Not All Interventions are Made Equal: Harnessing Design and Messaging to Nudge Bystander Intervention

Abstract

This study examined the influence of design “nudges” on bystanders’ willingness to intervene in online harassment using a social media simulation. Utilizing a 2 × 2 experimental design, we tested the ability of key design features (community guidelines and pop-up messaging) to induce a sense of self-efficacy (low/high) and personal responsibility (low/high) and thence to influence intervention levels. Participants (n = 206) were invited to “beta test” a new social networking site (SNS) for 15 minutes. All participants were exposed to four instances of online harassment against a victim. Bystanders in the low efficacy and high responsibility condition were most likely to intervene, although this finding only applied to “private” (e.g., direct, 1-2-1 messaging) rather than “public” (e.g., posting on a public feed) interventions. Overall, participants preferred “private” interventions that avoided public confrontation. Qualitative insights highlight a perceived lack of transparency in reporting options and a belief that interventions rarely made a difference as the “damage had been done.” Results are discussed in relation to the amplification of personal responsibility when the SNS does not provide clear guidelines and reminders. We recommend ways of “designing in” nudges in practice, to facilitate bystander intervention.

Introduction

Online harassment is a heavily debated term ¹ consisting of a range of behaviors varying in intent, severity, and harm. ² Here, we define it as targeted abuse or harmful behavior directed at another individual through computer-mediated communication (CMC). Exploring the role of the digital bystander is crucial since cyberbullying (a form of online harassment) has been linked with poorer mental health outcomes and increased suicide ideation in young adults ³ and has been found to be more detrimental than traditional in-person bullying. ⁴

Studies have identified numerous situational and personal factors that can facilitate intervention, such as prior relationship with the victim, ⁵ anonymity, ⁶ and level of personal responsibility.^7–9 Bystanders are more likely to intervene (both online and offline) when they are confident in their own abilities (a concept known as “self-efficacy”)^5,10,11 In contrast, those lower in self-efficacy act more passively in bystanding scenarios. ¹² In the context of CMC, a study of young adults (n = 1,180) found that low self-efficacy reduced the likelihood of intervening in online hate speech. ⁸ There is promising support that self-efficacy can be increased through digital intervention ¹³ ; however, there is large variance in the measurement of self-efficacy ranging from a stable, global sense of “efficacy” to a more nuanced, situation-specific efficacy with multiple facets. In this study, we examine both overall self-efficacy and bystander-specific efficacy in encouraging intervention when witnessing cyberbullying.

In recent years, there has been an innovative shift toward the use of simulation paradigms to measure bystander behavior^14–17 and “designing in” bystander intervention through “nudges.” ¹⁸ Digital nudges are interventions that steer people while also allowing them to “go their own way”. ¹⁹ (p3) Specifically, we test whether “reinforcement nudges” (increasing the salience of behavior in the mind of the user) and social nudges (inducing a social norm) ¹⁹ can increase bystander intervention. This study seeks to leverage the concepts of self-efficacy and personal responsibility specifically since they are well supported by existing literature. We take existing research ¹⁵ one step further by expanding the range of intervention options available to the bystander (as recommended by recent qualitative research ²⁰ ) in both the experimental design and analysis. The research has three interrelated research questions:

RQ1. To what extent do self-efficacy and personal responsibility “nudges” increase bystanders’ likelihood of intervening in online harassment?

RQ2. What is the interaction between personal responsibility and self-efficacy in relation to bystander intervention?

RQ3. How do efficacy and responsibility nudges influence the type of intervention (e.g. private vs. public) used to tackle online harassment?

Materials and Methods

A bystander simulation was conducted in 2023 using an adapted version of “SnapEatLove,” a high validity open-source platform, ¹⁵ which mimics the capabilities of the social networking site (SNS) Instagram (e.g., sharing, liking, posting) with a focus on food (see Supplementary Data S2). The platform was developed for studying behavior in social network sites in a naturalistic manner. We adapted the simulation to include a wider range of bystander intervention options ²⁰ and varied the location of “other users” to cities in the UK. All other user interactions remained the same as per the original simulation (see Procedure).

Design

The study adopted a 2 (low efficacy/high efficacy) × 2 (low responsibility/high responsibility) between-subjects factorial design. Self-efficacy and sense of personal responsibility were manipulated through two key design features: (a) the “community guidelines” prior to log on and (b) “pop-up” messaging on their newsfeed (see Supplementary Data S3). Participants were required to manually “close” the message, which ensured they attended to the information. To induce a sense of responsibility, we used a “social norm nudge” stating it was the responsibility of individual users to keep the community safe (high) or the responsibility of moderators (low). Self-efficacy was induced through “reinforcement nudges,” whereby users were reminded of all intervention options available for tackling online harassment (high) or omitting this information (low).

Participants

The participants (n = 206, 49.5 percent female, 49 percent male, and 0.5 percent nonbinary/third gender) were recruited using participant platform Prolific, under the guise of beta testing a new social media platform. Prolific screening was set that all participants were based in the UK and ensured a balanced gender composition. Under-16’s were excluded from the sample. Participants were removed from the study if any of the following criteria were met: (a) they did not create a profile on the site, (b) they did not create one post, and (c) they did not interact with others at least once (e.g., comment, like/share, report). Participants were paid above the national living wage at the time of the experiment (£9.50 per hour).

Ethics

Full ethical approval was granted through the School of Psychology's ethics committee (PREC) at the University of Bath. Given the deception involved in the study, mitigations were taken: (a) participants were fully debriefed on the aims following their participation, (b) they were provided with the opportunity to anonymously withdraw their data within 1 week, and (c) were signposted to cyberbullying resources.

Procedure

Participants were asked to provide information about their current social media use, demographics (e.g., age, gender), and directed to the SNS (“SnapEatLove”) where they created profile information for realism purposes. Over 15 minutes, participants were exposed to a feed of interactions occurring in apparently real time alongside other users (who were in reality preprogrammed bots) where they could share/flag/post as they would normally in their day-to-day lives. All participants were exposed to four instances of online harassment.

Tracking intervention rates and type

All interactions were logged and captured in individual CSV files (except images for ethical reasons). Overall bystander intervention was the sum of all “interventions” directed at the four harassment messages (our experimental stimuli) only (see Table 1).

Table 1.

Online Harassment Posts Used in the Simulation (Previously Used and Tested by Difranzo) ¹⁵

Post 1	“Stop posting this shit, nobody cares.”
Post 2	“Your life is sad. Look at what you eat.”
Post 3	“When will you get it into your head that nobody likes your stupid ass.”
Post 4	“This photo is uglier than you, and that’s saying something.”

Measures

After their allocated time on SnapEatLove, participants were directed to a survey site to complete a followup questionnaire before debriefing (full detail in Supplementary Data S1).

Experimental stimuli

Participants were exposed to four instances of online harassment against a single victim by a single offender. We used the same four posts as used by a previous study, ¹⁵ which had been tested for realism and severity (Table 1).

Analytic strategy

Multiple analysis of variance (MANOVAs) were used to test the impact of both main effects (responsibility and efficacy) as well as interactions and multiple intervention outcomes (and covariates) simultaneously. Using MANOVAs negates the need to adjust for multiple comparisons. The recommended sample size was calculated as 267 participants using GPower Software for a medium effect size according to Cohen (1998). Our final sample was slightly underpowered (N = 206) which should be factored into the interpretation of our results. We used the widely adopted p value of 0.05 throughout to determine the impact of all main effects and interactions.

Results

Manipulation check

We first conducted a manipulation check of the experimental conditions using a MANOVA. We included overall general self-efficacy score (M = 31.19, SD = 4.56) and a one-item responsibility score into the model as covariates.

First, we examined the self-efficacy manipulation on intervention-specific process and outcome efficacy. Level of confidence in one’s ability to intervene (“process efficacy” measured on a five-point Likert scale) was significantly different between the low and high self-efficacy conditions [F(1, 182) = 6.06, p = 0.015]. Specifically, those in the “low efficacy” condition scored significantly lower on confidence (M = 4.19, SD = 0.86) than those in the “high efficacy” condition (M = 4.43, SD = 0.71), suggesting that this manipulation was effective. There was no significant difference between groups in relation to confidence that this intervention would make a difference (outcome efficacy, F(1, 182) = 0.00, p = 0.96). Regarding the responsibility manipulation, self-reported responsibility was not significantly different between groups, suggesting this intervention alone was not effective.

However, when examining the interaction between both conditions, we found that there was a significant difference between groups in self-reported levels of responsibility [F(1, 182) = 5.92, p = 0.016]. Specifically, users’ reported sense of responsibility on the platform was the highest in the condition combining low efficacy/high responsibility (M = 4.07, SE = 0.15) compared with low efficacy/low responsibility (M = 3.50, SE = 1.44). Therefore, for the remainder of the analysis, we will focus on exploring interaction effects rather than the main effects of the conditions.

The role of responsibility and efficacy “nudges” on overall intervention rates

We explored whether the presence of personal responsibility nudges and efficacy “nudges” increased bystander intervention on SnapEatLove. A between-measures MANOVA (using the same covariates) revealed that induced efficacy alone was not a significant predictor [F(2, 206) = 0.07, p = 0.79], nor was induced responsibility [F(2, 206) = 2.64, p = 0.11]. However, the interaction between both nudges on intervention rates was approaching significance [F(2, 206) = 3.58, p = 0.06]. Specifically, participants intervened most in the low efficacy/high responsibility condition (M = 2.24, SE = 0.21) in comparison to the low efficacy/low responsibility condition (M = 1.44, SE = 0.21).

Other drivers of overall intervention

A linear regression examined intervention scores based on age. Age significantly predicted intervention scores [β = −0.18, p = 0.01], with intervention reducing as age increased. Gender did not have a significant impact on intervention, t(201) = 0.31, p = 0.6.

The impact of nudges on type of intervention—public vs. private

Rates of intervention were higher than expected, with only 28 percent of participants choosing not to intervene in any way (nonintervention rates were 75.4 percent in a previous study, ¹⁵ although a more limited range of intervention options was measured). Interestingly, the preferred intervention option was “liking” the victims original post (31.1 percent liked one post, 10.7 percent liked two posts) followed by flagging bully comments (21.4 percent flagged one post, 13.1 percent flagged twice).

To explore how design nudges impacted the choice of intervention, we created two new variables: (a) number of “public” interventions and (b) number of “private” interventions, for each participant. In line with previous research, ¹¹ private interventions (M = 1.016, SE = 0.41) were preferred over public interventions (M = 0.800, SE = 0.29). A between-measures MANOVA determined the impact of experimental condition on type of intervention. Taking each main effect in turn, there was no significant difference between the likelihood of conducting a public or private intervention. However, taken together as an interaction, the nudges significantly increased private interventions [F(1, 182) = 7.76, p = 0.01] but interestingly had no effect on public interventions [F(1, 182) = 0.22, p = 0.64]. Specifically, those in the low efficacy/high responsibility condition exhibited the highest rates of private intervention (M = 1.46, SD = 0.16) compared with those in the low efficacy/low responsibility condition (M = 0.71, SD = 0.16).

Qualitative insights

We examined all qualitative, free-text data collected in the followup survey through a process of bottom-up, semantic coding following the practices set out in thematic analysis. ²¹ (See Supplementary Data S4 for full analysis.) A strong theme was that social media companies themselves (or moderators) are responsible rather than individual users. In relation to self-efficacy, participants noted a lack of transparency and a perception that their intervention would not be effective. Exploring this further, participants described that the damage (social, emotional, reputational) to the victim had already been done.

Discussion

The present research explored how efficacy and responsibility design “nudges” influence bystander intervention in a realistic simulation. We found that self-efficacy “nudges” were effective and could be easily operationalized in the future, unlike responsibility “nudges.” This could speak to the difficulty of inducing a sense of responsibility when the bystander is immersed in a virtual network of acquaintances rather than close friends (a consistent finding in the broader literature ⁵ ).

Taken on their own, efficacy and responsibility “nudges” did not increase overall bystander intervention. More promising results were found when nudges were combined, although not statistically significant. In future studies, larger sample size and thus greater power may validate these findings. When examining the types of intervention, we found that a combination of the two design nudges (low efficacy/high responsibility) increased the likelihood of “privately” intervening (e.g., DM’s, reporting) but not public interventions (e.g., public reply). This finding suggests that a sense of responsibility can drive intervention under the right circumstances. One possible theoretical explanation is that the ambiguity created by not giving information on how to intervene (in the low efficacy condition) further amplified feelings of personal responsibility (in the high responsibility condition). In other words, if bystanders are not confident in intervention options offered by the SNS, they take intervention into their own hands (e.g., DM’s). In line with previous studies, participants preferred “private” interventions such as comforting victims that do not typically rely on the platform itself and often lack confidence in the SNS reporting process. ²⁰

In this study, we aimed to increase self-efficacy by clearly articulating and reminding participants of the intervention options. This manipulation alone was not sufficient to increase intervention (although it was a key covariate in our interaction analysis). Our qualitative findings suggest that bystanders often felt that interventions are unlikely to make a difference as the “damage had been done.” As such, we recommend future research and interventions could focus on the benefits of intervention (e.g., where intervention can lead to positive change), targeting the motivation to act in the first place. Our qualitative findings also suggest that bystander confidence (and the perceived social ramifications of “making it worse”) was a key barrier, regardless of intervention condition. Thus, we recommend future research and campaigns could specifically target “social” self-efficacy. Previous research has highlighted the specific role of “social” self-efficacy ²² —one’s confidence in social interaction and their perceived ability to resolve conflicts.

Conclusions

This study sheds light on the combined role of efficacy and responsibility on bystander intervention and supports the notion that such interventions can be “designed into” SNS using a realistic and controlled simulation. A key limitation is that participants were interacting with unknown users (bots) not their own social network community, which may well limit their overall sense of responsibility. Nevertheless, the study provides insight into how design initiatives can impact different types of digital intervention (e.g., public vs. private). Specifically, this study supports the idea that SNS should ensure bystanders have ample opportunities to privately intervene. ⁸ Our findings also suggest that interventions should target bystander confidence rather than bystander ability, for example, through mandatory bystander education.