Social bots weaken activist cohesion

• Prior work focused mainly on bots as content producers or amplifiers, while this study examines whether their activity is associated with erosion of human ties that sustain activism.Cohesion may erode before connectivity collapses, as local density and triadic closure can decline while indirect paths remain.Source: Introduction · pp. 1–2
• The study constructs longitudinal communication networks before, during, and after the 7–8 June 2020 protest peak, when diverse actors were most likely to enter the same communication space.In the setting of the 2020 Black Lives Matter protests following George Floyd’s death, it examines how bot exposure during the peak is associated with subsequent changes in human network cohesion (36).Source: Introduction · pp. 2–3
• The analysis measures cohesion as human-only ego-network clustering and human-to-human edge density within communities detected at the protest peak.These measures capture related but distinct dimensions: local triadic closure and group-level cohesion.Source: Introduction · p. 3
• The study tests whether prior support for Black Lives Matter moderates the association between bot exposure and subsequent cohesion change.This distinguishes a general association across the network from one concentrated among users more closely aligned with the movement.Source: Introduction · p. 5
• Greater bot exposure during the protest peak predicts steeper subsequent declines in human cohesion at both network scales, with the strongest erosion among movement supporters.The study examines triadic closure in individual ego networks and human-to-human edge density within detected communities.Source: Introduction · pp. 3, 5

Hybrid communication network

• At the peak, the 92,078-user protest core was fragmented and bot-heavy, and its post-peak interactions shifted further toward automation as overall volume fell.Bots accounted for 55% of the core; interactions were roughly 88% bot–bot or bot–human at the peak. Afterward, bot–bot share rose from about 42% to 68%, human–human share fell from about 12% to 6%, and total volume declined by more than an order of magnitude. Human-only density and clustering were low, while bot–human density across communities ranged from 0 to 1.Source: Hybrid communication network · pp. 5, 7, Figure 2 · p. 6

Bot interaction weakens cohesion

• Higher peak bot exposure was associated with larger subsequent declines in human-only clustering in ego networks and human–human edge density in communities.At the ego level, the no-exposure intercept was positive; at the community level, it did not differ significantly from zero, indicating no average change at that baseline.Source: Figure 3 · p. 8, Bot interaction weakens cohesion · p. 9

Greater cohesion loss among supporters

• After controls, bot exposure predicted greater cohesion loss among movement supporters at both ego and community levels, consistently across bot-classification thresholds.Supporters’ predicted cohesion declined as exposure increased, while detractors’ associations were flatter or near zero. The negative interaction terms indicate that the association was stronger among supporters.Source: Figure 4 · p. 10, Greater cohesion loss among supporters · pp. 11–12

Discussions

• Greater peak bot exposure predicted subsequent declines in human triadic closure and community edge density that persisted months after the 2020 BLM protest peak.The declines appeared at both the local ego-network and meso community scales.Source: Discussions · p. 12
• Cohesion loss concentrated among movement supporters, whose network-core position may expose them to bot targeting and leave local solidarity vulnerable as public attention recedes.The proposed disruption operates through noise that makes human-to-human communication less socially rewarding, rather than persuading users to adopt automated opinions.Source: Discussions · p. 12
• Evaluations of automated accounts in democratic spaces should shift beyond informational effects to consider their structural impact on human relationships.The findings point to erosion of the human network that sustains collective action.Source: Discussions · p. 12
• Platforms should monitor bot-human relational density and human-tie degradation alongside disinformation, protecting the relational scaffolding needed for digital activism.This concern is especially acute as generative AI lowers deployment barriers for automated accounts that mimic human interaction with exceptional context-sensitivity.Source: Discussions · p. 13

Data

• The peak window was 7–8 June 2020, immediately after 6 June’s more than 700 protest events across over 600 U.S. localities.The window was chosen to capture online discourse at the mobilisation peak.Source: Data · p. 13
• The keyword-based collection yielded approximately 2.5 million tweets from more than one million users, gathered through Twitter’s Academic Product track API.Tweets were posted between 7 June 2020 00:00 and 8 June 2020 00:00; user IDs supported retrieval of account information.Source: Data · p. 13

Bot detection

• Users were labelled bots if Botometer scores exceeded 0.65 and the trained classifier identified them as bots; scores of 0.5–0.65 were unknown.The method was developed in prior research (3), and models were re-estimated at bot-probability thresholds of 0.65, 0.70 and 0.75.Source: Bot detection · pp. 13–14

0.1 Network construction and filtering

• The directed, unweighted retweet network treated users as nodes and retweets or quotes as edges, collapsed repeated ties, retained self-loops and excluded inactive users.Edges ran from the original poster to the retweeter or quoter, following established online and political communication network practice (49,50,51).Source: §0.1 · p. 14
• The peak network’s k-core of at least two contained 92,078 users and 372,932 edges; cohesion was measured on undirected projections.The analysis compared peak-to-post-peak cohesion for the same core users; the pre-peak snapshot served only to construct baseline controls. The k-core metric identifies a maximal subgraph in which each node has at least k connections (53).Source: §0.1 · p. 14
• Pre- and post-peak snapshots used users’ full-timeline communication to test whether the mobilized core persisted as a social structure, rather than whether protest-topic attention endured; BLM-edge-only robustness results were substantively unchanged.Source: §0.1 · p. 14

Opinion categorisation using ChatGPT

• Prior stance came from users’ pre-peak tweets, with scores from -1 (opposition) to 1 (support) dichotomised at the median of 0.6 into supporters and detractors.Neutral and opposing users were grouped as detractors, contrasting movement supporters with all other users.Source: Opinion categorisation using ChatGPT · p. 15
• GPT-3.5 classified stance after human-in-the-loop prompt refinement; on Extinction Rebellion data, its scores correlated with coder consensus at r = 0.88.Two coders refined prompts over two rounds beginning with 100 tweets; the validation correlation exceeded the coders’ agreement with each other (r = 0.71). The movement-agnostic instrument was expected to transfer to BLM.Source: Opinion categorisation using ChatGPT · p. 15

Community detection

• Peak sub-communities are the primary units for studying collective cohesion, with individual-level analysis serving as a complement.This captures group interaction patterns that cannot be reduced to individual behavior.Source: Community detection · p. 15
• Clique Percolation Method (CPM) with k=4 detects overlapping, densely connected sub-communities suited to local cohesion in a large, sparse network.Its clique-based construction also avoids repeated global optimization, making it computationally efficient at this network scale (60).Source: Community detection · pp. 15–16
• From 4,042 CPM communities, excluding those with fewer than three human nodes left 642 communities for analysis.The threshold ensures each retained community contains at least one human triad.Source: Community detection · p. 16

Community network analysis

• Community cohesion is the peak-to-post-peak change in human–human edge density, measured on undirected projections within the same node sets.The pre-peak snapshot supplies baseline controls rather than a modeled endpoint; bot–human density is normalized by possible ties to avoid mechanical community-size effects.Source: Community detection · p. 16, Community network analysis · pp. 16–17

Statistical analysis design

• Complementary ego-network and community models capture distinct structural properties and sampling strategies, reducing reliance on a single network representation.The ego model follows users active across all three periods; the community model can include users present in only specific phases.Source: Statistical analysis design · p. 16

Community network analysis

• Community-level OLS relates peak bot–human density to peak-to-post-peak change in human–human density, controlling for baseline density, human-node count, and aggregated user characteristics.Controls include friend-to-follower ratio, account age, statuses, and favourites; received-to-sent ratio is omitted because it is strongly collinear with network-structural measures.Source: Community network analysis · pp. 16–17

Individual network analysis

• The individual model measures peak-to-post-peak change in human-only local clustering among non-isolated k-core-2 users active across all periods.Bot exposure is alter–bot edge count, excluding ego-incident edges to separate the predictor from ego-network structure; OLS controls for baseline position and user characteristics.Source: Individual network analysis · p. 17

Opinion dynamics

• A binary indicator of pre-peak movement support, defined as stance above the sample median, tests whether support moderates the association between bot exposure and cohesion change.The interaction tests whether the exposure–cohesion association differs between supporters and other users.Source: Opinion dynamics · p. 18

Robustness

• The authors report that their conclusions were stable across several checks, with full specifications and coefficients in SI Model Tables S4 to S12.Source: Robustness · p. 18
• Associations persisted across bot-probability thresholds of 0.65–0.75, indicating they were not an artefact of the bot/human classification cutoff.Thresholds below 0.65 were not tested because false positives were considered a greater concern than false negatives; higher thresholds guard against overstating bot effects.Source: Robustness · p. 18
• Restricting pre- and post-peak snapshots to BLM-related edges, matching the peak network’s edge definition, left the estimates substantively unchanged.The authors recomputed every predictor and outcome using these matched snapshots.Source: Robustness · p. 18
• After network-error correction, peak bot exposure remained negatively associated with change in human clustering (β = −0.040, SE = 0.019), and the bot-exposure × pro-BLM interaction remained negative (β = −0.190, SE = 0.050).Individual-level residuals were dependent under both ego-neighbourhood overlap and CPM co-membership weights; after correction, filtered residual Moran tests became non-significant. The opinion-interaction model retained a positive main term (β = 0.125, SE = 0.045).Source: Robustness · pp. 18–19

Supplementary materials

• The supplementary packet covers data collection, bot identification, ChatGPT-based opinion categorization, network construction and filtering, Tables S1–S12, references 7–66, Movie S1, and Data S1.Source: Supplementary materials · p. 25, This PDF file includes · p. 26

Data collection

• The study focused on June 7–8, 2020, identified as the peak of protest-related activity after George Floyd’s death.Source: Data collection · p. 27
• Tweets were gathered from an archive using BLM, George Floyd, and spelling-variant keywords.The complete keyword list is identified as Table S1.Source: Data collection · p. 27, Table S1 · p. 27
• The dataset contains over 2,500,000 tweets from more than 1,000,000 users, collected from June 6 through June 8, 2020.Tweets were collected using the Twitter Academic Product track API, with associated user IDs and additional user information retrieved through the user lookup API.Source: Data collection · p. 27
• Retrospective collection may omit deleted posts or accounts and suspended bots, potentially underestimating bot activity relative to human activity.The authors note that Twitter’s spam moderation and the ephemeral behavior of social bots may affect dataset completeness.Source: Data collection · p. 27
• The discontinued Academic Track API makes the dataset a unique snapshot for historical and comparative analyses of protest-related discourse.Source: Data collection · p. 27

• Botometer probabilities and thresholds of 0.65 and 0.5 classify accounts as automated, human, or unknown, following a methodology developed in prior work (3).Source: Bot Identification · p. 28
• Supervised models trained on features from multiple datasets are triangulated with Botometer; an account is classified as a bot only when both sources flag it.The models include Random Forest, Support Vector Machine, Logistic Regression, XGBoost, and Deep Learning, optimized to balance specificity, sensitivity, and F1 score.Source: Bot Identification · p. 28
• All users active at the protest peak (k-core ≥2) were classified as human, bot, or unknown using this methodology.A detailed description of the methodology and model performance is provided in earlier work (3).Source: Bot Identification · p. 28

Opinion categorization using ChatGPT

• ChatGPT 3.5 [55] categorized users’ opinions toward BLM protests on a continuous scale from -1 (extremely negative) to 1 (extremely positive), with 0 for neutral or irrelevant text.The prompt treated complaints about protesters’ behavior as negative, extreme claims about BLM’s aims as mildly negative, and instructed the model to account for sarcasm but not partisan preference.Source: Opinion categorization using ChatGPT · p. 28
• Two independent coders classified each sampled user’s combined timeline using the same prompt instrument as GPT, and their averaged judgments were compared with GPT-3.5 scores.The sample was drawn from a similar online activism event, rather than the BLM analysis itself.Source: Opinion categorization using ChatGPT · p. 28
• Coder judgments correlated at C=0.71, while averaged coder judgments correlated with GPT-3.5 at C=0.88, supporting GPT scores as a valid automated measure of user support.The validation used a random sample of N=300 users from a similar online activism event.Source: Opinion categorization using ChatGPT · pp. 28–29, Table S2 · p. 29

Network construction and filtering

• Retweets and quotes form directional edges from the original poster to the sharer or commenter, while the network aggregates connections as unweighted edges.Self-loops were included, but isolated nodes were removed; retweets have no added commentary, whereas quotes do.Source: Network construction and filtering · p. 29
• The analysis retained 92,078 users and 372,932 edges after filtering a 48-hour snapshot to users with k-core values greater than two.The snapshot focused on users who interacted with bots.Source: Network construction and filtering · p. 29
• Before and after snapshots cover May 1–2 and September 1–2, 2020; the peak sample includes only protest-relevant edges, so edge-count differences reflect sampling.The study focuses on changes between the before and after periods to assess shifts in network structure and bot interactions.Source: Network construction and filtering · pp. 29–30

• Individual-level OLS models are reported at bot thresholds 0.65, 0.7, and 0.75.Source: Table S4 · p. 31, Table S5 · p. 32, Table S6 · p. 33
• Community-level OLS models are reported at bot thresholds 0.65, 0.7, and 0.75.Source: Table S7 · p. 34, Table S8 · p. 35, Table S9 · p. 36
• Robustness tables compare unrestricted after-period outcomes with outcomes recomputed from BLM-related retweets only, holding peak exposure unchanged.The individual and community tables reuse the main Model 3/5 specifications at bot threshold 0.65.Source: Table S10 · p. 37, Table S11 · p. 38
• The ego-level network-autocorrelation robustness analysis uses network-error models, ego-overlap and community weights, retains zero-neighbour islands, and tests Moran statistics with 999 permutations.Ego overlap is Jaccard similarity between closed human ego neighbourhoods, retaining the top 20 neighbours per ego before symmetrization; CPM weights connect observations in the same peak community.Source: Table S12 · p. 39, Model tables · p. 39