Overview
Over the last several years, the study of implicit bias has taken the world by storm.Footnote 1 Implicit bias was even mentioned by the then candidate, Hillary Clinton, in a presidential debate in 2016. She went on to claim that implicit bias can have deadly consequences when Black men encounter law enforcement (for example, see Correll et al., Reference Correll, Park and Judd2002; Correll et al., Reference Correll, Park and Judd2007; Eberhardt et al., Reference Eberhardt, Goff and Purdie2004). The controversy over police shootings of Black men and women has only intensified as evidenced by public outcry over the murder of George Floyd on May 25, 2020 and increasing public support for the “Black Lives Matter” movement and its calls for liberty, justice, and freedom (Cohn & Quealy, Reference Cohn and Quealy2020). These current events are but one reason why the study of implicit bias has so captivated the attention of the larger public: reducing it seems to have the potential to solve real-world problems. One idea is that if police officers were made aware of their implicit bias or participated in training workshops to reduce implicit bias, then perhaps fewer Black people would end up dead, arrested, or disproportionately sentenced to receive the death penalty (Baumgartner et al., Reference Baumgartner, Epp and Love2014; Eberhardt, Reference Eberhardt2020).
Still, for all of its promise, contemporary scholars examining prejudice continue to struggle with the science of implicit bias. Most fundamentally, these struggles concern not only definitional issues (i.e., what is meant by implicit bias?) but also measurement issues such as how to best capture implicit bias. For example, researchers struggle with the reliability of implicit measures of attitudes, with the lack of correlations between alternate implicit measures of the same attitude, with low implicit to explicit attitude measure correlations (more so in some domains than others), and with the ability of scores on implicit measures to predict discriminatory behavior (e.g., Schimmack, Reference Schimmack2021). Although low correlations with explicit attitude measures are not in and of themselves a death knell (after all, implicit measures were developed precisely because it was thought they could capture something other than what explicit measures tap), the inability to predict or only weakly predict prejudicial behavior is more problematic. Although the link between attitudes and behavior has long been controversial (e.g., Ajzen & Fishbein, Reference Ajzen and Fishbein1977; Wicker, Reference Wicker1969), part of the enthusiasm around the notion of implicit bias was the belief that it would permit researchers to bypass the tendency for people to edit their attitude reports on sensitive issues, tapping into what respondents “really think” about the attitude object. That is, if people wish to avoid being viewed as bigoted, they may be motivated to conceal their prejudice (e.g., Dunton & Fazio, Reference Dunton and Fazio1997). Implicit measures of attitudes, because they are less susceptible to strategic control, were believed by some to be more authentic than their explicit counterparts. But if their ability to predict behavior is limited, this raises concerns about the usefulness of such measures (for examples of their predictive utility, see Greenwald, Poehman, et al., Reference Greenwald, Poehlman and Uhlmann2009; Greenwald, Smith, et al., Reference Greenwald, Smith and Sriram2009; Pasek et al., Reference Pasek, Tahk and Lelkes2009; Payne et al., Reference Payne, Krosnick and Pasek2010; Pérez, Reference Pérez2010; Ziegert & Hanges, Reference Ziegert and Hanges2005; and for evidence of their failure to predict, see Blanton et al., Reference Blanton, Jaccard and Klick2009; Kalmoe & Piston, Reference Kalmoe and Piston2013; Kinder & Ryan, Reference Kinder and Ryan2017).
Critically, in part because of the interest in changing prejudicial behavior, interventions based on current knowledge about implicit bias may be running ahead of the science. There is, at present, a push for anti-bias training (often conflated with implicit bias training) within government and private sector organizations (e.g., the Starbucks anti-bias training, Avila et al., Reference Avila, Parkin and Galoostian2019), when what we need is better evidence about the impact of implicit bias in society and how best to counter or diminish the likelihood of implicit bias operating. That is, we need an evidence base for making recommendations regarding whether and how implicit bias operates and how it can be reduced. Specifically, we need an evidence base for understanding how best to minimize its effects on decision and behavioral outcomes (see Dobbin & Kalev, Reference Dobbin and Kalev2018).
In this report we employ a multi-faceted definition of implicit bias and present relevant examples from the literature suggesting distinct types of implicit bias. We next present a schematic representation of attitudes, implicit and explicit measures of attitudes, and their link to behavior. One of the difficulties around the science of implicit bias is that different researchers use the same terms to mean very different things. The principal goal of the schematic is to make clear our terminology, as well as our working assumptions, so that readers of this report can have a clear understanding of these to help inform their own acceptance or rejection of the claims made here. Next, we provide a brief overview of the way that the field of social psychology has studied implicit bias. We then summarize some important points of agreement within the field regarding implicit bias and implicit measures of attitudes, as well as some unresolved issues.
Based on our understanding of where the science of implicit bias currently stands, we articulate important directions for research going forward. We end with some cautionary notes regarding the application of the construct of implicit bias in popular discourse and policy making, with the hope of allowing the science to catch up to the enthusiasm for the ideas and concerns that have led to so much interest in implicit bias. In the final analysis, researchers, grant funders, and policy makers must not lose sight of the key end goal of measuring, documenting, and ultimately reducing implicit bias – to improve outcomes for people as they navigate their lives in the real world, both as behavioral agents and as members of marginalized groups.
NSF Conference on Implicit Bias
On September 28–29, 2017, NSF convened a meeting to address the current state of knowledge regarding implicit bias as developed largely in social psychology. This report is organized around the key issues addressed at the meeting including: the definition of implicit bias and how it is similar to and different from related concepts such as implicit measures and implicit attitudes, what is known about the phenomenon (i.e., the general consensus in the field), what is uncertain (i.e., divergent views in the field or insufficient research available), and what new research is recommended.
Defining Implicit Measures, Attitudes, and Bias (as Compared to Explicit Measures, Attitudes, and Bias)
Before turning to a review of what is known and what remains to be studied, it is important to define what we mean by the major concepts and distinctions at use in the field of implicit bias. One key distinction is between implicit versus explicit attitudes and implicit versus explicit measures of attitudes. After defining these, we turn to the notion of implicit versus explicit bias.
Implicit versus Explicit Measures
An explicit measure is one that is transparent. It is clear to the person being assessed what is being measured. An implicit measure attempts to assess a belief, attitude, or behavior without the person’s knowledge. For example, an explicit measure of attitudes toward Black people might directly ask the respondent, “Do you like Black people?” An implicit measure of attitudes would aim to assess liking without directly asking the person, such as by measuring how closely the respondent sits next to a Black person. In this sense, contemporary implicit measures are in the same category as classic “indirect” measures, as they require the assessor to make an inference about the construct of interest from some other response (Petty & Cacioppo, Reference Petty and Cacioppo1981). If we think of a continuum of implicit to explicit measures, this continuum would map onto the extent to which people were aware of the fact that the measure was attempting to assess their belief, attitude, or behavior. An implicit measure is one for which people have relatively low awareness of what is being assessed, whereas for an explicit measure, awareness is relatively high. A perfect implicit measure would assess the relevant construct without the person’s awareness, and a perfect explicit measure would be one of which the person was fully aware. As we describe in some detail later, psychologists have developed a wide variety of imperfect implicit measures relevant to understanding bias. These measures have focused on assessing people’s overall attitudes and stereotypes toward various minority groups.
Implicit versus Explicit Attitudes
Just as an explicit measure is a measure of which people are aware, an explicit attitude is an evaluation that people consciously and willfully acknowledge (e.g., I like Black people). In contrast, an implicit attitude is often defined as one that people hold but do not recognize or endorse (e.g., Kihlstrom, Reference Kihlstrom, Sansone, Morf and Panter2004).Footnote 2 Explicit measures, by definition, assess explicit attitudes. However, it is not the case that implicit measures necessarily assess implicit attitudes. That is, even if people are not aware of what a measure is attempting to measure (e.g., such as measuring liking by unobtrusively assessing how close one person sits to another), what is assessed with this measure could still be an explicit attitude – an attitude that the person would have reported if they had been asked directly (i.e., people may sit closer to people that they explicitly like). Indeed, in many domains, explicit and implicit measures of attitudes are correlated with each other, though the magnitude of the implicit–explicit relationship varies with the subject matter. For example, correlations are relatively high for political candidates but quite low in socially sensitive domains such as race (e.g., Greenwald, Smith, et al., Reference Greenwald, Smith and Sriram2009).
The same factors that contribute to the link between implicit and explicit measures of attitudes being strong also contribute to the ability of implicit measures to predict behavior. For example, just as implicit measures on sensitive issues correlate less well with explicit measures than for non-sensitive issues, so too do they predict behavior less well in these domains. Indeed, meta-analyses show that the more strongly implicit measures of attitudes correlate with explicit measures, the more strongly implicit measures predict relevant behavior (Greenwald, Poehlman, et al., Reference Greenwald, Poehlman and Uhlmann2009; Kurdi, et al., Reference Kurdi, Seitchik and Axt2018). A situation of considerable interest in the field occurs when attitudes assessed with explicit measures and implicit measures fail to correspond (e.g., a person is positive on an explicit measure but negative on an implicit measure), a situation that has been called implicit ambivalence if people are unaware of the discrepancy (Petty et al., Reference Petty, Tormala and Brinol2006).
Implicit versus Explicit Bias
Having defined implicit versus explicit measures and attitudes, what about implicit versus explicit bias? Bias with respect to people occurs when an individual has more favorable or unfavorable beliefs, attitudes, and/or actions toward a person who is a member of a certain category (e.g., White, female) than another category (e.g., Black, male) in the absence of any individuating information about that person that would justify that favorability. Explicit bias occurs when people are aware of their bias and consciously and deliberately act on their acknowledged prejudicial attitude (e.g., I don’t like African Americans so I will recommend against hiring them). That is, people are aware of being biased and of acting on their bias. Importantly, if a person is fully aware of being biased in both evaluation and action, but the person simply aims to conceal this bias when asked about it, this person still has an explicit bias (i.e., has an explicit attitude and engages in explicit behavior). Though, due to social desirability concerns, it may be difficult to uncover this on explicit measures. That is, if people know that someone is attempting to assess their potentially biased attitudes or behaviors, they could deliberately conceal their responses. Indeed, getting around socially desirable responding was one reason that implicit measures of attitudes were developed in the first place (Fazio et al., Reference Fazio, Jackson and Dunton1995).
In contrast to a bias of which a person is aware, implicit bias also refers to the prejudicial judgments, decisions, and behaviors a person enacts, but there is something about the bias about which the person is unaware. Petty et al., (Reference Petty, Wheeler, Tormala, Millon and Lerner2003) noted that since the start of research on implicit bias in social psychology (Greenwald & Banaji, Reference Greenwald and Banaji1995), scholars in the area (and even the general public) have used the term implicit to refer to one of three things about which people could be unaware. First, people can be unaware of the biased attitude itself. In this case, one can say that the person has an implicit attitude. Second, people can be unaware of the impact of a biased attitude on other judgments and behavior. In this case, one can say that there is an implicit impact of the biased attitude. Third, people can be unaware of the underlying source (basis) of their biased attitude. In this case, one can say that the attitude has an implicit basis. We provide an example of each kind of implicit bias next.
1. Implicit attitude. As noted earlier, a person might not be aware of having a reaction to one group that is more negative than another group. For example, people might not realize that they harbor more negative feelings toward Black people than White people and would rate both groups equally on an explicit measure such as a direct self-report. When people are unaware of their attitudes, these attitudes are implicit. Yet, implicit (unaware) attitudes might still have an impact on behavior. However, there is relatively little evidence for the notion that people lack awareness of their attitudes. That is, there is little or no compelling research to date showing that people have no inkling of their evaluative responses. For example, a person reporting an explicit positive attitude toward a minority group might still recognize having an initial negative reaction that could be reported if the explicit measure had asked about “gut reactions” rather than an overall evaluation (Jordan et al., Reference Jordan, Whitfield and Zeigler-Hill2007; Ranganath et al., Reference Ranganath, Smith and Nosek2008). Or, people might be aware of an initial negative reaction but believe that this is invalid information because it reflects social stereotypes that they do not endorse, and thus do not report this on the typical explicit measure (Loersch et al., Reference Loersch, McCaslin and Petty2011; Petty et al., Reference Petty, Tormala and Brinol2006). What is more likely than people being completely unaware of their evaluative reactions is that they may not always appreciate that their automatic evaluative reactions can differ from their more deliberative evaluations and yet still have an impact on their behavior.
2. Implicit impact. In this definition, a person might be fully aware of a prejudicial negative reaction (whether automatic or deliberative) to a group member (i.e., the person has an explicit attitude), but the person might not be aware that this reaction is affecting his or her judgments or behavior (e.g., biasing perception of the person). This biasing impact can be completely unintended (e.g., see Greenwald & Banaji, Reference Greenwald and Banaji1995). Although people might be unaware of how various explicit attitudes affect their behavior, what people might be especially unaware of is how their relatively automatic reactions to other people can affect their behavior in addition to their more deliberative views. That is, as just noted, people might be aware of their tendency to have a quick negative reaction to someone (i.e., the reaction itself is not implicit), but they might believe that only their more carefully considered views affect their behavior and not their more automatic reactions. The fact that people can behave in a biased way without intending to be biased or “without even realizing it” (Kang et al., Reference Kang, Bennett and Carbado2011) is perhaps the most common use of the term implicit bias.
3. Implicit basis. Some theorists refer to implicit bias as including situations in which a person is aware of his or her attitude and is also aware of the effect it has (failing the first two definitions above), but the person is simply not aware of from where the attitude comes. For example, an attitude could be driven by underlying stereotypic expectations and the person does not realize this is the case (cf., Wilson et al., Reference Wilson, Lindsey and Schooler2000). This would include misattribution effects where, for example, the evaluation is negative (I am aware I dislike Hillary Clinton as a presidential candidate and that as a result I will not vote for her), but the person doesn’t realize that this negativity is a product of a group stereotype (e.g., a mismatch between one’s expectation for what is “presidential” and what a typical woman is like). Moss-Racusin et al. (Reference Moss-Racusin, Dovidio and Brescoll2012) show that science faculty judge a female applicant to be less competent and less hireable than an identical male applicant for a position managing a science laboratory, presumably not because they believe females to be incapable of succeeding in science (the effect was equally strong among female as male science faculty), but rather because the application when paired with a female name did not evoke the same sense of competence and hireability. In this version of implicit bias, the person fails to know the stereotypic basis of the negative attitude.
Accepting these three options implies that unawareness of the attitude is not the only or even a necessary characteristic of implicit bias since as explained, a person can accept that an attitude is biased but not realize that this attitude has an impact or what the basis of the attitude is. In short, the overarching characteristic of all three forms of implicit bias is that the phenomenon involves something of which the respective person is unaware.Footnote 3
These issues regarding lack of awareness of (1) the attitude, (2) its impact on behavior, and (3) its origins can raise complex issues with regard to personal responsibility. For example, if a person is completely unaware that a biased attitude is affecting his or her behavior, should the person be held accountable for this behavior? Notions of accountability in most societies are predicated on the belief that individuals are responsible for their own beliefs and actions, but numerous models in social psychology raise questions about this bedrock belief. For example, some scholars have proposed that society is so infused with stereotypes that they are activated automatically whenever group members are encountered. What differentiates prejudiced from non-prejudiced people in this model is whether the automatically activated stereotypes are subsequently used to guide judgment and behavior or are inhibited (Devine, Reference Devine1989). One implication of this notion is that people can no more prevent activation of their stereotypes than they can prevent themselves from reading a stop sign once they have seen it (for early refinements of this model, see Gilbert & Hixon, Reference Gilbert and Hixon1991; Spencer et al., Reference Spencer, Fein and Wolfe1998). Subsequent research took questions of accountability one step further by documenting that older adults have difficulty inhibiting their automatically activated stereotypes due to compromised frontal lobe functioning (Stewart et al., Reference Stewart, von Hippel and Radvansky2009; von Hippel et al., Reference von Hippel, Silver and Lynch2000). This finding raises the possibility that sometimes activation and application of stereotypes are difficult to prevent.
A Schematic Representation of the Current Conceptualization of Implicit and Explicit Measures of Attitudes
In an effort to be clear about the constructs we are describing, the following schematic is offered. It is rough, incomplete, and many at the conference and in the field more generally will likely disagree with some part of it. But some of the controversy regarding the body of research on implicit measures, attitudes, and biases comes from having different conceptions of what implicit versus explicit measures are assessing and especially the relationship between measures and behaviors. Our goal is to at least be clear about some definitional issues to help remove disagreements that stem from a basic misunderstanding or confusion about what researchers are talking about.
Next, we provide a few important points regarding the schematic and its implications for understanding implicit measures and bias.
1. As depicted in Figure 1.1, attitude measures, whether implicit or explicit, tap into the same contents in memory. This information can be evaluatively positive or negative and these two sources of evaluative information can influence both explicit and implicit measures of attitudes (that is, this need not be a bipolar evaluation, see Cacioppo & Berntson, Reference Cacioppo and Berntson1994). The memory contents can include attributes of the attitude object, specific past encounters with the attitude object, emotions, societal and media messages, past attitude reports, etc.
2. As implied by the figure, an attitude (or evaluative reaction) is not necessarily a “thing” that exists in memory as a unified entity and retrieved automatically and inevitably, but rather can be at least partially constructed “on the fly” in response to a particular attitude object in a given situation or in response to an explicit question (Schwarz, Reference Schwarz2007). Furthermore, evaluations, such as attitudes toward minority groups, can vary with the context such as the race of the experimenter taking the measurement (Lowery et al., Reference Lowery, Hardin and Sinclair2001) or the clothing worn by the target to be judged (Barden et al., Reference Barden, Maddux and Petty2004; Wittenbrink et al., Reference Wittenbrink, Judd and Park2001). Thus, it is not surprising that the measured attitude toward various objects as assessed with both implicit and explicit measures can vary somewhat from situation to situation.
3. Although as just noted, the implicit and explicit measures are informed by the same underlying memory contents, the two kinds of measures are not invariably related to each other and will not necessarily yield the same evaluative outcomes (i.e., the same attitudes) for a variety of reasons:
a. Moderators (e.g., motivation and opportunity to think) affect both kinds of measures and can do so to different degrees.
b. Different contexts or environments can make different memory contents more or less accessible automatically and deliberatively.
c. To the extent that the memory contents activated through automatic processes differ from the contents activated by controlled processes, coupled with the fact that some implicit measures are more influenced by automatic processes and less influenced by controlled processes relative to explicit measures, this can produce differences on the implicit versus explicit measures.
d. Automatic and controlled processes might access or prioritize different things such as positive versus negative memory contents or affectively versus cognitively based contents.
e. Often researchers are asking different questions with an implicit versus explicit task, and this contributes to the divergence of outcomes (for demonstrations of this, see Han et al., Reference Han, Czellar and Olson2010; Payne et al., Reference Payne, Burkley and Stokes2008; Wittenbrink et al., Reference Wittenbrink, Judd and Park2001).
4. Contemporary implicit measures primarily utilize automatic processes, whereas explicit measures are more influenced by controlled processes (Schneider & Shiffrin, Reference Schneider and Shiffrin1977), but neither measure is process pure (e.g., Calanchini & Sherman, Reference Calanchini and Sherman2013; Calanchini et al., Reference Calanchini, Lai and Klauer2021). Thus, although implicit measures are heavily affected by automatic processes and less so by controlled, and explicit measures are driven more by controlled than automatic processes, the processes themselves are separate from and not tied in an isomorphic manner to the measures.
5. The act of completing an implicit measure can influence responses on an explicit measure (e.g., one might respond differently to a thermometer rating judgment having just completed an Implicit Association Test (IAT)). Similarly, completing an explicit measure might affect responses on an implicit measure by making certain memory contents more salient.
6. The link between attitudes and behavior has a long history in social psychology and this linkage has been shown to be dependent on a variety of moderators such as the correspondence of measurement of attitude and behavior (Fishbein & Ajzen, Reference Fishbein and Ajzen1977), the strength of the attitude (e.g., its certainty, accessibility; Petty & Krosnick, Reference Petty and Krosnick1995), and other factors. Because of established moderators, neither explicit nor implicit measures of attitudes will necessarily predict behavior in all circumstances. Furthermore, sometimes, one kind of measure will predict behavior better than the other kind. For example, prediction of each kind of measure tends to be better the more the behavioral criteria match the attitude measure – in deliberative situations, explicit measures tend to predict more accurately, but in spontaneous situations, implicit measures tend to predict better (Dovidio et al., Reference Dovidio, Kawakami and Johnson1997). Although the field’s understanding of moderators of the ability of explicit attitudes to predict behavior is fairly well developed, the same is not true with respect to implicit measures.
Figure 1.1 A schematic representation of the current conceptualization of implicit and explicit measures of attitudes.
Implicit Measures That Have Been Developed to Study Implicit Bias
As noted earlier, social psychologists have developed implicit measures that are designed to assess people’s attitudes without asking them directly. When such measures were first introduced many decades ago, they were called indirect measures because a person’s attitudes needed to be inferred indirectly, typically from some behavioral response rather than from a direct self-report (Petty & Cacioppo, Reference Petty and Cacioppo1981). For example, rather than directly asking Person A how much they like Person B, liking could be inferred from spontaneous eye gaze or seating distance (Dovidio et al., Reference Dovidio, Kawakami and Beach2001). Or, rather than asking someone how much they liked religion, they could be asked to tell a story about a minister at the pulpit and the story be coded for positive versus negative content (Proshansky, Reference Proshansky1943).
In the past few decades, a new type of implicit measure was developed. What set these measures apart was that attitudes were inferred from how quickly people could give evaluative reactions to various stimuli. That is, these measures attempted to assess automatic evaluative reactions. Although it could be argued that the earlier indirect measures, such as seating distance and telling spontaneous stories, were successful because they also tapped into automatic reactions (e.g., people spontaneously sit closer to others they like than dislike without giving it much thought), the new implicit measures were quite explicit about the desire to assess automatically activated attitudes. These measures include the Implicit Association Test (IAT; Greenwald et al., Reference Greenwald, McGhee and Schwartz1998), the evaluative priming measure (Fazio et al., Reference Fazio, Sanbonmatsu and Powell1986), and the Affect Misattribution Procedure (AMP; Payne & Lundberg, Reference Payne and Lundberg2014), among others (for reviews, see Petty et al., Reference Petty, Fazio and Briñol2009; Wittenbrink & Schwarz, Reference Wittenbrink and Schwarz2007).
In their initial development during the 1980s, a primary goal of the new wave of implicit measures, like their predecessors, was to minimize strategic responding with respect to sensitive or socially undesirable attitudes, as it was assumed that quick responding would likely bypass social desirability concerns (e.g., Dovidio et al., Reference Dovidio, Evans and Tyler1986; Gaertner & McLaughlin, Reference Gaertner and McLaughlin1983). Logically, there are multiple ways to achieve this goal (e.g., the bogus pipeline method; Jones & Sigall, Reference Jones and Sigall1971). The newer reaction time-based implicit measures draw on paradigms from cognitive psychology that examine facilitation and inhibition effects in priming (Meyer & Schvaneveldt, Reference Meyer and Schvaneveldt1971; Stroop, Reference Stroop1935). For example, participants’ attitudes about race are inferred by assessing the degree to which race-related cues (e.g., pictures of Black and White faces) interfere with or facilitate some non-racial judgment (e.g., categorizing a word as positive or negative; Fazio et al., Reference Fazio, Jackson and Dunton1995).
By way of example, an IAT (Greenwald et al., Reference Greenwald, McGhee and Schwartz1998) examining racial attitudes might ask participants to alternate between classifying a set of valenced words as either good or bad (a task that is unrelated to race) and classifying a set of names (e.g., Jamal, George) as either stereotypically Black or White (a task presumably unrelated to attitudes). By forcing the participant to use the same keys for both tasks and by flipping the mapping of the keys for one categorization task in the middle of the larger task, the researcher can assess the degree to which classifying a word as good is easier when it shares a response key with White and harder when it shares a response key with Black. Similarly, an evaluative priming task (Fazio et al., Reference Fazio, Jackson and Dunton1995) might present a series of words on a computer screen and ask participants to classify each as either good or bad. This classification task has nothing to do with race. The words (e.g., happy) are simply categorized as good or bad. Immediately before each word, however, a prime appears – perhaps a face that is either a Black person or a White person. Bias is measured as the extent to which the race of the face interferes with or facilitates the judgment of the word as good or bad. Different measures adopt slightly different approaches (some rely on response competition, others on semantic priming, etc.), but almost all of the measures use some variation of this indirect strategy.
Most tasks also share a number of interesting, but likely unnecessary, characteristics. Perhaps due to their structure, perhaps due to the fact that these approaches were borrowed from cognitive psychology, most measures typically involve indices derived from reaction times or error rates. Accordingly, participants must generally perform the tasks on computers, ideally in a distraction-free environment. And most measures involve some degree of difficulty, presumably because if the focal task is too easy, responses might be impervious and therefore insensitive to race or gender or whatever dimension the researcher hopes to assess. Some implicit measures (such as the Go/No-go association test, or GNAT, Nosek & Banaji, Reference Nosek and Banaji2001) impose a strict response deadline, forcing the participant to respond quickly and inducing errors. Other tasks, such as a first-person-shooter task (Correll et al., Reference Correll, Park and Judd2002), present ambiguous stimuli, forcing the participant to interpret the information (e.g., is the target holding a gun or a cell phone) and increasing mean response times. Still other tasks, like the IAT, force the participant to perform cognitively demanding mental operations as they alternate between judgments.
In considering new approaches to implicit (indirect) measurement, it is worth revisiting the original goal of this class of measures. There are several reasons for doing so. First, it is important to consider whether or not the original goal (removing social desirability) is still relevant. Based on what researchers have learned over the past 30 years, there could be new considerations that should guide the field (or that may guide an individual researcher). And, for whatever new goals might be deemed relevant, scholars might consider whether the approach that has generally been adopted (difficult indirect measures that rely on response times or error rates) offers the best strategy for achieving those goals. For example, if the goal of contemporary implicit measurement is to get at gut reactions (rather than avoid social desirability), explicit measures that ask about gut feelings might accomplish a similar purpose (e.g., Ranganath et al., Reference Ranganath, Smith and Nosek2008).
Potential Goals of Implicit Measures
This discussion will only address a few of the potential goals of the currently popular implicit measures. First, it is noteworthy that despite social desirability concerns, people’s explicit attitudes can still show evidence of bias. For example, Piston (Reference Piston2010) found that 45 percent of White people interviewed in the face-to-face portion of the 2008 National Election survey rated Black people lower than their own racial group on the trait of “hard working.” The figure was 39 percent on the trait of “intelligence.” Nonetheless, because many people may be uncomfortable expressing racist or sexist attitudes in a public survey, measures that reduce socially desirable responding still offer value.
In addition to social desirability, there are other considerations that the field has come to recognize more fully over the last few decades of research. One, already discussed, is that it might be desirable to have measures that assess automatic as well as deliberative attitudes and that the new implicit measures focus on the former. This is a value above and beyond avoiding social desirability. Another factor is that participants might not have complete introspective access to all of the mental content that is associated with an attitude object. For example, as noted earlier, although depictions of different racial groups in the media might produce negative associations to racial categories, people do not consider this when they report their explicit attitudes. Thus, another goal for implicit measures could be to assess these “extra-personal” associations that exist in the mind of an individual because they exist in the culture in which that individual lives (Olson & Fazio, Reference Olson and Fazio2004). The importance of cultural associations on what is assessed by implicit measures is evident in data showing that IAT responses are associated with a person’s geographic location (Hehman et al., Reference Hehman, Flake and Calanchini2018; Payne et al., Reference Payne, Vuletich and Lundberg2017). Even in domains that are not particularly sensitive to social desirability concerns, participants might not actually be able to report all of the content that underlies their attitudes (i.e., as noted earlier, some attitudes can have an implicit basis). One of the virtues of the current class of implicit measures is that, because they never directly ask the question, they do not require that participants consciously access (or construct) the relevant attitude.
Once the field’s goals for implicit measures have been clarified, the particular characteristics of those measures should be revisited. Again, the vast majority of them involve indirect assessment. It should be clear that, if researchers choose to assess attitudes about race or gender in the context of some other task (e.g., classifying words as good or bad), the ability to assess the attitude in question is necessarily constrained to some extent by the features of that task. For example, the evaluative priming task described above requires a set of cognitive operations (perceiving the text, interpreting it, and classifying it according to valence) that are distinct from the attitude being measured but that can operate as a kind of filter for the associations that form that attitude. Is this approach necessary? Is it optimal? Are there other ways to achieve the goals of these measures without introducing an unrelated task? These tasks also generally involve some non-trivial level of difficulty. Again, does this approach serve the goals of these measures? These are important questions to ponder. In some cases, it might be that carefully designed and carefully administered explicit measures meet research goals as well as or better than implicit measures. Indeed, while the present committee’s work focuses on the state of the science on implicit bias, further study into the optimal measurement of explicit bias is important to pursue as well.
Atypical Measures
Although the previous discussion has focused on some of the most popular implicit measures currently in use, there are a variety of other measures that deviate, in one way or another, from the typical cognitively inspired structure that characterizes measures like the IAT, GNAT, evaluative priming, and the first-person-shooter task. These atypical paradigms might offer hints and inspiration that can facilitate development of new, different, and better measures. For example, consider the Affect Misattribution Procedure (AMP; Payne et al., Reference Payne, Cheng and Govorun2005). Like the other measures, the AMP relies on indirect measurement, but it does not require measurement of response time or error rates. Drawing on Murphy and Zajonc (Reference Murphy and Zajonc1993), the AMP presents a prime stimulus (e.g., a White or Black face) followed by a Chinese ideograph. Participants, who typically do not read Chinese, are asked to evaluate the complex but (to them) meaningless ideograph. Those evaluations are typically biased by the prime. That is, a prime about which a person feels positively biases responses to the Chinese ideograph in a positive way.
A second atypical example is one of the few tasks that does not rely on indirect measurement. This is actually an explicit measure in that it directly asks the participant to evaluate or classify the attitude object that is being studied. This measure relies on a mouse-tracking paradigm (Freeman & Ambady, Reference Freeman and Ambady2010), which has been used to study how people classify ambiguous stimuli. The paradigm presents two response labels (e.g., “good” and “bad,” “male” and “female,” or “caring” and “aggressive”) at the upper left and upper right corners of a computer screen. Participants must use a computer mouse to position the cursor at the bottom center of the screen. They are then shown an exemplar and asked to move the cursor to the appropriate response option. This work showed that, when presented with prototypic exemplars (e.g., a very masculine-looking male face), participants moved the cursor toward the appropriate option (e.g., “male”) in a relatively straight line. When presented with a less prototypic exemplar (e.g., a more feminine-looking male face), the movement was less direct: initially, the cursor might head toward the incorrect label (“female”) before veering off toward the correct response. This approach has also been applied to the measurement of attitudes (Vallacher et al., Reference Vallacher, Nowak and Froehlich2002). For example, when asked to classify an attitude object like sunshine, trajectories are relatively direct, but when asked to classify an attitude object like euthanasia, about which people might be more ambivalent, the mouse trajectories are less direct (Schneider et al., Reference Schneider, van Harreveld and Rotteveel2015; Vallacher et al., Reference Vallacher, Nowak and Kaufman1994). Thus, this task can assess ambivalence with respect to a target without ever asking about ambivalence.
Third, the recently proposed Judgment Bias Task (JBT; Axt et al., Reference Axt, Nguyen and Nosek2018) is an indirect measure that relies on much more deliberative processing than most of the other measures discussed here. Drawing on Beckett and Park (Reference Beckett and Park1995), this paradigm simulates a hiring or other selection task. It presents a series of applicant profiles, which the participant must evaluate. These profiles might include information about each applicant’s grades, test scores, recommendation letters, etc. The participant then must decide whether or not to hire (or admit, or date) each applicant. Critically, although some applicants are more qualified and some are less qualified, the differences are hard to detect. And, just as critically, the profiles also include information that interferes with the judgment task, such as photographs that vary in attractiveness, or information that the applicant is an ingroup or outgroup member. The goal of the measure is to see how the variable of interest (e.g., gender) affects some outcome of interest (e.g., a hiring decision). For example, how well can participants distinguish objectively more from less qualified female applicants for a job and how lenient or strict are they in making these decisions.
Finally, there is growing interest in using brain imaging in the study of implicit bias. The neurons in a number of brain regions routinely alter their activity during brain imaging studies of implicit racial attitudes (e.g., Phelps et al., Reference Phelps, O’Connor and Cunningham2000). Neuroscientists tend to refer to brain regions in terms of their own interests and preferences, so it is not uncommon for neuroscientists who study implicitly measured attitudes and prejudice to refer to these brain regions as the “prejudice network” (e.g., Amodio, Reference Amodio2014). Unfortunately, this can lure consumers of neuroscience research to the misperception that these regions are consistently and specifically associated with implicit racial bias. To date, brain imaging research does not yet clarify the debate regarding the power of implicit measures of racial attitudes to impact prejudicial behavior. The regions and activity patterns associated with implicit measures of prejudice contribute to many different psychological events, and therefore with current technology, brain images are only useful correlates of implicit measures of bias when they are combined with behavioral measures. Even then, however, brain imaging findings do not transparently explain the mechanisms behind what implicit measures assess. People might attend to other-race faces because those faces are novel. Similarly, people might better remember faces of the same race as a function of greater expertise. In fact, the novelty and lack of perceptual expertise (associated, as they are, with increases in arousal) might be a component of implicitly measured attitudes. Thus, linking implicit bias with brain imaging is a challenging task, but one that could ultimately yield important insights depending on the specific goals for such measures.
In sum, there are a variety of implicit measures of attitudes that have emerged, each with idiosyncrasies and each with particular strengths and weaknesses. The goal here is not to provide a catalog but merely to highlight a few of the most commonly used measures, as well as those that deviate from what has (for better or worse) become the standard set. Future measures need not rely on indirect assessment, they need not rely on speeded decision making, and they need not involve response times and error rates. Perhaps by thinking beyond the default approach, researchers can build measures that more effectively meet the desired goals, whether those goals are to: (1) avoid social desirability, (2) assess automatically activated attitudes, (3) bypass participants’ understanding of the basis of their attitudes, (4) tap evaluations of which someone is unaware, or some other goal.
What Does the Field Agree Upon Regarding Implicit Bias and Implicit Attitude Measures?
A number of points seem to generate agreement among researchers regarding the study of implicit bias and implicit attitude measures.
1. To optimally compare the impact of explicit (self-report) and implicit (e.g., IAT) evaluations, researchers should compare them at the same level of categorization (e.g., see Fishbein & Ajzen, Reference Fishbein and Ajzen1974). For example, if an implicit measure assessed attitudes toward Black people (good or bad associations), the explicit measure should not ask something more specific such as attitudes toward affirmative action policies. Similarly, if an implicit measure asks people to categorize specific exemplars of a group (e.g., pictures of famous powerful women), the explicit measure should not ask about the group at the general category level (e.g., judgments of what “women” are like).
2. Some evidence has emerged demonstrating a relationship between community-level implicit associations and demographic characteristics and outcomes of those geographic areas (e.g., Leitner et al., Reference Leitner, Hehman and Ayduk2016; Payne et al., Reference Payne, Vuletich and Lundberg2017). For example, one study showed that United States counties with higher levels of county-wide implicit racial bias also had greater racial gaps in infant health outcomes, even after controlling for relevant demographic and geographical factors (Orchard & Price, Reference Orchard and Price2017). This suggests that responses on implicit measures aggregated across individuals reflect behavioral characteristics of those communities, even though the correspondence of individual-level implicit measures to behavior has been relatively weak. At least in part this is likely because the community level measures aggregate across large numbers of observations and hence are much less “noisy” or unstable. A reasonable next step would be research examining the ability of measures of an individual aggregated across many contexts and situations to show enhanced test-retest reliability and better ability to predict that individual’s behavior (also across situations). Indeed, recent research is showing (in accord with psychometric theory) that when a person takes multiple versions of the same IAT across two different time periods and these assessments are averaged, test-retest reliability is increased compared to a single administration at each time period (Connor & Evers, Reference Connor and Evers2020; Lindgren et al., Reference Lindgren, Baldwin and Olin2018).
3. As just noted, implicit measures of attitudes might be more useful for tapping into group-level automatic evaluations than individual-level evaluations because, as currently developed, the typical single administrations have too much unreliability (and contextual impact) to be highly useful in assessing the automatic component of attitudes for particular people (unlike explicit attitude measures). Thus, if group A scores higher than group B in prejudice on an implicit measure, it is reasonable to predict that group A will demonstrate more prejudicial behavior than group B. It is less reasonable, however, to use the implicit measure to predict which particular members of group A will engage in prejudicial behavior. In this sense, implicit measures can be compared to imperfect disease predictors. Imagine a cancer diagnostic test that shows high scorers on the test are more likely to develop cancer overall than low scorers. But the screening test is imperfect so that for every ten who score high on the test, four (on average) will develop cancer, but for every ten who score low, only two (on average) will develop cancer. Such a test could be useful for screening groups and making conclusions such as those who score high are twice as likely to develop cancer than those who score low. But the test is not useful in saying which four out of the ten who score high will get cancer and which six will not. The same is true for implicit measures of prejudice as they currently stand. That is, they can be useful in predicting that certain groups of people (e.g., high scorers) are more likely to engage in prejudicial behavior, but they are not as useful for determining which particular individuals among the high scorers will engage in prejudicial behavior.
4. As articulated above, there are numerous moderators of the impact of both implicitly measured and explicitly measured evaluations on behavior.
5. Activation of an attitude should be distinguished from application of that attitude. That is, in any given situation, an evaluation can automatically come to mind, but that evaluation need not drive behavior (e.g., if situational pressures are more salient, or if deliberative thought processes override the evaluation that comes to mind).
What Issues Are in Some Dispute and Thus Point to Clear Research Questions to Resolve?
There are a number of issues that have achieved some research attention but are far from being fully resolved and thus are worthy of additional attention.
1. It is not clear how much contemporary implicit measures of attitudes assess transient momentary states versus relatively stable traits or what combination of each.
2. What is the internal reliability of various implicit measures and test-retest reliability, and how does this vary across different attitude objects, situations, and people (e.g., those more likely to rely on intuition, see Pacini & Epstein, Reference Pacini and Epstein1999)? In addition to aggregating over multiple assessments as noted earlier, how can reliability be increased for current measures? Can new measures be developed that are more reliable?
3. How highly related are various implicit measures to each other? Although methodological factors (e.g., using the same stimuli) can improve relationships, what else can improve consistency across measures? What level of consistency should be expected and what does lack of consistency imply (e.g., measurement error; measures are highly sensitive to the immediate context)?
4. How well do implicit measures correlate with explicit measures for different attitude objects and issues? As noted earlier, it is clear from meta-analyses that these correlations vary across attitude domains. They seem to be higher for political attitudes and mundane attitude objects (e.g., consumer products), and lower in socially sensitive (e.g., racial) domains. They are lower still for attitudes toward the self (Bar-Anan & Nosek, Reference Bar-Anan and Nosek2014). It is not entirely clear what is the underlying moderator. Is it the social sensitivity of the attitude object? Is it the complexity of the knowledge structure regarding the attitude object (with greater complexity lending itself to greater contextual variation in the activated attitude)? What are the primary causes and consequences of implicit-explicit attitude discrepancies?
5. How well do implicit measures predict behavioral outcomes alone, and over and above explicit measures? In what domains are implicit measures a useful supplement to explicit measures? In the domains in which they provide added predictive power, why does this occur?
6. Manipulations that alter implicitly measured associations don’t necessarily show parallel changes in behavior. This is not particularly surprising because manipulations that alter explicitly measured attitudes don’t always alter behavior and there is much research on the conditions under which this is more or less likely to occur (e.g., see research on attitude strength; Petty & Krosnick, Reference Petty and Krosnick1995). Three key questions for changing implicitly measured attitudes are: (1) When do such shifts also have downstream behavioral consequences? (2) How can the downstream consequences of such shifts be magnified? (3) When are changes in implicitly measured attitudes more impactful on behavior than changes in explicitly measured attitudes?
What Areas of Research Should Be Encouraged Going Forward?
In addition to the research suggestions just made, we next outline additional areas where even more knowledge has been accumulated but where existing research has not provided sufficient clarity.
Measurement Issues
1. A systematic assessment of the measurement properties (e.g., reliability, convergent, and predictive validity) of implicit measures of attitudes should be undertaken. Since its introduction more than twenty years ago, the IAT has increasingly dominated the field and mostly replaced other paradigms to detect implicit biases. In reviews on the psychometric qualities of different implicit measures, the IAT and AMP score best (e.g., Bar-Anan & Nosek, Reference Bar-Anan and Nosek2014). Yet, the appropriateness and reliability of different implicit measures will largely depend on the content of the to-be-detected bias and the particular procedures used. For example, when trying to investigate the impact of race on police officers’ propensity to use their weapon, using the shooter paradigm (e.g., Correll et al., Reference Correll, Park and Judd2002) obviously has a higher ecological validity than using a race-IAT. But how strongly will these two measures correlate?
2. There are now many different types of implicit attitude measures, and it seems highly likely that different measures have different strengths and weaknesses. It also seems highly likely that different measures are more or less suitable for different research questions and populations. To date, the field is still lacking in theoretical and empirical knowledge to better understand when and why different implicit measures are dissociated, and, in turn, which specific measures are best suited for specific aims, contexts, psychological processes, and behavioral predictions (see Brownstein et al., Reference Brownstein, Madva and Gawronski2020). In other words, a better understanding of similarities and differences among various psychometrically sound measures of implicit attitudes will help the field generate more straightforward suggestions on which measure might best fit which research question. A greater understanding of the current crop of implicit attitude measures would allow researchers to be more thoughtful and systematic in their choice of measures when initiating data collection.
3. Most measures of implicit bias rely on reaction time, but as explained earlier, not all of them (e.g., Affect Misattribution Procedure, Payne et al., Reference Payne, Cheng and Govorun2005; Partially Structured Attitude Measures, Vargas et al., Reference Vargas, von Hippel and Petty2004; Stereotypic Explanatory Bias, Sekaquaptewa et al., Reference Sekaquaptewa, Espinoza and Thompson2003; Linguistic Intergroup Bias, von Hippel et al., Reference von Hippel, Sekaquaptewa and Vargas1997). Greater attention should be given to the development of new measures that rely on properties of implicit measures of attitudes beyond strength of association. There has been almost no systematic study of the varying utility of existing measures nor has there been systematic efforts to develop new measures with known properties. An effort to tie the development of new measures to the different types of implicit bias (outlined at the beginning of this document) would be potentially valuable.
4. Future research should explore the conditions under which implicit measures predict (a) explicit measures of attitudes, (b) outcome measures and behaviors, and (c) outcome measures and behaviors over and above explicit measures. Moreover, to what extent do implicit and explicit attitude measures affect one another? Do implicitly and explicitly measured attitudes interact in causing behavior or do they exert orthogonal effects (e.g., Johnson et al., Reference Johnson, Petty and Briñol2017)?
Questions of Causation
5. What role do implicit attitudes and implicit bias more generally play in causing prejudicial behavior? To date, there has not been enough focus on studies of behavioral outcomes. Such studies should be prioritized due to their applied importance; the idea of implicit bias has captured the public imagination largely because of its presumed potential to explain discriminatory behavior outside the laboratory. For example, there are studies that reveal relationships between implicitly measured attitudes and important discriminatory behaviors outside the laboratory (e.g., Hehman et al., Reference Hehman, Flake and Calanchini2018); however, the causal role for the attitudes that implicit measures assess is unclear. Experimental work is necessary to establish the role of such attitudes in causing discriminatory behavior, but there are precious few laboratory and field studies that rely on experimental manipulation and measure discriminatory behavior. It will be particularly important for such experimental work to disentangle the influence of implicitly and explicitly measured attitudes, thereby establishing the unique predictive role of implicit measures in accounting for discrimination. It remains unclear whether the attitudes assessed with implicit measures are responsible for (a) individual differences in how members of outgroups are treated, and/or (b) differential outcomes that emerge as a function of group membership (different job or school performance among people of different genders, ethnicities, etc.).
6. Can interventions targeting implicit bias reduce prejudicial behavior? As noted earlier in our definition of implicit bias, there are several aspects of implicit bias of which people can be made aware. For example, people can be made aware of the fact that they hold prejudicial automatically activated attitudes, or they can be made aware of the fact that these automatic attitudes can influence their behavior without their awareness or intention. Some researchers claim that awareness of holding prejudicial automatic attitudes is helpful in reducing prejudicial behavior. Yet, this has not been examined carefully. Research is needed on the effectiveness of implicit bias workshops to better understand what psychological ingredients they contain, whether and under what conditions they work, and whether they could potentially produce boomerang effects. As governmental agencies and business organizations move to mandate implicit bias training, it is critical to establish a solid evidence base from which such interventions can draw. To what extent can scores on implicit or explicit attitude measures be changed by these interventions and what do such changes reflect? Is it driven by control or motivational processes in response to being made aware of biases? Is it necessary to change the underlying contents of the memory associations in order to reduce implicit bias, and is it reasonable that such trainings can do this? Or is a better strategy to educate perceivers about the conditions under which bias is most likely to operate (e.g., under high cognitive load, with ambiguous stimulus information), and to encourage them to monitor when such conditions are present in order to “check” on whether biases might be operating? For example, one technique is to ask perceivers to mentally simulate their response, changing the group characteristics of a target. If the Starbucks manager had asked herself, “would I call the police if these men were White?,” could this have helped her to identify the possible operation of implicit bias, and perhaps shift her behavior even if it didn’t fundamentally change the implicit bias itself?
Animus Versus Tacit Acceptance of Existing Inequalities
7. Although the literature on implicit bias understandably focuses on identifying, and ultimately reducing, biased outcomes, it is reasonable that the research would also include actions that reflect tacit acceptance of existing racial (or other group-based) inequities. That is, the study of implicit bias need not be confined to actions that are motivated by animus, unconscious or otherwise. The robust literature on racial prejudice in political science and sociology identifying “new racism” theories such as Symbolic Racism (Kinder & Sears, Reference Kinder and Sears1981; Sears, Reference Sears, Eberhardt and Fiske1998; Sears & Henry, Reference Sears and Henry2003; Sears & Kinder, Reference Sears and Kinder1971; Tesler & Sears, Reference Tesler and Sears2010), Racial Resentment (Kinder et al., Reference Kinder, Sanders and Sanders1996), or Modern Racism (McConahay, Reference McConahay1983), maintain that contemporary racial bias relies less on overt claims of racial inferiority and is instead borne out of the belief that racial inequality is due to the cultural inadequacies of Black people and other minorities. According to these theories, opposition on the part of White people to policies designed to address lingering racial inequity is derived from the widespread belief that African Americans and other minorities no longer face significant racial barriers following the passage of landmark civil rights legislation in the 1960s. Bobo, Kluegel, and Smith (Reference Bobo, Kluegel and Smith1997) adopt a similar perspective with their theory of Laissez Faire Racism. In brief, these authors argue that decades of overtly discriminatory policies during the period of Jim Crow have resulted in substantial racial disadvantage for African Americans in employment, educational attainment, household income, housing, and perhaps especially wealth. For example, they report that the average Black American family has only about one-tenth of the wealth of the average White American family. More recent studies confirm the persistence of this racial wealth gap for both Black and Latino families relative to White households (Darity et al., Reference Darity, Hamilton and Paul2018). As with the other new racism theories, one of the pillars of Laissez Faire Racism theory is the denial of significant racial barriers in contemporary American society. Consequently, Bobo and his colleagues (Reference Bobo, Kluegel and Smith1997) maintain that, “even if all direct racial bias disappeared, African Americans would be disadvantaged due to the cumulative and multidimensional nature of historic racial oppression in the U.S.” (Bobo et al., Reference Bobo, Kluegel and Smith1997, p. 4).
Importantly, then, biased outcomes can be a product of the belief that historically disadvantaged groups now operate on a level playing field, and thus any existing inequalities must be a product of shortcomings of individual members of those groups. Therefore, the question arises whether in contexts with widespread and persistent racial inequality, indifference to structural hierarchies can be as consequential as blatantly (explicit) racist beliefs and/or more subtle implicitly measured attitudes. While this perspective may be present in current research efforts, it has not been directly targeted as a factor that could underlie implicit bias, and going forward it would be a fruitful avenue of investigation.
Extensions of Existing Research
8. Effects of implicitly measured attitudes are sometimes stronger in the field than in the lab. It is not clear why this is the case. Is it strong situations versus weak situations? Is it the potency of the independent variables in the field versus the lab? Is it something else?
9. NSF and other granting agencies ideally can incentivize the field to do things that it won’t or can’t do on its own such as funding longitudinal field studies. Granting agencies should also consider funding consortium type research to pit different interpretations of important and robust effects against each other. As part of such an initiative, it would be important to establish a repository for all research, including and perhaps especially failed studies that were never published because they failed to find an effect. This information would helpfully inform researchers who begin a new and related research project and hopefully move the field along in a more efficient manner.
10. A substantial portion of the literature on implicit bias is based on convenience samples. Such samples play a critical role in the development stages of research. Still, results from these samples cannot be assumed to generalize to the larger population, especially in the size of effects obtained (see, e.g., Callegaro et al., Reference Callegaro, Villar, Yeager, Callegaro, Baker, Bethlehem, Goritz and Krosnick2014; MacInnis et al., Reference MacInnis, Krosnick and Ho2018; Malhotra & Krosnick, Reference Malhotra and Krosnick2007; Pasek & Krosnick, Reference Pasek and Krosnick2010; Traugott, Reference Traugott, Holtz-Bacha and Strömbäck2012; Yeager et al., Reference Yeager, Krosnick and Visser2019). There should be greater use of representative, probability-based samples in studying implicit bias. Such samples, constructed within the framework of inferential statistics, are necessary to make valid and reliable inferences to a broader population within a known and calculable margin of sampling error, including design effects. Comparisons of probability and convenience sampling in the literature are focused on discrepancies between measurements of opinions and knowledge of Americans or other population groups in separate administrations of similar questions. It stands to reason that if such issues are observed in this domain, they could also impact basic psychological research as described in this report. To our knowledge, only a handful of studies have attempted to explore implicit measures of attitudes in representative samples of Americans. Moreover, echoing themes in the overview, researchers using implicit measures to study attitudes should acknowledge that their work is being used by policy makers and the general public who assume the basic psychological findings would replicate in samples of American adults and that such findings are important enough to cause alarm about the prevalence and effects of implicit bias, and to inspire organizations to build education and attitude change campaigns in response to them. Representative samples are particularly important for identifying expected effect sizes of postulated outcomes in the general population.
11. The last two decades of neuroscience research have produced a growing number of studies that suggest that various psychological phenomena are produced by predictive processes in the brain (Hutchinson & Barrett, Reference Hutchinson and Barrett2019). Actions, and their accompanying experiences, begin as top-down representations in the brain, fashioned from past experiences that are tested against the state of the world. According to a predictive processing approach, neural predictions (constructed from past experiences) are thought to be a continuously changing filter through which sensory inputs are processed, influencing the relevance of those inputs, effectively deciding which sensory features warrant further processing and action. Once prediction errors are sufficiently minimized, these “inferences” become the brain’s account of what caused the sensations in the first place, effectively categorizing the sensations so that they are meaningful. This approach to understanding processing at the neural level places important constraints on, and suggests ways of understanding the nature of attitudes, assessed with both implicit and explicit measures, as well as their relation to one another. Conducting such work would importantly inform our understanding of these constructs.
12. Attention to and extension of the developmental literature on implicit attitude measures in children might inform the broader literature in the field. Research that explores changes as a function of age (with children or with adults) has the potential to help us better understand both the lower-level cognitive and higher-level sociological contributions to performance on these tasks. For example, using age-appropriate versions of the IAT, Baron and Banaji (Reference Dunham, Baron and Banaji2006; see also Dunham et al., Reference Dunham, Baron and Banaji2006) showed significant bias among children as young as six years old. In fact, the magnitude of bias in these children was statistically equivalent to bias in adults. Using a different implicit measure with children between the ages of nine and fifteen, Degner and Wentura (Reference Degner and Wentura2010) showed a very different pattern: at the age of nine, bias was not significant. But bias increased gradually over the course of adolescence (interestingly, during this time children might become more sensitive to social norms, e.g., Hirschfeld, Reference Hirschfeld1996). Degner and Wentura went on to show that among young children, the magnitude of bias on implicit measures depends heavily on the cognitive demands imposed by the task (this is a critical issue in developmental work, see Crookes & McKone, Reference Crookes and McKone2009). On tasks like the IAT, which force the participant to categorize faces by race, young children seem to show pronounced bias. On tasks that do not force categorization, bias might not emerge until later in development. At the other end of the age continuum, work with older participants reinforces the importance of cognitive ability in performance on implicit measures (e.g., Gonsalkorale et al., Reference Gonsalkorale, Sherman and Klauer2009; Stewart et al., Reference Stewart, von Hippel and Radvansky2009). As these studies demonstrate, developmental work might help the field disentangle the factors that influence measures of bias and should therefore be encouraged.
13. Prejudice and discrimination were initially conceived as the product of animosity toward an outgroup (e.g., Allport et al., Reference Allport, Clark and Pettigrew1954). In fact, however, while they can be driven by outgroup negativity, they can also emerge as the product of ingroup positivity, or both (Brewer, Reference Brewer1999). Unfortunately, the most widely used implicit measures of attitudes do not readily allow researchers to separate ingroup positivity from outgroup negativity. For example, the Implicit Association Test is inherently relative, pitting Us/Good and Them/Bad against the opposite set of associations. Nonetheless, there are numerous measures that allow the implicit assessment of bias toward a group in isolation – including variants of the IAT itself (the Single Category IAT; Karpinski & Steinman, Reference Karpinski and Steinman2006). Greater use of such measures would allow assessment of whether implicitly measured ingroup positivity is sufficient, in the absence of implicitly measured outgroup negativity, to generate discriminatory outcomes. Such findings could potentially play an important role in national debates about the underlying causes of prejudice and discrimination and the best ways to combat them. That is, it is necessary to understand the basis of the bias in order to develop methods to change that bias.
A Caution About the Widespread Use of Implicit Bias in Popular Discourse and Policy Making
Beyond the points made above, a consideration of implicit bias should include an assessment of the resonance and impacts of research claims in society at large. The concept of implicit bias has become firmly entrenched in American society. Government institutions, corporations, and professional associations require or recommend training on implicit bias for their employees and members to recognize and overcome their presumed implicit biases. Information disseminated by researchers, the media, government institutions, public officials, and companies assert that such biases are common, can be measured reliably, influence behavior, and are susceptible to intervention.
Yet, as discussed, these claims are less well supported by empirical evidence than many people think. Presenters at the NSF Conference on Implicit Bias raised questions regarding the definition and independent predictive power of implicit measures of attitudes, the validity and reliability of these measures (especially in assessing individuals), the consistency and extent of their influence on behavior, and whether and how training can help to overcome implicit bias.
A representative national survey conducted as a follow-up to the NSF conference found that broad majorities of Americans think unconscious biases are prevalent, influence behavior, and can be mitigated through training, in line with many representations in the public sphere. The public sees unconscious biases as more prevalent than biases that are consciously held and as worthy of mitigation efforts by businesses and government (see Langer et al., Reference Langer, Baskakova, Krosnick, Stark, Krosnick and Scott2021, for more detail). Self-reported exposure to information about implicit bias relates to these views. These attitudes have policy consequences. Despite uncertain empirical evidence of the effectiveness of training to overcome unconscious bias, three-quarters of Americans see such programs as worthwhile. Just as many Americans support spending public funds on unconscious bias training for the police in their communities. And six in ten think such training would in fact change police behavior (albeit without a great deal of confidence in this outcome). These optimistic public beliefs about the value of implicit bias training might ultimately prove to be correct, but unfortunately, at present, the evidence is weak. Thus, more research is needed.
Given the pernicious role of prejudice in society, it is essential for social scientists to devote their best efforts and practices to understanding the causes and effects of bias, both implicit and explicit; to developing, to the extent possible, empirically validated methods of addressing bias; and to communicating their findings accurately and effectively to policymakers and the public. The current disconnect between verified research claims and public understandings underscores the need for new, richer independent research into the meaning and measurement of implicit bias and its demonstrable impacts and treatability.
Conclusion
Over the past two decades, social psychologists have identified a new and exciting area of inquiry – implicit bias. We have defined this phenomenon above and provided a brief review of what knowledge has emerged on this topic, what is uncertain, and what remains to be done. The phenomenon of implicit bias in some form is real, but the research is still in its infancy. For example, it is unclear what kind of implicit bias (see definitions above) is most pervasive and problematic, how to best assess the presence of such bias, what conditions moderate implicit bias, and how best to address its negative consequences.
