- The paper combines retweet-network analysis of 38 million tweets from 20 European countries with manual annotation of 1,242 polarizing tweets to map wartime publics.
- The paper finds that rising homophily—from about 71% to over 82%—mainly reflects casual users leaving the debate, while persistent users form a more insular hard core rather than becoming more radical.
- The paper introduces “conditional publics,” showing that opposing sides share events on pragmatic issues such as weapons, sanctions, and war crimes but invoke different references on interpretive issues such as responsibility and the war’s desired outcome.
This paper analyzes the European Twitter debate surrounding the Russian invasion of Ukraine to determine whether opposing publics inhabit a shared informational reality (2604.05800). Drawing on a large-scale dataset of geolocated tweets from 20 European countries over the first eight months of the war, the authors combine retweet network community detection with manual stance annotation across six issue axes. Their central contribution is the concept of conditional publics: public formations whose relational structure—whether they share or fracture a common referential frame with their opponents—depends on the epistemic character of the issue under debate.
Data and methodological approach
The study builds on a streaming collection of 623,945,097 war-related tweets (February 27–October 12, 2022), gathered via keyword queries covering "Ukraine" in over 50 languages. Because Twitter's geo-tagging is rarely populated, users were geolocated by mapping the self-reported Location field against GeoNames; this approach was previously validated at 93.7% accuracy. After filtering to European countries with at least 10,000 users (plus countries bordering Russia), selecting each country's dominant tweet language, and excluding Russia and Ukraine for interpretive clarity, the final dataset comprises 38,044,266 tweets across 20 countries.
To isolate the most divisive content, the authors run the Leiden algorithm on each country's retweet network and extract, per community, the 20 tweets maximizing an in-minus-out polarization score (ni−no). This yields 1,242 unique tweets responsible for roughly 4.96 million retweets (3.75% of all retweets)—the most polarizing content in the corpus. All four authors independently annotated these tweets on six issues: responsibility for the war, desired outcome, preferred European action, weapons provision, sanctions, and war crimes attribution. Notably, rather than collapsing annotations into majority labels, the authors average individual scores, adopting a perspectivist stance toward ground truth consistent with their constructivist framework—a defensible but unusual choice that means stance scores are graded rather than categorical.
Structural polarization as a selection effect
The longitudinal analysis addresses RQ1 with a finding that runs contrary to the common narrative of platforms radicalizing users over time. Retweet volume peaks immediately after the invasion and declines steadily, while mean homophily—the fraction of within-community retweets—rises from approximately 71% to over 82%. The authors show that this increasing structural polarization is not driven by individuals becoming more homophilic, but by a compositional shift: median user lifespan rises as casual participants exit, leaving a persistent "hard core" of long-term, highly engaged, more insular users. Events such as the release of Bucha massacre footage temporarily reverse the trend by drawing casual users back into the debate. The implication is that polarization trajectories measured on social media may largely reflect audience attrition dynamics rather than opinion change—an important caveat for any study inferring attitude shifts from aggregate network metrics.
Two transnational clusters: hawkish and doveish
Hierarchical clustering of communities' daily posting dynamics reveals two distinct temporal clusters present within almost every country (with exceptions including Lithuania, Latvia, Finland, and Poland, where both top communities fall in one cluster). These clusters also diverge in expressed stances: score distributions differ between clusters on all six issues at p<10−4 (Mann-Whitney U). One cluster—"hawkish"—attributes responsibility and war crimes overwhelmingly to Russia and favors weapons and sanctions; the other—"doveish"—emphasizes NATO/EU responsibility, opposes escalation, and prioritizes peace over justice.
The labeled data show strongly asymmetric majorities: 95% of responsibility-labeled retweets blame Russia, and 96% of crimes-labeled retweets attribute atrocities to Russia, yet majorities prefer negotiating over fighting (88% vs. 12%), and weapons and sanctions split nearly evenly (48/52% and 51/49%). This coexistence of near-consensus on interpretive attribution with division on policy instruments is itself a substantive result about the structure of wartime opinion.
Asymmetric synchronization across Europe
On RQ3, the hawkish side exhibits strong pan-European temporal synchronization across all issues, while the doveish side is markedly more country-specific, with responsibility being its most synchronized issue. Shared language raises cross-country correlations significantly (p<0.05), but the hawkish/doveish difference exceeds what language alone explains. Country-level deviations are documented: Turkey is consistently idiosyncratic; Italy, Austria, Switzerland, and Germany show anti-sanctions spikes (including German debate around Nord Stream 2); France shows doveish activity on weapons. The authors interpret the asymmetry as evidence of horizontal Europeanization that depends on political alignment—the hawkish side constituting a pan-European movement defined antagonistically against Russia, the doveish side anchored in national debates. They candidly note that tight synchronization may partly reflect coordinated information operations or activist networks rather than organic attention, a caveat they do not resolve empirically since no bot or coordination detection was performed.
Conditional publics: shared events depend on issue type
The core analytical result (RQ4) comes from annotating 121 pairs of opposing-stance tweets drawn from temporally proximate attention peaks (within 7 days), coding whether each tweet concerns an event and whether the pair references the same event. Both sides are event-oriented overall (hawkish 73%, doveish 80% of peak tweets reference events), with exceptions on the pro-justice outcome side and both sides of the responsibility axis, where virality drives attention endogenously.
The decisive finding is the sharp split by issue type. On pragmatist issues—weapons, sanctions, war crimes—opposing sides converge on the same high-profile events (notably the Bucha massacre and European Parliament proceedings), forming an agonistic public sphere in Mouffe's sense: adversaries contesting meaning within a shared referential frame. On interpretive issues—responsibility, outcome, Europe's role—the sides foreground entirely different events and historical references, operating as affective publics and counterpublics constructing divergent meanings. This is surprising given that the peaks were selected for temporal proximity: simultaneity of attention does not imply shared referents. The paper's proposed concept, conditional publics, generalizes this observation: publics are not uniformly deliberative, affective, or oppositional, but switch between these modes depending on whether the issue imposes externally verifiable anchors.
Limitations and open questions
The paper concedes several constraints on interpretation. Retweets are treated as endorsement signals, though they can equally express amplification, irony, or strategic visibility; the networks therefore capture amplification patterns rather than stable beliefs. Non-retweeted original content is excluded, biasing the analysis toward high-visibility messages. Annotator positionality (all European-based researchers) may have introduced systematic interpretive tendencies. Coordinated inauthentic behavior is not detected, so synchronization results cannot be cleanly separated from manipulation. Finally, the findings are tied to Twitter's networked-public architecture, where event-binding travels through explicit social ties; the authors explicitly leave open whether "eventification" persists on algorithmically curated platforms such as TikTok, and whether the interpretive/pragmatist distinction holds outside exceptional wartime contexts where elite framing compresses legitimate debate.
Conclusion
By combining large-scale network analysis with careful manual annotation, this study demonstrates that structural polarization in the European Ukraine-war debate grew through user attrition rather than radicalization, that hawkish and doveish communities coexist within nearly every country while synchronizing asymmetrically across borders, and—most distinctively—that whether opposing sides attend to the same events is conditionally determined by issue type. The conditional publics framework offers a relational alternative to static typologies of digital publics, and the open questions it leaves—cross-platform generalizability and disentangling organic from coordinated synchronization—are concrete targets for subsequent empirical work.