Social bots weaken activist cohesion

Linda Li; Orsolya Vasarhelyi; Balazs Vedres

Source-linked AI summary

Social bots weaken activist cohesion

Linda Li, Orsolya Vasarhelyi, Balazs Vedres

arXiv:2609.00197v1Aug 31, 2026cs.SIcs.CY

TL;DR

Research has focused on bots as content producers, leaving less evidence about whether they alter the human ties that sustain activist movements. Using retweet networks from the 2020 BLM protests, this paper finds that greater peak bot exposure predicts steeper post-peak declines in human cohesion at both individual and community scales.

Problem
Research has focused on bots as content producers or amplifiers, leaving less evidence about their association with subsequent changes in human-network cohesion.
Method
The study tracks core participants’ retweet networks before, during, and after the 2020 BLM protest peak, measuring cohesion through ego-network closure and community edge density.
Results
Greater bot exposure during the peak predicts steeper subsequent declines in human cohesion through both individual triadic closure and community connective density, especially among core supporters.
Takeaways & Limitations
The findings suggest that bots may weaken activism by eroding the human-to-human ties through which participants coordinate and remain connected between mobilization waves.
Takeaways & Limitations
Retrospective data collection, including deleted posts and suspended accounts, may reduce dataset completeness and underestimate bot-generated messages.

Abstract

from arXiv · show

Social bots now make up a substantial share of online political communication, where they are studied mainly as producers of misinformation and amplified content. Far less is known about whether their presence reshapes the human relationships that hold movements together. We ask whether exposure to bots during a protest peak is followed by the erosion of cohesion in human networks. Tracking retweet networks of core participants in the 2020 Black Lives Matter (BLM) protests before, during, and after the peak, we measure change in cohesion at two scales: triadic closure in individual ego networks and edge density within detected communities. Greater bot exposure during the peak predicts steeper subsequent declines in human cohesion at both scales, and the loss concentrates among supporters of the movement. Bots may weaken activism less by changing what people believe than by dissolving the ties through which collective action is sustained.

Introduction

The study asks whether greater bot exposure during the 7–8 June 2020 BLM protest peak predicts subsequent erosion of human network cohesion. It examines this relationship at both individual ego-network and community scales, including differences by prior movement orientation.

Measures: The analysis measures cohesion through triadic closure in human-only ego networks and edge density within detected human communities.Ego-network cohesion tracks local clustering and alter–alter ties, while community cohesion tracks human–human density changes.
Research question: Greater bot exposure during the protest peak predicts steeper subsequent declines in human cohesion.The question is examined in longitudinal networks before, during, and after the peak.
Findings: Networks with heavier bot exposure show declines at both scales: individual activists’ neighborhood closure and whole communities’ connective density.These measures capture the erosion of local triangles and within-community human ties after the protest peak.
Heterogeneity: The cohesion loss is strongest among supporters of the movement’s network core.The study tests whether prior support for Black Lives Matter moderates the association between bot exposure and cohesion change.
Interpretation: The findings suggest bots may weaken activism by corroding relational ties rather than primarily changing what people believe.Human cohesion is presented as the structural foundation supporting recognition, coordination, and shared identity.

Findings

The protest core formed a persistent but fragmented human–bot network that thinned after the peak, while human cohesion remained sparse. Greater bot exposure predicted subsequent cohesion declines at both ego-network and community scales, concentrated among movement supporters.

Network composition: At the peak, the protest core contained 92,078 users and 372,932 edges, with 55% of users displaying automated-like behavior.The core included 27,221 human and 50,982 bot users; the automated-like share ranged from 42% to 55% across bot-probability thresholds of .65–.75.
Network composition: Bot–bot and bot–human exchanges comprised roughly 88% of peak interactions, while human–human exchanges comprised approximately 12%.After the peak, total interaction volume declined by more than an order of magnitude, while the network persisted and shifted further toward automation.
Baseline cohesion: Human-only cohesion was sparse: human–human density was 5.6 × 10−5 at peak and 9.0 × 10−5 after, while clustering coefficients were 0.029 and 0.030.At the ego-network level, mean local clustering was 0.055 at peak and 0.112 after among retained egos, with median values of 0 and 0.071.
Bot exposure and cohesion: Higher bot exposure was associated with larger decreases in human-only local clustering in ego networks and human–human edge density in communities.Ego exposure was measured as alter-bot edges, whereas community exposure was bot-human edge density; both fitted associations were negative.
Support as moderator: The negative association between bot exposure and cohesion loss was stronger among users and communities with greater prior support for the movement.Interaction coefficients were negative at both scales, and predicted declines were pronounced among supporters but flatter or near-zero among detractors and lower-support communities.
Robustness: The direction and relative magnitude of estimated effects remained stable across bot-probability thresholds and after accounting for baseline structure and user characteristics.This supports a conditional negative association between bot exposure and cohesion loss rather than one driven by a specific classification cutoff.

Discussions

The discussion argues that social bots affect political communication not only informationally but also through overlooked damage to human relational infrastructure. It calls for evaluating and mitigating bot-driven erosion of cohesion, especially among movement supporters.

Conceptual contribution: Social bots’ effects on human-to-human relational infrastructure have been largely overlooked despite extensive research on misinformation, sentiment manipulation, and divisive amplification.The paper identifies relational consequences as a critical structural dimension of automated political activity.
Implications for movements: Bot-driven cohesion erosion may weaken social-movement resilience by degrading local closure and making human communication less socially rewarding.The proposed mechanism is structural displacement through noise, rather than persuasion into automated opinions.
Implications for movements: The decay is structurally asymmetric, concentrating mostly among users and communities supportive of the movement.This asymmetry indicates that bot exposure does not affect movement-related networks uniformly.
Policy and research implications: Researchers, platform operators, and policymakers should evaluate automated accounts as threats to network structure in democratic spaces, not only as sources of harmful information.The discussion presents this as a necessary conceptual shift in threat assessment.
Policy and research implications: Platforms should track bot-human relational density and degradation of human ties alongside conventional monitoring of disinformation diffusion.The discussion frames relational monitoring as necessary to safeguard online civic life.

Data and methods … Opinion categorisation using ChatGPT

The study analyzes BLM protest discourse and retweet networks around the 7–8 June 2020 peak, identifying bots, restricting analysis to a network core, and measuring users’ movement stances. It combines computational classification with network construction and robustness checks to examine human cohesion.

Data: The peak window was defined as 7–8 June 2020, following more than 700 protest events across over 600 U.S. localities on 6 June.The study focuses on discourse following George Floyd’s death on 25 May 2020.
Data: Researchers collected all tweets in the peak window containing BLM-related keywords and retrieved user metadata through the Twitter Academic Product track API.Keywords included ‘BLM’, ‘George Floyd’, related terms, and common spelling variants.
Bot detection: Social bots were identified using Botometer and a self-trained algorithm, with the primary Botometer threshold set above 0.65.Sensitivity analyses re-estimated models at bot-probability thresholds of 0.65, 0.70, and 0.75.
0.1 Network construction and filtering: The researchers constructed a directed, unweighted retweet network in which users were nodes and retweets or quotes were edges, collapsing repeated interactions into one connection.This aggregation targets connections between human users rather than the intensity of individual ties.
0.1 Network construction and filtering: The peak network was restricted to its core using a k-core value of at least two, yielding 92,078 users and 372,932 edges.Cohesion measures were computed on undirected projections of the networks.
0.1 Network construction and filtering: Pre- and post-peak snapshots used users’ full-timeline communication rather than BLM-filtered edges to test whether the mobilised core persisted socially.The peak network remained restricted to BLM-related communication.
Opinion categorisation using ChatGPT: User stance was scored continuously from -1 for strong opposition to 1 for strong support, with 0 denoting neutrality or unrelated content, then dichotomised at the sample median.Stance was derived from each user’s tweets in the relevant snapshot.
Opinion categorisation using ChatGPT: GPT-3.5 classifications were refined through two rounds of human-in-the-loop review and validated against manual coding, with support scores correlating at r = 0.88 versus r = 0.71 between coders.Validation used Extinction Rebellion protest data with the same movement-agnostic prompt instrument.

Community detection · Statistical analysis design

The study treats protest-peak sub-communities as the main units for assessing collective cohesion, complementing this with an individual ego-network analysis. It estimates whether peak bot interaction predicts peak-to-post-peak declines in human cohesion using distinct community- and individual-level designs.

Community detection: Community-level analysis is the primary focus, with protest-peak sub-communities serving as the fundamental units for measuring collective structural change.Individual-level analysis provides a complementary perspective on cohesion.
Community detection: CPM detects overlapping, densely connected sub-graphs by linking adjacent k-cliques, isolating cohesive groups within a larger sparse network.Its local clique-based construction also avoids repeated global optimisation and is computationally efficient at network scale.
Community detection: 4,042 sub-communities containing both bot and human nodes were identified with k=4; excluding groups with fewer than three human nodes left 642 communities.The exclusion guaranteed at least one human triad in every retained community.
Community network analysis: Community cohesion is the peak-to-post-peak change in human–human edge density, computed on undirected projections with the same node set throughout.The pre-peak snapshot supplied baseline controls rather than a modelled endpoint.
Community network analysis: The community model uses OLS to relate peak bot–human density to changes in human–human density while controlling for pre-peak human–human density, human-node count, and aggregated individual characteristics.Human–human density is size-normalised, and bot–human connections are normalised by potential bot–human connections.
Individual network analysis: The individual model analyzes users with k-core values of at least two who were active across all three periods, excluding those isolated in the peak human-only k-core-2 network.This sampling strategy enables assessment of changes in individual connectivity, whereas the community model can include phase-specific users.
Individual network analysis: Individual bot interaction is the peak-period count of alter–bot edges, and the outcome is the peak-to-post-peak change in each ego network’s human-only local clustering coefficient.OLS controls for pre-peak ego degree and clustering plus aggregated user characteristics.
Opinion dynamics: Pre-peak movement support is a binary indicator above the sample median, interacted with bot exposure to test whether prior opinion moderates cohesion change.The moderation test is motivated by evidence that prior opinions shape susceptibility to bot influence and that bots may target particular viewpoints.

Robustness

The conclusions remain stable across bot-classification thresholds and alternative network-edge definitions. After correcting individual-level network dependence, the bot-exposure associations persist, while community-level residuals show no significant autocorrelation.

Bot-classification threshold: Associations hold across bot-probability thresholds of 0.65–0.75, indicating robustness to the bot/human classification cut-off.Thresholds below 0.65 were not tested because false positives could spuriously inflate estimated bot exposure, whereas false negatives are more conservative.
Edge definition: Reconstructing pre- and post-peak snapshots with BLM-related edges leaves every predictor and outcome estimate substantively unchanged.This matches the peak network’s edge definition and addresses asymmetry between filtered and full-timeline networks.
Network autocorrelation: At the community level, residuals show no significant autocorrelation in either model (all 𝑝≥0.67), supporting retention of OLS.At the individual level, overlapping ego networks produce dependent residuals under ego-neighbourhood overlap weights.
Network autocorrelation: After network-error correction, peak bot exposure remains negatively associated with change in human clustering (𝛽= −0.040, SE = 0.019).The opinion-interaction model retains a positive main term (𝛽= 0.125, SE = 0.045) and a negative bot-exposure × pro-BLM interaction (𝛽= −0.190, SE = 0.050) under both weight matrices.

Supplementary materials

The supplementary materials document the study’s data collection, bot identification, opinion categorization using ChatGPT, and network construction and filtering. They also include Tables S1 to S12, references, a movie, and a dataset.

The supplementary materials describe data collection, bot identification, opinion categorization using ChatGPT, and network constructing and filtering.
The materials include Tables S1 to S12.
The supplementary materials also list references (7-66), Movie S1, and Data S1, and identify Linda Li, Orsoloya Vasarhelyi, and Balazs Vedres.

Data collection

The study assembled a relatively comprehensive archive of BLM protest-related communication on X during the June 2020 activity peak. It combined keyword-based tweet collection with user metadata, while noting that retrospective collection and API discontinuation shaped the dataset’s limitations and value.

Collection period: The collection focused on June 7–8, 2020, identified as the peak of protest-related activity after George Floyd’s death.George Floyd died on 25 May 2020.
Collection method: Tweets were collected through the Twitter Academic Product track API using BLM- and George Floyd-related keywords, including spelling variations.The complete keyword list appears in Table S1 and includes “BlackLivesMatter,” “black lives matter,” “BLM,” “George Floyd,” and “georgefloyd.”
Dataset: The final dataset contains over 2,500,000 tweets from more than 1,000,000 users collected between June 6, 2020, 00:00 and June 8, 2020, 23:59.Associated metadata, including user IDs, was collected, and additional user information was retrieved through the Twitter user lookup API.
Limitations: Because the API provided a complete archive, the sample was considered relatively comprehensive, but retrospective collection may have been affected by moderation and deleted accounts or posts.The passage specifically notes Twitter’s spam-message censorship and possible deletion of accounts or posts between the protests and data collection.
Historical value: The Academic Track API’s discontinuation makes the dataset a unique and comprehensive historical snapshot of protest-related discourse.Its value is linked to preserving a complete view of social-media activity that current tools can no longer access.

Bot Identification

The study identifies protest-peak users with a prior methodology combining Botometer probabilities, threshold-based labels, and supervised machine-learning models. Users active at the peak were classified as human, bot, or unknown.

Bot Identification: Botometer calculates bot probabilities from user-profile and activity features, using thresholds 0.65 and 0.5 to classify accounts as automated, human, or unknown.The approach was developed in prior work.
Bot Identification: The methodology also trains Random Forest, Support Vector Machine, Logistic Regression, XGBoost, and Deep Learning models on derived features.Model performance is described in the earlier work.
Bot Identification: All users active at the protest peak, defined as k-core ≥2, were classified as human, bot, or unknown.The passage refers to the earlier work for a detailed methodological description and model-performance assessment.

Opinion categorization using ChatGPT

The study used ChatGPT 3.5 to classify users’ opinions toward the BLM protests on a continuous scale from -1 to 1. Validation against manual coding found that GPT’s scores closely matched human consensus, supporting its use as an automated measure of user support.

Method: ChatGPT 3.5 was used to categorize users’ opinions on the protests.The authors note that this LLM-based approach has been used in prior textual classification research.
Method: Opinions were scored continuously from -1 for extremely negative to 1 for extremely positive, with 0 assigned to neutral, irrelevant, or mixed views.The prompt instructed ChatGPT to interpret each user’s text using political and racial-justice context.
Validation: A random sample of N=300 users was independently coded by two human coders using the same combined timelines provided to GPT.The validation used data from a similar online activism event and the same prompt instrument as the BLM analysis.
Validation: C=0.88 was the correlation between GPT’s scores and the coders’ averaged judgement, compared with C=0.71 between the two coders.The authors therefore treated GPT’s support classification as a valid automated measure of user support level.

Network construction and filtering

The study used a directed, bi-directional, unweighted retweet network, removed isolated nodes, and retained a k-core-filtered community of 92,078 users and 372,932 edges. It compared communication snapshots before and after the protest peak to assess changes in human network cohesiveness and bot interactions.

Network representation: Retweet and quote interactions formed a bi-directional, unweighted network in which users were nodes and interactions were single edges.Retweets had no added commentary, whereas quotes included commentary; self-loops were retained and isolated nodes were removed.
Network filtering: 92,078 users and 372,932 edges remained after retaining nodes with a k-core value greater than two.The k-core filter selected the maximal subgraph in which each node connects to at least k other nodes.
Temporal snapshots: Two additional directed communication snapshots covered May 1–2, 2020, before the peak, and September 1–2, 2020, after the peak.The peak network was sampled for analytical purposes and included only protest-relevant edges, so the main comparison focused on the before and after periods.

Model tables

The supplementary model tables report individual- and community-level OLS specifications across three bot thresholds, alongside filtered-after-period and network-autocorrelation robustness analyses. The tables also document weighting, overlap, island, permutation, and significance conventions.

Individual-level OLS models are reported at bot thresholds of 0.65, 0.7, and 0.75.
Community-level OLS models are reported at bot thresholds of 0.65, 0.7, and 0.75.
Robustness tables recompute after-period outcomes using BLM-related retweets while holding peak exposure unchanged, for both individual and community models.Columns (3)/(5) reuse the main Model 3/5 specifications; “Full after” uses the unrestricted after timeline.
Ego-level network-autocorrelation robustness is assessed at bot threshold 0.65.
Network-error models use row-standardized weights; ego overlap uses Jaccard similarity after retaining each ego’s top 20 neighbours before symmetrization.
CPM weights connect observations in the same peak community, zero-neighbour observations remain islands, and Moran tests use 999 permutations.Significance markers are * p<0.05, ** p<0.01, and *** for the remaining stated significance threshold.