Whose Words Hurt? Contextual Determinants of Offensive Speech

Abstract

Tracing the boundaries of freedom of expression is a matter of wide societal and academic import—especially, as these boundaries encroach on the politics of inclusion. Yet, the elements that constitute offensive speech and determine its legal status remain poorly defined. In two studies, we examined how lay judges evaluate the offensiveness of apparently descriptive statements. Replicating prior work, we found that non-linguistic features (including speaker intent and outcomes on the audience) modulated the statements’ meaning. The speaker’s identity—and, in particular, their membership in the target group—independently influenced evaluations of offensive speech among conservatives and progressives alike. When asked to disclose their abstract principles, or jointly evaluate two contrastive cases, participants tended to deny the relevance of identity while primarily endorsing the intent principle. Taken together, our findings confirm that assessments of offensive speech are governed by contextual features, some of which are not introspectively deemed relevant.

Keywords

freedom of expression ingroup/outgroup social status verbal offense

Mathew Goodwin (MG):

More than 3 million EU nationals have applied to stay in the UK. 90% now have the right to Remain. Just 6 applications rejected (due to criminality). Intolerant Britain? Hmm.

Jon Henley (JH):

Misleading and actually offensive.

MG:

How and how?

JH:

It’s misleading because a large number have only been granted temporary status and have no guarantee they can remain. And if I really need to explain why it’s at best insensitive, and at worst offensive, to several million people, then I frankly give up.

MG:

For 3+ years we’ve been told doom-laden stories about deportations & intolerance. Vast majority of apps accepted, basically 60% permanent right to remain. 5 year residency requirement not shocking (e.g. Canada 3 yrs) & most will transfer to perm residency when criteria met.

JH:

I’m not going to argue with you Matt. A lot of people found your tweet misleading and offensive. That’s on you; you could have been more complete and a hell of a lot more sensitive. You chose not to be.

MG:

I’m tired of the endless catastrophising. It’s making things worse, not better.

JH:

Fine, I have no argument with that. But given this issue’s sensitivity, maybe try to phrase your tweets in a slightly less (deliberately?) provocative way. Just out of consideration for the very, very many people who do not find the UK a particularly tolerant country right now.

MG:

My tweet is a “good news” story, and I hope that our media would paint a balanced picture. Sure I push back against some narratives because some are very misleading. I suspect that we see things very differently, which is obviously fine.

JH:

Again, that’s fine. Push back all you like. I’m just trying gently to point out that that for quite a large number of people, the way in which you “pushed back” in your original tweet was misleading and offensive. Up to you whether you consider that worth knowing (Goodwin, 2020; Henley, 2020).

The above Twitter exchange between political scientist Matthew Goodwin and journalist Jon Henley illustrates a recurring conflict at the interface between freedom of expression and minority advocacy. At first glance, examining the literal content of Goodwin’s opening remarks suggests little in the way of offensive material. The tweet strings together a series of statistics regarding the EU Settlement Scheme. How could these numbers possibly be offensive?

To understand how, we must look beyond the plain words. A host of accompanying features could help establish the tweet’s meaning in context: the speaker’s position of privilege, coupled with his minimization of xenophobia and of the distress of immigrant groups, and the wide audience that Twitter provides. Each of these elements, above and beyond the words themselves, can help discern seemingly innocuous comments from hurtful provocations (Saul, 2018; Stanley, 2015). Relatedly, numerous empirical studies of language attest to the transformative effects of such contextual influences on meaning (Fasoli et al., 2015; Galinsky et al., 2013; Gibson et al., 2019; Jay & Janschewitz, 2008; McGowan, 2018; O’Dea et al., 2015).

Do Goodwin’s data merely report on the EU Settlement Scheme, or do they also constitute an affront—a denial of racial and cultural tension in the United Kingdom, and of those who denounce this climate? As the above anecdote reveals, these remarks can elicit quite distinct reactions, appearing innocuous and primarily informative to some, yet offensive and intolerable to others—and even linguists and legal scholars have struggled to define this boundary. Furthermore, profound uncertainty surrounding the elements that constitute offensive speech could in part aggravate conflicts at the interface of freedom of expression and social inclusion—a point we return to in the General Discussion. In the present work, we conducted a pair of exploratory and confirmatory studies investigating how lay judges distinguish purely descriptive uses of language from verbal affronts—specifically, by utilizing information beyond the plain meaning of words.

A Widespread Tension

In recent years, academic and public debates surrounding offensive speech have unearthed a deep tension between two fundamental pillars of democracy: freedom of expression and the protection of minority groups. A cursory glance at large-scale public opinion polls helps document this conflict.

On one hand, a Pew Research poll shows that most US Americans agree that people should be able to “say what they want” in public, whether these statements offend minority groups or others’ religion and beliefs (Wike & Simmons, 2015; see also CATO Institute, 2017). Meanwhile, other survey data paint a more cautious portrait: A majority of US Americans support government restrictions on free speech—for instance, on racist speech or on expressions of neo-Nazi ideology (Lystad, 2019).

Despite various cultural dissimilarities, including a distinct historical backdrop of authoritarian restrictions on freedom of expression, the same fundamental tension arises in Spain (where the studies we report were conducted): People adamantly defend free speech and refuse censorship (in very similar numbers¹; Wike & Simmons, 2015) while simultaneously supporting legal restrictions on sexist, racist, and anti-gay rhetoric. In sum, we value the right to speak and opine freely, while demanding censorship and punishment of opinions we find despicable.

In principle, offensiveness need not entail legal punishment. A listener could personally find certain language offensive and yet protect others’ right to use it. However, existing evidence to this effect is mixed: Even Supreme Court justices deciding freedom of expression cases tend to selectively suppress speech that is uncongenial to their personal ideology (Epstein et al., 2018). This ideological bias appears to be attenuated, however, among individuals higher in cognitive ability (De Keersmaecker et al., 2020). Our studies pursue this question, by asking whether convergent results arise across measures of a statement’s offensiveness and of its acceptability (/censurability).

The Intent-Plus-Outcome Framework

An extensive literature in moral psychology has investigated the factors that shape ascriptions of blame and punishment in the context of harm to third parties (Cushman, 2008; Gino et al., 2008; Greene et al., 2009; Van de Vondervoort & Hamlin, 2018; Young et al., 2007). This body of research conceptualizes harm as physical aggression and demonstrates that people spontaneously consider both an agent’s intent and the outcomes that befall victims when determining the former’s culpability. This cognitive template mirrors a fundamental tenet of criminal law (Mikhail, 2007): namely, that together harmful intent and negative outcome form the basis for legal responsibility. As a result, by comparison to intentional crime, both failed attempts to harm (e.g., attempting to poison a rival’s drink with sweetener believed to be poison), and accidental harms (e.g., attempting to sweeten a friend’s drink with poison believed to be sugar) elicit reduced condemnation (Young et al., 2007).

Furthermore, the perception of harmful intent not only shapes observers’ evaluations, but even modulates the victim’s own sensation of pain: For instance, in a laboratory task, electric shocks that were administered intentionally produced more intense pain than accidental shocks of the same voltage and duration (Gray & Wegner, 2008). Our present study inherits this two-part template emerging from research in moral cognition, and applies it to the study of verbally inflicted harm. Does the intent-plus-outcome framework explain our evaluations of verbal offense as well?

At least one study has already adopted the intent-plus-outcome perspective on verbal discrimination, discovering that negative intent and negative outcomes promoted the perception of prejudice and shaped judgments of a speaker’s conduct (Swim et al., 2003). Relatedly, a wider literature on the phenomenon of reappropriation illustrates how derogatory language can become innocuous when employed with a benign intent (Fasoli et al., 2015; Galinsky et al., 2013; Gibson et al., 2019; Leets, 2002; O’Dea et al., 2015). The studies reported below adopt a complementary focus: We draw attention to seemingly neutral and descriptive language, like Goodwin’s tweet, and experimentally probe the conditions under which they might constitute a verbal offense.

The Role of Speaker Identity

Empirical research on hate speech (Langton, 2012, 2018; McGowan, 2012) and, in particular, on the use of racial slurs (Croom, 2011; Jeshion, 2013), has uncovered dramatic effects that cannot easily be attributed to variation in speaker intent or outcome.

Broadly, this cluster of results point toward a series of identity-based norms that focus on the target, the speaker, or even their interrelationship (O’Dea et al., 2015; see also Fasoli et al., 2015). For instance, people perceive criticism aimed at high status targets (Goliath groups) as more permissible than identical criticism aimed at low status targets (David groups)—a pattern that emerges in both Western and Chinese cultures (Jeffries et al., 2012). Other identity norms revolve around the speaker, including their hierarchy and ingroup/outgroup status. For instance, criticism of a particular group is more welcome when put forth by an ingroup speaker than an outgroup speaker (see intergroup sensitivity effect; Hornsey & Imani, 2004; Miller & Effron, 2010).

These prescriptive norms governing the permissibility of speech may arise partly from more fundamental effects of speaker identity on linguistic meaning. For instance, in the study of indirect requests (Holtgraves, 1994), it is known that non-conventional requests (e.g., to shut a door by saying “It’s noisy in here,” or “Isn’t it noisy in here?”) are more likely to be interpreted as requests when issued by higher, than by lower, status speakers—echoing what philosophers have referred to as speech capacity (Ayala, 2016, 2018; Kukla, 2014). Also, experimental manipulations of speaker status affect listeners’ recall of their remarks (Holtgraves et al., 1989): For high status speakers, participants had greater difficulty distinguishing between transcripts from the conversation and assertive paraphrases of those remarks. By analogy, in our present case, an ambiguous remark may be encoded as an offense in one context but as a description in another, by consulting the speaker’s identity.

Motivated by the above evidence, our present research examines whether speaker identity norms shape judgments of offensiveness and acceptability. Some studies of offensive speech have focused on individual differences in participants’ identity as audience—while holding fixed the speaker’s identity (Leets & Giles, 2006; Swim et al., 2003). Other work has conceptualized a speaker’s identity as a proxy for their motives, and these motives, in turn, as a primary determinant of the expression’s offensiveness (Gibson et al., 2019). Our review of existing evidence, however, motivates a distinct prediction and experimental approach: i.e., that a speaker’s identity—independently of their intent and outcome—influences the offensiveness and acceptability of their remarks. To this end, the studies below treat outcome, identity, and intent as orthogonal factors.

Abstract Standards Versus Concrete Determinants

What standards do people believe should regulate offensive speech? A second purpose of our research was to contrast people’s case-based judgments with the general standards they explicitly endorse. Abundant research across the psychological sciences has revealed that people’s abstract avowals are incongruent with their concrete judgments and behavior (Nisbett & Wilson, 1977; Wilson, 2002): For instance, research on punitive policies (Carlsmith et al., 2002), legal adjudication (Struchiner et al., 2020), and racist attitudes (Dovidio et al., 2002) shows that we endorse standards in the abstract which our concrete behavior fails to reflect. Relatedly, inductively recognizing general patterns in our behavior is particularly error-prone—which may be attributable to overconfidence (Alter et al., 2010; Fernbach et al., 2013; Rozenblit & Keil, 2002) and desirable self-presentation (Paulhus & Reid, 1991).

Research on prejudice has amply documented this form of intrapersonal conflict: Symbolic racism (see also modern racism and racial resentment; e.g., Tarman & Sears, 2005) is grounded in the idea that people engage in racist conduct and endorse noxious beliefs about race while explicitly eschewing racism as a label or a creed. Period change in the salience and scope of cultural norms governing prejudice can, however, bring implicit and explicit attitudes in line (Crandall et al., 2013).

To understand whether implicit-explicit incongruence arises in the context of offensive speech, we also recorded participants’ abstract views about offensive speech. These measures allow us to examine whether the impact of each contextual determinant is statistically moderated by its professed relevance in the abstract. Simply put, is there a disconnect between (a) the criteria that people explicitly support and (b) their judgment patterns while evaluating the offensiveness of particular statements in concrete cases? Evidence of moderation in the aforementioned analyses could suggest otherwise—that is, that people’s abstract standards govern their concrete assessments, thus helping to dispel this looming concern.

Objectives

We approached the studies in this article with three general objectives. Inspired by prior evidence that moral evaluations integrate intent and outcome information, we ask whether this conceptual template explains evaluations of verbal offense. First, we seek to conceptually replicate the finding that (H₁) speaker intent and outcome affect the offensiveness and acceptability of people’s speech (see Swim et al., 2003). Second, we evaluate whether (H₂) speaker identity independently affects offensiveness and acceptability, after controlling for the impact of intent and outcome. Finally, we ask whether (H₃) lay judges’ abstract beliefs about offensive speech are congruent or incongruent with patterns in their concrete, case-based evaluations.

Studies 1 and 2 employ a set of eight hypothetical scenarios, each involving a statement about a subordinate group (Scenario 1: economic differences across regions in Spain; Scenario 2: islamophobia in school setting; S3: gender disparity in the film industry; S4: STEM & gender equality; S5: cultural stereotypes within Europe; S6: national stereotypes in a global context; S7: research on Muslim migration; S8: national stereotypes in business). Materials, data, and analysis scripts are available on the Open Science Framework at: https://osf.io/jmn2c/.

Experiment 1

Method

Materials

For each of the eight scenarios, we wrote a total of eight variants—one for each factorial combination in the 2 (Identity: subordinate ingroup, dominant outgroup) by 2 (Intent: negative, neutral) by 2 (Outcome: negative, neutral) matrix.

By way of example, in one of the scenarios, the speaker refers to the region of Ceuta in the context of a debate about Catalonian independence. Ceuta is a small and economically disadvantaged Spanish city, located on the African continent, bordering with Morocco. In the dominant outgroup condition, the speaker is from Catalonia, a socially and economically developed region in Spain. Meanwhile, in the subordinate ingroup condition, the speaker is a fellow Ceuta native:

[SPEAKER IDENTITY]: Enric (/Ahmed) is a Catalan (/Ceutí) politician who is being interviewed about Catalonia’s independence.

Next, we introduce information about the speaker’s intent, followed by the statement (which, critically, was held constant across conditions):

[INTENT]: In the past, Enric has refused numerous invitations to participate in activities organized by (/has regularly collaborated with) Ceutí groups. During the interview, Enric says the following:

“For Spain, losing Catalonia is not the same as losing Ceuta.”

Finally, the vignette concluded with information about the outcome to the audience:

[OUTCOME]: Some of the Ceutí people who are following the interview start to complain through social networks after hearing his statement (/Enric’s claim has no repercussions among the Ceutí people who are following the interview).

Procedure

In a randomized block design, participants were assigned to one of eight groups and viewed a battery of eight trials in a random order. In each group, participants viewed every factorial combination in the 2 (Speaker identity: ingroup, outgroup) × 2 (Intent: offensive, neutral) × 2 (Outcome: offense, no offense) matrix paired with a different scenario on each trial. Thus, no participant viewed the same scenario or factorial combination twice. Collapsing across groups, we achieved balance in the scenario-by-condition matrix (n per cell: M = 40, Min = 36, Max = 44).

After each trial, participants evaluated the offensiveness and the acceptability of the statement. Offensiveness was assessed through a pair of items ([o₁] the speaker “is offensive,” and [o₂, reverse-scored] the speaker “simply provides information about. . .” the situation) rated on continuous scales from 1 “Strongly disagree” to 7 “Strongly agree.” The composite score, (o₁ − o₂)/2, exhibited good reliability (Cronbach’s α = .73). Acceptability was assessed through a single dichotomous measure, indicating whether the statement was acceptable (1: e.g., “It was OK for . . . to say this”) or not (0: e.g., “. . . should not have said this”).

After the block of eight trials, we included a post-test questionnaire about the relevance of each factor. Participants made relevance judgments for each other using continuous scales from −3: “No importance/relevance” to 3: “Absolute importance/relevance”:

1. Identity relevance (“whether or not the speaker belongs to the group he or she talks about”),

2. Intent relevance (“whether or not the speaker holds a pejorative attitude toward the group he or she talks about”).

3. Outcome relevance (“whether or not people in the group the speaker talks about are offended”).

Finally, participants provided the following demographic information: age (in years), gender, church attendance (from “Several times a week” to “Never”), political orientation (from “Far left” to “Far right”), monthly income (from “No income” to “Over €4500 per month”), and educational attainment (from “Less than high school” to “Master’s or PhD Degree”).

Participants

We partnered with Netquest (www.netquest.com), a survey research firm, to recruit a sample of 320 Spanish natives, stratified by gender (160 men, 160 women). Mean sample age was 40.3 years old and ranged from 19 to 64. The sample was largely non-religious, with 62% reporting no church attendance, and politically somewhat left of center (M = 3.30, SD = 1.41, on a 1: “Left” to 7: “Right” scale). This ideological skew is consistent with recent representative surveys of the adult Spanish population (Centro de Investigaciones Sociológicas, 2019; Fundación BBVA, 2019).

Results

Summary statistics for each experimental condition are reported in Table 1.

Table 1.

Offensiveness and Acceptability by Condition: Mean and 95% Confidence Interval.

	Intent	Outcome	Offensiveness		Acceptability
Condition	Intent	Outcome	Ingroup	Outgroup	Ingroup	Outgroup
Description	Neutral	Neutral	–0.84 [–1.27, –0.42]	–0.57 [–1.00, –0.15]	0.78 [0.65, 0.87]	0.70 [0.56, 0.81]
Accident	Neutral	Negative	–0.65 [–1.08, –0.23]	–0.47 [–0.90, –0.05]	0.69 [0.55, 0.81]	0.61 [0.45, 0.74]
Attempt	Negative	Neutral	–0.67 [–1.10, –0.24]	–0.22 [–0.65, 0.20]	0.71 [0.57, 0.82]	0.57 [0.41, 0.71]
Offense	Negative	Negative	–0.41 [–0.84, 0.02]	–0.06 [–0.49, 0.37]	0.64 [0.50, 0.77]	0.54 [0.39, 0.68]

Note. Offensiveness ranges from −3: completely informative to 3: completely offensive. Acceptability judgments range from 0: unacceptable to 1: acceptable.

Experimental effects

Offensiveness

We analyze our data using a multilevel model with random effects of participant and scenario, applying the Kenward-Roger approximation to calculate degrees of freedom (Bates et al., 2015; Luke, 2017). In the fixed effects portion of the model, we enter Identity, Intent, and Outcome, along with every two- and three-way interaction. Furthermore, the model employs an effect coding system, such that factors with two levels are coded as {−0.5, 0.5}. As a result, in models involving interactions, the lower level terms represent main effects and not simple effects.

Speaker identity, F(1, 2226) = 24.76, p < .001, intent, F(1, 2226) = 22.18, p < .001, and outcome, F(1, 2226) = 8.19, p = .004, each affected offensive meaning. No higher order interactions were observed, all Fs < 1.90, all ps > .16. Model comparisons confirmed that, by comparison to the additive model (Model 1; AIC = 9977), entering the two-way interactions (Model 2; AIC = 9980) did not improve model fit, χ²(3) = 2.75, p = .43, and neither did the full factorial model (with the three-way interaction term, Model 3; AIC = 9982), χ²(4) = 2.75, p = .60. See Table 2 for further details regarding each of the three models.

Table 2.

Multilevel Regression Models: Offensiveness and Acceptability (Study 1).

	Fixed effects	Offensiveness			Acceptability
Models	Fixed effects	B[95% CI]	t	p	OR[95% CI]	t	p
	Intercept	–0.49[–0.84, –0.14]	–2.71	.023	1.92[1.09, 3.37]	2.26	.024
Model 1: main effects only	Identity	0.31[0.19, 0.44]	4.98	<.001	0.63[0.52, 0.75]	–5.09	<.001
	Intent	0.30[0.17, 0.42]	4.71	<.001	0.70[0.59, 0.84]	–3.86	<.001
	Outcome	0.18[0.06, 0.30]	2.86	.004	0.73[0.61, 0.88]	–3.40	.001
Model 2: +two-way interactions	Identity × Intent	0.17[–0.07, 0.42]	1.38	.168	0.86[0.60, 1.24]	–0.81	.42
	Identity × Outcome	–0.10[–0.34, 0.15]	–0.60	.46	1.11[0.77, 1.59]	0.56	.58
	Intent × Outcome	0.07[–0.18, 0.31]	0.55	.59	1.23[0.86, 1.77]	1.14	.26
Model 3: full factorial	Identity × Intent × Outcome	–0.01[–0.50, 0.49]	–0.02	.99	1.18[0.57, 2.41]	0.44	.66

Note. These models employ an effect coding system, such that factors with two levels are coded as {–0.5, 0.5}. As a result, in models involving interactions, the lower-level terms represent main effects and not simple effects. OR = odds ratio; CI = confidence interval.

Marginal effects revealed that (a) dominant outgroup speakers (M = −0.33, 95% CI = [−0.74, 0.08]) were more offensive than subordinate ingroup speakers (M = −0.65, 95% CI = [−1.05, −0.24]), t = 4.98, p < .001; (b) speakers with negative intent (M = −0.34, 95% CI = [−0.75, 0.07]) were more offensive than speakers with neutral intent (M = −0.64, 95% CI = [−1.05, −0.23]), t = 4.71, p < .001; and (c) harm to the audience rendered speech more offensive (M = −0.58, 95% CI = [−0.99, −0.17] vs. M = −0.40, 95% CI = [−0.81, 0.01]), t = 2.86, p = .004.

Acceptability

We again analyze our data using a multilevel model with the same random- and fixed-effects structure as used in the analyses of offensiveness. Given the dichotomous nature of the acceptability dependent variable, however, we use a logistic multilevel model. Marginal effects describe the average effect of one independent variable on the predicted probability of acceptability (i.e., that a respondent reports the statement was OK; denoted as ŷ), while holding the remaining independent variables constant.

In this model, speaker identity, χ²(1) = 26.08, intent, χ²(1) = 15.02, and outcome, χ²(1) = 11.26, each affected judgments of acceptability, all ps < .001. Once again, no interaction effects were observed, all χ²s < 1.34, all ps > .24. Model comparisons confirmed that, by comparison to the additive model (Model 1; AIC = 3033), entering the two-way interactions (Model 2; AIC = 3037) did not improve model fit, χ²(3) = 2.15, p = .54, and neither did the full factorial model (with the three-way interaction term, Model 3; AIC = 3039), χ²(4) = 2.33, p = .68.

As displayed in Figure 1, the analysis of marginal effects revealed that statements were more acceptable: (a) when made by ingroup speakers (ŷ = .71, 95% CI = [0.58, 0.81]) rather than dominant outgroup speakers (ŷ = .60, 95% CI = [0.46, 0.73]), z = 5.11; (b) when made with neutral intent (ŷ = .70, 95% CI = [0.56, 0.80]) rather than negative intent (ŷ = .62, 95% CI = [0.48, 0.74]), z = 3.88; and (c) when not causing offense to an audience (ŷ = .69, 95% CI = [0.56, 0.80] vs. ŷ = .62, 95% CI = [0.48, 0.74]), z = 3.37, all ps < .001.

Figure 1.

Main effects of speaker identity, intent, and outcome: Studies 1 and 2.

Abstract principles

Participants reported that intent is the most relevant criterion (M = 1.54, SD = 1.52), followed by harm (M = 0.11, SD = 1.83) and speaker identity (M = −0.07, SD = 1.98). Intent was rated as significantly more relevant than either speaker identity, t = 13.28, Cohen’s d = 0.74, 95% CI = [0.62, 0.87], or outcome, t = 12.64, Cohen’s d = 0.71, 95% CI = [0.58, 0.83], both ps < .001. The latter criteria did not themselves differ in perceived relevance, t = 1.53, p = .13, Cohen’s d = 0.09, 95% CI = [−0.02, 0.20]. Figure 2 displays endorsement of abstract principles separately for progressives and conservatives.

Figure 2.

Endorsement of speaker identity, intent, and outcome principles: Studies 1 and 2.

Moderation analyses

We were also interested in understanding whether participants’ relevance judgments moderated the effect of the corresponding experimental manipulations. To this end, we expanded upon the additive model (Model 1), by entering the three measures of participants’ abstract principles, and allowing each rating to interact with the corresponding factor.

This analysis provided evidence that the effect of intent on offensiveness was moderated by subjective relevance, F(1, 2,228) = 4.70, p = .030. Specifically, participants who considered the intent criterion relevant were more likely to perceive speakers with negative intent (B = 0.04, 95% CI = [−0.04, 0.12]), but not neutral intent (B = −0.05, 95% CI = [−0.13, 0.03]), as offensive, t = 2.17, p = .030—though neither simple effect was statistically significant. In contrast, the effects of speaker identity and outcome were not moderated by beliefs about their explicit relevance; identity: F(1, 2,228) = 0.10, p = .75, outcome: F(1, 2,227) = 0.09, p = .77.

Respectively, we conducted moderation analyses for the acceptability measure—by including participants’ ratings of abstract principles, as well as their interactions with the corresponding factor, in our multilevel logistic model of acceptability. No interaction effects arose in this model, all χ²s(1) < 2.38, all ps > .12.

Ideological differences

Previous research has unveiled mixed evidence of political differences in attitudes toward freedom of expression (Wike & Simons, 2015; but see Epstein et al., 2018). Did conservatives and progressives differ in their emphasis on the contextual determinants of offensive speech?

Experimental effects

Significant interactions between political orientation and the outcome factor emerged in both multilevel models: that is, offensiveness, F(1, 2,227) = 7.21, p = .007, and acceptability, χ²(1) = 4.78, p = .029.

To interpret these results, we calculated Johnson–Neyman intervals (Johnson & Fay, 1950; Long, 2019) for both two-way interactions. First, the effect of outcome on offensiveness was significant for values of political orientation below 3.73 (and above 9.33, which falls outside the scale range). Second, the effect on acceptability was significant for values below 4.04 (and above 24.54—again outside the scale range). In other words, harmful outcomes rendered speech more offensive and censurable for progressives, but not for conservatives. No such differences emerged for either the speaker identity or intent manipulation, all ps > .25.

Abstract principles

We observed no ideological differences in the abstract relevance of speaker identity (r = .09, 95% CI = [−0.02, 0.20], p = .11) or outcome (r = .01, 95% CI = [−0.10, 0.12], p = .79). However, progressives treated the speaker’s intent as more relevant than did conservatives (r = −.15, 95% CI = [−0.25, −0.04], p = .008).

Discussion

Study 1 provided exploratory evidence in support of all three hypotheses. In line with H₁, we conceptually replicated Swim and colleagues’ (2003) study and found that information about a speaker’s intent and ensuing outcomes impact people’s ratings of a statement’s offensiveness and acceptability. Second, regarding H₂, speaker identity played an important role in both acceptability (greater among subordinate ingroup speakers) and offensiveness (greater among dominant outgroup speakers). Third, participants’ abstract principles appeared to diverge from their concrete, case-based judgments—providing tentative support for H₃.

In the abstract, participants deemed a speaker’s identity irrelevant, despite the large effect that speaker identity had on their own case-based evaluations. Relatedly, individual ratings of the importance of each standard failed to predict the magnitude of the corresponding experimental effects. Thus, Study 1 documented a degree of incongruence between the standards that people report as constitutive of offensive speech, and the factors that shape their evaluations in the context of specific statements.

Our initial evidence of speaker identity effects was based on a coarse instrumentalization. For each scenario pair, we compared speakers who differed in both their membership in the target group and their placement in the socioeconomic hierarchy. As a result, Study 1 leaves open whether differences in the speakers’ meaning were driven by their social status, or their membership in the target group. Dissociating the contribution of these two characteristics was among the main objectives of Study 2a.

Furthermore, evidence of incongruence between participants’ judgment pattern and their abstract criteria is compromised by two limitations of Study 1: the unknown reliability of our single-item measures, and the fixed order of the experiment. So, in the following study, we added further items to the assessment of abstract principles and also established a 2-week delay between the case-based evaluations (in Study 2a) and the abstract belief task (in Study 2b). This delay could be expected to weaken, and perhaps entirely annul, order effects resulting from the sequential assessment of concrete cases and abstract standards (please see supplementary file).

Experiment 2a

In Study 2a, we conducted a pre-registered replication and extension of the findings in Study 1. The extension aspect of Study 2a involved ascertaining whether the effect of speaker identity was driven by differences in social status, or target group membership.

Method

Pre-test

We recruited 39 volunteers to evaluate perceived differences in social status between the speakers in each vignette. Participants completed an adaptation of the MacArthur Scale of Subjective Social Status (Adler et al., 2000), ranking each of the speakers on a pictorial scale from depicting a ladder, anchored at 1: “Lowest rung” and 10: “Highest rung.”

High Status Outgroup members (M = 6.98, 95% CI = [6.24, 7.72]; employed in Study 1) were ascribed higher social status than both Ingroup members (M = 5.29, 95% CI = [4.55, 6.02]), B = 1.69, t = 12.55, and the new set of Low Status Outgroup members (M = 4.90, 95% CI = [4.16, 5.64]), B = 2.06, t = 15.40, both ps < .001. Low Status Outgroup members were ascribed slightly lower status than Ingroup members, B = −0.38, t = −2.83, p = .013.

Pre-registration

To estimate our required sample size, we conducted a bootstrap power analysis (Efron & Tibshirani, 1994; Walters & Campbell, 2004) by sampling with replacement from the data in Study 1. We reran our primary multilevel model for each of 5,000 resamples and computed the proportion of significant main effects (p < .05) as our estimate of statistical power. A target N of 320 resulted in very high power to observe effects of speaker identity (>.99) and intent (.99), but insufficient power to detect an effect of outcome (.74). To ensure sufficient statistical power to draw comparisons between three levels of speaker identity, we sought to recruit at least 480 participants. Further details concerning our analysis plan were pre-registered at https://aspredicted.org/t7jw4.pdf.

Procedure

In a balanced incomplete block design, participants were assigned to one of 12 blocks. In each block, participants viewed eight conditions drawn from the 3 (Speaker: ingroup, non-dominant outgroup, dominant outgroup) × 2 (Intent: negative, neutral) × 2 (Outcome: negative, neutral) factorial design. On each trial, we paired a unique experimental condition with a different scenario—such that no participant viewed the same scenario or experimental condition twice. Collapsing across groups, we achieved balance in the scenario-by-condition matrix (n per cell: M = 42, Min = 40, Max = 47).

After each trial, participants made four judgments on continuous scales, anchored at −3: “Strongly disagree” and 3: “Strongly agree.” Two items concerned the offensive nature of the statement, just as in Study 1. A further two items concerned the acceptability (i.e., “I don’t see a problem in saying that sort of thing.”) and censurability (i.e., the speaker “should not say that sort of thing.”) of the statement. All four items were averaged to form a composite acceptable-offensive index, which exhibited very good reliability (Cronbach’s α = .85). This level of reliability somewhat undermines the prediction that people might conceptualize offensiveness and permissibility as orthogonal: for example, that offensive speech might nonetheless be acceptable, or that informative speech might be censurable.

As in Study 1, we included a post-test questionnaire about the relevance of each criterion, composed this time of two items per criterion. The first part of the task asked how relevant each of four features should be on 7-point scales from −3: “No importance/relevance” to 3: “Absolute importance/relevance”: membership (“whether or not the speaker belongs to the group he or she talks about”), status (“whether or not the speaker has a higher social status than the group he or she talks about”), intent (“whether or not the speaker is derogatory toward the group he or she talks about”), and outcome (“whether or not people in the group the speaker talks about are offended”).

In the second part of the task, participants were asked whether the four target features “tended to make a statement more offensive” and then rated each feature independently, from −3: “Strongly disagree” to 3: “Strongly agree”: “if the speaker does not belong to the group he or she talks about,” “if the speaker has a higher social status than the group he or she is talking about,” “if the speaker intends to offend the target group,” and “if the statement harms the target group.” The two-item averages constituted our measures of abstract beliefs about offensive speech and revealed poor to satisfactory reliability (Cronbach’s αs from .57 to .74).

Participants

We recruited a sample of 500 native Spanish speakers from the Netquest panel (www.netquest.com), stratified by gender (1:1 ratio). A total of 515 participants were invited to participate, of whom 3 did not complete the survey and 7 failed our attention check, resulting in a final sample of 505 participants (252 women, 253 men).

Mean sample age was 44.6 years, ranging from 18 to 84 years. The sample was largely non-religious, with 51% reporting no church attendance, and somewhat left of center (M = 3.39, SD = 1.40, on a 1: “Left” to 7: “Right” scale).

Results

Summary statistics for each experimental condition are reported in Table 3.

Table 3.

Offensiveness by Condition: Means and 95% Confidence Intervals.

	Intent	Outcome	Member	Non-member
			Low status		High status
Condition			Ingroup	Outgroup	Outgroup
Description	Neutral	Neutral	–0.67[–1.20, –0.14]	–0.54[–1.07, –0.01]	–0.46[–0.99, 0.07]
Accident	Neutral	Negative	–0.27[–0.81, 0.26]	–0.11[–0.64, 0.43]	–0.16[–0.69, 0.37]
Attempt	Negative	Neutral	–0.25[–0.78, 0.28]	–0.19[–0.72, 0.34]	–0.12[–0.65, 0.41]
Offense	Negative	Negative	–0.17[–0.70, 0.36]	0.18[–0.35, 0.71]	0.32[–0.21, 0.85]

Note. Offensiveness ranges from −3: completely informative/acceptable to 3: completely offensive/unacceptable.

Pre-registered analyses

We pre-registered a replication and an extension analysis. As in Study 1, we analyze our data using multilevel models with random effects of participant and scenario, with effect coding for each of the fixed effects, and applying the Kenward-Roger approximation to calculate degrees of freedom.

Experimental effects: Replication

In our first analysis, we enter Identity, Intent, and Outcome, along with every two- and three-way interaction as fixed effects (just as in Study 1). This multilevel model revealed main effects of speaker identity, F(2, 3,528) = 9.74, as well as of intent, F(1, 3,517) = 53.41, and outcome, F(1, 3,517) = 55.85, all ps < .001. Again, no two- or three-way interactions emerged, all Fs < 1.92, all ps > .14.

Replicating Study 1, post hoc tests indicated that dominant outgroup speakers were more offensive than ingroup speakers, B = 0.24, 95% CI = [0.13, 0.35], t = 4.25, p < .001. In addition, negative intent resulted in more offensive speech than did neutral intent, B = 0.33, 95% CI = [0.24, 0.42], t = 7.31, and harmful speech was rated more offensive than innocuous speech, B = 0.34, 95% CI = [0.25, 0.43], t = 7.47, both ps < .001.

Experimental effects: Extension

To distinguish the effects of group membership and social status, we defined two additional dummy-codes, indicating whether each speaker was (a) an Ingroup or Outgroup member and (b) of Higher or Lower social status. Then, in our next analysis, we replaced the Identity term with these Membership and Status terms in the fixed effects portion of the model. We allowed both the Membership and Status terms to interact with Intent, with Outcome and with the Intent × Outcome interaction (see Table 4).

Table 4.

Multilevel Regression Models: Study 2a.

Models	Fixed effects	B[95% CI]	t	p
	intercept	–0.22[–0.66, 0.22]	–0.98	.36
Model 1: main effects only	Membership	0.18[0.07, 0.29]	3.19	.001
	Status	0.06[–0.05, 0.17]	1.07	.29
	Intent	0.34[0.23, 0.45]	5.97	<.001
	Outcome	0.31[0.20, 0.41]	5.48	<.001
Model 2: +two-way interaction terms	Member × Intent	0.06[–0.17, 0.28]	0.48	.64
	Status × Intent	0.09[–0.14, 0.31]	0.76	.45
	Member × Outcome	0.17[–0.06, 0.39]	1.47	.14
	Status × Outcome	–0.03[–0.25, 0.19]	–0.29	.78
	Intent × Outcome	–0.09[–0.30, 0.14]	–0.75	.45
Model 3: full factorial	Member × Intent × Outcome	0.25[–0.20, 0.70]	1.09	.28
Model 3: full factorial	Status × Intent × Outcome	0.20[–0.25, 0.65]	0.87	.39

Note. These models employ an effect coding system, such that factors with two levels are coded as {–0.5, 0.5}. As a result, in models involving interactions, the lower-level terms represent main effects and not simple effects. CI = confidence interval.

In addition to the main effects of intent and outcome, this multilevel model revealed an effect of Membership, F(1, 3,528) = 10.16, p = .001, but no effect of Status, F(1, 3,528) = 1.13, p = .29. As before, no two- or three-way interactions emerged, all Fs < 2.17, all ps > .14. Model comparisons confirmed that, by comparison to the additive model (Model 1; AIC = 15,012), entering the two-way interactions (Model 2; AIC = 15,017) did not improve model fit, χ²(5) = 4.82, p = .44, and neither did the full factorial model (Model 3; AIC = 15,017), χ²(7) = 8.68, p = .28.

Post hoc tests indicated that outgroup speakers (M = −0.13, 95% CI = [−0.65, 0.39]) were more offensive than ingroup speakers (M = −0.31, 95% CI = [−0.83, 0.21]), t = 3.19, p = .001, whereas higher (M = −0.25, 95% CI = [−0.77, 0.27]) and lower (M = −0.19, 95% CI = [−0.71, 0.33]) status speakers were seen as comparably offensive, t = 1.07, p = .29 (see Figure 1).

Supplementary analyses

Abstract principles

Next, we turned to the post-test questionnaire on abstract principles. Participants tended to endorse the intent principle (M = 1.71, SD = 1.26) and also agreed moderately with the outcome principle (M = 1.15, SD = 1.39). Meanwhile, features of the speaker’s identity elicited more ambivalent attitudes: Participants weakly agreed that the speaker’s membership in the target group is relevant (M = 0.25, SD = 1.60) and weakly disagreed that the speaker’s social status is relevant (M = −0.23, SD = 1.67). Every pairwise comparison was statistically significant (all ps < .001; see Figure 2).

Moderation analyses

Did people’s abstract relevance beliefs moderate the effects of the corresponding experimental manipulation? A series of model comparisons indicated that abstract beliefs about status, F(1, 3,550) = 0.02, p = .88, intent, F(1, 3,513) = 0.02, p = .88, and membership, F(1, 3,547) = 1.39, p = .24, were unrelated to the magnitude of the corresponding experimental effects. In contrast, the effect of the outcome factor was moderated by explicit beliefs about the relevance of outcomes, F(1, 3,514) = 5.60, p = .018. Specifically, endorsing the relevance of outcomes predicted greater offensiveness for negative outcomes (B = 0.19, 95% CI = [0.11, 0.26]) than for neutral outcomes (B = 0.11, 95% CI = [0.03, 0.18]), t = 2.37, p = .018.

Ideological differences

We were also interested in assessing whether political orientation moderated each of the experimental effects. We observed a main effect of political orientation F(1, 502) = 17.83, p <.001, such that liberals were more likely to ascribe offensiveness than were conservatives, B = −0.12, 95% CI = [−0.18, −0.06], t = −3.98, p < .001. However, political orientation did not moderate any of the experimental manipulations, all Fs < 1.59, all ps > .20.

Ideological differences in abstract principles were generally non-significant (membership: r = −.06, 95% CI = [−0.15, 0.03], p = .16; status: r = −.05, 95% CI = [−0.13, 0.04], p = .28; outcome: r = −.07, 95% CI = [−0.16, 0.01], p = .10), with the exception of speaker intent which was more relevant to liberals than to conservatives, r = −.10, 95% CI = [−0.19, −0.02], p = .020.

Discussion

Study 2a successfully replicated all three main effects observed in Study 1: i.e., of speaker identity, intent, and outcome. Our pre-registered extension further enabled us to dissociate the contributions of social hierarchy versus mere membership to the overall effect of speaker identity. This latter analysis revealed that, overall, outgroup speakers were more offensive than ingroup speakers, regardless of their social status.

Once again, in the abstract, people privileged the intent criterion (and to a lesser extent the outcome criterion) and were ambivalent toward standards concerning the speaker’s identity. Abstract beliefs about the relevance of these criteria were unrelated to participants’ susceptibility to the corresponding experimental manipulation. However, the absence of evidence indicating “top-down” moderation could stem in part from the limited reliability of the abstract measures we employed.

Finally, a noteworthy discrepancy arose between the results of Studies 1 and 2a. In particular, the effect of speaker identity was smaller in Study 2a than in Study 1.

Experiment 2b

In both studies, the post-test assessment of participants’ beliefs about offensive speech suggested that people endorse the intent and outcome criteria, but are divided regarding the relevance of features of speakers’ identity—such as their status or their membership in the target group. Yet, given the fixed order of the tasks, we cannot rule out that participation in the preceding experimental portion of the study distorted reports of participants’ abstract principles. For instance, participants may have retrospectively discerned a bias in their own evaluations of the concrete scenarios and felt motivated to conceal it. Alternatively, they may have been largely uncertain of the determinants of their concrete judgments and simply advanced what they consider a sensible justification (see, for example, Haidt, 2001).

To examine this question, participants in Study 2a were invited to take part in a re-test following a 2-week delay. After this delay, the concrete scenario portion should no longer be vivid in participants’ memory—substantially diminishing the motivation to justify or otherwise compensate for it in their reassessment of abstract beliefs.

Second, we obtained mixed evidence of a disconnect between abstract principles and concrete judgments. In Study 1, speaker identity was the largest determinant of offensiveness though participants denied its relevance in the abstract. Then, in Study 2a, speaker identity (specifically, membership) was the weakest determinant and also the most questioned principle. So, in Study 2b, we employed a new method to assess whether abstract principles and concrete judgments converge or diverge, by incorporating a “joint evaluation” task (see Hsee et al., 1999).

Method

Pre-registration

We pre-registered three hypotheses. First, to examine whether demand effects tarnished participants’ post-test endorsement of abstract principles, we sought to compare post-test to re-test beliefs in a series of paired t-tests. Second, we analyze whether abstract principles moderate concrete judgments, employing the re-test measures. Third, to further assess the correspondence between participants’ standards for offensiveness and their actual evaluations of offensive speech, we probed their attitudes toward each standard using a joint evaluation method.² Details concerning our analysis plan are pre-registered at https://aspredicted.org/cw3cb.pdf.

Procedure

After a two-week delay, we invited participants in Study 2a to take part in a follow-up study. This session consisted of a re-test of the eight-item questionnaire concerning beliefs about offensive speech (with slightly modified instructions), and an additional joint evaluation task. Once again, the two-item averages constituted our measures of abstract beliefs about offensive speech, and again showed variable reliability ranging from unacceptable to good (Cronbach’s αs from .54 to .78).

In addition, participants completed a novel, joint evaluation task. In this task, they viewed one of the eight statements in our scenarios out of context. Then, for each of the four determinants in Study 2a, they were asked a comparative question, for example:

Suppose that A belongs to the group he/she is talking about, while B does not belong to that group. In other respects, they are identical. The only difference between them is whether they belong to that group.

Is this statement more offensive when A makes it, when B makes it, or equally offensive in both cases?

They then reported the comparative importance (of each factor) on a bipolar scale, ranging from −3: “Much more offensive when A makes it,” through 0: “Equally offensive/inoffensive in both cases,” to 3: “Much more offensive when B makes it.”

Participants

All 505 participants in Study 2a were invited to take part, of whom 408 completed the follow-up study (192 women, 216 men).

Results

Pre-registered analyses

Longitudinal change in abstract principles

First, we examined longitudinal change in abstract beliefs about offensive speech through a series of paired t-tests. We observed no change in the relevance of membership (t = −0.88, p = .38, Cohen’s d = −0.04, 95% CI = [−0.14, 0.05]), social status (t = 1.03, p = .30, Cohen’s d = 0.05, 95% CI = [−0.05, 0.15]), or outcome (t = −1.56, p = .12, Cohen’s d = −0.08, 95% CI = [−0.17, 0.02]. We did, however, observe a significant decline in the relevance of intent, t = 3.47, p < .001, Cohen’s d = −0.17, 95% CI = [−0.27, −0.07]. Reduction in the abstract relevance of intent could indicate that post-test endorsement of the intent standard was inflated—perhaps in the service of justifying one’s pattern of responses. Furthermore, test–retest reliability was low (.34 < rs < .51), indicating instability in these attitudes.

Moderation analysis

Next, we examined whether abstract beliefs (measured during the retest) were associated with participants’ susceptibility to the corresponding experimental manipulations. As before, we observed no principle-by-effect interactions—whether for membership, F(1, 2,718) = 0.15, p = .70, status, F(1, 2,720) = 0.61, p = .44, intent, F(1, 2,688) = 1.12, p = .29, or outcome, F(1, 2,689) = 1.02, p = .31.

Joint evaluation task

To examine the comparative relevance of each factor, we conducted one-sample t-tests against the midpoint, that is, against the belief that a statement is “equally offensive/inoffensive in both cases.” Participants reported that a statement would be more offensive: (a) “on occasions when the audience was offended” (t = 6.91, Cohen’s d = 0.34, 95% CI = [0.24, 0.44]), (b) “when made by a speaker with the intention to offend” (t = 16.50, Cohen’s d = 0.81, 95% CI = [0.70, 0.93]), and (c) “when made by a speaker with a high socioeconomic status” (t = 8.26, Cohen’s d = 0.41, 95% CI = [0.31, 0.51]), all ps < .001.

Meanwhile, in a joint evaluation context, participants did not view the statement as more offensive ‘when made by a speaker who does not belong to the group he or she is talking about’: Comparative judgments about the relevance of membership did not significantly differ from “equally offensive/inoffensive in both cases,” t(407) = −0.85, p = .40, Cohen’s d = −0.04, 95% CI = [−0.14, 0.05] (see also Figure 3).

Figure 3.

Comparative judgments in joint evaluation task.

Supplementary analyses

We observed no ideological disagreements concerning the abstract relevance of membership (r = −.04, p = .41) or status (r = −.02, p = .64). However, the intent (r = −.18, 95% CI = [−0.27, −0.08], p < .001) and, to a lesser extent, outcome (r = −.09, 95% CI = [−0.19, 0.00], p = .056) criteria were more relevant for progressives than for conservatives.

Discussion

Following a 2-week delay, participants’ abstract principles had shifted. In particular, their endorsement of the intent criterion declined over time. Together with evidence that abstract criteria revealed low stability, and failed to predict variation in the magnitude of experimental effects, this result provides new insight into their post-test beliefs: Perhaps, uncertain of the determinants that caused their concrete judgments, participants tended to offer what they perceived to be a sensible justification—namely, that condemnation of offensive speech was motivated by concerns about the speakers’ harmful intent. A parallel prediction that participants might have downplayed their sensibility to features of the speakers’ identity was not supported—since we observed no longitudinal increase in the endorsement of these criteria.

Studies 1 and 2a produced mixed evidence of a disconnect between abstract principles and concrete judgments. Drawing on a comparative method (inspired by research on joint evaluation; see Hsee et al., 1999), we obtained new evidence that participants reject the ingroup membership criterion, even though ingroup membership shaped their concrete evaluations in the experimental portion of Study 2a.

General Discussion

The goal of the current research was to examine the effect of three factors that may contextually modulate the meaning and permissibility of a contentious statement. Looking across a diverse set of eight statements, we manipulated speaker identity, intent, and outcome to assess three hypotheses about the nature of verbal offense. To initiate the discussion, we provide a synthesis of our primary results:

Dovetailing with prior research on physical harm, evaluations of verbal harm integrated information about the speaker’s intent and the outcomes to the audience.

Controlling for the effects of intent and outcome, outgroup speakers were seen as more offensive than ingroup speakers—even when matched on socioeconomic status.

Although speaker identity norms—and, specifically, the membership norm—exerted a robust influence on participants’ case-based evaluations, participants tended to discount the relevance of these norms when asked abstractly or comparatively.

In support of H₁, intent and outcome information modulated participants’ characterization of speech—both in terms of its offensiveness and its acceptability (see also Swim et al., 2003).

Our results also provided support for H₂, in that speaker identity—in particular, whether the speaker belonged to the target group—played a robust role in judgments of offensive speech. Namely, offensiveness was greater among outgroup speakers, whereas acceptability was greater among ingroup speakers.

But did participants introspectively recognize the impact of speaker identity on their case-based judgments of offensive speech? Overall, participants were ambivalent toward speaker identity norms—ranking them the least relevant—while endorsing speaker intent as the most relevant criterion. Furthermore, when participants were asked to consider the relevance of speaker identity through an abstract exercise in comparing ingroup speakers to outgroup speakers, they overwhelmingly denied that their statements would differ in offensiveness. Thus, our study documented a degree of incongruence between the guidelines that people report as constitutive of offensive speech, and the eliciting factors that shape their evaluations of specific statements.

Somewhat surprisingly, our study revealed relatively modest, and inconsistent, effects of political orientation. We observed a main effect of political orientation, such that progressives were more likely to view speech as offensive than were conservatives. This result converges with survey evidence gathered in the United States (Fingerhut, 2016)—where Republicans typically report that “Too many people are easily offended these days over the language that others use,” while Democrats report that “People need to be more careful about the language they use to avoid offending people with different backgrounds.” However, progressives and conservatives did not systematically differ in the weight they placed on speakers’ identity or intent. Regarding the outcome factor, we found inconsistent evidence that progressives placed greater weight on outcome considerations than did conservatives (i.e., in Study 1 but not Study 2).

The discrepancy between explicit, abstract principles and case-specific evaluations of offensive speech may yield insight into recurring conflicts surrounding the limits of freedom of expression. At the explicit or abstract level, people appear to reject speaker identity norms: that is, many believe that ingroup and outgroup speakers are equally offensive or inoffensive in their remarks and should receive the same degree of reprobation. This explicit cognition is likely to guide numerous action plans: such as prospectively deciding what type of statements we may make in public, and the free speech guidelines we endorse. Yet, at the implicit level, people react to outgroup speakers and ingroup speakers differently—and these implicit cognitions are likely to govern other behaviors, such as our retrospective interpretation of concrete events. Together, these attitudes foreshadow a paradoxical outcome: namely, that people might favor a liberal policy toward free speech but retrospectively disapprove of its implementation.

Having diagnosed this problem, it is worth noting two potential resolutions. On one hand, we might conclude that our concrete judgments regarding particular utterances are “noisy” and more susceptible to bias and motivated reasoning than our abstract avowals—which are the product of careful deliberation over general moral principles. From this perspective, it may follow that we ought to favor the latter, and exercise control in suppressing the effects of speaker identity on our concrete appraisals.

However, we can read this problem in a very different way: Our concrete judgments are the more genuine kind, since they approximate the way we might respond to a range of potential circumstances. Abstract principles do not share this quality (see also Hsee & Zhang, 2004); they are instead poor reports of the general pattern in our concrete judgments, or perhaps prescriptive ideals that we routinely fail to apply. If so, there is a strong case to be made that agents in the public sphere should heed the opposite recommendation: namely, to acknowledge the relevance of speaker identity when deliberating abstractly about the standards that ought to regulate verbal offense.

One way to explain this disanalogy between our concrete and abstract judgments could be that we tacitly follow the rules that govern linguistic exchanges, without declaratively recognizing them—an idea that stretches back to Wittgenstein (1922): Humans “possess the capacity of constructing languages, in which every sense can be expressed, without having an idea how and what each word means” (§4.002). If so, this incongruence could be an instance of a more general introspective failure—that is, to declaratively recognize our tacit beliefs—which have been previously observed across diverse domains of prejudice research (Crandall et al., 2013; Tarman & Sears, 2005).

Limitations

First, our study was conducted in a single country, Spain. The effects we observed—concerning who may say what—could hinge on prevailing cultural norms which themselves are prone to substantial fluctuation. If so, we cannot confidently assert that our findings will generalize to future time periods and/or cultures (Simons et al., 2017)—particularly if speaker identity effects are brought to the light and made the focus of sustained academic and public debate. At the same time, our results dovetail with previous studies carried out in China, Australia, and the United States (see Gibson et al., 2019; Jeffries et al., 2012)—indicating some degree of robustness across cultures.

Second, we obtained limited reliability on our measures of abstract beliefs about offensive speech. Limited reliability in these measures of abstract beliefs also somewhat tempers our conclusion regarding their incongruence with case-based evaluations—since measurement error must have attenuated the effects in our moderation analyses (Schmidt & Hunter, 1996). Thus, future work should pay closer attention to the psychometric aspects of this task to strengthen our understanding of the correlates and downstream effects of abstract views about verbal offense.

Third, though our studies made considerable progress in distinguishing the determinants of offensive meaning, they provided limited insight into questions of psychological mechanism. Which psychological representations causally mediate the impact of identity norms on determination of meaning? Future research may benefit from discriminating between intuitive candidate explanations: for instance, (a) representations of intent, implicating mentalization about the speaker’s intent and motives (Baker et al., 2009; Hannikainen, 2019) and (b) representations of meaning, via contextual shifts in the linguistic meaning of particular terms (Holtgraves et al., 1989).

Conclusion

At the most general level, our findings demonstrate that the contours of offensive speech are not informed exclusively—or perhaps even primarily—by linguistic cues, but rather by a wider contextual assessment (see also Galinsky et al., 2013; Gibson et al., 2019). This contextual analysis draws retrospectively on information about the speaker’s background and identity, and even prospectively on knowledge about downstream repercussions.

Research in moral psychology has made great strides toward understanding the cognitive processes that undergird evaluations of physical harm (Cushman, 2008; Cushman et al., 2009; Gino et al., 2008; Greene et al., 2009; Malle, 1999; Mikhail, 2007; Young et al., 2007; Van de Vondervoort & Hamlin, 2018). In contrast, research focused on understanding the determinants of verbal offense is limited (but see Swim et al., 2003). The present research constituted an attempt to fill this critical gap in our understanding of moral cognition. To this end, we reveal that offensive speech appears to be built upon a similar psychological template, whereby people integrate intent and outcome information to derive a moral appraisal. However, we also find that the nature of offensive speech is determined by a third criterion, the speaker’s identity—an effect that arises among conservatives and progressives alike. Lay judges underestimate the relevance of speaker identity, but not intent, on their own assessments of a series of putatively offensive remarks. We suggest that this introspective failure may contribute to the persistent controversy regarding the boundary between freedom of expression and intergroup offense.

Supplemental Material

sj-docx-1-psp-10.1177_01461672211026128 – Supplemental material for Whose Words Hurt? Contextual Determinants of Offensive Speech

Supplemental material, sj-docx-1-psp-10.1177_01461672211026128 for Whose Words Hurt? Contextual Determinants of Offensive Speech by Manuel Almagro, Ivar R. Hannikainen and Neftalí Villanueva in Personality and Social Psychology Bulletin

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: This work was funded by the Spanish Ministry of Economy (Project FFI2016-80088-P, FPI Predoctoral Fellow BES-2017- 079933, IJC2018-037682-I), the Spanish Ministry of Science (PID2019-109764RB-I00), Junta de Andalucía (B-HUM-459-UGR18), and the FiloLab Group of Excellence at the Universidad de Granada.

ORCID iD

Manuel Almagro

Supplemental Material

Supplemental material is available online with this article.

Notes

References

Adler

N. E.

Epel

E. S.

Castellazzo

Ickovics

J. R.

(2000). Relationship of subjective and objective social status with psychological and physiological functioning: Preliminary data in healthy, White women. Health Psychology, 19(6), 586–592.

Alter

A. L.

Aronson

Darley

J. M.

Rodriguez

Ruble

D. N.

(2010). Rising to the threat: Reducing stereotype threat by reframing the threat as a challenge. Journal of Experimental Social Psychology, 46, 166–171. https://doi.org/10.1016/j.jesp.2009.09.014

Ayala

(2016). Speech affordances: A structural take on how much we can do with our words. European Journal of Philosophy, 24(4), 879–891.

Ayala

(2018). A structural explanation of injustice in conversations: It’s about norms. Pacific Philosophical Quarterly, 99, 726–748. https://doi.org/10.1111/papq.12244

Baker

C. L.

Saxe

Tenenbaum

J. B.

(2009). Action understanding as inverse planning. Cognition, 113(3), 329–349.

Bates

Mächler

Bolker

Walker

(2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 1–48. https://doi.org/10.18637/jss.v067.i01

Carlsmith

K. M.

Darley

J. M.

Robinson

P. H.

(2002). Why do we punish? Deterrence and just deserts as motives for punishment. Journal of Personality and Social Psychology, 83(2), 284–299. https://doi.org/10.1037/0022-3514.83.2.284

CATO Institute. (2017, August 15–23). Free Speech and Tolerance Survey. https://www.cato.org/sites/cato.org/files/survey-reports/topline/cato-free-speech-tolerance-toplines.pdf

Centro de Investigaciones Sociológicas. (2019). CIS Survey Number 3269: December Barometer. Post-electoral General Elections. http://www.analisis.cis.es/cisdb.jsp

10.

Crandall

C. S.

Ferguson

M. A.

Bahns

A. J.

(2013). When we see prejudice: The normative window and social change. In Stangor

Crandall

C. S.

(Eds.), Frontiers of social psychology. Stereotyping and prejudice (pp. 53–69). Psychology Press.

11.

Croom

A. M.

(2011). Slurs. Language Science, 33, 343–358.

12.

Cushman

(2008). Crime and punishment: Distinguishing the roles of causal and intentional analyses in moral judgment. Cognition, 108, 353–380. https://doi.org/10.1016/j.cognition.2008.03.006

13.

Cushman

Dreber

Wang

Costa

(2009). Accidental outcomes guide punishment in a “Trembling Hand” game. PLOS ONE, 4, Article e6699. https://doi.org/10.1371/journal.pone.0006699

14.

De Keersmaecker

Bostyn

D. H.

Van Hiel

Roets

. (2020). Disliked but free to speak: Cognitive ability is related to supporting freedom of speech for groups across the ideological spectrum. Social Psychological and Personality Science, 12, 34–41.

15.

Dovidio

J. F.

Kawakami

Gaertner

S. L.

(2002). Implicit and explicit prejudice and interracial interaction. Journal of Personality and Social Psychology, 82(1), 62–68. https://doi.org/10.1037/0022-3514.82.1.62

16.

Efron

Tibshirani

R. J.

(1994). An introduction to the bootstrap. Chapman & Hall.

17.

Epstein

Parker

C. M.

Segal

J. A.

(2018). Do justices defend the speech they hate? Journal of Law and Courts, 6(2), 237–262.

18.

Fasoli

Carnaghi

Paladino

M. P.

(2015). Social acceptability of sexist derogatory and sexist objectifying slurs across contexts. Language Science, 52, 98–107.

19.

Fernbach

P. M.

Rogers

Fox

C. R.

Sloman

S. A.

(2013). Political extremism is supported by an illusion of understanding. Psychological Science, 24(6), 939–946.

20.

Fingerhut

(2016, July 20). In “political correctness” debate, most Americans think too many people are easily offended. Pew Research Center. https://www.pewresearch.org/fact-tank/2016/07/20/in-political-correctness-debate-most-americans-think-too-many-people-are-easily-offended/

21.

Fundación

BBVA

. (2019). European Values Study 2019. https://www.fbbva.es/wp-content/uploads/2019/09/Presentacion_Estudio_Valores_2019.pdf

22.

Galinsky

A. D.

Wang

C. S.

Whitson

J. A.

Anicich

E. M.

Hugenberg

Bodenhausen

G. V.

(2013). The reappropriation of stigmatizing labels: The reciprocal relationship between power and self-labeling. Psychological Science, 24(10), 2020–2029.

23.

Gibson

J. L.

Epstein

Magarian

G. P.

(2019). Taming uncivil discourse. Political Psychology, 41, 383–401. https://doi.org/10.1111/pops.12626

24.

Gino

Moore

D. A.

Bazerman

M. H.

(2008). No harm, no foul: The outcome bias in ethical judgments (Working Paper No. 08-080). Harvard Business School.

25.

Goodwin

[Goodwin MJ]. (2020, February 8). More than 3 million EU nationals have applied to stay in the UK. 90% now have the right to Remain. Just 6 applications rejected (due to criminality). Intolerant Britain? Hmm [Tweet]. https://twitter.com/GoodwinMJ/status/1226087204453310464

26.

Gray

Wegner

D. M.

(2008). The sting of intentional pain. Psychological Science, 19(12), 1260–1262.

27.

Greene

J. D.

Cushman

F. A.

Stewart

L. E.

Lowenberg

Nystrom

L. E.

Cohen

J. D.

(2009). Pushing moral buttons: The interaction between personal force and intention in moral judgment. Cognition, 111(3), 364–371.

28.

Haidt

(2001). The emotional dog and its rational tail: A social intuitionist approach to moral judgment. Psychological Review, 108(4), 814–834. https://doi.org/10.1037/0033-295X.108.4.814

29.

Hannikainen

I. R.

(2019). Ideology between the lines: Lay inferences about scientists’ values and motives. Social Psychological and Personality Science, 10(6), 832–841.

30.

Henley

[johnhenley]. (2020, February 8). Misleading and actually offensive [Tweet]. https://twitter.com/jonhenley/status/1226203786978496512

31.

Holtgraves

(1994). Communication in context: Effects of speaker status on the comprehension of indirect requests. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(5), 1205–1218. https://psycnet.apa.org/doi/10.1037/0278-7393.20.5.1205

32.

Holtgraves

Srull

T. K.

Socall

(1989). Conversation memory: The effects of speaker status on memory for the assertiveness of conversation remarks. Journal of Personality and Social Psychology, 56(2), 149–160. https://doi.org/10.1037/0022-3514.56.2.149

33.

Hornsey

M. J.

Imani

(2004). Criticizing groups from the inside and the outside: An identity perspective on the intergroup sensitivity effect. Personality & Social Psychology Bulletin, 30(3), 365–383. https://doi.org/10.1177/0146167203261295

34.

Hsee

C. K.

Loewenstein

G. F.

Blount

Bazerman

M. H.

(1999). Preference reversals between joint and separate evaluations of options: A review and theoretical analysis. Psychological Bulletin, 125(5), 576–590.

35.

Hsee

C. K.

Zhang

(2004). Distinction bias: Misprediction and mischoice due to joint evaluation. Journal of Personality and Social Psychology, 86(5), 680–695.

36.

Jay

Janschewitz

(2008). The pragmatics of swearing. Journal of Politeness Research Language, Behaviour, Culture, 4, 267–328. https://doi.org/10.1515/JPLR.2008.013

37.

Jeffries

C. H.

Hornsey

M. J.

Sutton

R. M.

Douglas

K. M.

Bain

P. G.

(2012). The David and Goliath principle: Cultural, ideological, and attitudinal underpinnings of the normative protection of low-status groups from criticism. Personality and Social Psychology Bulletin, 38, 1053–065. https://doi.org/10.1177/0146167212444454

38.

Jeshion

(2013). Slurs and stereotypes. Analytic Philosophy, 54(3), 314–329.

39.

Johnson

P. O.

Fay

L. C.

(1950). The Johnson-Neyman technique, its theory and application. Psychometrika, 15, 349–367.

40.

Kukla

(2014). Performative force, convention, and discursive injustice. Hypatia, 29(2), 440–457.

41.

Langton

(2012). Beyond belief: Pragmatics in hate speech and pornography. In McGowan

M. K.

Maitra

(Eds.), Speech and harm: Controversies over free speech (pp. 72–93). Oxford University Press.

42.

Langton

(2018). The authority of hate speech. In Gardner

Green

Leiter

(Eds.), Oxford studies in philosophy of law (Vol. 3, pp. 123–152). Oxford University Press.

43.

Leets

(2002). Experiencing hate speech: Perceptions and responses to Anti-Semitism and antigay speech. Journal of Social Issues, 58(2), 341–361. https://doi.org/10.1111/1540-4560.00264

44.

Leets

Giles

(2006). Words as weapons—When do they wound? Investigations of harmful speech. Human Communication Research, 24(2), 260–301.

45.

Long

J. A.

(2019). interactions: Comprehensive, user-friendly toolkit for probing interactions (R Package Version 1.1.0). https://cran.r-project.org/package=interactions

46.

Luke

S. G.

(2017). Evaluating significance in linear mixed-effects models in R. Behavior Research Methods, 49(4), 1494–1502.

47.

Lystad

R. D.

(2019, September 3–5). Campaign for free speech. https://www.campaignforfreespeech.org/wp-content/uploads/2019/10/Free-Speech-Survey-Standard-Banners.pdf

48.

Malle

B. F.

(1999). How people explain behavior: A new theoretical framework. Personality and Social Psychology Review, 3, 23–48. https://doi.org/10.1207/s15327957pspr0301_2

49.

McGowan

M. K

. (2012). On “Whites only” signs and racist hate speech: Verbal acts of racial discrimination. In McGowan

M. K.

Maitra

(Eds.), Speech and harm: Controversies over free speech (pp. 121–157). Oxford University Press.

50.

McGowan

M. K.

(2018). On covert exercitives speech and the social world. In Fogal

Harris

Moss

(Eds.), New works on speech acts (pp. 185–201). Oxford University Press.

51.

Mikhail

(2007). Universal moral grammar: Theory, evidence and the future. Trends in Cognitive Sciences, 11, 143–152. https://doi.org/10.1016/j.tics.2006.12.007

52.

Miller

D. T.

Effron

D. A.

(2010). Psychological license: When it is needed and how it functions. Advances in Experimental Social Psychology, 43, 115–155. https://doi.org/10.1016/S0065-2601(10)43003-8

53.

Nisbett

Wilson

(1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84(3), 231–259.

54.

O’Dea

C. J.

Miller

S. S.

Andres

E. B.

Ray

M. H.

Till

D. F.

Saucier

D. A.

(2015). Out of bounds: Factors affecting the perceived offensiveness of racial slurs. Language Science, 52, 155–164.

55.

Paulhus

D. L.

Reid

D. B.

(1991). Enhancement and denial in socially desirable responding. Journal of Personality and Social Psychology, 60(2), 307–317. https://doi.org/10.1037/0022-3514.60.2.307

56.

Rozenblit

Keil

(2002). The misunderstood limits of folk science: An illusion of explanatory depth. Cognitive Science, A Multidisciplinary Journal, 26(5), 521–562.

57.

Saul

(2018). Dogwhistles, political manipulation and philosophy of language. In Fogal

Harris

Moss

(Eds.), New works on speech acts (pp. 360–383). Oxford University Press.

58.

Schmidt

F. L.

Hunter

J. E.

(1996). Measurement error in psychological research: Lessons from 26 research scenarios. Psychological Methods, 1(2), 199–223.

59.

Simons

D. J.

Shoda

Lindsay

D. S.

(2017). Constraints on Generality (COG): A proposed addition to all empirical papers. Perspectives on Psychological Science: A Journal of the Association for Psychological Science, 12(6), 1123–1128. https://doi.org/10.1177/1745691617708630

60.

Stanley

(2015). How propaganda works. Princeton University Press.

61.

Struchiner

De Almeida

G. D. F.

Hannikainen

I. R.

(2020). Legal decision-making and the abstract/concrete paradox. Cognition, 205, 104421.

62.

Swim

J. K.

Scott

E. D.

Sechrist

G. B.

Campbell

Stangor

(2003). The role of intent and harm in judgments of prejudice and discrimination. Journal of Personality and Social Psychology, 84(5), 944–959. https://doi.org/10.1037/0022-3514.84.5.944

63.

Tarman

Sears

(2005). The conceptualization and measurement of symbolic racism. The Journal of Politics, 67(3), 731–761.

64.

Van de Vondervoort

J. W.

Hamlin

J. K.

(2018). Preschoolers focus on others’ intentions when forming sociomoral judgments. Frontiers in Psychology, 9, Article 1851. https://doi.org/10.3389/fpsyg.2018.01851

65.

Walters

S. J.

Campbell

S. J.

(2004). The use of bootstrap methods for analysing health-related quality of life outcomes (particularly the SF-36). Health and Quality of Life Outcomes, 2(70). https://doi.org/10.1186/1477-7525-2-70

66.

Wike

Simmons

(2015, November 18). Global support for principle of free expression, but opposition to some forms of speech. Pew Research Center. http://www.pewglobal.org/2015/11/18/global-support-for-principle-of-free-expression-but-opposition-to-some-forms-of-speech/

67.

Wilson

(2002). Strangers to ourselves: Discovering the adaptive unconscious. Harvard University Press.

68.

Wittgenstein

(1922). Tractatus Logico-Philosophicus ( Ogden

C. K.

, Trans.). Routledge & Kegan Paul.

69.

Young

Cushman

Hauser

Saxe

(2007). The neural basis of the interaction between theory of mind and moral judgment. Proceedings of the National Academy of Sciences of the United States of America, 104, 8235–8240. https://doi.org/10.1073/pnas.0701408104

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.05 MB