Negativity bias in the spread of voter fraud conspiracy theory tweets during the 2020 US election

The study examines why and how conspiracy theory messages—specifically voter-fraud claims during the 2020 US election—spread on Twitter. Conspiracy theories, defined as explanations alleging secret plots by powerful actors that persist without reliable evidence, have notable societal impacts, including reduced engagement with mainstream politics, increased support for political violence, and heightened prejudice. The 2020 US election saw extensive dissemination of voter-fraud allegations on Twitter, amplified by prominent figures and hashtags such as #stopthesteal, despite official findings of no widespread fraud. The research question is whether the observed retweet frequencies among proponents are consistent with frequency bias (conformity/anticonformity) and/or content bias (intrinsic message features such as emotional valence). Using a cultural evolution framework and generative inference, the study aims to identify which transmission biases best explain the population-level retweet distribution within a large proponent community.

Prior work on social media diffusion highlights the role of message content, including negativity, emotionality, and out-group derogation, in driving engagement and spread. Cultural evolution theory posits several social learning biases—frequency, content, and demonstrator biases—that can shape population-level distributions. Earlier generative-inference work on Twitter suggested retweet patterns resemble random copying more than conformity but omitted follower influence and content bias parameters due to computational constraints. Studies on negativity bias show negative content often gains an advantage and spreads more in digital environments, particularly around political events, though some research finds positivity can characterize viral false rumors. Emotional intensity can also influence diffusion independent of valence. These mixed findings motivate modeling that jointly estimates multiple potential drivers—including content, frequency, and follower effects—while accounting for platform dynamics.

Preregistered analyses used the VoterFraud2020 dataset (Abilov et al.), comprising tweets and retweets about voter-fraud claims collected via Twitter’s streaming API between October 23 and December 16, 2020. Analyses focused on cluster #2 (the English-speaking “proponent” community) to capture within-community spread; resulting data included 3,982,990 tweets from 341,676 users. An agent-based model (ABM) simulated a fully connected population of N=341,676 users over 216 six-hour timesteps. Each user was assigned real-data-derived attributes: follower count (T, scaled mean 1 SD 1), activity level (r), and probability of posting original tweets (μ≈0.45). At each timestep, a subset of users became active and either posted original content or retweeted existing tweets. Each tweet had an attractiveness M drawn from a truncated normal (mean 1, SD 1, lower bound 0). Retweet probability for tweet x was proportional to F_x^a T_i^d M_x^c e^{-g x}, where F is prior retweets; parameters estimated via ABC were: content bias (c), follower influence (d), frequency bias (a), and age dependency (g). Uniform priors: c∈[0,12], d∈[0,4], a∈[0,2], g∈[0,8]. Summary statistics from simulated and observed retweet distributions included: proportion of singletons, proportion of the most common tweet, and Hill numbers q=1 and q=2. Random forest ABC (1,000 trees per parameter) with logit-transformed parameters produced posterior estimates, supported by posterior predictive checks; robustness was assessed with rejection ABC and single-parameter simulations. Secondary analyses used GLMMs (Poisson) to examine content targets of bias while controlling for tweet length, presence of media, follower count (preferred over verification status), and user random intercepts. Sentiment was measured with VADER (compound score; proportions of negative/positive words). A separate GLMM compared quoted tweets vs. their targets to ensure treatment of quotes as original tweets did not bias content-related results.

• Generative inference (ABC) indicated a strong content bias: median c=4.612, 95% CI [3.479, 5.679]; a tweet 1 SD more attractive (M=2) was ~24× more likely to be retweeted.
• Follower influence (d) showed a weak effect with a right-skewed posterior converging toward zero: median d=0.362, 95% CI [0.007, 1.663].
• Frequency bias (a) had median 0.295, 95% CI [0.004, 0.916]. Under a neutrality interpretation of a=0 (decoupling from popularity), results suggest frequency information is largely irrelevant to future retweet probability; the alternative interpretation as extreme anticonformity was deemed unrealistic by authors.
• Age dependency (g) had an uninformative posterior spanning the prior, with high error; no conclusion drawn.
• GLMM (best-fitting model with user random intercept) showed significant fixed effects (IRR, 95% CI): • Tweet length: IRR 1.478 [1.477, 1.480] • Presence of media: IRR 1.515 [1.507, 1.523] • Follower count (scaled): IRR 1.505 [1.496, 1.515] • VADER compound score: IRR 0.959 [0.958, 0.960] (lower/more negative increases retweets)
• Alternative content specifications: 1 SD more negative words → +8.0% retweets; 1 SD more positive words → −6.4% retweets.
• Variance explained: fixed effects R²≈0.101; fixed + random intercept R²≈0.676.
• Quote tweets exhibited reduced negative valence relative to their targets, suggesting that treating quotes as originals likely made the content-bias estimate conservative.

The findings indicate that among proponents of voter-fraud conspiracy theories, retweet frequencies are best explained by a strong content bias favoring messages with more negative emotional valence. While follower count strongly predicts retweet counts in GLMMs, its effect appears weak in the ABM once temporal and dynamic factors (e.g., changing activity, aging, concurrent retweets) are accounted for. Frequency information is largely decoupled from future retweet likelihood, suggesting little to no conformity effect in this context. Reduced negativity in quote tweets compared to their targets suggests users may not amplify negativity when commenting, possibly reflecting norms within like-minded communities. Interpretation of follower influence and frequency bias is complicated by the opaque and engagement-focused Twitter recommendation algorithm; disentangling algorithmic amplification, network structure, and user bias remains difficult. Overall, results support a negativity-focused content bias driving spread within this community and align with broader evidence of negativity bias in social transmission, while highlighting domain and individual variability and equifinality in diffusion processes.

Using a cultural-evolutionary, generative-inference approach, the study shows that the spread of voter-fraud conspiracy tweets among proponents during the 2020 US election is consistent with a strong content bias favoring negatively valenced content. Compared against frequency and follower effects under dynamic conditions, content features are central to explaining the observed retweet distribution. The work demonstrates the utility of ABM+ABC for disentangling social learning biases on platforms and underscores the need for algorithmic transparency to build realistic null models. Future research should apply this methodology comparatively across topics and communities, integrate more granular network and behavioral data, assess algorithmic contributions via natural experiments, and incorporate individual-level factors (e.g., ideology, emotional state) to better understand when and why negativity biases shape online information spread.

• Algorithmic confounding: Twitter’s recommendation system, driven by author and engagement features, complicates separation of user-level biases from algorithmic amplification.
• Network data unavailability: Lack of follower network (partly due to account suspensions) limits disentangling demonstrator bias from network structure.
• Age dependency parameter (g) was not reliably estimated (uninformative posterior).
• Sentiment measurement limitations: VADER lexicon may not fully capture emergent phrases (e.g., “stop the steal”); dataset’s seed terms (e.g., “fraud”) are inherently negative, inducing slight skew.
• Scope and generalizability: Focus on a single proponent community and topic; not a comparative study across domains or general Twitter population.
• Data coverage: Source dataset estimated to include ~60% of relevant tweets; may omit some activity.
• Modeling choices: Fully connected ABM baseline; attractiveness assumed normally distributed; six-hour timestep resolution; treatment of quote tweets as originals (likely conservative for content-bias estimate).
• Inability to include many individual-level covariates (e.g., anxiety, demographics).