Longitudinal Predictors of Generative AI-Involved Cyberbullying: A Two-Wave Longitudinal Study

Generative artificial intelligence as a facilitator of cyberbullying

Cyberbullying has emerged as a critical issue that increasingly attracts the attention of educators and mental health professionals. ¹ Victims of cyberbullying often suffer severe emotional and psychological consequences, including anxiety, depression, and diminished self-esteem. Olweus ² defines cyberbullying as bullying behavior carried out through electronic means such as email, instant messaging, chat rooms, websites, or mobile text messaging. Typically regarded as an unacceptable form of aggression, cyberbullying involves perpetrators intentionally causing harm to others via social media platforms, instant messaging, online games, and other digital communication technologies.^3,4 Historically, the scale of such cyberbullying was often constrained by the perpetrator’s cognitive bandwidth, temporal resources, and technical proficiency. Traditional cyberbullying relies heavily on the active engagement of the perpetrator, who sets conversational traps and monitors the victim’s emotional responses to derive sadistic pleasure.^5,6

However, the growing accessibility of modern generative artificial intelligence (GAI) tools—such as ChatGPT, Claude, Midjourney, and Sora—appears to lower many of these conventional barriers, potentially transforming this emerging form of aggression from a primarily manual endeavor into a more scalable and automated risk. ⁷ In the past, specific forms of cyberbullying (e.g., fabricating hyper-realistic forged images) required specialized technical skills. ⁸ GAI can substantially reduce these prerequisites, enabling perpetrators to generate deceptive or harmful content more rapidly and efficiently. ⁸ For instance, individuals can now employ deepfake technologies to fabricate compromising media to shame or defame victims, or utilize conversational AI to mass-produce harmful rumors. ⁷

This shift in technological accessibility may present new opportunities for malicious actors. A comprehensive qualitative study involving 43 Trust and Safety experts across domains such as child safety, hate and harassment, and violent extremism identified this lowered barrier to entry as one of the mechanisms of AI empowerment for perpetrators. ⁹ The widespread availability and intuitive, conversational interfaces of GAI tools allow users with limited technical proficiency to potentially create and propagate harmful material. This phenomenon is highlighted by research examining the capabilities of publicly available large language models. Studies have demonstrated that models such as GPT-4o, Gemini, Claude, and DeepSeek can be manipulated to assist individuals with minimal experience in engaging in sophisticated cyberbullying. ¹⁰

Consequently, GAI empowers perpetrators by facilitating cyberbullying with reduced manual effort and increased scalability. To address this emerging threat, it is essential to understand how aversive personality traits drive individuals to engage in GAI-involved cyberbullying (GAICB).

Dark Tetrad

Barlett et al. ⁵ conducted a meta-analysis of 211 studies, finding that after controlling for traditional bullying, dark personality traits were uniquely associated with cyberbullying perpetration rather than victimization. The Dark Tetrad comprises Machiavellianism, narcissism, psychopathy, and sadism. ¹¹ Prior research has consistently demonstrated strong associations between these traits and various forms of cyberbullying.^12–14 Although these traits share antisocial tendencies and beliefs, they differ in expression and mechanisms, prompting many researchers to examine them separately.^15,16 Individuals high in psychopathy are particularly prone to cyberbullying due to deficits in empathy and impulse control. ¹⁷ Narcissists may retaliate against perceived ego threats by attacking others, while Machiavellians tend to strategically manipulate others to serve personal interests. ¹⁴ Sadistic individuals derive pleasure from inflicting pain and have been identified as robust predictors of antisocial behaviors in online settings.^18,19 In particular, sadists often exploit the anonymity and psychological distance of digital environments to inflict verbal and behavioral harm without remorse. ²⁰

Demographic variables

Beyond personality traits, this study examines demographic variables due to their practical utility. Unlike psychological constructs, demographics are easily accessible through institutional records, making them invaluable for educators and practitioners to identify at-risk individuals without extensive psychometric testing.

The first variable of interest is gender. Research on face-to-face bullying has consistently shown that boys tend to engage in bullying behaviors more frequently than girls, ²¹ with male aggression more often taking a direct form, while girls are more likely to employ indirect forms of aggression. ²² Because cyberbullying can involve indirect, relational, or technologically mediated aggression, some studies have reported higher involvement among girls.^6,23 Although certain studies support this hypothesis (e.g., Kowalski and Limber ²³ ) other research has found no significant gender differences in cyberbullying involvement (e.g., Smith and colleagues^24–26). Therefore, gender was included as a demographic predictor in the present study since prior research suggests that gender may be associated with cyberbullying involvement in complex and context-dependent ways.

The second variable is age. Research on traditional bullying shows that prevalence rates tend to peak during middle school, as youth strive to establish their place within the social hierarchy. ²⁷ Similarly, cyberbullying is particularly common among minors, especially children and adolescents. ²⁸ A meta-analysis of 131 studies, most of which involved participants between the ages of 10 and 18, found that the likelihood of engaging in cyberbullying increases with age. ⁶

Present study

While the link between Dark Tetrad and traditional cyberbullying is well-established, research on GAICB remains scarce. It is unclear whether these personality traits retain their predictive validity within this novel technological landscape. To address this gap, we employed a two-wave longitudinal design with a 6-month interval. At Time 1 (T1), we assessed demographics (gender and age) and Dark Tetrad; GAICB was measured at both T1 and T2 to account for behavioral stability. Using hierarchical regression, we first identified predictors of T2 GAICB and subsequently controlled for T1 GAICB to evaluate their incremental predictive power. Specifically, this study addresses the following research question: After accounting for longitudinal stability, do demographics and Dark Tetrad uniquely predict GAICB at Time 2?

Participants and procedure

Data were collected via a convenience sampling method through the Chinese social media platform WeChat from participants in Mainland China. Individuals were invited to complete two identical online surveys administered through Tencent Questionnaire, a professional survey platform. This anonymous online survey format provided participants with a private environment that reduces the interpersonal pressure to provide socially acceptable answers compared to face-to-face settings. Prior to participating, all individuals provided informed consent, which explicitly guaranteed the confidentiality of their responses and clarified that the data would be used strictly for academic research purposes. These procedural safeguards could reduce social desirability bias when assessing sensitive antisocial behaviors. The first wave of data collection occurred between December 6 and December 30, 2024. The second wave was administered 6 months later, between June 6 and June 30, 2025. At the end of the first wave, participants were asked to provide their mobile phone numbers for follow-up contact. The second-wave survey was distributed via text message. Participation in both surveys was voluntary and compensated. The study received approval from the University Research Ethics Committee for Human Subject Protection and was conducted in accordance with ethical research guidelines.

A total of 1,019 participants completed the first wave of the survey, including 744 females (73.0 percent) and 275 males (27.0 percent). Participants ranged in age from 16 to 52 years, with a mean age of 23.58 years (SD = 4.06). Regarding educational background, most participants were undergraduates (n = 734, 72.0 percent), followed by those with a graduate degree or higher (n = 159, 15.6 percent), junior college education (n = 103, 10.1 percent), and high school or below (n = 23, 2.3 percent).

Of the original 1,019 participants, 514 (50.4 percent) completed the second wave of the survey. This longitudinal sample consisted of 116 males (22.6 percent) and 398 females (77.4 percent), with ages ranging from 16 to 43 years (M = 23.59, SD = 3.77). The educational distribution—comprising high school or below (0.4 percent), junior college (7.8 percent), undergraduate (72.8 percent), and graduate degrees or higher (19.1 percent)—demonstrates that the follow-up sample remained demographically consistent with the baseline population. Although a follow-up rate of 50 percent is generally considered adequate, with 60 percent regarded as good, ²⁹ achieving such retention levels is often challenging in longitudinal cohort studies. ³⁰ For example, the Malaysian Cohort study reported a follow-up rate of only 42.7 percent, illustrating that substantial attrition is not uncommon in long-term follow-up research. ³⁰

To address missing data due to attrition, we used missForest, a nonparametric imputation method based on the Random Forest algorithm. This approach can accommodate nonlinear relationships and interactions among variables and has been shown to perform competitively with other modern imputation techniques. In the present study, missForest was used to retain the full baseline sample and reduce information loss associated with listwise deletion. However, given the substantial attrition from Time 1 to Time 2, imputation should be viewed as a sensitivity-enhancing strategy rather than a complete remedy for missing data. Comparative evaluations have demonstrated that missForest performs competitively with other modern imputation techniques. ³¹ Meanwhile, the validity of imputed values depends on the extent to which missingness can be explained by observed variables, and the possibility of attrition-related bias cannot be ruled out.

Data analysis

Data analysis was conducted in R (version 4.4.2). We performed confirmatory factor analysis (CFA) to assess model fit based on several indices recommended by Hair et al. ³² : Comparative Fit Index (CFI; ≥0.90), Root Mean Square Error of Approximation (RMSEA; ≤0.07), and Standardized Root Mean Square Residual (SRMR; <0.08). Internal reliability was examined via Cronbach’s alpha. Hierarchical linear regression was then employed to test the research hypotheses. To address potential multicollinearity, we monitored the variance inflation factor (VIF), with values below 10 indicating no serious multicollinearity issues. ³²

Measures

Descriptive statistics

Table 1 presents the correlations among the study variables. Machiavellianism was significantly and positively correlated with narcissism (r = 0.324, p < 0.001) and sadism (r = 0.086, p < 0.01), but showed no significant association with GAICB at Time 2 (r = 0.053, p = 0.088). Narcissism was significantly related to psychopathy (r = 0.136, p < 0.001), sadism (r = 0.128, p < 0.001), GAICB at Time 1 (r = 0.145, p < 0.001), and GAICB at Time 2 (r = 0.103, p < 0.01). Psychopathy showed strong positive correlations with sadism (r = 0.558, p < 0.001) and was significantly associated with GAICB at both Time 1 (r = 0.332, p < 0.001) and Time 2 (r = 0.326, p < 0.001). Similarly, sadism was significantly correlated with GAICB at Time 1 (r = 0.424, p < 0.001) and Time 2 (r = 0.381, p < 0.001). Finally, GAICB at Time 1 was strongly associated with GAICB at Time 2 (r = 0.728, p < 0.001), indicating good test–retest reliability of the measure over the 6-month interval.

OPEN IN VIEWER

Table 1.

Descriptive Statistics

Variable	1	2	3	4	5	6
1. Machiavellianism	—
2. Narcissism	0.324**	—
3. Psychopathy	−0.005	0.136**	—
4. Sadism	0.086**	0.128**	0.558**	—
5. GAICB (Time 1)	0.071*	0.145**	0.332**	0.424**	-
6. GAICB (Time 2)	0.053	0.103**	0.326**	0.381**	0.728**	—
Mean	15.186	13.171	9.123	8.885	2.085	1.514
SD	2.644	3.318	3.784	3.394	4.121	2.903

*p < 0.05.

**p < 0.01.

GAICB, GAI-involved cyberbullying.

Hierarchical linear regression

As shown in Table 2, the hierarchical linear regression analyses were conducted to examine the predictors of GAICB at Time 2.

OPEN IN VIEWER

Table 2.

Hierarchical Linear Regression

Variable	GAICB (Time 2)
	Model 1				Model 2
	B	β	t	p	B	β	t	p
Gendera	−1.272	−0.195	−6.689	<0.001	−0.311	−0.048	−2.094	0.037
Age	−0.002	−0.003	−0.105	0.916	−0.014	−0.019	−0.875	0.382
Machiavellianism	0.014	0.013	0.425	0.671	0.002	0.002	0.074	0.941
Narcissism	0.005	0.006	0.195	0.845	−0.017	−0.019	−0.827	0.408
Psychopathy	0.126	0.164	4.813	<0.001	0.055	0.072	2.758	0.006
Sadism	0.224	0.262	7.666	<0.001	0.043	0.051	1.868	0.062
GAICB (Time 1)					0.473	0.672	27.363	<0.001
F	42.698 (p < 0.001)				170.498 (p < 0.001)
R2	0.202				0.541

^aBoys = 0, girls = 1.

Model 1 included demographic variables (gender and age) as well as the four Dark Tetrad. The model was statistically significant (F = 42.698, p < 0.001), accounting for 20.2 percent of the variance in GAICB (R² = 0.202, adjusted R² = 0.197). VIF values ranged from 1.057 to 1.618, all well below the conventional threshold of 10, indicating no concerns regarding multicollinearity.

Within Model 1, gender emerged as a significant predictor (B = −1.272, β = −0.195, p < 0.001), suggesting that males were more likely to engage in GAICB than females. Regarding the Dark Tetrad, psychopathy (B = 0.126, β = 0.164, p < 0.001) and sadism (B = 0.224, β = 0.262, p < 0.001) were both significant positive predictors of GAICB. In contrast, Machiavellianism (B = 0.014, β = 0.013, p = 0.671), narcissism (B = 0.005, β = 0.006, p = 0.845), and age (B = −0.002, β = −0.003, p = 0.916) were not significant predictors.

Model 2 further incorporated GAICB at Time 1 as a control variable. The inclusion of this autoregressive predictor substantially increased the explanatory power of the model (F = 170.498, p < 0.001), with R² rising to 0.541 (adjusted R² = 0.537). Consistent with expectations, GAICB at Time 1 strongly predicted GAICB at Time 2 (B = 0.473, β = 0.672, p < 0.001). After controlling for prior GAICB, the effects of gender (B = −0.311, β = −0.048, p = 0.037) and psychopathy (B = 0.055, β = 0.072, p = 0.006) remained statistically significant, albeit weaker than in Model 1. Sadism (B = 0.043, β = 0.051, p = 0.062) no longer reached conventional levels of statistical significance, although it approached the 0.05 threshold in the presence of the autoregressive control. Machiavellianism, narcissism, and age continued to show no significant effects. VIF values for Model 2 ranged from 1.057 to 1.618, again confirming that multicollinearity was not a concern.

Together, these findings highlight that while psychopathy and sadism were initially strong predictors of GAICB at Time 2, the stability of GAICB over time accounted for the majority of the variance, attenuating the predictive power of sadism but leaving psychopathy and gender as significant predictors.

Dark Tetrad and GAICB

Our findings reveal that not all Dark Tetrad are equally implicated in GAICB. Specifically, psychopathy consistently emerged as a significant predictor across models, whereas sadism showed a strong effect initially but lost its significance once prior levels of GAICB were taken into account. These findings contribute to the literature by showing that established predictors of cyberbullying do not operate uniformly in the emerging context of GAI-facilitated aggression.

Individuals high in psychopathy—marked by impulsivity, emotional detachment, and lack of empathy ¹¹ —were especially relevant in GAICB because GAI lowers the behavioral friction involved in producing harmful content. Compared with traditional cyberbullying, which often requires the perpetrator to manually craft messages or repeatedly interact with victims, GAI can rapidly generate insulting comments, defamatory rumors, humiliating images, or other harmful content. This accessibility may make it easier for individuals high in psychopathy to convert aggressive impulses into online behavior. Moreover, GAI-mediated content creation may increase psychological distance from the victim, further reducing the empathic inhibition that might otherwise discourage harmful behavior.

By contrast, sadism, characterized by deriving pleasure from others’ suffering, ²⁰ was a strong predictor of GAICB in the baseline model but became only marginally nonsignificant (p = 0.062) after accounting for prior GAICB. One interpretation is that sadism may be more closely associated with initial involvement in GAICB than with changes or maintenance over time. This possibility is theoretically plausible because GAI tools can intensify the victim’s humiliation or distress through realistic synthetic media, impersonation, or scalable defamatory content. Such affordances may be attractive to individuals who derive pleasure from others’ suffering. However, once prior GAICB was included in the model, sadism no longer explained unique variance beyond behavioral stability.

The nonsignificant effects of Machiavellianism and narcissism should also be interpreted in relation to the affordances of GAI. Theoretically, Machiavellianism may be relevant to strategic forms of GAI misuse, such as impersonation, coordinated rumor-spreading, or reputation manipulation. Narcissism may be relevant when GAI-assisted cyberbullying is used to retaliate against perceived ego threats or restore threatened self-image. However, the present measure captured overall GAICB rather than distinguishing among specific behavioral subtypes. It is therefore possible that Machiavellianism and narcissism are more strongly related to particular forms of GAICB than to general involvement. Whether different Dark Tetrad traits predict different modes of GAI-facilitated aggression, such as deepfake humiliation, automated harassment, and impersonation, remains to be seen.

Demographic variables and GAICB

Among demographic variables, gender consistently emerged as a significant predictor across models, with males more likely than females to engage in GAICB. Notably, the effect size decreased once prior GAICB was controlled for; however, gender remained significant in the longitudinal model, underscoring its role as a robust demographic correlate of GAICB. This finding aligns with longstanding research on traditional bullying, which shows that males typically display higher levels of direct aggression. ²¹ Although cyberbullying is sometimes conceptualized as a form of indirect aggression—leading to expectations of greater female involvement^6,23—our results did not support this notion. A more nuanced interpretation by Smith et al. ²⁴ suggests that although cyberbullying lacks face-to-face confrontation, its technological aspects may particularly appeal to males. Given the technical skills often required to manipulate GAI tools, this explanation may partly account for the observed gender disparity. However, this result should be interpreted cautiously because the present sample was female-skewed at both waves. Therefore, although gender remained statistically significant after controlling for baseline GAICB, the magnitude and stability of this effect should be tested in future studies with more gender-balanced samples. The present finding should be understood as preliminary evidence that gender may be associated with GAICB involvement, rather than as a definitive estimate of gender differences in the broader population.

By contrast, age was not a significant predictor in either model. However, this null effect should be interpreted with caution. It does not necessarily imply that age is irrelevant; rather, it may reflect the characteristics of our sample, which ranged from 16 to 52 years and thus potentially masked developmental trends. Bullying behaviors are known to peak during early adolescence, a period marked by heightened social competition and identity exploration. ²⁷ Aggregating across such a broad age range may have obscured these developmental patterns. Future research should therefore consider age-specific sampling strategies to better capture developmental nuances in GAICB involvement.