Source-linked AI summary
Russian propaganda on social media during the 2022 invasion of Ukraine
Dominique Geissler, Dominik Bär, Nicolas Pröllochs, Stefan Feuerriegel
TL;DR
Research on Russian propaganda during the 2022 invasion is largely anecdotal, leaving systematic evidence about its social-media spread limited. The paper analyzes pro-Russian message diffusion and bot behavior, finding that bots amplified propaganda rather than creating it and did not specifically target influential users.
Problem
Evidence about Russian propaganda campaigns during the 2022 invasion remains largely anecdotal, while systematic empirical research on pro-Russian social-media spread is limited.
Source: Main · p. 3, Discussion · p. 28Method
The study analyzes pro-Russian and pro-Ukrainian Twitter datasets and identifies bots using Botometer, a classifier based on account, network, and linguistic features.
Source: Dataset with pro-Ukrainian support · p. 8, Bot detection · p. 9Results
Bots acted as amplifiers rather than creators of pro-Russian propaganda and did not specifically target influential users.
Source: Discussion · p. 29Takeaways & Limitations
The authors suggest that amplification can increase the prominence and activity of organic accounts aligned with the campaign’s objectives.
Source: Discussion · p. 29The analysis covers one platform and mostly English messages found through English hashtags, limiting its platform and language scope.
Source: Discussion · p. 29
Abstract
from arXiv · showhide
The Russian invasion of Ukraine in February 2022 was accompanied by practices of information warfare, yet existing evidence is largely anecdotal while large-scale empirical evidence is lacking. Here, we analyze the spread of pro-Russian support on social media. For this, we collected N = 349,455 messages from Twitter with pro-Russian support. Our findings suggest that pro-Russian messages received ~251,000 retweets and thereby reached around 14.4 million users. We further provide evidence that bots played a disproportionate role in the dissemination of pro-Russian messages and amplified its proliferation in early-stage diffusion. Countries that abstained from voting on the United Nations Resolution ES-11/1 such as India, South Africa, and Pakistan showed pronounced activity of bots. Overall, 20.28% of the spreaders are classified as bots, most of which were created at the beginning of the invasion. Together, our findings suggest the presence of a large-scale Russian propaganda campaign on social media and highlight the new threats to society that originate from it. Our results also suggest that curbing bots may be an effective strategy to mitigate such campaigns.
Main
Research on Russian propaganda during the 2022 invasion remains largely anecdotal, despite concerns about social-media influence and bot amplification. The study measures pro-Russian message reach, dissemination, and bot activity on Twitter.
Main
- Rigorous empirical evidence about Russian propaganda campaigns during the 2022 invasion was missing, despite concerns about efforts to shape narratives and weaken support for Ukraine.The paper notes concerns about influence efforts in countries that abstained from the UN resolution deploring the invasion.Source: Main · p. 3
- Earlier research described Russian influence operations, including IRA tactics that shifted from trolls toward more refined campaigns, motivating new large-scale evidence.Prior work also documented campaigns tied to the 2014 conflict and elections, while the influence on U.S. voting behavior has been questioned.Source: Main · pp. 3–4
- Bots can amplify early diffusion and attract attention despite receiving fewer retweets than humans, raising concern they could fuel pro-Russian propaganda during the invasion.Prior research describes bots as low-cost tools for producing many profiles and as potential amplifiers of content that might otherwise not go viral.Source: Main · pp. 4–5
- 349,455 pro-Russian messages received about 251,000 retweets and reached about 14.4 million users; 20.28% of spreaders were bots, with pronounced activity in abstaining countries.Most bots were created at the beginning of the invasion, and the findings point to bot amplification in early diffusion.Source: Main · p. 5
Methods
The methods combine hashtag-based Twitter collection, filtering and human validation, bot classification, geographic inference, and retweet-network construction. The resulting pro-Russian dataset contains 349,455 messages, while location and bot-classification procedures receive separate validation.
Data collection
- The Twitter search used four pro-Russian hashtags from February through July 2022, iteratively expanding from #istandwithrussia to capture other common pro-Russian tags.Hashtags were chosen over keywords to focus on likely coordinated-campaign messages and distinguish pro-Russian support from broader invasion discussion.Source: Data collection · pp. 6–7
Preprocessing
- After filtering removed 19,307 messages (5.24%), the resulting dataset contained N = 349,455 pro-Russian messages from 132,131 users, of which 250,853 (71.78%) were retweets.Filtering discarded messages containing identified anti-Russian or anti-Putin hashtags and messages from 44 identified Western news media outlets, as well as retweets of those messages.Source: Preprocessing · pp. 7–8
Dataset with pro-Ukrainian support
- The pro-Ukrainian comparison dataset contains 9,818,566 messages from 2,079,198 users after parallel preprocessing and removal of anti-Ukrainian hashtag messages.Unless stated otherwise, the paper’s main analyses use the pro-Russian dataset.Source: Dataset with pro-Ukrainian support · p. 8
Human validation
- Human validation sampled 50 removed and 50 retained messages per dataset, with three screened workers per dataset; preprocessing labels had substantial agreement when neutral or unclear labels were excluded.For the Russian dataset, Cohen’s κ was 0.7 without neutral or unclear labels; for the Ukrainian dataset, it was 0.71.Source: Human validation · pp. 8–9
Bot detection
- Botometer scores above 0.5 classified users as bots; among 2,661 accounts, Botometer and Bot Sentinel classified 25.29% and 26.53% as bots, respectively.The classifiers flagged different account concepts, but the main bot-target findings remained unchanged in the Bot Sentinel validation subset.Source: Bot detection · pp. 9–10
Location analysis
- 70.19% of users received location estimates from geotags, geocoded profile locations, or a follower-country proxy; validation agreement was κ=0.92 and κ=0.81, respectively.The follower proxy assumes users live in the same country as their followers; the two reported agreement values compare geocoding and follower-based estimates with validation locations.Source: Location analysis · pp. 10–12
Retweet network
- The retweet network represents users as nodes and retweets as edges, colors them by account country, and filters for giant components and weighted degree above 10.Because the network is undirected, weighted degree counts incoming and outgoing retweets.Source: Retweet network · p. 12
Results
The analysis tracks pro-Russian support from February through July 2022, finding substantial reach, differences from pro-Ukrainian spread, and a prominent role for bots. Bot activity and retweet patterns varied across countries, with India and South Africa showing locally clustered networks.
Pro-Russian support on social media
- About 14.4 million users were reached by pro-Russian messages, which generated nearly 1 million likes in the dataset.The analysis estimated readership from unique users following message authors; the dataset contained 349,455 pro-Russian messages.Source: Pro-Russian support on social media · p. 13
- On March 2, 2022, message volume peaked as the UN General Assembly adopted Resolution ES-11/1; India accounted for about 41.7% of messages that day.On March 3, the largest shares came from the U.S. (about 14.1%), Nigeria (about 10.5%), and India (about 10%).Source: Pro-Russian support on social media · pp. 13–14, Figure 1 · p. 14
Spreading dynamics of pro-Russian support
- 20.28% of accounts spreading pro-Russian messages were classified as bots, and bot-account creation peaked around the invasion’s start on February 24, 2022.Bots tended to have been created more recently than human-classified accounts.Source: Spreading dynamics of pro-Russian… · p. 15, Figure 2 · p. 15
- Pro-Russian support had more bots than pro-Ukrainian support (20.28% vs. 14.25%) and fewer mean followers (1,690 vs. 2,248); source tweets averaged fewer likes, replies, and retweets (12.97, 1.16, 3.38 vs. 28.35, 1.22, 6.56).Both supports had broad distributions spanning several orders of magnitude, and a substantial proportion of pro-Russian messages went viral.Source: Spreading dynamics of pro-Russian… · pp. 15–16
Cross-country heterogeneity in the exposure to pro-Russian support
- Abstaining countries had the highest relative frequency of bots (20.3%), compared with 14.9% in countries voting against and 16.6% in countries approving ES-11/1.India, the United States, South Africa, and Nigeria had the most accounts spreading pro-Russian messages; bot shares were 24.2%, 23.9%, 10.2%, and 7.9%, respectively.Source: Cross-country heterogeneity in the… · pp. 17–19, Figure 4 · p. 20
Retweet network
- Humans primarily retweeted humans, while bots mainly retweeted humans and rarely retweeted other bots; India and South Africa formed geographically homogeneous network clusters unlike the more scattered U.S. accounts.The authors interpret the retweet patterns as bots exposing humans to human-generated pro-Russian messages; the geographic clusters were relatively isolated.Source: Retweet network · p. 22, Figure 6 · p. 23
Amplification of pro-Russian support spreading through bots
- Bots produced 20.82% of source tweets but accounted for 25.72% of retweets, and bots retweeted other bots with disproportionately short time lags.Humans were retweeted equally fast by bots and humans; bots tended to mention accounts with fewer followers than humans did, and those accounts had lower mean PageRank (0.002 vs. 0.0022).Source: Amplification of pro-Russian support… · pp. 24–26, Figure 8 · p. 27
Discussion
The discussion interprets the findings as evidence of a coordinated Russian propaganda campaign, with bots amplifying pro-Russian messages and distinctive country clusters. It also emphasizes limits on what the Twitter data can establish and proposes reducing bot influence as a possible countermeasure.
Discussion
- The authors interpret bot-heavy dissemination, highly connected retweet networks, country clusters, and bot creation before the UN vote as evidence of coordinated manipulation.They note that messages accumulated on the day of Resolution ES-11/1 voting and that many bots were created shortly beforehand.Source: Discussion · p. 28
- The authors distinguish this campaign from bots spreading low-credibility or inflammatory content, which had targeted influential users.Here, bots mentioned users with less social influence than humans, and bots were rarely retweeted by humans.Source: Discussion · p. 29
- Bots acted more as amplifiers than creators, retweeting more than producing original content and not specifically targeting influential users.The authors suggest this strategy sought broad exposure and prominence for organic accounts; they describe trending hashtags as one route to visibility.Source: Discussion · pp. 28–29
- The study is limited to Twitter and mostly English-language messages, and it cannot establish whether exposure changed public opinion.The authors also note possible message removal, reliance on bot-detection accuracy, and that pro-Russian support was a minority of war-related tweets analyzed.Source: Discussion · pp. 29–30
- The authors propose reducing bot influence as a countermeasure, while noting that intervention effectiveness requires further research.They call for continued monitoring and active efforts to counter online propaganda.Source: Discussion · pp. 30–31
Declarations
The declarations identify the authors’ contributions, acknowledge plotting-code support, and include a data-and-materials availability statement.
Declarations
- The declarations include an availability statement referring to data and code supporting the study’s findings.The supplied statement does not specify their availability status.Source: Declarations · p. 33
- The authors declare no competing interests.Source: Declarations · p. 33
- All authors contributed to conceptualization, results interpretation, and writing; DG contributed to data analysis.All authors approved the manuscript.Source: Declarations · p. 33
- The authors acknowledge Shao et al. [20] for plotting code.Source: Declarations · p. 33
Supplementary tables
The supplementary tables provide examples and filtering materials, list frequent hashtags and influential accounts, and identify excluded Western news outlets.
Supplementary tables
- Table S1 provides examples of pro-Russian messages, while Table S2 provides examples of pro-Ukrainian messages.Source: Table S1 · p. 44, Table S2 · p. 45
- Table S3 lists the most frequent pro-Russian hashtags, with account counts and total occurrences.Source: Table S3 · p. 46
- Tables S4 and S6 list anti-Russian and anti-Ukrainian hashtags used to filter messages by stance.Source: Table S4 · p. 47, Table S6 · p. 49
- Table S5 lists verified Western news-media accounts excluded because they reported on pro-Russian hashtags or platform actions without disseminating propaganda.Source: Table S5 · p. 48
- Table S7 ranks the ten accounts with the most followers and describes three accounts’ differing communication strategies.Follower count is used as a proxy for social influence; the examples include an account promoting a hashtag to trend in India and one actively disseminating pro-Russian messages.Source: Table S7 · p. 50
Supplementary Figures
The supplementary figures examine message volume over time, account creation and location patterns, retweet networks, and bot–human virality distributions. They report temporal peaks, robust geographic patterns, and comparatively isolated India and South Africa retweet clusters.
Supplementary Figures
- Bot and human message volumes peaked on March 2 and 3, coinciding with the overall message peaks.The inset shows message volume across the full dataset period.Source: Figure S1 · p. 51
bot information
- Accounts without bot-score information were created only after 2016, and most were created shortly before the invasion.Figure S2 presents account creation dates on both a long-range timeline and a timeline focused near the invasion.Source: Figure S2 · p. 52
- A location analysis using profile self-reports and message geolocations found patterns similar to the main analysis.The comparison shows users per country and the relative frequency of bots, excluding bots whose locations could not be inferred.Source: Figure S3 · p. 53
counts without bot information
- Humans, bots, and accounts without bot information showed no deviating geospatial patterns across countries.Source: Figure S4 · p. 54
- India and South Africa formed relatively isolated retweet clusters, while U.S. accounts were more broadly scattered.The network encodes account types by color and retweet direction and relative frequency by edges; the virality figure compares human and bot CCDFs across source tweets, retweets, followers, likes, and replies.Source: Figure S5 · p. 55, Figure S6 · p. 56
Supplementary Materials: Content analysis of the Russian
The content analysis finds that the campaign focused on discrediting Western countries and institutions while targeting countries prone to abstain from the UN vote. It also documents historical narratives and messages connecting Russia to specific countries.
propaganda campaign
- NATO appeared in 7.1% of messages, the West in 5.4%, and the USA in 4.1%.These relative shares came from keyword matching for geographic regions and international bodies.Source: propaganda campaign · p. 57
- A manual sample of 100 messages showed frequent efforts to discredit Western countries and organizations.Examples invoked Russia’s support for South Africans during Apartheid and accused the West of oppression and exploitation.Source: propaganda campaign · pp. 57–58
- The campaign frequently reused existing and historical narratives, including Apartheid, slavery, oppression, imperial times, and mass murders.The authors note a similar observation in research on Russian traditional media.Source: propaganda campaign · p. 58
- Iraq (3.3%), Syria (2.7%), and India (2.5%) were commonly mentioned, and manual examples linked South Africa and India to Russia through historical ties and the UN vote.The authors describe this as further quantitative and qualitative evidence connecting the campaign with the vote.Source: propaganda campaign · p. 58
- The campaign focused on eliciting negative opinions toward Western countries and institutions and targeted countries prone to abstain from the UN vote.The authors interpret its activity around the UN vote and in specific countries as suggesting an intent to steer public opinion there.Source: propaganda campaign · pp. 57–59
B Supplementary Materials: Sentiment analysis
The sentiment analysis found comparatively small differences between human and bot source tweets, and did not support the hypothesis that bots strategically favor negative language. The analysis used the NRC lexicon to classify sentiment and emotions.
Sentiment analysis
- The analysis removed numbers, mentions, hashtags, and links, then used the NRC lexicon [82] to score sentiment and eight emotions in source tweets.Tweets were classified by predominant positive or negative sentiment, or as neutral when scores were equal.Source: B Supplementary Materials: Sentiment… · p. 60
- Differences in source-tweet sentiment between humans and bots were comparatively small, and the hypothesis that bots strategically favor negative language was not supported.This differs from bots spreading inflammatory content, where negative language is more common [34].Source: B Supplementary Materials: Sentiment… · p. 60, Figure S7a
- Figure S7 compares humans’ and bots’ source tweets by relative frequency of positive, negative, or neutral sentiment and by emotion category.The emotion categories are fear, anger, trust, surprise, sadness, disgust, joy, and anticipation.Source: B Supplementary Materials: Sentiment… · p. 60, Figure S7 · p. 61