‘I love women’: an explicit explanation of implicit bias test results

Abstract

Recent years have seen a surge of interest in implicit bias. Driving this concern is the thesis, apparently established by tests such as the IAT, that people who hold egalitarian explicit attitudes and beliefs are often influenced by implicit mental processes that operate independently from, and are largely insensitive to, their explicit attitudes. We argue that implicit bias testing in social and empirical psychology does not, and without a fundamental shift in focus could not, establish this startling thesis. We suggest that implicit bias research has been conducted in light of inadequate theories of racism and sexism. As a result, such testing has not sufficiently controlled for subjects’ prejudiced explicit beliefs and emotions, and has not ruled out the possibility that explicit prejudice best explains test subjects’ discriminatory associations and behavior.

1 Introduction

The concept of implicit bias has generated intense interest among philosophers and the public. Yet the empirical research upon which this enthusiasm is founded has recently come under scrutiny. Critics have worried that implicit bias tests do not reliably predict subjects’ performance of prejudiced actions better than explicit bias tests; they have emphasized the conflicting and anomalous data such tests yield; and they have argued that academic and political focus on implicit explanations for prejudice draws attention and resources away from efforts to understand and mitigate structural racism.¹

In this paper, we raise a different kind of criticism, one that researchers and philosophers studying implicit bias must confront even if they can meet these formidable challenges. We will argue that the empirical research has not demonstrated, and without radical changes could not demonstrate, that the rapid associative behaviors it measures are anything more than straightforward expressions of subjects’ explicit attitudes.

Fascination with implicit bias is predicated on the basic claim, offered by psychologists and accepted by philosophers, that the rapid associations and behavior that tests such as the Implicit Association Test (IAT) measure are not merely expressions of subjects’ explicit attitudes. We will call this fundamental assumption, which drives the need to posit some other set of exotic attitudes to explain the data, the Implicit Explanation. The Implicit Explanation implies conclusions that are both surprising and morally significant, and that further fuel public and philosophical interest in implicit bias. Perhaps the most striking and troubling of these is the claim that even egalitarian people, who hold egalitarian explicit attitudes and beliefs, are often influenced by implicit mental processes that operate independently from, and are largely insensitive to, their explicit attitudes. Depending on how one interprets this conclusion, one may deduce from the empirical research that, disturbingly, no matter our moral convictions, the sub-personal processes of cognition that help us navigate the world will lead us to perform sexist and racist actions.²

The promise that empirical psychologists, armed with a new set of tools, would now be able to measure the operation of these sub-personal processes and ground this body of research in empirical fact has been another source of enthusiasm. We will argue here, however, that the data from the most prominent existing tests for implicit bias do not support the Implicit Explanation. Our argument will rely on the intuitive idea that some rapid mental associations, and the behaviors that such associations cause, express and reflect our explicit attitudes. For example, a teenager’s immediate and reflexive lunge for a slice of pizza might express his love of pizza; a father’s instantaneous grimace of pain when he sees his child skin her knee reflects his concern for her; a classical pianist’s tendency to associate the concept [MUZAK] with [BAD] is an expression of her belief that Muzak is a grating abomination. These agents are open books: their rapid associations and behaviors tell us something about their explicit attitudes. If the rapid associations and behavior measured by existing implicit bias tests were also straightforward expressions of agents’ explicit attitudes in this way, then the Implicit Explanation would be false. In this paper, we argue that implicit bias research has not undermined this banal Explicit Explanation of implicit bias test results.

To demonstrate that empirical research has failed to establish the Implicit Explanation, we will begin by showing that researchers have conducted implicit bias testing in light of inadequate theories of, and therefore without adequate controls for, explicit prejudice. As a result, their data may merely reflect subjects’ explicit prejudice that extant tests simply fail to measure. Next, we will offer suggestions for how researchers might improve their tests to attempt to address the criticisms we raise, while highlighting the underappreciated difficulties of interpreting the data that even the best possible empirical testing would yield. Finally, we will expand upon the philosophical significance of our criticism. We will show that in their eagerness to chart its philosophical and ethical implications, philosophers have not been sufficiently skeptical of the Implicit Explanation. As a result, they have encouraged a larger social shift in our understanding of how to relate to prejudiced agents that has serious moral costs. We conclude with the positive suggestion that philosophers interested in implicit bias shift their focus toward developing and defending views of explicit prejudice that could serve as the basis of successful controls in implicit attitude testing.

2 Measuring implicit bias

Before making the case that implicit bias research has failed to vindicate the Implicit Explanation, we note two caveats that will frame our discussion:

• 1. Most of the current work within social and empirical psychology, as well as much of the philosophical literature on implicit bias, relies on data obtained from several dominant experiments conducted and described in frequently cited papers. Most prominent is the “Implicit Association Test” (IAT), developed by Greenwald et al. (1998). Other dominant tests include the “Go/No-Go Association Task” (“GNAT”; described in Nosek & Banaji, 2001), the “Sorting Paired Features Task” (“SPF”; described in Bar-Anan et al., 2009), the “Brief Implicit Association Test” (BIAT; described in Sriram et al. 2009), the “Affect Misattribution Procedure” (“AMP”; described in Payne, 2009), and the “Affective Lexical Priming Score” (“ALPS”; described in Lebrecht et al., 2009). We also focus our discussion on these tests.

• 2. We use the term “implicit bias” to refer to bias that is at odds with a subject’s explicit attitudes.³ In pressing our criticism of implicit bias research, we treat “explicit attitudes” as attitudes that subjects could verbally report if prompted to do so. Common theoretical interpretations of “implicit attitudes” tend to assume that because subjects cannot verbally report their presence, or assess their influence on behavior, such attitudes must be “unconscious,” “uncontrollable,” or “arational.”

To demonstrate the truth of the Implicit Explanation, one would need to provide good evidence that subjects’ explicit attitudes did not underlie and explain their rapid associations and behavior. The most straightforward way to do so would be to show that even test subjects who harbored no explicit prejudice still exhibited biases in their rapid associations and actions. Indeed, this is what standard interpretations of implicit attitude testing claim that such testing does.⁴ Once a test established that a subject did not harbor explicit biases, one could safely conclude that her biased associations and actions in test conditions were best explained by implicit mental processes. But, as we will now argue, the dominant implicit bias tests fail to demonstrate such a disparity and therefore fail to demonstrate the existence of implicit bias. The reason is straightforward: such testing has been conducted with inadequate measures of subjects’ explicit prejudices.

The most common measures of explicit prejudice the dominant tests employ are the “Feeling Thermometer” and the “preference” survey. The first involves subjects’ self-reports of “temperature” on a numerical scale. For example, when filling out a Feeling Thermometer for implicit racial prejudice, participants are asked, “How warm or cold do you feel toward Black people?” and “How warm or cold do you feel toward White people?” On preference tests for implicit racial bias, participants are asked which of a set of statements best describes them: “I strongly/moderately/slightly prefer White people (or White Americans) to Black people (or African Americans)” or, “I strongly/moderately/slightly prefer Black people to White people,” or “I like White and Black people equally.”⁵

The preference and Feeling Thermometer surveys are the standard and dominant explicit attitude tests typically cited by both empirical psychologists and philosophers who write about implicit bias. In fact, preference and temperature surveys are the only means the canonical versions of the SPF (Bar-Anan et al., 2009, p. 333) and the GNAT (Nosek & Banaji, 2001, p. 651) use to measure explicit prejudice.⁶ Moreover, it is worth emphasizing that many studies draw exclusively on data obtained from the IAT as administered by the Project Implicit “virtual laboratory” (Rachlinski et al., 2009; Beaman et al., 2009, Sabin et al., 2009, for example). More than 20 million people have taken the IAT via the Project Implicit “Virtual Laboratory,” where racial implicit attitude tests have no controls for explicit prejudice beyond standard Feeling Thermometer and preference measures.⁷

These extant tests of “explicit prejudice” are insufficient to detect, and so to control for, explicitly prejudiced attitudes. This is because a person may sincerely feel and report “warmth” toward members of the group he is prejudiced against. Consider, for example, a form of prejudice that involves holding explicit beliefs about a person’s proper place in a social or natural hierarchy. Imagine, for example, a man who explicitly believes that women are goddesses who should be put on a pedestal, who should be pampered at home, but who lack the natural aptitude necessary to hold public office. Such a man could coherently, and sincerely, report feeling the same “temperature” toward both men and women, or even that he “feels warmer” toward women. Or imagine a person who explicitly believes stereotypes about white intellectual superiority and black physicality, who coherently and sincerely reports having no “preference for,” or feeling any difference in “warmth” toward, members of one racial group over the other.

To see just how poor these measures of explicit prejudice are, consider Jones, who explicitly believes that each race has its place in the world, with whites occupying the top rung. While Jones believes that non-whites are by nature inferior, he also holds that God loves people of all races. Jones is aware of the content of his beliefs, how he came to have these beliefs (e.g. he knows he developed them during childhood), and which of his actions result from these beliefs (e.g. he knows that he is harsher and sterner with young Black boys than with young white boys because the former “need more discipline” to “learn their place”). We can imagine, too, that Jones’s beliefs about the races play some role in shaping the patterns of emotions and desires that he feels. Moreover, we can imagine that Jones fills out the thermometer and preference questionnaires slowly and deliberately, and that he reasons to certain conclusions using basic rules of logical inference, exhibiting characteristics which indicate that his actions are the result of his explicit attitudes. When Jones is asked how warmly he feels about whites and Blacks, he reasons, “White Man was made in God’s image; Black Man was made by God as a lesser being whom White men are spiritually and morally obligated to help. Ultimately, we are all God’s creatures. 10/10 for both.” We can imagine, too, that Jones’s beliefs about the races play some role in shaping the patterns of emotions and desires that he feels.

Let us assume that Jones forms prejudiced conceptual associations that involve the concepts [BLACK] and [WHITE]. Because these prejudiced conceptual associations are totally consistent with his explicit attitudes, it would be surprising if Jones’s rapid associations, and whatever actions they cause, were anything more than expressions of his explicitly racist beliefs. But given the inadequacy of the dominant tests’ controls for explicit prejudice, Jones would be classified as a subject who holds “egalitarian beliefs” but holds “implicitly” prejudiced attitudes toward Blacks—a clearly absurd conclusion.

It should go without saying that one’s explicit racism need not be as obvious as Jones’s for the point to hold. Subjects whose prejudices are subtler than Jones’s, but nevertheless still explicit, will easily evade the standard controls as well. To infer the Implicit Explanation from results on the dominant implicit bias tests would not be justified.

3 Meeting the theoretical challenge?

To develop a test that could serve as a control for subjects’ explicit biases, one would need to know what kinds of explicit attitudes could potentially underlie and explain subjects’ biased associations and behaviors. This involves commitment to theoretical claims about what explicit prejudice consists in. Use of the Feeling Thermometer, for example, implies that feelings of relative “coldness,” and no other explicit attitudes, could explain biased rapid associations and behavior in test conditions. We have argued that the Feeling Thermometer is too crude to rule out the Explicit Explanation of such behavior, but it is not obvious what an adequate supplement to that rudimentary model of prejudice might be. In this section, we will present several conceptual and practical difficulties researchers would have to overcome to construct and use an acceptable alternative. We do not claim that building a satisfactory model would be impossible, but, as we will show, the current research is not close to meeting the challenges we will outline.

One difficulty is that a successful theory of prejudice must vary depending on the form of prejudice one is controlling for and the characteristic operations of that prejudice in the local context one is studying. Explicit racism toward Blacks in the southern United States, for example, may characteristically consist in a set of beliefs that differs from the beliefs held by explicitly prejudiced subjects in northern states. And for any general form of prejudice (racism, sexism), there will be competing conceptions of what that form of prejudice consists in. Is racist prejudice, for example, fundamentally a matter of holding different attitudes toward different racial groups (as the Feeling Thermometer seems to assume), a matter of holding attitudes that evince acceptance or endorsement of racial domination, or something else altogether? Does sexist prejudice consist in holding different attitudes toward men and women, holding attitudes that implicate one in patterns of objectification or sex-based domination, or in something else altogether?

Moreover, and to state the obvious, the mechanisms and psychological manifestations of racist and sexist prejudice in each context are complex and evolving. For example: Would recording the negative stereotypes participants held about members of a marginalized group be sufficient to supplement the Feeling Thermometer, or would one have to ask about “positive” stereotypes about members of that group as well?⁸ A good answer would draw on a theory of prejudice and its operations, and one could press similar questions about other attitudes and beliefs. Should researchers ask subjects about their positive beliefs about members of the dominant group? Their political ideals? Their appreciation of historical injustices? Their hopes and fears? Their senses of humor? Answering these questions, and determining the extent to which rapid associations and behaviors may be straightforward expressions of explicitly prejudiced attitudes, requires an intellectually serious theory of sexist and racist moral psychology and a plausible conception of the functioning of sexism and racism in the society under scrutiny.

Though they are rarely used to control for explicit attitudes in implicit bias research, there are psychological tests that reflect more theoretically sophisticated views of racism and sexism and thus seem to acknowledge some of these difficulties.⁹ These include John McConahay’s (1981) Modern Racism Scale, the Symbolic Racism Scale (Henry & Sears, 2002), the Diversity Scale (Wittenbrick et al. 1997), and the Discrimination Scale (Payne et al., 2009). Similarly, the Ambivalent Sexism Inventory (Glick & Fiske, 1996) clearly reflects a conception of sexism that goes beyond “temperature” and preference.

To be clear, this is not to say that these measures of explicit bias are “better” in the sense that they more accurately predict discriminatory behavior than temperature or preference scales. Rather, they are conceptual improvements.¹⁰ These more complex measures of explicit racism and sexism would make for better explicit bias controls because they strive to identify and measure ways in which a subject could harbor forms of explicit prejudice that would evade cruder surveys but that could still underlie and explain the subject’s rapid associations and behavior. For example, the Ambivalent Sexism Inventory presumes, rightly, in our view, that sexism is not merely a matter of “temperature” and preference, but, rather, that it can be constituted by more complex attitudes, many of which could be “positive,” “warm,” or, as the developers of the test write, “benevolent” (“A good woman should be set on a pedestal”; “Women have a quality of purity few men possess”). In selecting from among possible controls, researchers are, in effect, taking a position on the question of which explicit attitudes could underlie and explain subjects’ rapid associations. Again, this implies a commitment to a substantive theory of racism or sexism.

To be fair, some implicit bias researchers have used the more theoretically sophisticated measures we mentioned above.¹¹ But we will argue that even implicit bias tests that do employ these more robust controls fail to vindicate the Implicit Explanation.

To fully evaluate whether even the best existing tests are reliable indicators of explicit prejudice, one would first need to successfully present a justifiable and plausible theory of explicit prejudice. Then, one would need to show that some set of existing tests could, with a reasonable degree of accuracy, identify those respondents with views marking them as explicitly prejudiced according to the theory. Implicit bias researchers have not done this.

Even if implicit bias researchers did show that the most sophisticated existing surveys (or some combination of them) reflected the best theories of the biases they were studying, they would still have to overcome serious difficulties of interpretation in order to use them to control for explicit bias. Crucially, they would have to attend to the distinction between measuring explicit prejudice and controlling for it on an implicit bias test. Controlling for explicit prejudice requires ruling out the Explicit Explanation for implicit bias test results, not determining “how racist” or “how sexist” respondents are. All of the most sophisticated tests aim to measure subjects’ degrees of prejudice, but the Implicit Explanation is a claim about the foundations of certain associative and behavioral patterns. This means that while one might be justified in concluding, on the basis of a low overall score on the Modern Racism Scale, for example, that a subject harbored only a small degree of explicit racial bias, one would not be justified in concluding that the handful of prejudiced explicit attitudes the subject did harbor did not best explain his biased associations and behavior in test conditions.

The way psychologists tend to analyze the data they obtain from these tests virtually guarantees that they will not be able to rule out the Explicit Explanation. For example, in Greenwald et al. (1998), researchers scored subjects’ responses to questions about their racial attitudes on a scale of 1 to 5, with lower scores indicating less explicit prejudice. Thus, even a person receiving a very “low” overall score might, for example, strongly agree that “Blacks should not push themselves where they are not wanted.”¹² If such a person then registered biased rapid associations, would his answers to the explicit bias control rule out the Explicit Explanation? We think not. Despite his low overall score, it is plausible that this single prejudiced explicit attitude could best explain his rapid associative patterns. Or suppose a subject who took the Discrimination Scale (Payne et al., 2009) reported that he did not associate Black people with any of the listed stereotypes except the stereotype that they are violent, or the stereotype that they are not “intelligent at school.” Endorsing one or two stereotypes of this kind would not prevent him from registering a low overall score on the scale, but it could easily explain a host of biased associative patterns and actions.

Consideration of this issue raises a more general question of interpretation: What do subjects’ answers to these more sophisticated questions mean? What is the ethical significance, for example, of someone “agreeing somewhat” or “disagreeing slightly” with statements such as “Once a man commits [to a woman], she puts him on a tight leash,” (Glick & Fiske, 1996) or reporting that he has “felt sympathy for Blacks about half the time?” (Payne et al., 2009) What does it say about a person if he believes that “There is a real danger that too much emphasis on cultural diversity will tear the United States apart” (Wittenbreck et al., 1997), that “Black leaders have been moving at about the right speed” (Henry & Sears, 2002), or that “Women have a superior moral sensibility” (Glick & Fiske, 1996)?

In highlighting these possible questionnaire responses, we do not mean to suggest that the questions they answer are the wrong kinds of questions to ask. On the contrary, they may yield just the kind of data one would need to assess a person’s explicit racist or sexist attitudes. But the very complexity and nuance that distinguishes tests that ask such questions from their simpler counterparts make them difficult to analyze. A successful interpretation of a respondent’s answers to these loaded questions would require an excellent understanding not only of how prejudice functions in the context one is studying, but also of how the subject’s responses relate to his own psychology and background. Someone like Jones, for example, might feel sympathy for Black people all the time. He sees them as lowly and pitiable, after all. To score this response as evidence of anti-racism, or egalitarianism, when it is in fact just the opposite, would be a mistake. And a participant who agreed that women tended to keep men on “tight leashes” might be offering a report of her life experience, endorsing a misogynistic stereotype, or both. All of the most sophisticated tests for explicit racist and sexist prejudice are subject to these worries.

At this point, one might ask whether, in the absence of effective controls for explicit prejudice, and in light of these interpretive difficulties, there is other evidence in favor of the Implicit Explanation, or if there are alternative measures researchers could use to avoid some of the challenges we’ve highlighted here. For example, it may seem significant that, anecdotally, many people who take implicit association tests are surprised by their results, and one may wonder if such surprise is itself evidence for the Implicit Explanation. A genuine egalitarian, after all, would have reason to be taken aback upon learning that he formed prejudiced conceptual associations under test conditions.

Researchers have not attempted to use surprise as a control for explicit attitudes, but we might still ask whether such an approach could yield data that supported the Implicit Explanation.¹³ We doubt, however, that such a strategy would succeed, given that there are many reasons why one might be surprised by implicit bias test results that do not imply a discrepancy between explicit attitudes and rapid associations. We will offer just two here.

• 1. Subjects who “fail” implicit bias tests such as the IAT do so because they fail to perform a task the tests instruct them to perform, such as quickly sorting photographs into categories. Those who are surprised at their results may simply be surprised at their failure to complete the tasks successfully, which is compatible with being surprised that they were unable to prevent their explicit attitudes from influencing their rapid associative behavior. We might see such surprise outside of test conditions, as well. A person might be surprised, for example, that he could not suppress tears of joy at his child’s wedding, even though he tried very hard to keep his composure. But his surprise would not be evidence that his tears were anything but an expression of his explicit love and happiness.

• 2. A person may possess an inadequate conception of what “egalitarianism” requires and believe, wrongly, that he holds egalitarian explicit attitudes when, in fact, he does not. This misconception could explain a test subject’s sense of surprise at his results without appealing to a mismatch between his rapid associations and explicit attitudes. Imagine a man who believes that women have the right to work outside the home and to control their own bodies, but that as a matter of “scientific fact,” men make for better leaders. Such a man might be surprised by his propensity to associate [WOMAN] and [HOME] on an implicit attitude test, but this test result is not at odds with his explicit attitudes.

We are not claiming that these are in fact the best explanations of subjects’ surprise at their test results. Rather, these alternate explanations aim to undermine the inference from surprise to the Implicit Explanation. Measures of surprise cannot take the place of robust controls for prejudiced explicit attitudes that reflect sophisticated conceptions of explicit prejudice and its operation in particular contexts.

To be clear: our position is not that it is in principle impossible to effectively control for explicit bias.¹⁴ Talented researchers may be able to overcome the obstacles we have described. But we do believe that the challenges are formidable. Meeting them, and vindicating the Implicit Explanation, would require conceptual insight and ingenious test construction that the current empirical research has not achieved.

4 Implications and clarifications

We will now discuss the implications of this failure, both for philosophy and for society at large. The two are, of course, related, and it is easy to see why implicit bias research has excited both theorists and laypersons. Implicit bias scholarship purports to use empirical methods to demonstrate a surprising moral-psychological claim with enormous ethical and social implications: Even good people with good explicit attitudes (“committed egalitarians”) are subject to rapid associations and behavior that cause them to behave in prejudiced ways. If true, this would indeed be a disturbing conclusion—one that would call for us to rethink the ways in which we understand the operations of prejudice and our own psychology. We have argued that this claim, which has driven both academic scholarship and public discourse, has not been established.

At this point, one might protest that at least some of the recent criticism of implicit bias research, including the critique that we are raising here, is misplaced. One might suggest that we are reacting to a misrepresentation of the significance of implicit bias research by media members unfamiliar with the underlying science, by opportunistic public figures, or by institutions eager to employ “implicit bias training” in lieu of addressing deeper issues. We do not deny that such figures have sometimes exaggerated and misinterpreted the significance of the empirical research. But our criticism does not target the embellishments and “hype” in public discourse that psychologists and others familiar the empirical literature would reject.¹⁵ Rather, we are casting doubt on the basic claim that the biases the empirical tests reveal are implicit.

In doing so, we are also responding to the widespread uncritical acceptance of the Implicit Explanation among philosophers. Philosophical articles on implicit bias tend to remind their audiences of the significance of the topic by echoing the core claim of the empirical research, that it has provided us with good evidence of the Implicit Explanation, before going on to theorize about what could explain the discrepancy between explicit and implicit attitudes, or about the moral significance of this discrepancy. Here are some prominent examples:

• 1. “Research on implicit bias demonstrates that individuals can act in discriminatory ways even in the absence of explicitly prejudiced motivations” (Brownstein and Madva, 2018, p. 1).

• 2. “Evidence has been building that implicit attitudes are at least moderately good at predicting real-world behavior, independent of the effects of people’s explicit (verbally reported) attitudes” (Carruthers, 2018, p. 1).

• 3. “What explains the apparent disparity between self-reported attitudes and behaviour?… While self-presentation effects undoubtedly play a role, it is very likely that what psychologists call implicit attitudes explain some of the disparity between reported attitudes and behaviour” (Levy, 2017, p. 535).

• 4. “There is abundant evidence that most people, often in spite of their conscious beliefs, values, and attitudes, have implicit bias. ‘Implicit bias’ is a term of art referring to evaluations of social groups that are largely outside of conscious awareness or control” (Brownstein & Saul, 2016, pp. 1–2).

If our argument is sound, then even this basic framing of the topic, which begins with and relies on the conclusion that implicit bias tests have demonstrated a discrepancy between rapid associations and explicit attitudes, is misleading.

Once accepted, the Implicit Explanation gives rise to at least three kinds of philosophical questions that philosophers have been eager to take up. The first set of questions is metaphysical: What are these implicit biases that help to shape our behavior? Are they best understood as emotions, beliefs, “aliefs,” “behavioral clusters,” or something else altogether? And given what they are, are they unconscious, arational, or uncontrollable? One’s answers to these metaphysical inquiries lead to the second and third topics, which are both ethical: First, are we morally responsible, and therefore potentially blameworthy, for our biased rapid associations and behavior? Second, how should we go about altering our pernicious associative patterns, given our moral and political hopes of living in a more just society? We will consider these three areas, metaphysics, responsibility, and intervention, in turn.

Metaphysics: Credulous acceptance of the Implicit Explanation by the public and philosophers has been mutually reinforcing and has helped shape public thought about human psychology and the operations of individual prejudice. More specifically, it has helped establish and strengthen a conception of human psychology and prejudice that understands rapid associative behavior as a mental phenomenon divorced from, and potentially unresponsive to, the explicit attitudes that one might hope would form the basis of our self-understandings and our everyday interactions with other people. This conception depends on a particular kind of metaphysical interpretation of what underlies and explains rapid associative behavior, one that understands it as the product of a cognitive architecture that operates automatically, arationally, or outside of our conscious control.¹⁶

Moral responsibility: This understanding of the relationship between rapid associative behavior and human psychology has ethical implications. Notably, it is relevant to inquiry about our moral responsibility for rapid associative behavior and its consequences. Insofar as one’s metaphysical interpretation of rapid associative behavior distinguishes it from attitudes and beliefs that we rationally endorse or control, one may be moved to take up what P.F. Strawson called the “objective,” rather than “participant,”¹⁷ stance toward “implicitly” biased agents (including oneself), emphasizing treatment and management as opposed to reasoning, conversation, empathetic engagement, and the feeling and expression of “reactive attitudes.”

We realize that this characterization of the pressure to move away from the interpersonal ideals of the participant stance is abstract. There are, of course, competing views of responsibility for “implicit” bias in the philosophical literature, each of which has its own subtleties. To mention a few representative examples: Natalia Washington and Daniel Kelly (2016) argue that individual moral responsibility for implicit bias turns on whether or not knowledge of the empirical science surrounding implicit bias is available in that person’s community. If not, then the person should not be held fully accountable and blameworthy for his behavior. Maureen Sie and Nicole van Voorst Vader-Bours (2016) urge a re-orientation from individual responsibility toward collective responsibility for implicit bias. And Robin Zheng (2016) argues that we may hold “implicitly” biased agents accountable for their behavior in some sense (for example, by demanding that they compensate victims, make efforts to change their behavior in the future, or make amends), but that blaming them “would be like blaming a person for a behavior that they acquired as the result of some trauma, which gets triggered under certain circumstances; while such a disposition is something to be managed by her and others, it is not something for which she deserves blame or deep moral criticism” (79).¹⁸

Our aim here is not to evaluate the comparative merits of these metaphysical and ethical positions. Rather, we are focusing on what they share in common to make two general points. First, that one’s metaphysical interpretation of rapid associative behavior can make a concrete difference to our everyday interactions and relationships. Second, that acceptance of the Implicit Explanation, combined with standard views about moral responsibility, tends to encourage at least some shift away from the participant stance and its standard modes of interaction.

If the empirical research provided strong evidence for the Implicit Explanation, then one would need to carefully consider the merits of these views of responsibility for rapid associative behavior. If, on the other hand, the Explicit Explanation turned out to be correct (or largely correct), then we could reject all of them, and therefore avoid the serious moral and social costs of a shift toward the objective stance. If the Explicit Explanation is true, then to tell victims that transgressors’ behavior cannot be attributed to them, or that they are not proper targets of “deep” criticism and blame, would not only be misleading; it would deprive victims of the opportunity to confront and engage with offenders in a way that fully reflects their shared humanity. And to claim that offenders’ prejudiced associations are merely regrettable products of their cognitive architecture and their surroundings, rather than expressions of their rational agency, would not only be a distortion but an insult to the transgressors themselves.

Intervention: A similar point applies to efforts to alter prejudiced behavior. Acceptance of the Implicit Explanation may lead one to suspect that rational changes to rapid associative behavior are difficult or even impossible. And indeed, some recent work on implicit bias in empirical psychology has focused on how to mitigate or intervene on implicit bias using non-rational methods that rely on re-conditioning agents’ associations or encouraging subjects to exercise more control over their actions in order to align them with their egalitarian beliefs.¹⁹

But while one might be able to “recondition” one’s association of dark skin with the concept [DANGEROUS], this change would have nothing to do with one’s understanding, perception, or outlook toward people with dark skin. Similarly, it would be one thing to eliminate the stereotype that women are bad at math through retraining evaluative-semantic associations between the phrase “Women are good at” and “math.” It would be another thing to eliminate one’s belief that women are naturally bad at math by coming to a realization that women can be good at math, and to experience confidence in their mathematical abilities as a result.

To be clear, our concern is not primarily prudential but ethical, motivated by a moral commitment that is deeply embedded within the broadly liberal outlook that inspires standard objections to sexism and racism in the first place. The point is not that “rational” interventions would be more effective than non-rational interventions. Rather, our concern is that these non-rational methods, when aimed at agents whose rapid associations do in fact reflect their explicit attitudes, are inconsistent with respect. They merely manipulate a person’s behavior, rather than encourage rational changes in how he views and responds to others.

If the Implicit Explanation does turn out to be true, and re-conditioning is needed, this would be a disturbing fact, and one we should not take lightly. The conclusion that such means are necessary should be understood not as an opportunity to toast the dawning of an exciting new era of science-based approaches to moral improvement, but rather as a sobering concession that we cannot confront a new, pernicious form of racism and sexism without abandoning our traditional understanding of persons as responsible agents who could come to understand one another as moral equals. This should be seen for the pessimistic conclusion it is, regardless of the cheery progressive attitude that can sometimes accompany it. Our argument, that psychologists and philosophers in this debate have not shown that the biases in question are indeed implicit, gives us reason to eschew, or at least delay, such pessimism.

5 Conclusion

It is important that philosophers, especially moral psychologists, engage with empirical observation, and not only because their work should be informed by real life, and what real human beings are like. Philosophical reflection and conceptual refinement allow us to better interpret experimental data. In the case of implicit bias, philosophers have been eager to interpret empirical results, but they have largely focused their attention on answering the three kinds of questions we outlined above. In the process, they have produced a vast body of technical literature in a short period of time, offering critical and competing views of what implicit attitudes are and the implications that follow from these metaphysical conclusions. We have not compared the relative merits of these positions because our contention is that the philosophical drive to interpret needed to express itself at an earlier stage.²⁰

Thus, though we are critical of the extant work on implicit bias, our thesis is not entirely negative. We believe that our understanding of the current research’s conceptual shortcomings suggests a way in which both empirical psychologists and philosophers working on implicit bias might productively change course. Psychologists should devote attention toward finding better ways to measure (and not solely predict) explicit prejudice, rather than focusing primarily on refining their methods of tracking implicit associations and behavior. In doing so, they should bear in mind the points we raised in section III, where we discussed the difficulties of developing adequate controls for explicit prejudice and stressed that doing so would require both a sophisticated theory of the way that prejudice operates in the context under study and a sense of the agent’s psychology and history. Developing effective controls for explicit prejudice will involve drawing on normative moral-psychological theories—theories that help us identify morally good and bad states of mind. Ethicists and normative moral psychologists must be seated at the table alongside the empirical psychologists and philosophers of mind who take on the challenge.^21,22

Appendix

Where the full content was not made available by authors, we include the “Representative Items” listed.

Glick, Peter and Susan T. Fiske. “The Ambivalent Sexism Inventory: Differentiating Hostile and Benevolent Sexism.” Journal of Personality and Social Psychology, 70(3), (1996).

Greenwald, A., D. McGhee, and J. Schwartz. “Measuring Individual Differences in Implicit Cognition: The Implicit Association Test”. Journal of Personality and Social Psychology 74 (1998): 1464–80.

Henry, P.J. and David O. Sears, “The Symbolic Racism 2000 Scale,” Political Psychology, 23:2 (2002).

McConohay, John, Betty B. Hardee and Valerie Batts, “Has Racism Declined in America? It Depends on Who is Asking and What is Asked,” The Journal of Conflict Resolution 25(4), 1981.

The “Modern Racism” Scale:

Payne, B. (2009). “Attitude Misattribution: Implications for Attitude Measurement and the Implicit-Explicit Relationship.” In A. Black and W. Prokasy (Eds.) R. Petty, R. Fazio, and P. Brinol (Eds.), Attitudes: Insights from the new wave of implicit measures. Hillsdale, NJ: Erlbaum.

1.1 Sample 1 (ANES Panel Study)

Feelings: Do you feel warm, cold, or neither warm nor cold toward blacks? [Response options: Warm, cold, neither warm nor cold].

If warm: Do you feel a little warm, moderately warm, or extremely warm toward blacks?

If cold: Do you feel a little cold, moderately cold, or extremely cold toward blacks?

Sympathy: How often have you felt sympathy for blacks? [Always, most of the time, about half the time, once in a while, or never].

Admiration: How often have you felt admiration for blacks? [Always, most of the time, about half the time, once in a while, or never].

Influence: Would you say that blacks have too much influence in American politics, just about the right amount of influence in American politics, or too little influence in American politics? [Response options: Too much influence, Just about the right amount of influence, Too little influence].

1.2 Sample 2 (ANES Time Series)

Feelings: Do you feel warm, cold, or neither warm nor cold toward blacks? [Response options: Warm, cold, neither warm not cold].

If warm: Do you feel a little warm, moderately warm, or extremely warm toward blacks?

If cold: Do you feel a little cold, moderately cold, or extremely cold toward blacks?

Sympathy: How often have you felt sympathy for blacks? [Always, most of the time, about half the time, once in a while, or never].

Admiration: How often have you felt admiration for blacks? [Always, most of the time, about half the time, once in a while, or never].

Stereotypes:

Where would you rate BLACKS on this scale? [ 1 = Hardworking; 7 = Lazy].

Where would you rate BLACKS on this scale? [ 1 = Intelligent; 7 = Unintelligent].

Symbolic Racism:

• 1. Irish, Italians, Jewish, and other minorities overcame prejudice and worked their way up, blacks should do the same without special favors” [Disagree, Somewhat disagree, Neither agree nor disagree, Somewhat agree, Strongly agree].

• 2. Generations of slavery have created conditions that make it difficult for blacks to work their way out of the lower class” [Disagree, Somewhat disagree, Neither agree nor disagree, Somewhat agree, Strongly agree].

• 3. It’s really a matter of some people just not trying hard enough; if blacks would only try harder they could be just as well of as whites. [Disagree, Somewhat disagree, Neither agree nor disagree, Somewhat agree, Strongly agree].

• 4. Over the past few years, blacks have gotten less than they deserve.” [Disagree, Somewhat disagree, Neither agree nor disagree, Somewhat agree, Strongly agree]

1.3 Sample 3 (Associated Press/Yahoo! News/Stanford University study)

Liking: “How much do you like or dislike each of the following groups? Whites … Blacks …” [Response options: dislike a great deal, dislike a moderate amount, dislike a little, Neither like nor dislike, like a little amount, like a moderate amount, like a great deal].

Admiration: “How often have you felt admiration for blacks?” [Extremely often, Very often, Moderately often, Rarely, Never].

Sympathy: “How often have you felt sympathy for blacks?” [Extremely often, Very often, Moderately often, Rarely, Never].

Stereotypes: Respondents were asked “How well does each of these words describe most blacks?” and were shown a list of 14 adjectives (Friendly, Determined to succeed, Law abiding, Hard-working, Intelligent at school, Smart at everyday things, Good neighbors, Dependable, Keep up their property, Violent, Boastful, Complaining, Lazy, Irresponsible). [Extremely well, very well, moderately well, slightly well, not well at all].

Wittenbreck B., CM Judd, and B Park. “Evidence for racial prejudice at the implicit level and its relationship with questionnaire measures.” Journal of Personality and Social Psychology. 72(2):1997.