DEMOGRAPHIC TRAITS AND ASSESSMENTS IN ONLINE ENVIRONMENTS

The Social Role Theory (SRT) contends that social norms and stereotypes can have significant implications for evaluations of minority groups and women, even when conducting the same activities or delivering the same messages (Diekman and Eagly Reference Diekman and Eagly2000; Schneider and Bos Reference Schneider and Bos2019). Minority groups often are associated with specific tasks, activities, and behaviors that are perceived as expected or appropriate for a given group (Luisi, Adams, and Kilgore Reference Luisi, Adams and Kilgore2021; Negrón-Muntaner et al. Reference Negrón-Muntaner, Abbas, Figueroa and Robson2014). In mass media, the absence of women in male-dominated fields or their infrequent appearance as experts has led to lower evaluations of women, even when they deliver the same message as men (Armstrong and McAdams Reference Armstrong and McAdams2009; Luisi, Adams, and Kilgore Reference Luisi, Adams and Kilgore2021). Moreover, according to SRT, men are rewarded for exhibiting behaviors that are congruent with an agentic personality and women are rewarded for exhibiting behaviors that are associated with communion—and likewise punished for those that are incongruent with these expected personality profiles (Koenig and Eagly Reference Koenig and Eagly2014). Similarly, the overrepresentation of African Americans as “lawbreakers” or “servants” and US Latinos as “unauthorized immigrants” or “non-English speakers” in the marketing, news, and entertainment industries is associated with negative assessments of these groups (e.g., lower levels of perceived credibility, leadership, and intelligence). Hence, the evaluation of the behavior or messages from minority groups tends to be contingent on existing norms and stereotypes.

Building on SRT, the Social Information Processing Theory (SIPT) contends that the implications of social norms and the backgrounds of messengers and receivers also transfer to online settings (Marmat Reference Marmat2022). This occurs because the web environment provides users with cues to “perceive and evaluate their communication partners with a similar level of accuracy to that of face-to-face communicators by utilizing and accumulating whatever information is available within the environment” (Lee and Lim Reference Lee and Lim2014, 556). Online cues include names, profile photographs, and videos, which allow computer-mediated communication participants to examine multiple sociodemographic characteristics of the messengers (both elites and the public), including sex, age, and racial and ethnic backgrounds (Settle Reference Settle2018). Some of these traits have been shown to influence receivers’ attitudes or evaluations not only in face-to-face interactions but also in printed and online environments. Based on SRT and SIPT, we expected that social media users would be more likely to report posts from women and ethnic minorities due to stereotypes associated with these groups (Hypothesis 1).

THE MEDIATING EFFECTS OF SHARED BACKGROUNDS

Shared demographic characteristics between messengers and receivers, however, can mediate the negative implications of strict social norms including gendered and racial stereotypes (Armstrong and McAdams Reference Armstrong and McAdams2009). In diverse contexts such as the United States, message receivers—particularly those from underrepresented groups—often acknowledge similarities between themselves and the message sources. These perceived similarities have the potential to influence thoughts, decisions, and evaluations because “we tend to compare ourselves to others, and the greater the similarity we perceive, the more likely we are to attend to and trust what [messengers] say” (Andsager and Mastin Reference Andsager and Mastin2003, 59). In other words, shared demographic characteristics such as race and ethnicity can be salient in a social media user’s mind when evaluating inflammatory content. As a result, we expected that users would be less likely to report posts from people with whom they share demographic characteristics (Hypothesis 2).

EXPERIMENTAL DESIGN

To examine whether some users are more likely to be reported than others for their inflammatory but rule-abiding content, we conducted an experiment embedded in the Knight Foundation–Ipsos Freedom of Expression Study, which was fielded online using the probability-based Ipsos KnowledgePanel on July 30, 2021, to a sample of 2,500 panel members (Feezell et al. Reference Feezell, Conroy, Gomez-Aguinaga and Wagner2022). Each subject was shown a short list of produced social media posts designed to mimic a mini-Twitter feed and they were asked to report any posts that they believed violated the terms of agreement (see online appendix A for examples). Specifically, the experiment pre-text read as follows:

Participants then were exposed to four mini-feeds on separate pages, each featuring one inflammatory post and two decoy posts.Footnote ¹ Each inflammatory post was designed to be provocative but not to violate the rules of common social media platforms, which prohibit threats of credible violence or hate speech directed at protected classes.Footnote ² As a result, this experiment allowed us to examine the likelihood that a post was reported, which—in the real world—would subject the post to review up the chain of command and cause it to experience a higher probability of mistakenly being censored.

For our treatment, each inflammatory post was randomized to vary the gender (Male; Female) and race (White Non-Hispanic; Black or African American) of the poster through the photograph and name of the person making the comment.Footnote ³ This created a 2X2 factorial design plus a control group (i.e., Black Male, Black Female, White Male, White Female, Control). Subjects were asked to check a box next to any posts in the list that they would “report” for violating community standards. Online appendix B describes the factorial design.

To test Hypothesis 1, we compared the average reporting rate for each of the treatment conditions to the control condition across the four inflammatory posts. Averaging the four posts together allowed us to better isolate the influence of the treatment (i.e., race and gender) on the likelihood of reporting regardless of the idiosyncrasies of the content for each post. Figure 1 shows the mean values and 95% confidence intervals for each treatment condition compared to the control group.Footnote ⁴ Using four t-tests, we found that none of the treatment conditions were significantly different from the control group in their likelihood of posts being reported. To ensure that the findings were not biased by the wording of each post or the oversampling of young and diverse populations, we also ran four separate logistic regressions to predict whether respondents indicated that they would flag each post (0/1). We included a categorical variable for each experimental treatment as the main independent variable, which allowed for a comparison of each treatment group to the control group, as well as control variables for age, gender, race, income, education, and party identification (see online appendix C for descriptive statistics). Again, the treatment conditions were not significant predictors of post reporting when compared to the control group. We found through post hoc analysis, however, that when compared to Republicans, Democrats had a higher likelihood of reporting all posts (p<0.05 for all four posts) (see online appendix D for full models and appendix E for an analysis by partisanship).

Figure 1

Effect of Treatment Assignment on Post Reporting

To test Hypothesis 2, we examined the moderating role that the respondent’s demographic characteristics play in deciding whether to report a post. We anticipated that people would be less likely to report others who they consider to be in their in-group. We ran a random-effects model interacting the respondent race and gender with the treatment condition race and gender to predict the probability that the respondent would report a post. Although it is likely that there is important nuance to be learned from more discrete categories of race and ethnicity, we restricted our analysis to concurrence between Black/White/Other respondents and Black/White treatment categories.

We ran separately two models with interactions for race and gender, including control variables for age, income, education, and party identification. Figure 2 (left side) presents the predicted probability of reporting a post comparing the race of the poster and the race of the participant. When interacting a post made by a Black poster with the participant’s own race, there was no significant effect and thus no support for Hypothesis 2 concerning race. Individuals were no more or less likely to report a poster of the same race. Figure 2 (right side) presents the predicted probability of reporting a post comparing the gender of the poster and the gender of the participant. The interaction between poster gender and participant gender was not significant. Thus, whereas women were more likely to report posts across both models, they did not appear to be more likely to report a post based on the perceived gender of the poster (see online appendix F for full models). Therefore, we found no support for Hypothesis 2. Individuals were no more or less likely to report a poster based on shared demographic characteristics.

Figure 2

Predicted Probability of Reporting a Post

Note: Bars show 95% confidence intervals.