Self-Awareness and Stereotypes: Accurate Prediction of Implicit Gender Stereotyping

Participants

We originally aimed at recruiting 140 to 160 participants at the University of Cologne, Germany, via direct approach, posters, and flyers in return for course credit or monetary compensation. However, our preregistered plans were foiled by the exacerbation of the COVID-19 pandemic, which led to a halt of all data collections in our laboratory in mid-March 2020. At this point, we had yielded a sample size of 90 participants. Given the within-subject design of our study, we decided that this sample size was large enough to test our hypotheses with adequate power. Of the 90 subjects who participated, four people indicated that they were non-native German speakers and were excluded in line with our pre-registration protocol. The remaining 86 participants were aged between 18 and 69 years (median age was 24 years, 73% female, 27% male), and 85 out of 86 participants were current university students.

Design and Measures

Materials consisted of explicit endorsements of stereotypes, followed by explanatory texts about the concept of “implicit attitudes,” IAT score predictions, and IATs. They were the same for all participants except for two minor variations: the block order of the IATs and the layout of the prediction slides varied between participants (see below for details). This resulted in a 2 (IAT block order, IBO) × 2 (female pictures were presented above vs. below the male pictures on the prediction slide) design. As preregistered, we did not expect these minor differences in the layout and procedure to influence the results.

Predictions of implicit stereotyping

A sample prediction slide, translated to English, is shown in Figure 1. Participants were asked to predict how they would score on IATs that assessed associations of women and men with the five attribute pairs that were also compared on the explicit ratings. Participants saw all words that would be used on the IATs in word clouds to the right and left on top of the prediction slides, as well the 5 pictures each of men and women that would be used on the IATs in two rows in the top center of the screen. The exact words can be found in the appendix. We carefully chose unambiguous, nonnegated, evaluatively balanced, and grammatically gender-neutral words that are commonly used in German. Whether pictures of men or women were shown on top or below on the prediction slides was counterbalanced across participants, but was constant within each participant. Which of two contrasted concepts would show on the left or the right of the screen was also counterbalanced across participants. For half of the participants, the stereotypical response would be on one side in three of the predictions (language vs. mathematics, arts vs. sports, family vs. career) and on the other side for the other two (self- vs. other-interest, rational vs. emotional). For the other half of the participants, this was exactly reversed. We implemented these variations to increase vigilance and variance in participants’ predictions, to prevent participants from making the same constant prediction across the five IATs (leaving no variance in the predictions that could be used for analysis), and to make sure results were independent of these variations. Hence, there were four different versions of the design (pictures of men above vs. below pictures of men, and attributes on the right vs. the left). Participants completed one prediction for the dog-cat independent-loyal IAT and then completed a short version of the IAT. They completed the remaining predictions only after these explanations and examples.

OPEN IN VIEWER

Figure 1.

A sample prediction slide, translated to English from German.

Procedure

Participants were seated at individual cubicles with stationary computers. After signing informed consent, participants were randomly assigned to one of the four methodological conditions (IBO and sidedness). Participants started the study with the explicit ratings, completed in orders that were individually randomized for each participant. Next, they received instructions and completed a prediction task with a subsequent mini dog-cat-IAT for practice, and to understand how an IAT measures stereotyping. This was followed by the prediction procedure that ended on five predictions for five IATs, also completed in individually randomized orders. Participants then went on to complete the five IATs (in new randomized orders), before they repeated the explicit ratings (once again in new randomized orders). The instruction for the second explicit ratings said that we were interested in whether participants’ attitudes had changed after completion of the IATs. Finally, participants answered some demographic questions and had the option to provide feedback in free-response questions. Upon completion, participants were debriefed and reimbursed.

Standardization of variables

The dependent variables and all predictor variables that were entered into the multilevel analyses were within-person standardized. Standardizing the variables beyond just centering them allowed us to look only at the accuracy with which participants predicted the patterns of their implicit stereotyping scores, independent of the variation in their scores and predictions. ³ In addition, this standardization simplifies the interpretation of the regression coefficients: the regression weights of the models’ fixed effects of predictions correspond to the Pearson correlation coefficients between the IAT score predictions and the IAT scores per person. As a result of the standardization, all intercepts were zero and removed from the regression models (no-intercept models).

IAT Scores

On average, participants were faster in the stereotype-congruent IAT-blocks than in the stereotype-incongruent blocks across all five stereotype domains. The IAT-D-scores, scored such that higher numbers meant stronger associations of the first attribute with men and the second with women, were: D = –0.39 for emotional vs. rational, D = 0.39 for sports vs. arts, D = 0.33 for career vs. family, D = 0.28 for mathematics vs. language, and D = –0.21 for other-interest vs. self-interest. According to common conventions, these implicit gender stereotyping scores can be labeled as slight to moderate.

Awareness

To estimate the prediction accuracy for individual participants across the five different IATs, we regressed the person-standardized IAT-D-scores onto the predicted IAT scores in a multilevel model with all data points nested within each individual. The multilevel models included a random slope for each participant to account for Level 2 variance in participants’ accuracy. The results from the multilevel model can be seen in the first column of Table 1.

OPEN IN VIEWER

Table 1.

Multilevel Regression Models Predicting IAT-D-Scores From IAT Score Predictions and Explicit Stereotyping.

Predictors	Prediction model estimates	Implicit–explicit model estimates	Simultaneous model estimates
Fixed effects
IAT score predictions	0.63 (0.04)***		0.60 (0.07)***
Explicit ratings		0.52 (0.05)***	0.07 (0.07)
Random effects SD
Residuals	0.69***	0.76***	0.64***
IAT score predictions	0.13		0.30**
Explicit ratings		0.18	0.41***
Observations
No. of obs.	430	430	430
No. of groups: participants	86	86	86
Goodness of fit
Marginal R2/Conditional R2	0.393/0.410	0.250/0.280	0.401/0.494
AIC	919.77	1,009.76	913.63
BIC	931.96	1,021.95	938.01
Log-likelihood	–456.88	–501.88	–450.82

Note. The dependent variables in all models are IAT-D-scores. All models are random-slopes no-intercept models estimated with the REML method. All variables are standardized for each individual participant, rendering all intercepts zero. This way, regression coefficients can be interpreted like Pearson correlation coefficients in one-predictor models, or part-correlations in simultaneous predictions, respectively. Standard errors are in parentheses. IAT = Implicit Associations Test; AIC = Akaike information criterion; BIC = Bayesian information criterion.

**

99% Confidence Interval does not include zero.

***

99.9% Confidence Interval does not include zero.

As can be seen, the average prediction accuracy was significant and large, b = 0.63, 95% CI [0.56, 0.71], t(83.00) = 15.80, p < .001. As described before, the size of the regression weight corresponds to the average correlation coefficient between predicted and actual IAT scores. The standard deviation of the random slopes was small, τ = 0.13 and nonsignificant, 95% CI [0, 0.26]. This indicates that the best fitting slopes were similar across participants.

Because the distribution of the correlations per person is right-skewed, with many high correlation coefficients and few low or negative correlation coefficients, we also computed correlations manually for each participant and looked at the resulting distribution across participants. The mean of the distribution was equal to the regression weight of the multilevel model, and the median was r = 0.76. ⁴ Fisher-z-transforming these values revealed a mean value of 0.96 which back-transforms into a correlation coefficient of .75, closely matching the median accuracy of .76.

Neither sidedness (of men vs. women in both the predictions and the IATs), nor IBO showed any significant impact on prediction accuracy. Predictors for IBO, sidedness, and their interaction, did not interact with prediction accuracy; all Fs < 1, all ps > .27, all R² ~ 0.

Explicit–Implicit Correspondence

Theoretical models claim that explicit and implicit measures capture the results of two different attitude retrieval processes, namely propositional and associative ones (Gawronski & Bodenhausen, 2006). If an attitude target activates another concept in the associative network but is propositionally rejected, implicit and explicit measures differ. A correspondence between IAT score predictions and actual IAT scores can only be interpreted as evidence for awareness if the results of implicit measures cannot be explained by explicit measures to the same degree. Hence, explicit measures need to be less predictive of IAT results than IAT score predictions to justify the conclusion that they capture different information, and that the spontaneous mental processes captured in IAT scores are often rejected for traditional explicit ratings. In contrast, the relationship between explicit measures and IAT scores should be explained entirely by IAT score predictions. As a result, a simultaneous regression where IAT scores are predicted by both IAT score predictions and explicit ratings should show no noticeable reduction in the predictive power of IAT score predictions, while the predictive power of explicit ratings should drop substantially.

To test these points, we ran the same multilevel analyses described previously, replacing IAT score predictions with explicit ratings as the main independent variable. This model revealed a significant correspondence between the patterns of participants’ explicit scores and their IAT-D-scores on a within-person-level, b = 0.52, 95% CI [0.43; 0.62], t(77.00) = 10.98, p < .001 (see Table 1, Column 2). However, when both predictors were included into the model, only the IAT-predictions, but not the explicit stereotype ratings, significantly explained variance in the dependent variable, IAT score predictions: b = 0.60, 95% CI [0.44; 0.74], t(46.06) = 8.82, p < .001, explicit ratings: b = 0.07, 95% CI [–0.07, 0.21], t(56.18) = 0.90, p = .37 (see Table 1, Column 3). Confirming predictions, the predictive power of IAT score predictions remained almost unchanged with the inclusion of explicit measures, whereas explicit measures were no longer predictive of IAT scores when IAT score predictions were included in the model (compare different columns of Table 1). Gauging principles of parsimony and goodness-of-fit, the model including IAT score predictions as the only predictor prevails over the other models (cf. Table 1).

Do Participants Have Insight Into Their Own Unique Stereotyping Patterns, or Do They Draw Upon Shared Knowledge?

Which aspects of their IAT scores did participants predict? As outlined above, we programmed two of the gender stereotypes such that two of the IATs would always be completed in opposite order concerning stereotype-consistency (participants would complete the stereotype-consistent sorting first and the stereotype-inconsistent critical block second for three of the IATs, but reversed for the remaining two, or vice versa; see materials section), and this order was matched by the prediction slides such that higher numbers on the scales would match with higher numbers on the IATs. In addition to the advantages of this procedure in creating variance and increasing attention by switching sides as described above, however, this scoring allows participants to predict a significant portion of their scores based on just echoing the commonly known stereotypical assignment of traits, without any particular knowledge about the specific strength of these associations in their own minds. That is, by resorting to common cultural knowledge on gender roles, any participant should be able to correctly predict the sign of their IAT-D-scores as positive or negative (i.e., they will pair “men” with “rational” and “women” with “emotional” more readily than “men” with “emotional” and “women” with “rational”), and given these signs were reversed for two of the IATs compared with the other three, such a prediction would already look reasonably accurate in our models. However, if participants were able to gauge the idiosyncratic strength of these gender associations in their own minds, they should be able to predict variance beyond this particular contrast. We were interested to disentangle the relative contribution of these two aspects that both contribute to the prediction accuracy: direction (shared) and strength (idiosyncratic) of their automatic response. We conducted three analyses to investigate this point. They were not preregistered.

First, we ran another multilevel regression model including a binary variable “stereotype-consistent.” It coded those IATs where higher numbers indicated less stereotypical associations (emotional/rational and other-interest/self-interest) as –3 and the others as 2. This predictor captures the proportion of accuracy that can be explained by indicating the stereotype congruent direction of the implicit score. Thus, the remaining, partial regression weight for the predicted implicit stereotyping only captured if nuances beyond the mere direction of the stereotype were accurately predicted. The predictor for stereotype consistency was significant, b = 0.19, 95% CI [0.13, 0.25], t(8.09) = 6.36, p < .001. Importantly however, even when accounting for mere directionality of the implicit scores, participants’ IAT score forecasts remained a significant predictor, b = 0.21, 95% CI [0.09, 0.33], t(308.99) = 3.39, p < .001. This indicates that even when controlling for stereotype congruence, participants correctly predicted the relative strength of the respective gender stereotypes within their own patterns of scores. Hence, the strategy to simply reproduce existing cultural gender stereotypes cannot fully explain the high prediction accuracy that we found in our data.

As a second test of how much insight participants had into their own unique patterns beyond intersubjectively shared patterns, we looked at how well any participant’s predictions would explain any other participant’s IAT scores. Specifically, if all participants simply predicted the patterns of IAT scores based on the same knowledge of the same common stereotypes, this time including variance beyond a simple contrast, then any participant recruited from the same population should be as accurate in predicting any other participant’s IAT scores as they are accurate in predicting their own scores. Based on this reasoning first presented by Hahn et al. (2014), we expanded and refined an analysis the authors presented in their paper. Specifically, we ran a simulation in which every participant’s predictions now predicted the IAT results of a random other participant of the sample. This procedure was repeated for 1,000 iterations. For each of these 1,000 permutations, we ran a multilevel regression model and calculated the average regression weight of each iteration for the predicted IAT scores. The multilevel models were specified exactly like the Prediction Model (cf. Table 1, Column 1) that is described in detail above.

The average accuracy of all permuted datasets was b = 0.56. The highest estimate across the 1,000 permutations was b = 0.62, while the lowest was b = 0.50. This suggests that a large part of the predicted patterns could be explained by knowledge that was shared between participants and not idiosyncratic to any particular participant’s pattern of associations, even though the prediction accuracy remained lower in all 1,000 iterations than the prediction accuracy in the original dataset with participants’ own predictions, b = 0.63. Importantly, however, participants’ own predictions outperformed the random other predictions in all of 1,000 iterations.

Specifically, we ran another 1,000 iterations with random pairings, but this time, every participant’s IAT scores were regressed onto their own predictions and the predictions of a random other participant simultaneously. Results showed that participants’ own predictions outperformed the predictions of random others in 100% of the cases. That is, in all 1,000 cases of random permutations the regression, weight b was larger for the participant’s own predictions than for a random other person’s prediction, with the difference ranging from .002 to .441. On average, participants’ own predictions had a regression weight of b = 0.46, 99% CI = [0.40, 0.54], whereas the regression weight of the random other participant’s prediction over and above each participant’s own prediction was b = 0.25 on average, 99% CI = [0.14, 0.36]. Again, this demonstrates that while there is a shared pattern in every participant’s IAT scores that can be applied across participants, participants also have unique insights into the patterns of their own scores beyond this intersubjectively shared pattern, such that their own predictions always outperform the predictions made by random others on average, even in 1,000 pairing iterations.

Together, these three analyses show that although a notable proportion of the patterns of people’s IAT scores were shared across participants, echoing recent theorizing that automatically activated biases reflect cultural patterns (Payne et al., 2017), participants seemed to have unique insights into their own unique patterns of implicit gender stereotyping.

Social Calibration

Previous studies sometimes concluded from low between-subjects correlations a lack of awareness of automatic attitudes (e.g., Nosek, 2005, 2007) As argued before (Hahn et al., 2014; Hahn & Goedderz, 2020), this level of analysis measures not only conscious accessibility of mental processes but rather the social calibration, that is, the intersubjective agreement of how to label the strength of an attitude. That is, in order for a between-subjects correlation between predicted and actual IAT scores to be high, the person with the strongest stereotype would have to know to make a more extreme prediction than the rest of the sample. This requires not only knowing one’s spontaneous stereotypes but also the ability to calibrate one’s strength in comparison to the remainder of the sample. In contrast, in order for within-subjects correlations to be high, participants need only know which of their stereotypes is stronger than which other stereotype (see Hahn et al., 2014; Hahn & Goedderz, 2020). To demonstrate the conceptual difference between social calibration and subjective awareness, we ran between-subjects analysis to assess the correlation between predicted IAT scores and actual IAT-D-scores. All predictions and IAT scores were standardized by IAT type. As a first overall analysis, we computed a multilevel-model regressing IAT-D-scores onto IAT score predictions by IAT type, controlling for IBO as a covariate. The overall between-subjects calibration was β = 0.24, 95% CI [ 0.14, 0.33], t(70.01) = 4.93, p < .001. All random effects were negligibly small, indicating that this average calibration accuracy did not vary meaningfully for the five IATs. The individual part-correlations per target pair can be seen in Table 2. These moderate-to-low relationships are in line with previous research documenting lower levels of social calibration than awareness for implicit bias measures (Hahn et al., 2014; Hahn & Goedderz, 2020).

OPEN IN VIEWER

Table 2.

Social Calibration: Between-Subjects Correlations of Predicted IAT Results and Actual IAT Scores.

	Arts vs. sports	Family vs. career	Mathematics vs. language	Self-interest vs. other-interest	Rational vs. emotional	Average between-subject accuracy
β weight	0.17	0.24*	0.17	0.34**	0.21 †	0.24***
95% CI	[–0.06, 0.39]	[0.02, 0.45]	[–0.05, 0.38]	[0.13, 0.54]	[0.00, 0.42]	[0.14, 0.33]

Note. Regression weights for participants’ IAT score predictions (standardized by IAT type). The dependent variable is the IAT-D-score. To control for block order effects, we also included a binary variable coding whether participants completed the IAT with the stereotype-consistent block or stereotype-inconsistent block first. These weights are not reported, none of them were significant, all Fs < 1, all ps > .99. Due to standardization, the regression weights correspond to the part-correlations between predicted and actual IAT scores. Last column shows the average prediction accuracy (β-weight of predicted IAT score in the multilevel model with IAT type as random effect). IAT = Implicit Associations Test; CI = confidence interval.

†

p = .051. *p < .05. **p < .01. ***p < .001.

Adjustment of Explicit Stereotyping After Completing the IATs

As a last analysis, we conceptually replicated Hahn and Gawronski’s (2019) acknowledgment of bias effects in the domain of gender stereotyping. We ran a multilevel regression model with random slopes testing whether the patterns of explicit stereotypes changed to be more in line with implicit stereotyping scores after predicting and then completing the five gender stereotyping IATs. The IAT-D-scores as well as the explicit ratings before completing the IATs were entered as predictors for the post-prediction explicit stereotyping ratings. The results can be seen in Table 3.

OPEN IN VIEWER

Table 3.

Multilevel Regression Model Measuring How Much Explicit Stereotyping Changed After Completion of the IATs as a Function of IAT-D-Scores.

DV: Explicit Stereotyping Post
Predictors	Estimates
Fixed effects
Explicit stereotyping pre	0.67 (0.04)***
IAT-D-score	0.20 (0.03)***
Random Effects SD
Residuals	0.49***
IAT-D-score	0.09
Explicit stereotyping pre	0.13
Observations
No. of observations	430
No. of groups: participants	89
Goodness of fit
Marginal R2	0.661
AIC	637.64
BIC	662.03
Log-likelihood	–312.82

Note. *** 99.9% Confidence Interval for the effect does not include zero. All variables are standardized for each individual participant. Standard errors are in parentheses. AIC = Akaike information criterion; BIC = Bayesian information criterion.

Participants generally demonstrated stability in the explicit stereotyping they reported. This can be seen in that the post-procedure explicit stereotypes were strongly related to the same stereotypes reported before the prediction and IAT completion procedure, b = .67, 95% CI = [0.60; 0.74], t(69.69) = 18.74, p < .001. However, a significant slope of the IAT value showed that beyond this stability, participants also adjusted their explicit stereotypes significantly toward the cognitions reflected on the IATs, b = .20, 95% CI = [0.14; 0.27], t(89.21) = 6.15, p < .001. Participants did not receive any feedback on their performance in the IAT. Hence, these results replicate Hahn and Gawronski’s (2019) findings that predicting one’s IAT scores leads to alignment of explicit with implicit measures.

Which Aspects of the IAT Score Patterns Did Participants Predict?

We interpret our results such that the IAT score predictions captured both shared normative patterns of gender stereotyping and unique knowledge of one’s own idiosyncratically learned gender associations. That is, on one hand, to increase vigilance in our participants and variance in IAT scores, the design of our study was such that knowledge of culturally normative patterns of stereotyping would already lead to accurate predictions. On the other hand, however, mere reproduction of normative stereotype patterns could not explain prediction accuracy in its entirety. Specifically, in our analyses, we consistently found that the variance in IAT scores that could be explained by participants’ predictions consisted of two portions. One portion was intersubjectively shared across participants. Thus, every participant’s predictions were informative about the patterns of IAT scores of a random other participant, and a binary predictor coding just the direction of the stereotypes explained a sizable proportion in IAT scores and prediction accuracy. At the same time, however, another portion was person-specific. Specifically, predictions of a random other participant were always less predictive of a person’s IAT scores than their own predictions in 1,000 random pairings of participants. Own predictions furthermore out-predicted both random other participants’ predictions in all of 1,000 pairing iterations, as well as a binary predictor for stereotype consistency.

Interestingly, while this last analysis confirmed that a portion of every participant’s accurate prediction was idiosyncratic, the portion that was shared was higher than we expected, and higher than what Hahn et al. (2014) found in their studies on intergroup prejudice. One explanation for this large amount of shared variance in stereotyping patterns might be restricted between-subjects variance. Our participants consisted of a homogeneous student sample with similar cultural backgrounds. Given that stereotypes are likely largely culturally learned, and demographic characteristics are known to predict results on indirect attitudinal measures (Nosek et al., 2007; Payne et al., 2017), this cultural homogeneity might have caused our sample to produce similar patterns of automatic responses on the IATs. Further research is needed to confirm if indeed random permutations will result in less-accurate other-predictions for more diverse samples. Taken together, the results suggest that the prediction patterns have both a culturally shared component and an individual component.

Another question concerns the processes that allowed participants to detect and quantify their own stereotyping patterns. Hahn et al. (2014) as well as Hahn and Gawronski (2019) argue that the cognitions reflected on IAT scores produce spontaneous affective reactions to which people must pay attention to notice their own biases (see also Goedderz & Hahn, 2022; Hahn & Goedderz, 2020; Phills et al. 2020; see also Kurdi et al., 2019) demonstrated that group-stereotyping scores and group evaluations are related, such that more negative evaluations predict more negative stereotyping, and more negative stereotyping predicts more negative evaluations. Hence, one may wonder whether people simply predicted their stereotyping scores based on how positive or negative their reactions toward women and men were. The design of the study with predictions of five different stereotyping IATs precluded this possibility.

Specifically, to predict an accurate pattern of stereotyping based on valence alone, a person would have to either always prefer the stereotypically female or the stereotypically male attribute. This may have been true for individual participants, but unlikely for enough participants to achieve a corrected mean accuracy coefficient of r = .75, considering different sources of measurement error in predictions and IAT scores, and considering the attribute pairs were chosen to be relatively neutral. That is, we assume that most people believe one should balance emotion and reason; that both arts and sports, and language and math skills are important; and that life is about balancing self and other-interest, and family and career aspirations. While individual participants may exist who prefer all of the attributes they associate with men (e.g., reason, sports, self-interest, careers, and math skills) over all of the attributes they associated with women (e.g., emotion, arts, other-interest, family, and language skills), and some of those participants may also show a preference for men over women, this is unlikely to be true for all participants. And even if it were, it could not explain why we found patterns of stereotyping scores beyond positive-negative. That is, rather than finding binary sets of equivalent scores where women were always associated with one (presumably positive) attribute to the exact same degree on all IATs, we found nuanced patterns of both predictions and IAT scores, and these predictions were highly accurate.

Group evaluations and stereotyping are related to the degree that the attributes in stereotypes are evaluative (Phills et al., 2020). In our study, too, participants with stronger preferences for women over men would certainly be more likely to stereotype women as emotional over rational to the degree that they consider being emotional more positive than being rational. The question, however, was whether they knew how much they associate women with emotion as opposed to other-orientation, family, language, and arts. The affective reaction to women alone wouldn’t allow for this prediction. To obtain accurate pattern predictions, participants had to predict the portion of variance that goes beyond comparing affective reactions to women with affective reactions to men. Hence, these affective reactions alone would not suffice to make these predictions with the degree of accuracy that we found.

This invites the question of how participants predicted their patterns of semantic associations. One possibility would be to construe the reaction upon which participants’ predictions are made more broadly as a general intuition, as the first thought that comes to mind. In this case, the process upon which people base their responses would not be a spontaneous affective reaction, but an observation of any spontaneous thought or association with gender, evaluative or not.

Another possibility is that participants might have derived their insights into their automatic reactions from the fluency that certain word-image pairs generated. Fluency creates a sense of positivity (Claypool et al., 2015; Wänke & Hansen, 2015). This could mean that stereotypic combinations would create more positive feelings than counter-stereotypic combinations. Alternatively, participants may have inferred from perceptions of fluency which combinations would be easier to complete on the IATs. Interestingly, this latter interpretation would indicate that affective reactions may have played a role in accurate semantic pairing after all, albeit in a less straight-forward way than in predictions of prejudice scores.

At present, both interpretations are compatible with the present data and the data presented by Hahn and colleagues. Participants might either (a) pay attention to the first thought that comes to mind when looking at the concepts or (b) pay attention to which combination “feels” better and more “fluent”; or both. Future research is needed to further clarify the role of affect in observation of one’s own spontaneous thoughts.

In sum, participants in this study were able to predict the patterns of their stereotyping scores, and this demonstrates high levels of awareness of the cognitions reflected on these IATs. That is, it is important to remember that participants readily applied these cultural patterns to themselves. This study did not ask participants what their culture would show or what other participants would show. Instead, we asked them to predict their own results on IATs. They may have drawn upon shared cultural knowledge and other sources to make these predictions. But that does not diminish the finding that they accurately predicted which biases they themselves would show on implicit tests, even when those went against their explicit beliefs. As such, this study provides empirical support for the idea that people can be aware not only of their automatic evaluative reactions but also of their automatic gender stereotypes.

Limitations and Future Directions

This study was intended as a “proof-of-concept” to demonstrate that scores on semantic gender-IATs can be predicted accurately. However, as such, it poses the question of whether these findings generalize to other populations, other types of stereotypes (that are less culturally ingrained or have different valence), and other measures.

Our sample consists of a highly demographically homogeneous sample of university students. We see at least two aspects of this sample that might limit generalizations to other samples. First, young university students might be especially interested in the topic of gender stereotyping, and this may have led them to reflect on these stereotypes more before coming to our study. If this is true, other samples might show lower accuracy in predicting gender stereotyping. Second, the degree to which our participants showed similar patterns of gender stereotyping may be lower in other samples. Stereotypes, as well as cognitions reflected on implicit measures generally, are likely reflective of a person’s surrounding culture (Payne et al., 2017). Accordingly, the degree to which our sample showed overlap in stereotyping patterns might be smaller in more diverse samples. If this is true, the degree of unique insight into one’s own patterns may in fact be larger than our findings indicate. Importantly, however, both of these limitations concern minor aspects of our findings. That is, given that people around the world have ample contact with people of other genders, we believe the general finding that it is possible to predict the unique patterns of one’s own implicit gender stereotyping scores will likely generalize.

Concerning the measures we used, our attempt was to investigate whether people would be able to predict patterns of implicit measures that primarily assess semantic rather than primarily evaluative relationships between concepts. And in doing so, we tried to minimize the affective connotation of our semantic attributes. One may question whether we succeeded—the attribute pairs we chose may not be evaluatively neutral to all participants.

On one hand, we have explained above how we believe the within-subjects mapping of groups to five pairs of attributes require at least some processing of semantic information beyond listening to one’s affective reactions toward social groups. Hence, we believe that predicting patterns of stereotyping with one and the same group is nevertheless an extension over Hahn et al. (2014). On the other hand, if we believe that the attributes chosen in this study were not affectively neutral to all participants, this raises the question of how important the role of affect is even in the mapping of semantic relations in one’s mind. Future research is needed to test if truly valence-neutral stereotypes (e.g., those that are formed in a lab about fictitious social groups) are accessible to awareness as well.

Finally, we extended Hahn et al.’s (2014) findings to stereotyping by leaving many other aspects of the original authors’ paradigm—such as the IAT as the main instrument—similar. Many more and broader extensions to other measures, groups, and domains are necessary to gauge the generalizability of these findings. As it stands, the current findings show that accurate prediction of the patterns of one’s IAT scores is possible in the domain of mere evaluations and semantic stereotypes. Future studies will have to show how broadly this conclusion applies.

Implications

Reframing spontaneously activated biases, stereotypes, and evaluations as accessible to consciousness has important ramifications both in theory and in practice. Theoretically, these findings contribute to a better understanding of implicit measures by adding another piece of evidence that refutes the idea that they capture inaccessible mental content. This is compatible with dual-process models of mental processes (e.g., Fazio, 2007; Gawronski & Bodenhausen, 2006) and may support more precise theorizing on different aspects of automaticity.

The finding that the IAT measures a construct available to conscious awareness has sometimes been used as evidence against the construct validity of implicit measures (e.g., Machery, 2022; Mitchell & Tetlock, 2017). In contrast to such reasoning, we believe that the fact that the IAT can correlate so highly with a self-report measure—given entirely different sources of measurement error—may be seen as support for its validity rather than the opposite (Kurdi et al., 2021). While we agree that there is a lot to be learned about the nature of the cognitions reflected on implicit measures and what they represent, the current studies show that what the IAT measures matches their perceptions of their own spontaneous thoughts. More research is needed to further elucidate the nature of such spontaneous cognitions.

In addition, our findings may be interpreted such that they question the usefulness of implicit measures. One may, after all, apparently just ask participants about their spontaneous reactions. Note, in this respect, that there are still a variety of aspects of people’s implicit stereotyping scores that predictions did not reveal. First, recall that participants were substantially less accurate in ranking the strength of their spontaneously activated biases in comparison to the other participants of their study than compared with their own other biases, as is illustrated in the social calibration analysis. The latter is in line with a battery of findings that people can only poorly rank the strength of their own automatic reactions among others’ automatic reactions (see Hahn & Goedderz, 2020; Hahn et al., 2014). Hence, even though our studies show that people may be able to observe their own gender biases, they seem to have limited knowledge on how much more or less stereotypical their thoughts are compared with the stereotypical thoughts of others.

Second, this study and the studies reported by Hahn et al. (2014) and Hahn and Gawronski (2019) involved confronting participants with a variety of concrete stimuli to reach awareness, and acknowledgment findings indicate that, to many participants, the discovery of their biases was new and unexpected. Together, these findings suggest that there are aspects about people’s spontaneous reactions toward social groups captured by the IAT that do remain outside of awareness, and future research is needed to elucidate factors that enable participants to make accurate predictions (Hahn & Goedderz, 2020).

Finally, the findings on awareness can also be useful in applied settings, for instance, to the design of anti-bias training or to tackle subtle sexist or racist reflexes more generally. In our last analysis, we looked at the acknowledgment effect of awareness and found that our data closely replicated findings by Hahn and Gawronski (2019). Taken together, these results can be interpreted as encouraging. It is much easier to tackle problems of which people can become aware, than to convince someone of harboring an “unconscious” bias they cannot perceive. As such, the current findings support the idea that people should be encouraged to observe their own biases instead of believing that their biases are guided by unconscious forces.