experiment 1, that stimulus valence constitutes a stronger cue than consistency and therefore solely influences participants’ evaluations, another possible explanation is that the presented stereotype label simply overshadowed the information given by the attributes. Specifically, it is not clear whether the current evaluative judgments draw on the valence of the stereotype label or on the net valence of the attributes, or both. In terms of the continuum model by Fiske and colleagues (Fiske and Neuberg, 1990; Fiske et al., 1999; Fiske, 2012), the (in)consistent attributes are individuating information, and inconsistency should have motivated piecemeal processing of the single attributes in inconsistent trials. In consistent trials, participants should have relied on category information when making their evaluative judgment (Fiske, 2012). The lacking impact of coherence might suggest that inconsistency was not sufficiently motivating participants to process individuating information but to rather rely on the provided stereotype label. To investigate this possibility, Experiment 3 changed the experimental set-up. Targets were presented without stereotype labels, thus forcing participants to use individuating information in forming their judgments.

The 2 (valence: positive, negative) × 2 (consistency: consistent, inconsistent) repeated measures ANOVA on the preference judgments again found only a main effect of valence, F(1,41) = 231.22, p < 0.001, ηp2 = 0.85, 90% CI = 0.77–0.89, with more positive evaluations for positive (M = 5.78, SD = 1.00) compared to negative stereotypes (M = 2.02, SD = 0.76), and still neither an effect of consistency, F(1,41) = 0.12, p = 0.72, ηp2 = 0.00, nor an interaction between consistency and valence, F(1,41) = 0.72, p = 0.40, ηp2 = 0.02. Results are illustrated in Figure 3.

Again, apart from a strong valence effect of implied stereotype, no effect of semantic consistency was found, although in the current set-up, no stereotype category labels were shown and thus participants had to rely on piecemeal processing of individuating information of the attributes (e.g., Fiske, 2012). This lack of a consistency effect stands in contrast to earlier findings on coherence-driven preference (Gordon and Holyoak, 1983; Newell and Bright, 2001; for reviews, see Winkielman et al., 2012). For instance, in Topolinski and Strack (2009a, Experiment 3) participants were presented with three semantically consistent (e.g., DEEP FOAM SALT, implying SEA) or inconsistent (e.g., DREAM BALL BOOK, no common associate) words and were asked for their overall evaluation of these word triads, which resulted in more positive evaluations of coherent compared to incoherent word triads. In these experiments, however, the evaluation task was not framed as an impression formation task. Instead, participants incidentally read over words that were not framed as a particular entity or compound (see also Gordon and Holyoak, 1983; Newell and Bright, 2001; Topolinski et al., 2009).

Thus, to realize an experimental set-up as close as possible to earlier consistency-preference findings (Gordon and Holyoak, 1983; Newell and Bright, 2001; Topolinski and Strack, 2009b), Experiment 3 was replicated without an impression formation mind-set. For this, Experiment 4 left out the impression information instruction and first names of the target persons.

The 2 (valence: positive, negative) × 2 (consistency: consistent, inconsistent) repeated measures ANOVA on the preference ratings for the word groups again found a main effect of valence, F(1,27) = 382.00, p < 0.001, ηp2 = 0.93, 90% CI = 0.88–0.95, with more positive evaluations of groups of positive attributes (M = 6.04, SD = 0.76) than of groups of negative attributes (M = 1.85, SD = 0.59). But still neither an effect of consistency, F(1,27) = 0.61, p = 0.44, ηp2 = 0.02, nor an interaction between consistency and valence emerged, F(1,27) = 1.71, p = 0.20, ηp2 = 0.06. Findings are illustrated in Figure 4.

Experiment 4 still found no impact of consistency on evaluative judgments, even though only semantically consistent and inconsistent word groups were evaluated. Moreover, the manipulation check in Experiment 2 shows that these word groups do indeed constitute a consistency manipulation. This shows that the lack of consistency effects in the previous experiments is not due to a motivation to make an unbiased judgment elicited by impression formation instructions.

In contrast to earlier successful manipulations (Mendes et al., 2007, who let participants interact directly with the target persons) the present consistency manipulations might have been too subtle. Earlier manipulations used material that was less strongly valenced (e.g., Mendes et al., 2007 used nationalities). As already speculated in the discussion of Experiments 1 and 2, it is possible that the strong stereotype valence affected evaluative judgments so strongly that consistency-driven fluency might be unable to modulate the effects on top of that. Thus, the final experiments use rather neutral stereotypes.

A 2 (consistency: consistent, inconsistent; within) × 3 (valence: negative, neutral, positive; within) repeated measures ANOVA on the preference ratings of the targets found a main effect of valence, F(2,226) = 2231.00, p < 0.001, ηp2 = 0.95, 90% CI = 0.94–0.96. Post hoc tests confirmed that positive sets were evaluated more positively than neutral sets, which, in turn, were evaluated more positively than negative sets (all ps < 0.001). There was no main effect of consistency, F(1,227) = 0.33, p = 0.57, ηp2 < 0.01, and no interaction of consistency and valence, F(2,226) = 0.32, p = 0.72, ηp2 < 0.01. Results are illustrated in Figure 5.

Manipulating the consistency of strongly valenced and neutral stereotype attributes, we again found that consistency of word groups did not influence evaluations. Thus, consistency did not affect preference judgments even for neutral stereotypes, whereas previous research with moderately valenced material did find consistency to influence preferences (Mendes et al., 2007; Topolinski and Strack, 2009a).

However, in Experiment 5, neutral stereotypes were randomly intermixed with strongly valenced stereotypes. Thus, the neutral stereotypes were presented in a strongly valenced context, and this experimental context could still have masked a consistency effect. Accordingly, we performed one more experiment using only neutral stereotypes. Without strong valence, subtle valence effects, such as consistency-induced fluency, should have a better chance of influencing judgments.

A 2 (consistency: consistent, inconsistent; within) × 2 (label: with stereotype label, without stereotype label; between) mixed model ANOVA on the preference ratings of the targets found a main effect of label, F(1,202) = 7.60, p = 0.01, ηp2 = 0.04, 90% CI = 0.01–0.09, with attributes presented without a stereotype label being evaluated more positively than attributes presented with a label. The main effect of consistency was marginally significant, F(1,202) = 3.07, p = 0.08, ηp2 = 0.015, 90% CI = 0.00–0.06. Consistent stereotypes tended to be evaluated more positively than inconsistent stereotypes. There was no interaction of consistency and label, F(1,202) = 0.69, p = 0.41, ηp2 < 0.01. Results are illustrated in Figure 6.

Manipulating the consistency of attributes of neutral stereotype exemplars, we found that consistent target persons were liked more than inconsistent ones. However, this consistency effect on preference was only marginally significant and very small.

From a more general perspective of affective-cognitive undercurrents of preference, the current evidence is highly informative concerning the relative contribution of different sources of affect; specifically the relative contribution of subtle affect due to fluency compared to stimulus-inherent valence. To our knowledge, the current study is the first to orthogonally manipulate stimulus valence and stimulus fluency (see also the mixed evidence on stimulus valence and repetition as fluency induction, Grush, 1976; Dijksterhuis and Smith, 2002; Schellenberg et al., 2008). Although processing fluency has been shown to be a pervasive influence on a variety of judgments (Reber et al., 2004), and semantic consistency has been shown to be a powerful tool to effectively manipulate fluency (Whittlesea, 1993; Topolinski and Strack, 2009a,b,d), in the current studies consistency-driven fluency had only a small impact on preferences for neutral stereotypes (and this only under optimal experimental conditions), but no impact at all on valenced stereotypes.

This clearly shows a boundary condition of fluency effects: when the stimulus itself is strongly valenced and socially meaningful, the affective variations triggered by fluency dynamics (Winkielman and Cacioppo, 2001; Topolinski et al., 2009) are too subtle to influence preference judgments. This is probably due to the fragile experiential status of fluency-variations. Only strong variations of fluency can be experienced consciously, and only under certain conditions of awareness (for perceptual fluency, see Reber et al., 2004; for semantic fluency, see Topolinski and Strack, 2009d). When additional phasic affective input is too strongly valenced, it seems to overshadow the subtle fluency affect.

In a more general sense, the present research examined the influence of a biasing factor in the context of more relevant information. In several domains, biases have been found to be reduced if more relevant information is available. For example, the trustworthiness of a person’s face influences behavior toward that person to a lesser degree when information about the person’s relevant behavior is available (Chang et al., 2010). This is highly adaptive—behavior should be most influenced by the most relevant information. In the present research, we examined the same idea. We examined whether consistency-induced fluency, a biasing factor in many laboratory paradigms, influences judgments about people even when more diagnostic information is available (here, traits). With strongly valenced information, the influence of consistency completely vanished. Even for neutrally valenced information, consistency influenced preference judgments at best to a very small degree. Thus, it seems that consistency-induced fluency’s role in applied settings is small. While this does not invalidate fluency as a cognitive mechanism, its role in applied social settings with rich information available seems minor at best.

How information from different stimuli is integrated into evaluative judgments is explained by early additive accounts of information integration (e.g., Anderson, 1962, 1971; for a review, see Massaro and Friedman, 1990). Additive accounts postulate that the overall evaluation of a target person is simply determined by the evaluation of single pieces of information about the target person that are then averaged. In a pioneering study, participants evaluated hypothetical persons on the basis of a random set of three adjectives they learned about the person (e.g., good-natured, bold, humorless). All adjectives had been pre-rated for their valence. The evaluation of the person could be predicted by the arithmetic mean of the evaluation of the individual attributes (Anderson, 1962). Later, additive accounts were refined by a weighing factor, assigning weight to the value of each piece of information depending on its importance or salience (Anderson, 1971), or due to motivational reasons. For instance, negative information was shown to be weighed stronger in evaluative judgments than positive information, because negative information can have a warning function (negativity bias, for a review, see Rozin and Royzman, 2001).

Most of the evidence presented here can be accounted for by additive theories of information integration (e.g., Anderson, 1962, 1971). Experiments 1–5 (and particularly Experiment 1 confounding information valence with consistency) showed that the higher the number of positive attributes presented with the target person, the more positive was the evaluation of the target person.

In contrast, in Experiment 6 evaluative judgments differed contingent on consistency, even though stimulus valence was constant between consistent and inconsistent target persons. Additive accounts of information integration have difficulties in explaining the differences in the evaluation of consistent compared to inconsistent neutral target persons. As consistent and inconsistent stimuli were sampled from the same pool of attributes, their average valence could not differ. Thus, the observed small difference in the evaluative judgment of consistent and inconsistent target persons cannot stem from differences in stimulus valence. The only difference between the two conditions was the semantic consistency of the attributes presented with the target person, which triggered cognitive feelings of ease above and beyond stimulus valence (cf., Schwarz, 1990, 2002, 2012).

In sum, the present Experiments 1–5 support additive accounts of information integration in an impression formation task using strongly valenced stereotypes. However, the results of Experiment 6 cannot be explained fully by additive information integration. Here, consistency-triggered fluency seem to have influenced preference judgments.

Stereotypes have been shown to be pervasive (Devine, 1989; Devine and Sharp, 2009) and hard to change (Weber and Crocker, 1983). Because activation and application of stereotypes can have negative effects (Stangor, 2009), the attempt to reduce the application of stereotypes or to even change stereotypic representations is one of the oldest and most prominent research aims in social psychology (Allport, 1954; Bodenhausen et al., 2009). One of the factors frequently discussed as a possible means to stereotype change is the confrontation with stereotype-inconsistent behavior or atypical members of stereotyped groups (Allport, 1954; Weber and Crocker, 1983; Tausch and Hewstone, 2010).

Will an individual’s stereotype about hooligans be altered by encountering a child-loving hooligan? This is exactly what theories of intergroup contact would predict and what empirical research on stereotype change has found (Allport, 1954; Kunda and Oleson, 1995, 1997; Brown and Hewstone, 2005; for a review, see Pettigrew and Tropp, 2006; but see for boundary conditions, e.g., Weber and Crocker, 1983; Fyock and Stangor, 1994; Trope and Thompson, 1997). As changing stereotypes on a large scale promises the solution to many intergroup conflicts, it has received immense attention over the last 100 years (Stangor, 2009). However, a widely ignored factor in this equation is the stereotype-inconsistent individual. What does the child-loving hooligan gain or suffer from displaying stereotype-inconsistent behavior? The answer that can be derived from the present findings is twofold. For strongly valenced stereotypes about the group, inconsistent information only had a strong effect on evaluative judgments of the individual when this information was affectively mismatching (Experiment 1).

However, the situation was different when stereotypes were not strongly valenced in the first place. Stereotype-inconsistent individuals from neutral groups tended to be evaluated less positively than stereotype-consistent individuals. For example, a literate butcher tended to be evaluated less positively than a strong butcher (Experiment 6). This shows that the effects of stereotype-inconsistent information that can possibly help to alter stereotypic beliefs about the stereotyped group as a whole can be at the expense of the inconsistent individual (cf., Mendes et al., 2007; Phelan et al., 2008; Rudman et al., 2012). However, as the present inconsistency effects were small and occurred only for neutral stereotypes and under certain experimental conditions, negative consequences for the inconsistent individual should not be overestimated and are likely to be irrelevant in enriched social interactions.