Growing polarization around climate change on social media

The study investigates how discussions around climate change on social media intersect with political polarization, focusing on Twitter conversations tied to the UN Climate Change Conferences (COP) from 2014 to 2021. Social media's role in climate discourse is contested—seen both as democratizing and as fostering polarization—making monitoring polarization vital due to its potential to drive antagonism and political deadlock. By centering on COP events—discrete, recurring, thematically focused, and highly salient to climate politics—the authors aim to enable robust, multi-year quantification of polarization and interaction patterns. They pose the research problem of measuring ideological polarization in Twitter climate discourse, identifying the groups involved, understanding the political dimensions, and characterizing topic themes (including climate-contrarian narratives and cross-ideological content such as allegations of political hypocrisy). The paper contends that since late 2019, a prominent opposition to the dominant pro-climate discourse has established itself on Twitter, producing a highly polarized climate debate.

Prior work highlights Twitter’s importance for climate communication, political polarization, and misinformation research, while acknowledging it is not a direct proxy for public opinion. Studies have examined polarization dynamics on social platforms, echo chambers, and the spread of misinformation, as well as event-focused analyses of climate communication (e.g., IPCC reports, national elections). Broader literature addresses the politicization and polarization of climate change across media and public opinion. The paper builds on latent ideology approaches for estimating ideological spectra from Twitter interactions, and draws on the infodemics literature to connect structural online behaviors with content themes. It also leverages classifications of climate-contrarian claims to contextualize minority narratives. The authors situate COP as a focal event due to its prominence in climate diplomacy and strong media/public attention, enabling a connected interaction network suitable for structural analysis.

Datasets: Twitter data (tweets and user metadata) were collected via the Twitter Academic API using the keyword query 'cop2x' (x ∈ {0,...,6}) for each COP edition from June 1 of the conference year to May 31 of the following year (except COP26, ending November 14, 2021). Politician accounts were labeled using the Twitter Parliamentarian Database (26 countries; 2017–2021) with manual additions for missing 2021 influencers. Network construction: From the corpus, the authors focused on retweets (as signals of endorsement) and filtered to English-language tweets. They constructed a directed, weighted retweet network: nodes are users; edges from A to B indicate A retweeted B; weights reflect unique retweet counts. Measuring polarization: Polarization is operationalized as distribution multimodality (bimodality), aligning with political science literature. Because ground-truth opinions are unavailable on Twitter, the authors infer a synthetic opinion distribution from network structure using latent ideology derived from retweet interactions. Latent ideology estimation: They built a user-by-influencer matrix A with entries a_ij equal to the number of times user i retweeted influencer j. Users retweeting fewer than two influencers were pruned. They performed correspondence analysis on the standardized residuals S = D_r^-1/2 (P − r c) D_c^-1/2, where P = A / sum(a_ij), r and c are row and column sums, and D_r, D_c are diagonal matrices. Singular value decomposition yields row coordinates X = D^-1/2 U; the first dimension (largest singular value) is used as the ideology axis. User scores are rescaled to [-1, 1], with majority mapped to −1 and minority to +1; influencer positions are the median of weighted retweeter positions. Influencer selection: For COP21 and COP26, the top 300 most retweeted accounts (minus a small filtered set <3% that conflated geography/language/out-of-domain topics) were used as influencers; additional rankings and robustness checks are reported in Supplementary materials. Hartigan’s dip test: Multimodality was assessed with Hartigan’s diptest, yielding statistic D and p-value. Statistical uncertainty was estimated via bootstrapping (randomly sampling 70% of users and influencers; 1,000 repeats). Topic modeling: BERTopic was used to extract discussion topics from tweet text and to position topic cohorts along the latent ideology axes (political and non-political dimensions computed by separating influencer sets into politicians-only vs. excluding politicians). News media reliability: The authors cross-referenced URLs in tweets with NewsGuard trust scores (0–100) and political lean ratings to compare news source usage between majority and minority groups. Comparative datasets: Complementary analyses on broader 'climate change' and climate-skeptic keyword datasets, and limited comparisons on YouTube and Reddit activity, were conducted to assess representativeness.

1. Polarization surge at COP26: The latent ideology distribution of users was unimodal for COP21 but became significantly multimodal by COP26. Hartigan’s diptest: COP21 D = 0.0023 (95% CI: 0.0020–0.0026; p = 0.003); COP26 D = 0.049 (95% CI: 0.048–0.050; p < 2.2×10^-16). Similar low polarization (unimodal) was observed for all COPs prior to COP26. 2) Minority growth and right-leaning activity: The increase was driven by growing right-wing activity, with minority influence increasing markedly (a fourfold increase since COP21 relative to pro-climate groups). COP26 minority influencers include accounts aligned with US Republicans, UK Brexit/UKIP, and Canadian Conservatives; COP21’s minority was small (e.g., @BjornLomborg, @Tony_Heller, @JunkScience). 3) Political mapping: A two-dimensional latent ideology (political vs. climate axes) showed minority clusters dominated by right-leaning parties (US Republicans; UK Brexit/UKIP; Canadian Conservatives), while the majority included most mainstream parties (e.g., Australian Liberals appearing majority-aligned due to pro-climate rhetoric attracting majority retweets). Some left-leaning parties appeared closer to the minority on the climate axis due to cross-ideological criticism of political hypocrisy. 4) Topic dynamics: Majority topics emphasized core COP themes (e.g., women’s day, transport day, climate finance), climate justice, youth/indigenous activism, and decolonization. Minority topics covered all five principal climate-contrarian claim types (global warming denial; non-anthropogenic causation; benign impacts; ineffective solutions; unreliable science/movement) and broader right-leaning issues (COVID-19, vaccines, illegal immigration), along with criticism of Biden, Johnson, and Trudeau. 5) Cross-ideological bridge: Political hypocrisy (e.g., private jets, fossil fuel use, sewage dumping) emerged as a cross-ideological theme receiving substantial engagement; half of majority tweets referencing hypocrisy were posted since December 2020. 6) News consumption: Majority users preferentially linked to high-trust news domains, whereas the minority frequently referenced low-trust domains (robust to country-specific NewsGuard scores). 7) Echo chambers and interactions: The formation of ideological echo chambers during COP was observed. In broader ‘climate change’ data, cross-ideological mentions showed no significant increase over time (linear slope not different from zero; p = 0.121). 8) Representativeness: COP Twitter discourse tracked broader climate discussions. Correlations between COP-labeled users and general activity were high: hypocrisy-related tweets R = 0.93; climate-skeptic patterns among COP26 minority vs. others R = 0.76; ‘climate change’ activity R = 0.94. 9) Event salience: COP21 (Paris Agreement) and COP26 (Glasgow Climate Pact) had the highest levels of content creation and engagement; Google Trends indicated strong host-country effects (France for COP21; UK for COP26).

The analyses indicate that climate discourse polarization on Twitter was relatively low across COP20–COP25 but rose sharply by COP26, with evidence that this trend began around 2019 (e.g., global climate strikes). The surge appears associated with increased right-leaning engagement and amplification of contrarian narratives, possibly catalyzed by anti-science sentiments during the COVID-19 pandemic. Despite expectations around earlier political events (Trump’s election, Brexit), a major shift toward climate skepticism on Twitter emerged post-2019, suggesting evolving political strategies where opposition to climate action gains perceived popular appeal (e.g., 'Net Zero Referendum'). The cross-cutting theme of political hypocrisy functions as a bridge between ideological groups, increasing the risk that accusations of hypocrisy may mobilize or normalize contrarian stances. The preferential consumption of low-trust media by the minority and the presence of echo chambers may further entrench polarization, potentially contributing to political deadlock that could hinder climate action. The study underscores that while Twitter does not perfectly mirror public opinion, its structural and topical signals are informative for understanding how climate politics unfold online. Given that right-wing climate views may be more susceptible to change than left-wing views, the observed opposition could be reversible with targeted communication strategies that address perceptions of hypocrisy and trust.

This multi-year, event-focused analysis shows a substantial increase in ideological polarization in Twitter’s COP-related climate discourse by COP26, after years of low polarization. The rise is driven by increased right-leaning minority activity, wider non-specialized influencer participation, and the spread of diverse climate-contrarian claims. Political hypocrisy emerges as a salient cross-ideological theme. Majority users rely more on high-trust news sources, whereas minority users often reference low-trust outlets, consistent with echo chamber dynamics. The COP discussion is broadly representative of wider climate conversations on Twitter. These findings emphasize the need to monitor polarization and its effects on public climate discourse, particularly as they may influence the political context for climate action. Future research should track: (1) the trajectory of ideological minority influence, (2) the broader impact of social media polarization on public debates and policy processes, and (3) the evolution of echo chambers and cross-ideological engagement across platforms and events.

Twitter behavior is not a direct proxy for public opinion and may be subject to platform-specific dynamics that foster polarization. The keyword-based data collection (‘cop2x’) focuses on COP-related content and may omit climate conversations outside that frame. Analyses prioritize English-language tweets, potentially introducing geographic and linguistic biases; influencer filtering was required to mitigate confounds (e.g., geographic clustering, non-core topics), but residual biases may remain. The latent ideology approach infers synthetic opinion distributions from network structure, which, despite validation via content themes, may still reflect structural separations for reasons other than opinion (e.g., geography). The selection of top influencers (and removal of ~3%) may affect ideological positioning. Retweets are treated as endorsements, though this is imperfect. Results from Twitter may not generalize to other platforms (YouTube/Reddit showed lower activity for COP events).