Definitional boundaries of discrimination: Tools for deciding what constitutes discrimination (and what doesn't)

What counts as discrimination? Sometimes an event has to be a deliberate act of hate before it is described as discrimination. Sometimes “ discrimination ” can include much more subtle actions (e.g., microaggressions). There is good evidence that “ what counts ” as discrimination is mired in controversy, uncertainty, or ambivalence. We present a novel approach that bridges sociocultural and social cognitive accounts of “ discrimination. ” Definitional boundaries of discrimination are the widely shared, common sense rules that people deploy when arguing whether an event constitutes discrimination or not (e.g., an emphasis on the personality of the alleged perpetrator; the importance of intention; and claims that the problem of discrimination is small and/or mostly limited to the past). These rules are culturally situated but also deployed by individuals in specific local contexts. We introduce a 15 ‐ item measure of the extent to which participants are deploying broader or narrower definitional boundaries of discrimination (DBDs). We demonstrate that the measure has good convergent and discriminant validity (Study 1); that participants who are deploying narrow DBDs are less likely to make attributions to discrimination (Study 2); that participants' DBDs predict judgments for some intergroup contexts but not others (Study 3), and that participants who identify as racialized majorities have significantly narrower racism DBDs compared to participants who identify as racialized minorities (Study 4). We conclude with suggestions about how DBDs could be used in future research.


| INTRODUCTION
What is discrimination?What does it look like?Consider the constellation of actions in Figure 1. 1 Most people would agree that acts of hate constitute discrimination (but see also Crandall et al., 2013).There is much less consensus about subtle actions (e.g., microaggressions), or actions that are subjectively positive (e.g., complimentary stereotyping, helping behavior, and interactions between friends; Greenland et al., 2020;Sommers & Norton, 2006).These actions are frequently not constructed as discrimination, even though they have discriminatory impacts (Czopp et al., 2015;Douglass et al., 2016;Nadler, 2020;West, 2019).Definitional boundaries of discrimination address where and how people demarcate the boundary between discrimination and not discrimination.Theoretically, people can define discrimination in very narrow terms to exclude everything but overt acts of hate.Conversely, people can include a much broader range of actions.Importantly, people may also define discrimination in narrower/broader terms when in different contexts (e.g., for perpetrator-/target-combinations that implicate the self; West et al., 2022;Carter & Murphy, 2015).
The concept of definitional boundaries of discrimination bridges sociocultural and social cognition research: it considers how cultural constructions of "discrimination" are deployed by an individual (see also Crandall & Eshleman, 2003).We synthesize the extant research literature to suggest that there are widely shared cultural resources that construct "discrimination" in particular ways.We introduce a measure (DBDs) that records the extent to which participants are deploying these resources (Study 1). 2 We demonstrate that these DBDs have meaningful consequences; participants who deploy more narrow DBDs are less likely to make attributions to discrimination (Study 2); and participants' DBDs predict judgments in some intergroup contexts but not others (i.e., prototypical and reverse discrimination; Study 3).Finally, we demonstrate that racialized majorities are significantly more likely to deploy narrow racism DBDs (compared to racialized minorities, Study 4). 3 These constructions therefore work to reduce the probability that people will construct an event as discrimination.Definitional boundaries of discrimination therefore constitute a tool with which to represent sociocultural phenomena in a form that is usable within social cognition: they measure the deployment of cultural rules at an individual level and within a specific local context.

| Social cognition and sociocultural approaches to "discrimination"
The methods by which ordinary people make sense of "discrimination" have been examined within both social cognitive and sociocultural research traditions (e.g., Barreto, 2015;Salter et al., 2018).Social cognition research explores the cognitive processes that contribute to participants' decision making when making attributions to discrimination.Research in this tradition might describe the factors (individual, motivational, situational) which, when present, will cause a significant proportion of participants to make an attribution to discrimination (e.g., Major & Crocker, 1993).In contrast, sociocultural research explores the cultures, practices, and institutions that produce and reproduce "discrimination."Research in this tradition might explore the particular arguments that participants deploy in conversations about "discrimination" and the themes that emerge repeatedly in these arguments (e.g., Greenland et al., 2018;Perez & Salter, 2019).These two approaches ("in the head" and "in the world"; Salter et al., 2018) are not in conflict, and indeed there are examples of research that bridge both (Adams et al., 2008;McCoy & Major, 2003;Nelson et al., 2013;O'Brien, Crandall, et al., 2010;White et al., 2021).
Interestingly (and importantly for our argument), there are close parallels between research findings within each tradition.Social cognition indicates that there are widely lay beliefs about/lay beliefs about (for example) racism and racists.Findings suggest that the most prototypical form of discrimination is White on Black racism and involving an act of physical or symbolic violence (Flournoy et al., 2002;Inman & Baron, 1996;Inman et al., 1998;Sommers & Norton, 2006).
"Discrimination" is often associated with a particular kind of personality: violent, irrational, and unreasonable, or else ignorant and socially inexperienced (Esses & Hodson, 2006;Miglietta et al., 2014;Sommers & Norton, 2006).Evidence that someone intended to discriminate is considered particularly diagnostic, especially when they intended to cause harm (Bastart et al., 2021;Swim et al., 2003).Finally, there is good evidence that members of majority-, dominant-, or high-status groups tend to underestimate contemporary levels of discrimination and think of it as something that is largely limited to the past (Kraus et al., 2017(Kraus et al., , 2019;;Miron et al., 2011;M. I. Norton & Sommers, 2011;West & Eaton, 2019).
Sociocultural research finds very similar themes: participants often argue that discrimination is done by a particular kind of person (Andreouli et al., 2016;Billig, 1988;Figgou & Condor, 2006;Greenland & Taulke-Johnson, 2017;Howarth, 2009;Kadianaki, 2014;Kirkwood et al., 2013;Schwalbe et al., 2000;A. Wilkins, 2012); that the best evidence constitutes what someone really thought or intended (Greenland et al., 2018); and that discrimination is a problem that is either over or very nearly over (Andreouli et al., 2016).People also argue that you need a strong evidence base to make a claim of discrimination and that the relationship between actors is key (e.g., the contrast between friends and strangers; Andreouli et al., 2016;Greenland et al., 2018Greenland et al., , 2020)).
There is therefore a triangulation of themes that appear in both the social cognition and sociocultural literature.Importantly, these are not just repeated in White participants' understandings of racism: similar results are found in the context of sexism and heterosexism, and by both majority and minority participants (e.g., Bastart et al., 2021;Greenland et al., 2018).We also hear comparable arguments in other contexts, for example, the use of intention arguments within apologies and excuses (Burford-Rice & Augoustinos, 2018;Sliwa, 2019) and within the law (e.g., the distinction between murder and manslaughter in the United Kingdom).Indeed, these arguments tend to sound like common sense to anyone who lives in a WEIRD 4 culture (Henrich et al., 2010;Moscovici, 1984Moscovici, /2000;;Salter et al., 2018).Greenland et al. (2018) suggested that this repetition indicates the presence of broad ideological-cultural resources that have become "sedimented" into everyday practices (Gibson, 2015;Potter & Hepburn, 2005).In other words, this repetition does not arise from individual information processing, but rather from the deployment of widely shared cultural resources that are reproduced at an individual level (Adams et al., 2019;Crandall et al., 2013;Dixon et al., 2020;Salter et al., 2018).
Definitional boundaries of discrimination circulate as culturally available arguments, but the extent to which they are deployed (and in what contexts) is not fixed: the boundary between "discrimination" and "not discrimination" can change very markedly and within a relatively short period of time (Crandell et al., 2013;Kleinpenning & Hagendoorn, 1993;van Dijk, 1992).People's definitional boundaries also seem to be strongly influenced by the actions of others around them (Blanchard et al., 1991(Blanchard et al., , 1994;;Condor, 2006;Crandall et al., 2002Crandall et al., , 2013;;Durrheim, 2017;Thomas et al., 2020;Williams et al., 2019) and by intergroup processes (Carter & Murphy, 2015;Miron et al., 2011;Platow et al., in press;Platow et al., under review;Wang et al., 2022;White et al., 2021).This malleability indicates the extent to which definitional boundaries of discrimination are both culturally constructed and sensitive to local/intergroup contexts.In this respect they are closer to social representations or social norms than to prototypes: Definitional boundaries of discrimination can be understood as "justification ideologies" or "legitimizing myths" (Crandall, 2000;Crandall & Eshleman, 2003;Pratto et al., 2006), that is, tools by which people can rationalize their actions and the status quo.
It should be clear that definitional boundaries of discrimination are a topic of heated, polarized, and increasingly politicized public debate (e.g., Dost et al., 2019;Durrheim et al., 2018;Ellemers & van den Bos, 2012).These arguments are consequential in that they set the standards by which people judge themselves and others: they set estimates about the prevalence of discrimination, and what (if anything) should be done about it (Adams et al., 2008;Andreouli et al., 2016Andreouli et al., , 2020;;Greenland et al., 2021;Crandall & Eshleman, 2003;Crandall et al., 2002;DiAngelo, 2010;Dixon et al., 2015;Frosh et al., 2000;O' Brien, Garcia, et al., 2010;Rucker & Richeson, 2021;Unzueta & Lowery, 2008).We should therefore be able to predict people's attributions to discrimination and associated judgments if we know the definitional boundaries that they are deploying.Moreover, contemporary instantiations of definitional boundaries tend toward the maintenance of the status quo: they are more effectively deployed to resist claims of discrimination and narrow the range of actions that can be understood as "discrimination" (Greenland et al., 2018(Greenland et al., , 2020;;Perez & Salter, 2019;Salter & Adams, 2013;Salter et al., 2018).
We would therefore expect that people who are deploying narrow contemporary instantiations of definitional boundaries of discrimination (as measured with DBDs) should be less likely to make attributions to discrimination (Study 2).In as much as definitional boundaries of discrimination are cultural tools that are deployed flexibly according to the local (intergroup) context, then we would also expect that participants' DBDs will predict judgments differently in different intergroup contexts (i.e., prototypical and reverse discrimination; Study 3).Finally, there is good evidence that majority, dominant, or high status groups tend to be less likely to make attributions to (identity-relevant) discrimination compared to minorities (Carter & Murphy, 2015; but see also Dost et al., 2019).We expect that people who identify as members of racialized majority groups will have significantly narrower racism DBDs compared to participants who identify as members of racialized minority groups (Study 4).

| Current research
We have outlined the concept of definitional boundaries of discrimination as resources that construct "discrimination" in broader or narrower terms.We have synthesized the extant research literature to outline how discrimination is conceptualized in WEIRD cultures.In this paper we present a measure of DBDs and use it to test the three hypotheses outlined above.In Study 1, we test the relationship between DBDs and extant measures.In Study 2, we test the hypothesis that participants who deploy narrower DBDs are less likely to make attributions to discrimination, and that this is consequential for other kinds of judgments.In Study 3, we test the extent to which DBDs are deployed differently within different contexts (i.e., prototypical and reverse discrimination).In Study 4, we compare the extent to which different racialized groups (majorities and minorities) deploy narrower or broader racism DBDs.

| STUDY 1: DEVELOPMENT OF A MEASURE
We present the 15-item DBDs measure, which records the extent to which participants are deploying contemporary definitional boundaries of discrimination (i.e., as narrow or broad).We explore convergent and discriminant validity with a barrage of established measures including measures associated with racism; measures associated with attributional complexity; and socially desirable responding.

| Rationale for items
We have outlined the range of different processes (within social cognition) and arguments (within sociocultural research) that are implicated when people make sense of different -isms (racism, sexism, heterosexism).These processes/arguments are diverse, interdependent, and not necessarily logically consistent.Crucially, however, they are repeated across a range of different methods, samples, and target groups.This repetition was central to the development of the measure, and consequential in two ways.
First, the final set of items contain a combination of behavioral, attitudinal, and more general questions.While the distinction between "prejudice" (attitudinal) and "discrimination" (behavioral) components has a long tradition in psychology, it is not reflected in public discourse.In practice, people use a combination of attitudinal and behavioral judgments when making attributions (e.g., Bastart et al., 2021;Swim et al., 2003); locate "racism" (for example) in the bodies of particular actors (Flournoy et al., 2002;Inman & Baron, 1996;O'Brien, Crandall, et al., 2010), but also in institutions and structural patterns of inequality (e.g., Macpherson, 1999).In some cases, "racism" also has specific geographical and historical properties (Andreouli et al., 2016;M. I. Norton & Sommers, 2011;Onyeador et al., 2020).A measure that focused on only one of these elements would be missing the different arguments in current public discourse.To put it another way, the items on the measure were designed to be mixed, as a reflection of how arguments that exist in public discourse are also mixed.
Second, we designed the measure so that it should be possible to "swap out" one form of discrimination for another (e.g., racism, sexism, heterosexism).As already outlined, these processes/arguments are repeated across a range of different -isms.There is therefore good reason to believe that these arguments will be heard in a range of intergroup contexts (although we used racism in all studies presented here).This is one reason that we refer to definitional boundaries of discrimination (rather than definitional boundaries of Racism): we believe the measure has a broader application.We also use "discrimination" (rather than "racism") because racism is a messy construct (within public discourse), whereas discrimination is relatively clearly understood (i.e., as an action that is consequential).Focusing on "discrimination" yields a more precise conceptual space because alternative words (e.g., "prejudice," "racism" or "stereotyping") already contain hearable attributions (i.e., in terms of attitudes or beliefs; Salter et al., 2018).

| Predictions
As outlined above, participants who score high on DBDs (corresponding to more narrow definitional boundaries of discrimination) should be less likely to make attributions to discrimination.
Participants would be drawing on constructions which enable them to resist claims of discrimination for all but the most egregious acts of hate.As such, we predicted that there would be a positive association between narrow DBDs and measures associated with hostile intergroup relations and/or preferences for social hierarchies (e.g., modern racism, systems justification, social dominance orientation).Greenland et al. (2018) further suggested that narrow DBDs are relatively simple, culturally available tools for making sense of the world (Crandall et al., 2002;Inman & Baron, 1996;Moscovici, 1984Moscovici, / 2000)).We therefore predicted that there would be a negative relationship between narrow DBDs and measures associated with preferences for cognitive complexity (e.g., attributional complexity; need for cognition).
We designed items to have good face validity without being too reactive.As such, we predicted that there would be no relationship between DBDs and a measure of socially desirable responding.

| Design
We used an online questionnaire to assess the correlation between narrow DBDs and established questionnaires associated with hostile intergroup relations, preferences for social hierarchies, preferences for cognitive complexity, and socially desirable responding.

| Participants
Participants were recruited via an online community panel ( n = 177).
They were not paid.Of these participants, 165 identified as British and 168 identified as White: 63 described their gender as male, and 107 described their gender as female.Participants were aged between 19 and 79 years (M age = 53.7,SD = 15.3).The sample size was opportunistic, that is, geared toward achieving a sufficient number to examine correlations, and dependent on the number of participants available for data collection.
The software randomly allocated participants to one of the three packets of materials described below.The number of participants for packets one, two and three were 62, 57, and 58, respectively (participant numbers vary due to missing data). 5

| Materials
An initial pool of 74 DBD items was developed.Items represented themes from existing research and that have been observed in a range of different target and sample groups.The themes in the measure included: that discrimination is performed by a particular kind of person (e.g., "A sensible person is not likely to be racist"); the need for a strong evidence base (e.g., "To say that someone is racist, you need to show that race and only race made them do what they did"); the importance of understanding intention (e.g., "The core of racism is that it is malicious: if you are not being malicious, then it can't be racism"); that claims of discrimination are often illegitimate (e.g., "Some groups are far too willing to make accusations of racism when they don't know the facts"); and that the problem of discrimination is small and/or mostly limited to the past (e.g., "Racism will probably die out completely within the next generation or two").
Participants responded to the items using a 7-point scale anchored at "strongly disagree" to "strongly agree."Items were coded so that high score would indicate that participants were deploying narrow constructions of discrimination (with three items thus reversed).
We selected items on the basis (1) that each had good variance (SD > 1; 2 < M < 6); (2) that they loaded with an eigenvalue > 0.45 during principle components analysis; (3) that they had an item-total correlation > 0.30 within reliability analysis.Thirty-eight items did not meet these criteria.After some reflection, 10 items were removed based on face validity (because each made explicit reference to minority groups, and we were concerned that they confounded beliefs about racism with actual racism).We reviewed the remaining  (.90;Brigham, 1993); importance to identity (.74; Luhtanen & Crocker, 1992); self-other intergroup anxiety (.91 and .93;Greenland et al., 2012); generic intergroup anxiety (.86; Plant & Devine, 1998); and internal and external motivation to control prejudice (.83 and .78;Plant & Devine, 1998).

| Analysis
We predicted that there would be a positive association between narrow DBDs and measures associated with hostile intergroup relations and/or preferences for social hierarchies.Results supported this hypothesis (see Table 1).Narrower DBDs were positively associated with both measures of racism (modern racism and racial attitudes) and preferences for social hierarchies (systems justification beliefs, meritocracy ideology, and social dominance orientation).The more narrow participants' DBDs, the higher they scored on measures of racism and measures that justified the status quo.However, there was no relationship between DBDs and ethnic identification (as measured by the importance to identity subscale), and no relationship between DBDs and any of the intergroup anxiety measures (self-, other-and generic intergroup anxiety).We had expected a positive relationship between narrow DBDs and identification, given the wellestablished relationship between identification and ingroup bias (but see also Brown et al.,1992;Mummendey et al., 2001).We used an alternative measure (i.e., Leach et al., 2008) in subsequent studies.We had considered that there might be a potential relationship between DBDs and intergroup anxiety, since anxiety often includes concerns about appearing to be racist (Greenland et al., 2012).Possibly individuals may be similarly anxious about showing prejudice but at different definitions of discrimination (i.e., those with narrow definitions might be anxious about using an offensive word, while those with broader definitions might be anxious about microaggressions).Narrow DBDs were negatively associated with internal motivation to control prejudice (which is associated with more progressive attitudes, Devine et al., 2002;Plant, 2004), but there was no relationship between DBDs and external motivation to control prejudice.This was not surprising as high levels of internal motivation to control prejudice predict lower levels of prejudice and outgroup negativity.Higher levels of external motivation to control prejudice tend to indicate perceived pressure from others to be or appear less prejudiced.
We predicted that there would be a negative relationship between narrow DBDs and measures associated with preferences for cognitive complexity.The findings provided partial support for this hypothesis.Participants with more narrow DBDs also tended to score lower on preference for attributional complexity, that is, people who deployed narrow DBDs also tended to like simpler attributions.
However, there were no statistically significant correlations with intolerance for ambiguity, need for cognition, or need for closure.
Finally, and as we hoped, there was no relationship between DBDs and socially desirable responding.This confirmed that the measure was not too reactive.

| Discussion
Study 1 provided good preliminary data on convergent and discriminant validity.Participants who deployed more narrow DBDs (as evidenced by a high score) also tended to score higher on variables that are known to be associated with reduced attributions to discrimination (e.g., modern racism and meritocracy ideology).That said, some of the correlations were very high (i.e., r > .5;Cohen, 1988): it was therefore possible that DBDs contained racial attitudes in addition to representations of discrimination.There was also a noticeable overlap in wording between one of the themes within the DBDs measure (i.e., that the problem of discrimination is small and/or mostly limited to the past) and an item on the modern racism measure ("Discrimination against blacks is no longer a problem in the United States").In Studies 2 and 3 we therefore included modern racism (as an example of one of the correlates, and one that we particularly wanted to exclude) as a covariate to test whether DBDs can indeed be considered to be empirically distinct.

| STUDY 2: PREDICTIVE VALIDITY OF DBDs FOR ATTRIBUTIONS TO DISCRIMINATION VERSUS A COMPARABLE MORAL VIOLATION
We tested the hypothesis that participants who deployed narrow DBDs would also be less likely to make attributions to discrimination (and associated judgments), but that there would be no relation between DBDs and a comparable moral violation (i.e., embezzlement from a charity).
Participants were asked to complete the DBDs measure, followed by one of two vignettes.The discrimination vignette described an employer who had been accused of failing to promote an employee.The embezzlement vignette described a manager who had been accused of failing to manage an expenses account.Each vignette was designed to be moderately ambiguous as to whether this accusation was reasonable or not.Participants were asked to make attributions to discrimination/embezzlement, and how much they thought that the alleged perpetrators and victims were to blame for what happened.Modern racism and ingroup identification were included as covariates (since both have been shown to be associated with attributions to discrimination: Pinel, 1999;Sellers & Shelton, 2003;Sommers & Norton, 2006).
We predicted that for participants in the discrimination condition, narrow DBDs (indicated by a high score) would be negatively associated with attributions to discrimination, negatively associated with blaming the alleged perpetrator, and positively associated with blaming the alleged victim.This relationship would be statistically significant even when modern racism and ingroup identification were included as a covariates.We expected there would be no relationship between DBDs and the dependent variables in the embezzlement condition.Finally, participants completed the 7-item modern racism scale (α = .85;McConahey, 1986), and a 14-item identification measure for their heritage group (α = .93;Leach et al., 2008), with higher scores indicating higher racism and stronger identification respectively.

| Analysis
Our principle aim was to demonstrate that DBDs predicted attributions to a moral violation (discrimination or embezzlement) and associated judgments of blame, but in the discrimination condition only.We also wanted to show that this relationship remained statistically significant even when controlling for modern racism and ingroup identification.We tested this hypothesis using the PROCESS macro version 3.5, Model 1 (Hayes, 2018).In each analysis, the model was defined as follows: X = narrow DBDs, Y = dependent variable, W (moderating variable) = condition, and the covariate.
Modern racism and identification were included as covariates in separate analyses to preserve degrees of freedom.
Consistent with our expectations, there was a significant moderating effect of condition on the relationship between narrow DBDs and attributions to discrimination/embezzlement, even when modern racism was included as a covariate; ß = −.50,SE = 0.17, Thus, the more narrow participants' DBDs, the less likely they were to make an attribution to discrimination.This effect was significant even when controlling for modern racism and identification.There was no relationship between DBDs and attributions to embezzlement (Figure 2).
The same pattern emerged regarding blame to the alleged perpetrator.There was a significant moderating effect of condition The narrower participants' DBDs, the less they blamed the alleged perpetrator.As above, these relationships were only observed in the discrimination condition, and even when modern racism and identification were covaried out of the analysis.
We conducted supplementary analysis in which we explored the interaction between the covariates (modern racism and identification) and condition to predict the dependent measure, and tested for any three-way interactions between DBDs, the covariates, and condition to predict the dependent measures (see Supporting Information).
Results suggested that participants' scores on modern racism did interact with condition to predict the dependent measures, and in the same pattern (reduced attributions to discrimination, reduced perpetrator blame, and increased victim blame) that we have evidenced above.There were no three-way interactions between DBDs, modern racism, and condition to predict the dependent measures.There was some evidence of a three-way interaction involving ingroup identification, such that the difference between conditions (in terms of the relationship between DBDs and the dependent measures) became stronger as ingroup identification increased (see Supporting Information).

| Discussion
Results supported the hypothesis.Participants who deployed more narrow DBDs were less likely to make attributions to discrimination.
DBDs were also consequential for judgments about blame: participants who deployed more narrow DBDs were less likely to blame the alleged perpetrator.There was no clear relationship between DBDs and blaming the alleged victim.
As expected, there was no relationship between DBDs and attributions for a comparable moral violation (i.e., embezzlement): the DBD measure was recording something specific to discrimination rather than morality per se.Further, the relationship between DBDs and the dependent variables was statistically significant even when modern racism and ingroup identification were included in the model.& Murphy, 2015).We would argue that a shift from attitudes (as measured by modern racism) to constructions of discrimination (as measured by DBDs) is important both culturally (in capturing public discourse) and practically (in suggesting ways of achieving change).
We return to this in Section 6.We would also note that modern racism and identification were measured after the manipulation, which (although there was no significant difference according to condition) constitutes a flaw in the design.In Study 3, the covariates were measured before (so that they were truly independent of the manipulation).

| STUDY 3: PREDICTIVE VALIDITY OF DBDs FOR DIFFERENT INTERGROUP CONTEXTS
Study 2 demonstrated that DBDs can be used to predict attributions to discrimination, but (as expected) not a comparable moral violation.
Study 3 extended Study 2 to explore the extent to which DBDs can predict attributions to discrimination for different intergroup contexts (i.e., prototypical and reverse discrimination).The prototypical form of discrimination is White on Black racism (Inman & Baron, 1996).Increasingly, however, we also hear claims of reverse discrimination, that is, discrimination against White people (e.g., Johnson & Goodman, 2013;Monteith & Hildebrand, 2020;Mutz, 2018;M. I. Norton & Sommers, 2011). 9To what extent are these claims of reverse discrimination treated according to the same criteria as the prototypical form of White on Black discrimination?
If attribution to discrimination is a logical process (as is often claimed in public discourse), then people should apply the same widely shared arguments irrespective of the context.If, however, attribution to discrimination is also motivated by intergroup processes, then we might expect that White majority participants will mobilize arguments in different way when the role of the ingroup shifts from being the alleged perpetrator to being the alleged victim (Hewstone, 1990;Mayrl & Saperstein, 2013;O'Brien, Crandall, et al., 2010;White & Crandall, 2017;C. L. Wilkins & Kaiser, 2014;C. L. Wilkins et al., 2017).
There is also evidence that members of stigmatized groups are expected to be more tolerant compared to those who have not been stigmatized (Branscombe et al., 2015;Fernández et al., 2014;Sambaraju, 2021).We examine these predictions in Study 3.
Participants were asked to complete the DBDs measure in a pretest.
Later (minimum 1 month), they were asked to read a vignette which was interpretable as either an act of prototypical discrimination by a majority ingroup against a minority outgroup (Belgians against Turks in Belgium) or an act of reverse discrimination by a minority outgroup against a majority ingroup (Turks against Belgians).The dependent measures were the same as in Study 2 (attributions to discrimination, blame to the alleged perpetrator, and blame to the alleged victim).
We predicted that there would be a negative relationship between narrow DBDs and attributions to discrimination in the prototypical discrimination condition, that is, the more narrow participants' DBDs, the less likely they would be to make an attribution to discrimination.
We also predicted that they would be less likely to blame the alleged perpetrator and more likely to blame the alleged victim.These predictions are consistent with Study 2. In contrast, we predicted a positive relationship between narrow DBDs and attributions to discrimination in the reverse discrimination condition.We theorized definitional boundaries of discrimination as a tool which can be deployed to maintain the status quo (Crandall & Eshleman, 2003;Crandall, 2000;Pratto et al., 2006).Participants who mobilize narrow DBDs are constructing "discrimination" in a specific way that minimizes its prevalence and its impact.We expected that participants who do this might also be more alert to threats to the racialized social order, that is, reverse discrimination.We therefore predicted a backlash effect such that participants with more narrow DBDs would be more likely to make attributions to discrimination when this was discrimination by a minority outgroup against the majority ingroup (i.e., reverse discrimination).We predicted that participants with narrow DBDs would also blame the alleged perpetrator of reverse discrimination more, and blame the alleged victim of reversed discrimination less.
As in Study 2, we included modern racism as a covariate.Modern racism is known to be associated with claims of reverse discrimination, and indeed the sense that minorities are making illegitimate claims is part of the raison d'etre of the measure (McConahey, 1986).
We also included socially desirable responding (Paulhus, 1991).As we have outlined, attributions to discrimination are a topic of heated and polarized public debate, and this will be particularly the case during attributions where the alleged perpetrator is themselves a racialized minority.Although Study 1 did not indicate a correlation between socially desirable responding and DBDs, it nevertheless seemed conceivable that socially desirable responding might impact on participants' responses in this particular design.We expected that the predicted results would be observed even in the presence of these covariates.

| Design
A two-cell (prototypical vs. reverse discrimination), betweenparticipants experimental design was administered online.The independent variables were DBDs (administered as a pretest), and the version of the vignette (prototypical discrimination vs. reverse discrimination, recorded in the main study).The dependent variables were participants' attributions to discrimination, and the extent to which they blamed the alleged perpetrators and victims.Participants' scores on modern racism and socially desirable responding were recorded as a covariates.

| Participants
Undergraduate students at the University of Leuven completed the measures online for course credit.The pretest consisted of 214 participants.The main study was completed by 177.Of these participants, 83 generated usable data that we could track across both times (45 in the prototypical condition and 38 in the reverse condition). 10Of these, 73 were female, 9 were male, and 1 declined to say.Participants were aged between 17 and 27 years (M age = 18.6, SD = 1.77).The majority (77) described their heritage as White Belgian or White Dutch.The exclusion of participants who did not identify as White Belgian or Dutch had no systematic effect on the results: these participants were included in the analysis.The sample size was opportunistic (i.e., based on the size of the database and proportion of participants who responded with usable data).

| Materials and procedure
The independent measures were the 15-item DBDs measure (Cronbach's α = .76)and the condition (prototypical vs. reverse discrimination vignette).The vignette was adapted from the discrimination condition of Study 2 (see Appendix C).In the prototypical discrimination condition, Lucas Peeters was described as a manager who had failed to promote any of his Turkish-origin employees, and Mehmet Yildiz was the employee who was suing him for discrimination.In the reverse discrimination condition, Mehmet Yildiz was the manager and Lucas Peeters was the employee.Both versions were constructed to look like clippings from a newspaper article.
The dependent variables were the same as Study 2: a single item measuring attribution to discrimination; blame attributed to the alleged perpetrator (α prototypical = .84;α reverse = .73);and blame attributed to the alleged victim (α prototypical = .74;α reverse = .79). 11  Participants completed the pretest materials (DBDs, modern racism, and socially desirable responding).The main study (vignette and dependent variables) was administered later in the academic year (minimum 1 month).Both were completed online and in participants' own time.

| Analysis
We tested the extent to which narrow DBDs interacted with condition to predict responses to two different kinds of discrimination (i.e., prototypical and reverse), while including modern racism or socially desirable responding as a covariate.We used the same PROCESS macro described in Study 2, i.e., Model 1, The results showed that, as expected, more narrow DBDs were associated with lower attributions to discrimination in the prototypical discrimination condition (consistent with Study 2), but we did not see the backlash effect that we had expected in the reverse discrimination condition.
This pattern of results (expected negative relationship in the prototypical condition, but no relationship in the reverse condition) was repeated when the dependent variable was blame toward the alleged perpetrator (and modern racism was covariate) ß = .75,As expected, participants' DBD scores predicted their responses to an example of prototypical discrimination.This replicated Study 2.
However, DBDs were not related to participants' responses to an example of reverse discrimination: there was no evidence of the backlash effect that we had predicted (but note that the slopes were always in opposite directions across the two conditions).These results suggest that the rules and standards that participants applied when thinking about prototypical instances of discrimination (i.e., majority discrimination against a minority) were not relevant when considering a comparable, but nonprototypical intergroup context (i.e., minority discrimination against a majority).
As in Study 2, we ran a supplementary analysis in which we explored the interaction between the covariates (modern racism and socially desirable responding) and condition to predict the dependent measures, and tested for any three-way interactions between DBDs, the covariates, and condition to predict the dependent measures (see Supporting Information).Modern racism interacted with condition to predict blame towards the alleged perpetrator, but in the reverse condition (rather than the prototypical condition).In this condition, participants who scored higher on racism also tended to blame the alleged perpetrator more.There were no other significant relations involving modern racism or socially desirable responding in predicting the dependent variables.

| Discussion
The results therefore partly supported the hypothesis: DBDs were a good predictor of attributions to discrimination and judgments of the blame for prototypical forms of discrimination (i.e., majority discrimination against a minority target).These relationships remained when including modern racism and socially desirable responding as covariates.This therefore replicates the discrimination condition of Study 2. However, there was no relationship between DBDs and the dependent measures in a different, non-prototypical intergroup context (i.e., minority discrimination against a majority target), and where we had expected to see a backlash.There is some evidence elsewhere that DBDs seem to be deployed strategically in contexts that implicate the ingroup as perpetrators but are not affected by contexts when the ingroup are targets (West et al. 2022): in this interpretation, it may not be the (non)prototypicality of the vignette as much as its relevance to the ingroup that caused an effect in one condition but not the other.Unfortunately, our design confounded prototypicality, group memberships, and minority/majority status, so this remains a tentative hypothesis.
Ultimately, Study 3 demonstrated that the arguments that were associated with decision making in one context (DBDs and prototypical forms of discrimination) did not have the same consequence for a comparable form of decision making (DBDs and nonprototypical forms of discrimination).This is consistent with our argument that the deployment of deinitional boundaries of discrimination does not rest on logical and even-handed deliberation of what is discrimination (as is often claimed in public discourse), but rather something that is shifted and reconstructed as the context changes (see also Biernat et al., 1991).

| STUDY 4: COMPARING THE DBDs OF MEMBERS OF MAJORITY AND MINORITY GROUPS
This final study pools the data from Studies 2 and 3 to ask: to what extent do racialized majorities deploy more narrow racism DBDs compared to racialized minorities?
Data in Studies 1−3 are a de facto analysis of White people's DBDs about racism: we did not actively recruit participants who did not identify as White, and the number of participants who did not identify as White was small in each study.However, research evidence indicates that racialized majorities are generally less likely to make attributions to racism compared to racialized minorities (for a review see Carter & Murphy, 2015).Racialized minorities witness and experience racism to a much greater extent than racialized majorities, and are more likely to be familiar with relevant history and events F I G U R E 3 Interaction between condition and narrow DBDs to predict dependent measures (Study 3).DBDs, measure of definitional boundaries of discrimination.(Nelson et al., 2013;Reinka & Leach, 2018;van Veelen et al., 2020;West, 2019).Racialized majorities, in contrast, may be more motivated to resist making attributions to racism (Adams et al., 2006).
On this basis, we predicted that participants who identified as racialized majorities (i.e., White Belgian or White Dutch origin) would report more narrow racism DBDs compared to participants who identified as racialized minorities (i.e., from any other heritage group).
Majorities would therefore score higher on the DBD measure compared to minorities, indicating more narrow DBDs.

| Design
We conducted a secondary analysis of data from Studies 2 and 3, using a two-cell, between-participants design (majority vs. minority participants), where the dependent variable is DBDs.

| Participants
The DBDs measure was completed by 572 participants in Studies 2 and 3. Of these, 525 identified as White Belgian or White Dutch, and 46 identified in some other way (one declined to say). 12Of these participants, 82 identified as men and 487 as women (three declined to say).Participants were aged between 16 and 27 years (M age = 18.29,SD = 1.24).

| Materials
Participants completed the 15-item DBDs measure as described in Studies 2 and 3 (Cronbach's α = .77).A higher score indicates that participants were deploying more narrow constructions of discrimination.
At the end of each study, participants were asked to indicate their heritage in a series of closed items ("Belgian," "Dutch," "Turkish," etc.).Where participants felt that none of these categories fitted, there was a free text field in which participants could use their own words (e.g., "Lebanese" "half Belgian, half Nigerian").Data were recoded into two categories: participants who selected "Belgian" or "Dutch" with no other qualification (self-identified majority participants), and participants who selected any other category and/or added a free text qualification (self-identified minority participants).

| Analysis and discussion
We tested the hypothesis using one-way ANOVA where the independent variable was majority vs. minority status and the dependent variable was DBDs.The results supported the hypothesis: there was a significant difference in DBDs between majority and minority participants F(1, 569) = 9.14, p = .003,η 2 = 0.016.Participants who identified as majorities reported significantly more narrow DBDs (M = 3.22, SD = 0.65) compared to participants who did not (M = 2.92, SD = 0.66).
Majority participants were thus deploying more narrow DBDs (compared to minority participants).Theoretically, this would mean that racialized majorities would be less likely to make attributions to understand the intentions of any alleged perpetrators; that claims of discrimination are often illegitimate; and that the problem of discrimination is small and/or mostly limited to the past).We predicted and found that the more participants deployed narrow DBDs, the less likely they were to make attributions to discrimination and the less they blamed the alleged perpetrator (Study 2); that participants' DBDs predicted attributions to discrimination in some contexts but not others (Study 3); and that participants who identified as racialized majorities had significantly more narrow racism DBDs compared to participants who identified as racialized minorities (Study 4).These effects remained even when covariates included modern racism, ingroup identification, and socially desirable responding.Contemporary instantiations of definitional boundaries of discrimination therefore work to narrow the range of actions (illustrated in Figure 1) that can be considered to be "discrimination." Data indicated that both DBDs and modern racism can predict attributions to discrimination and associated dependent variables (Study 2, Study 3, and Supporting Information; but see also West et al., 2021 for an example when DBDs was a stronger predictor than modern racism).We suggest that there are two advantages to a DBD approach (which measures social constructions of discrimination) over a modern racist approach (which measures attitudes).
First, DBDs constitute a tool to capture a real and current phenomenon, that is, "culture wars" and public discourse (Hunter, 1991).and ideological (e.g., Dost et al., 2019), but they appear to be an intergroup phenomenon rather than a political one per se (Carter & Murphy, 2015;Durrheim et al., 2018;Ellemers & van den Bos, 2012;Miron et al., 2011;Platow et al., in press;Wang et al., 2022;White et al., 2021).West et al. (2022), for example, has demonstrated how participants can shift their DBDs from broad to narrow when it serves their group interests to do so.Anecdotally, we see the same strategic deployments of DBDs among both the political left and the political right: in the United Kingdom, for example, the political left can be narrow on anti-Semitism but broad on transphobia, while the right can be narrow on Islamophobia but broad on "reverse racism".A DBD approach enables us to capture the ways that these public discourses are deployed, but also how they shift and change in different intergroup contexts.In short, DBDs operationalize a cultural phenomenon in a form that is usable to social cognitive research.
The second advantage of a DBD approach (compared to modern racism) relates to malleability.We do not consider DBDs to be either an individual difference measure, nor a consequence of cognitive processes per se, but rather an ad hoc assemblage of culturally available arguments that are deployed in specific local contexts.
People appear to be flexible and strategic in that specific DBDs arguments may be deployed in some contexts but not others (Study 3), and/or they may deploy broader or narrower versions of DBDs in different contexts and in ways that are consistent with ingroup interests (e.g., West et al. 2022).Importantly, these constructions can also change over time and culture: arguments that work in one historical or cultural moment can therefore become cliches and red flags in another (e.g., "I'm not racist but"; "some of my best friends"; Billig, 1988;van Dijk, 1992); and new arguments will emerge (e.g., Danbold et al., 2022;O'Brien, Crandall, et al., 2010;White & Crandall, 2017;White et al., 2021).Malleability (whether at an available arguments in current public discourse, but public discourse is not fixed.DBDs therefore do not measure a logical set of structures with a clear boundary, but rather represent a social reality that is continuously being re-interpreted, re-thought, and represented (Moscovici, 1984(Moscovici, /2000)), that is, an interdependent assemblage of arguments rather than a logical and consistent set of beliefs.To this extent, we are less interested in the content of DBD arguments than the fact that they exist in the first place (see also White et al., 2021).This is also why we treat DBDs on a single dimension (narrow− broad), rather than as distinct factors: we expect that specific arguments will ebb and flow over time.Indeed, we expect that the measure will need to be revised repeatedly or else become redundant as DBD arguments change over time.
A  (Biernat & Manis, 1994;Fernandez et al., 2014;Miron et al., 2011;Sullivan et al., 2012).We might also expect that definitional boundaries tighten or loosen under conditions of group threat (Knowles et al., 2014;Lowery et al., 2007Lowery et al., , 2012)), for example, to resist attributions of discrimination that implicate the ingroup (West et al. 2022;Young & Sullivan, 2016).DBDs may be relevant in understanding the principle-implementation gap (i.e., when people agree that discrimination is a problem in principle, but resist making attributions to discrimination in concrete situations).If we conceive definitional boundaries as social norms then they may be implicated in intergroup contact (direct and indirect; Christ et al., 2014;Greenland, 2021;Pettigrew et al., 2007;Ramiah et al., 2015).DBDs might even be considered as performative, that is, a way in which to communicate and regulate inter-and intragroup relations (Durrheim et al., 2016;Ellemers & van den Bos, 2012;Schrock et al., 2022).
items with a view to (a) representing the themes outlined above, (b) reducing repetition, and (c) keeping a reasonable balance of positively and negatively coded items.This yielded a final set of 15 items (Cronbach's α= .83;see Appendix A).There were three packets containing established questionnaires.Putting all of the measures of interest into a single questionnaire would have resulted in an excessively long questionnaire for a participant to complete in a single sitting.Asking a participant to complete the measures over a number of sittings would have resulted in excessive attrition.Accordingly, we divided the measures into three (outlined below) and asked each participant to complete one packet only.The first packet contained the DBD measure and measures of modern racism (Cronbach's α = .86;McConahey, 1986); Brigham's racial attitudes scale

3. 1
| Method 3.1.1| Design This was a two-cell (discrimination vs. embezzlement), betweenparticipants experiment administered via online survey.The independent variables were participants' narrow DBDs score and T A B L E 1 Correlations between narrow DBDs and established measures (Study 1) experimental condition (i.e., whether they read a vignette relating to discrimination or embezzlement).The dependent variables were participants' attributions, and the extent to which they blamed the alleged perpetrators and victims.Participants' scores on modern racism and ingroup identification were included as covariates.3.1.2| ParticipantsUndergraduate students at the University of Leuven (n = 373) completed the study online for course credit (186 in each condition).Three hundred and sixteen were female, 53 were male, and 4 declined to say.Participants were aged between 16 and 23 years (M age = 18.2, SD = 0.97). 6Most (332) described their heritage as White Belgian or White Dutch.The sample size was opportunistic (i.e., based on the size of the database and proportion of participants who responded with usable data).7The exclusion of participants who did not identify as White Belgian or Dutch had no systematic effect on the results and hence these participants were included in the analysis.3.1.3| Materials and procedureParticipants completed the online study in their own time.All items were scored on 7-point Likert scales anchored at "strongly agree" and "strongly disagree." 8Participants first completed the 15-item DBDs measure (Cronbach's α = .77).A high score indicated that participants were deploying more narrow constructions of discrimination.The experimental condition was administered in the form of short vignette (around 120 words), mocked up to look like a newspaper article (see Appendix B).Participants read a vignette that was either designed to be interpretable as an act of discrimination (discrimination condition), or an act of embezzlement (embezzlement condition).In the discrimination condition, Lucas Peeters (the manager of a Belgian children's charity) had failed to promote any of his Turkish-origin employees, despite them being qualified.One of these employees (Mehmet Yildiz) was suing Mr Peeters and the charity for discrimination.In the embezzlement condition, Lucas Peeters had been responsible for overseeing the charity expenses account, but had turned a blind eye to its systematic misuse.A family whose funding from the Charity had been cut were suing Mr Peeters and the charity for embezzlement and misuse of company funds.Both vignettes were therefore designed to generate a level of moral disgust(Abitan & Krauth-Gruber, 2015) and to have an alleged perpetrator and associated complainant.The extent to which the alleged perpetrator was responsible (either through action or inaction) was deliberately ambiguous, as was the extent to which the complainant could legitimately claim material harm.The first dependent variable was a single attribution measure ("In your opinion, to what extent could Lucas Peeters' actions be described as discrimination?")."Discrimination" was replaced with "embezzlement" in the embezzlement condition.alleged perpetrator was measured with eight items (three reversed), for example, how much is Lucas Peeters responsible for what happened?; how much are Lucas Peeters' actions excusable?(Cronbach's α discrimination = .86;α embezzlement = .73).Higher scores indicate ascribing more blame to the alleged perpetrator.Blame for the alleged victim was measured with a parallel set of eight items (three reversed), for example, how much is (Mehmet Yildiz/the family who are suing) responsible for what happened?; how much are (Mehmet Yildiz/the family who are suing)'s actions excusable?(α discrimination = .74;α embezzlement = .77).Higher scores indicate ascribing more blame to the alleged victim.
Debates around what constitutes discrimination (and what does not) are increasingly heated, pervasive, and consequential for social action because they construct competing versions of what is permissible (and what must be suppressed).DBDs are clearly political individual, group, or cultural level) indicates a possible location for interventions: understanding contemporary DBD arguments enables us to consider the arguments that might persuade people to shift their DBDs (where those DBDs are demonstrably polarized, inappropriate, or unhelpful).For example, many DBD arguments are potentially falsifiable (e.g., that you cannot discriminate unless you intend to do so).One could conceive an intervention that names and then challenges these arguments.Definitional boundaries of discrimination therefore constitute an alternative approach that is less individualized, more context-based, more cultural, and more available for interventions.DBDs record the extent to which participants are deploying specific, culturally limitation of the current results is that the DBDs are consistently measured, but not manipulated.Although we have demonstrated a relationship between DBDs and attributions, and differential effects of DBDs depending on the context, we have not manipulated DBDs to demonstrate conclusively that DBDs cause attributions to discrimination.It is possible that a third variable (e.g., implicit associations or hierarchy justifying beliefs) causes both.In treating DBDs as a single dimension, we have also negated the possibility that the observed effects might be disproportionately driven by one factor only (e.g., intention arguments).Finally, we have suggested that DBDs are culturally specific, but we have not provided evidence to support this, nor that DBDs would predict attributions to other -isms (e.g., sexism or heterosexism).DBDs raise several additional empirical questions and links to extant research that have yet to be explored.The extent to which DBDs may be applied or strategically according to/or strategically according to context (Study 3; West et al. 2022) speaks to research on shifting standards and competitive victimhood