Stig Hebbelstrup Rye Rasmussen

Hostility and toxicity permeate political debates on social media, but less is known about which users engage in hostile political expression online. While many early theories of political behavior highlighted the positive democratic role of political engagement, recent research suggests that such engagement may also be associated with more polarized and antagonistic behavior. Combining two years of survey and behavioral data from X/Twitter, we examine associations between online political hostility and indicators of political engagement and its antecedents among American users. Online political hostility is positively associated with these indicators, with the strongest associations for political interest, political knowledge, and online participation. These associations are more pronounced for some indicators, especially online and offline participation, among users with more polarized affect directed towards US political leaders. Finally, the results show that engaged but hostile social media users do not simply cater to fringe audiences; they post more frequently and have more friends and followers. The greater engagement of hostile social media users may thus provide them with a higher reacher than other users.

Large language models (LLMs) are increasingly used to evaluate people in domains such as hiring, education, and housing. A central concern is whether these systems reproduce human intergroup bias, reduce it, or express it in new forms. Here, we compare human and LLM bias using matched stereotype, decision, and text-evaluation tasks. We fielded a nationally representative Danish survey (N = 1,000), administered parallel instruments to a five-model LLM panel, replicated published audit studies in hiring, housing, and school access, and conducted a naturalistic text-evaluation experiment using real Danish social media posts from demographically identified authors whose identities were removed before evaluation. We demonstrate that LLMs reproduce the broad human stereotype hierarchy: people with a non-Western background are consistently seen as least warm and competent. We also find that this bias is often hidden in standard audit designs. When models make decisions from explicit identity cues, such as names, they generally do not reproduce human discrimination gradients. This is often because they accept or reject nearly all cases rather than make graded individualized decisions. Bias reappears when identity is implicit. Unlabeled texts written by non-Western authors receive lower ratings on warmth, competence, and employability and more negative trait descriptions. Summarizing the text before evaluation reduces these gaps, suggesting that some bias is carried by surface features of language. These results show that aligned LLMs are not simply unbiased versions of humans: They can suppress visible discrimination while retaining stereotype structures that emerge in associative judgments and evaluations of natural language.

Reducing hostility in social media interactions is a key public concern. Most extant research emphasizes how online contextual factors breed hostility. Here, we take a different perspective and focus on the offline roots of hostility, that is, offline experiences and stable individual-level dispositions. Using a unique dataset of Danish Twitter users (N = 4,931), we merge data from administrative government registries with a behavioral measure of online hostility. We demonstrate that individuals with more aggressive dispositions (as proxied by having many more criminal verdicts) are more hostile in social media conversations. We also find evidence that features of childhood environments predict online hostility. Time spent in foster care is a strong correlate, while other indicators of childhood instability (e.g., the number of moves and divorced parents) are not. Furthermore, people from more resourceful childhood environments—those with better grades in primary school and higher parental socioeconomic status—are more hostile on average, as such people are more politically engaged. These results offer an important reminder that much online hostility is rooted in offline experiences and stable dispositions. They also provide anuanced view of the core group of online aggressors. While these individuals display general antisocial personality tendencies by having many more criminal verdicts, they also come from resourceful backgrounds more often than not.

Hate is widespread online, hits everyone, and carries negative consequences. Crowd moderation—user-assisted moderation through, e. g., reporting or counter-speech—is heralded as a potential remedy. We explore this potential by linking insights on online bystander interventions to the analogy of crowd moderation as a (lost) public good. We argue that the distribution of costs and benefits of engaging in crowd moderation forecasts a collective action problem. If the individual crowd member has limited incentive to react when witnessing hate, crowd moderation is unlikely to manifest. We explore this argument empirically, investigating several preregistered hypotheses about the distribution of individual-level costs and benefits of response options to online hate using a large, nationally representative survey of Danish social media users (N = 24,996). In line with expectations, we find that bystander reactions, especially costly reactions, are rare. Furthermore, we find a positive correlation between exposure to online hate and withdrawal motivations, and a negative (n-shaped) correlation with bystander reactions.

Reducing hostility in social media interactions is a key public concern. Most extant research emphasizes how online contextual factors breed hostility. Here, we take a different perspective and focus on the offline roots of hostility, i.e., stable individual-level dispositions. Using a unique dataset of Danish Twitter users (N = 4,931), we merge data from administrative government registries with a behavioral measure of online hostility. We demonstrate that individuals with more aggressive dispositions (as proxied by the number of criminal verdicts) are more hostile in social media conversations. We also find evidence that features of childhood environments predict online hostility. Time spent in foster care is a strong predictor, while other indicators of childhood instability (e.g., the number of moves and divorced parents) are not. Furthermore, people from more resourceful childhood environment -- those with better grades in primary school and higher parental socioeconomic status -- are more hostile on average, potentially because such people are more politically engaged. These results offer an important reminder that much online hostility is rooted in offline experiences and stable dispositions. Accordingly, efforts to reduce online hostility may need to broaden the focus to remedy the effects of problematic offline environments.

Hostile interactions permeate political debates on social media, but what is driving the long-term developments in online political hostility? Prior research focuses on individual-level factors such as the dispositions of users or network-level factors such as echo chambers. Moving beyond these accounts, we develop and test an event-oriented explanation and demonstrate that over the course of the 2020 election year in the United States, all major shifts in political hostility on the social media platform Twitter were driven by external offline events. Importantly, these events were magnified by Twitter users within the most politically hostile and most ideologically homogeneous networks. Further contributing to the individual and network-oriented accounts, we show that divisive offline events mobilized individual users not already disposed for hostility and may have helped facilitate the formation of echo chambers. The dynamics of online interactions—including their level of hostility—seem crucially dependent on developments in the offline world.

We live in a world of text. Yet the sheer magnitude of social media data, coupled with a need to measure complex psychological constructs, has made this important source of data difficult to use. Researchers often engage in costly hand coding of thousands of texts using supervised techniques or rely on unsupervised techniques where the measurement of predefined constructs is difficult. We propose a novel approach that we call ‘super-unsupervised’ learning and demonstrate its usefulness by measuring the psychologically complex construct of online political hostility based on a large corpus of tweets. This approach accomplishes the feat by combining the best features of supervised and unsupervised learning techniques: measurements of complex psychological constructs without a single labelled data source. We first outline the approach before conducting a diverse series of tests that include: (i) face validity, (ii) convergent and discriminant validity, (iii) criterion validity, (iv) external validity, and (v) ecological validity.

By most accounts, an important prerequisite for a well-functioning democracy is engaged citizens. A very prominent explanation of variation in political engagement suggests that parental transmission through socialization accounts for individual-level differences in political engagement. In this paper, we show, using a large Danish twin survey ( N = 2,071), that classic formulations of parental transmission theory can be supplemented by findings from the biopolitics literature, allowing us to disentangle when heritable factors are important and when socialization factors are important predictors of political engagement. We show that as the level of family politicization and consistency increases, the influence of genes decreases. We take this to imply that family socialization can compensate for (genetic) individual differences and foster increased political engagement. By only focusing on the “causal” effect of education, we are missing the forest for the trees.

Political hostility and ideological polarization permeate political debates on social media, but what is driving long-term developments in ideological polarization and political hostility? Prominent theories of social media argue that social media networks are best conceptualized as echo chambers, i.e., systems insulated from external input. Contrary to this view, we demonstrate that over the course of the 2020 election year in the US, all major ebbs and flows in political hostility on the social media platform Twitter were driven by external offline events. Importantly, these events were magnified for Twitter users within the most politically hostile and least ideologically diverse networks. Accordingly, social media networks are often best conceptualized as semi-open resonance chambers that are not insulated from external events but magnify them. We demonstrate that resonance chambers facilitate event-driven hostility via feedback loops that makes Twitter both more hostile and more ideologically polarized.