key: cord-0560046-5epb4wul authors: Falkenberg, Max; Galeazzi, Alessandro; Torricelli, Maddalena; Marco, Niccolo Di; Larosa, Francesca; Sas, Madalina; Mekacher, Amin; Pearce, Warren; Zollo, Fabiana; Quattrociocchi, Walter; Baronchelli, Andrea title: Growing polarisation around climate change on social media date: 2021-12-22 journal: nan DOI: nan sha: a759c0543a478665a51e3715b7ce9c0373743f09 doc_id: 560046 cord_uid: 5epb4wul Climate change and political polarisation are two critical social and political issues of the 21st century. However, their interaction remains understudied. Here, we investigate the online discussion around the UN Conference of The Parties on Climate Change (COP) using social media data from 2014 to 2021. First, we highlight that cross-platform engagement peaked during COP26. Second, focusing on Twitter, we reveal a large increase in ideological polarisation during COP26, following low ideological polarisation between COP20 and COP25. Finally, we show that this increase is driven by growing right-wing engagement (a 4-fold increase since COP21), while we find no evidence of polarisation driven by individuals moving between ideological groups. With future climate action reliant on negotiations at COP27 and beyond, our results highlight the importance of monitoring polarisation in the public climate discourse and how this may impact political action. Social media platforms provide important locations for the everyday discussion and debate of climate change [1] . The nature of this role is highly contested in the literature, with some scholars pointing to its democratising potential, while others argue social media is accelerating political polarisation [2] . In this paper, we study polarisation around climate change by analysing the online discussion around the Conference of the Parties (COP). COP has been the supreme decision making body of the United Nations Framework Convention on Climate Change (UNFCCC) since its establishment in 1992. This makes the online discussion around COP particularly well suited to gauging public views on climate action. Various studies have considered the polarisation of climate ideologies [3, 4] . Previous work has found that polarisation is particularly high amongst science-literate individuals [5] , and that views can be strongly influenced by corporate action [6] and politicised content from the media [7] . Other studies investigate attitudes to climate change on social media [1, [8] [9] [10] [11] . In most cases, they focus on interactions around a specific event, e.g., on the 2014 IPCC report [12] , or on COP21 [13] , while others study particular time periods, e.g., four months in 2013 [14] . Simultaneously, there has been major growth in the study of online polarisation, misinformation, and so-called "infodemics" [15] [16] [17] [18] [19] [20] [21] [22] . Historically, the dominant focus of these studies has been political polarisation in the United States and Europe [20, 23] , with a recent shift towards studying the Covid-19 discussion [17, 21, 22] . Three gaps stand out when assessing the literature on climate polarisation: (1) longitudinal studies which span a multi-year period, (2) studies analysing very large datasets (with the recent exception of [24] ), and (3) studies where polarisation is measured quantitatively. Here, we address these gaps by studying the online discussion around COP across a seven year period, from 2014 to 2021, focusing on Twitter, Youtube, and Reddit. The motivation for our focus is twofold. Firstly, the COP discussion can be characterised as a discrete, regularly repeated online event which lends itself to a quantitative, longitudinal analysis of climate polarisation. Secondly, COP is the pre-eminent international forum for climate diplomacy, directing considerable public attention towards climate change [25, 26] and fostering debates about climate politics across a range of locations, including social media [13] . Social media content is not representative of public opinion, but is important both as an information source for climate change journalists [27] and more broadly as a window into the contemporary politics of climate change [1] . Our results show that the COP discussion can be partitioned into an ideological majority, including most politicians, climate activists, and journalists, and an opposed ideological minority. From COP20 to COP25, the ideological minority are climate focused and self-identify as falling outside the climate mainstream. However, only a small number of users engage with this group. In COP26, polarisation increased dramatically, driven largely by growing right-wing engagement. We start by showing the evolving nature of COP related content creation and engagement, see Figure 1 . Panel (a) shows the number of posts for Twitter and Reddit, and video uploads for Youtube, from 2014 to 2021. The inset shows general online engagement with COP measured using Google Trends, revealing that social media engagement closely reflects wider online attention. Within our study period, COP21 and COP26 are of particular significance, with the Paris Agreement signed at COP21, and the Glasgow Climate Pact agreed at COP26. Consequently, content creation and engagement are larger for COP21 and COP26 than in the intermediate years. Our data also shows the influence of local engagement, see Figure 1 inset, where overall Google Trends scores are presented alongside country specific scores for France (the host of COP21) and Great Britain (the host of COP26). Previous work has shown that Google Trends is reliable for large geographic regions over extended time periods [28] , and can complement, but not replace, traditional analysis [29] . Content engagement on Youtube, Twitter and Reddit increased from COP21 to COP26 (see Figure 1 Panels (b) -(d)). Total engagement on Twitter is around two orders of magnitude larger than on the other platforms. Unfortunately, the low volume of data on Youtube and Reddit limits our ability to robustly study polarisation on these platforms. However, we do find a range of Youtube channels and Subreddits from across the ideological spectrum, see Supplementary Information (SI). We now analyse the evolving nature of ideological polarisation between COP20 and COP26 on Twitter. Before proceeding, we note that choosing suitable terminology to describe the views of millions of online users is difficult. Manually labelling users as pro-climate, sceptics, or otherwise is practically impossible in such cases. Here, we refer to the two opposed groups as the ideological majority and minority. To measure polarisation quantitatively, we first assume that the climate ideology of an individual, i, can be expressed as a single number, x i [18] . Then, polarisation refers to the properties of the probability distribution, P(x), of ideological leanings across a population. We assume ideological leanings fall in the range x ∈ [−1, +1], and require that users with similar views should be close to each other on the spectrum. Within this framework, an ideological issue is polarised if the distribution P(x) contains multiple well-defined peaks [30] . To extract ideological scores from the Twitter retweet network, we use the "latent ideology" method introduced in [20, 31, 32] , which partitions accounts into a set of influencers, and a set of users who retweet those influencers. A precise mathematical definition is provided in the Methods. As a loose analogy, we can think of the method as starting with an arbitrary ordering of influencers. The method then looks at all the influencers retweeted by a user and attempts to shuffle the influencer ordering to bring those accounts closer together. By repeating this shuffling process iteratively, influencers which share a large set of common retweeters move progressively closer to each other. The method produces an ordering where neighbouring influencers have the greatest overlap in their set of retweeters, whereas influencers at opposite ends of the ordering are most extreme in their differences. For consistency, we specify that the ideological majority are mapped to +1 and the minority to −1. The extent of retweeter overlap between two influencers is captured by their numerical separation on the ideological scale. This method produces ideological scores for individual Twitter accounts. Polarisation is then quantified using Hartigan's diptest for unimodality [33] (see Methods). This outputs the test statistic D and a corresponding statistical significance, p. If D is small and not statistically significant, there is little to no polarisation in our network; increasing D signifies increasing polarisation. Figure 2 shows the latent ideology applied to the COP21 and COP26 networks. At the top of each panel we show the distribution of user and influencer ideology scores. Influencers are selected as the top 300 most retweeted accounts, excluding a small number (< 2%) which conflate the results (see Methods). Statistical analysis is robust to variable influencer number (see SI). Note that the latent ideology reveals the most extreme axis of ideological polarisation. Consequently, the latent ideology can hide within group polarisation. Qualitatively, the user ideology for COP21 appears unimodal, whereas for COP26 the user ideology appears multimodal. This is because only three minority influencers are detected for COP21. In contrast, 18% of the top 300 most retweeted accounts are part of the minority for COP26. The increase in polarisation appears to be driven by users who were not engaged in the COP discussion between COP20 and COP25, but engaged with minority influencers during COP26 (see SI). Selecting the top Twitter accounts who take part in both the COP21 and COP26 discussions, we find that minority influencers from COP21 remain in the minority group, and majority influencers remain in the majority group. However, engagement with this fixed set of minority influencers increases significantly between COP21 and COP26. Finally, Figure 3 shows that polarisation was largely flat between COP20 and COP25, before a large increase in COP26. Bot activity and deleted accounts are unlikely to have a substantial impact on this result (see SI). It is possible that the increase in polarisation is driven by local engagement which enhanced English language engagement with COP26, but not with COP21. However, an analysis of French language Twitter (see SI) reveals similar (although smaller) increases in polarisation between COP21 and COP26, and comparable growth in rightwing engagement. This suggests that observed increases in polarisation cannot be explained purely by enhanced local engagement. We now deepen the analysis on the relationship between the observed ideological groups and climate ideology. For COP21, 3 of the top 300 influencers fall in the ideological minority: @BjornLomborg, @Tony Heller, and @JunkScience. Manual inspection of these user profiles suggests that these individuals self-identify as outside the climate mainstream: @JunkScience quotes a Nature Climate Change article referring to him as "the most influential climate science contrarian" [34] , @BjornLomborg references his book "False Alarm: How Climate Change Panic Costs Us Trillions, Hurts the Poor, and Fails to Fix the Planet", and @Tony Heller links to his blog "realclimatescience.com" which includes titles such as "There is no climate crisis". Focusing on COP26, we find that 56 of the top 300 influencers are in the ideological minority. Of these, 6/56 selfidentify as climate focused, based only on their Twitter biography and pinned posts. A large number of the remaining influencers are news media organisations or journalists (e.g., @newsmax, @nypost, @GBNEWS, @talkRADIO, @spikedonline, @PrisonPlanet, @darrengrimes , @Nigel Farage, @bennyjohnson, @zerohedge), politicians (@SteveBakerHW, @MartinDaubney, @laurenboebert, @davidkurten, @lavern spicer), or accounts campaigning against Covid-19 restrictions (@BernieSpofforth, @JamesMelville). Accounts campaigning against Covid-19 restrictions may not have strong view on climate, however, their presence in the ideological minority remains important given that Twitter algorithms recommend accounts to follow based on similarities in user interactions [35] . Right-wing engagement with the ideological minority is confirmed by analysing the URLs shared in minority group tweets. Using NewsGuard to classify domains (see Methods) gives the following distribution of political leanings: Far Right: 10.2%, Slightly Right: 70.2%, Slightly Left: 19.2%, Far Left: 0.4%. This is in contrast to leanings for the ideological majority: Far Right: 0.2%, Slightly Right: 9.7%, Slightly Left: 84.3%, Far Left: 5.8%. These observations confirm that the latent ideology is effective at encapsulating climate ideology and, more broadly, wider political ideology. Despite this, the method does, rarely, mislabel some accounts. One example is the journalist @C4Ciaran (see Figure 2 ) who tweeted that the side streets of Glasgow were "choked up with chauffeur-driven cars and vans, many with their engines idling". This was interpreted as evidence of hypocrisy by political elites, leading to cross-ideological engagement. However, marginally higher engagement from the ideological minority results in @C4Ciaran being labelled as part of the minority. Here we validate our results using two alternative methods: (1) by applying community detection to the retweet network (excluding influencers), and (2) by comparing the latent ideology with the credibility of news sources used by each ideological group. In network science, a community refers to a subset of a network whose nodes are more connected to one another (intra-community links) than to the rest of the network (inter-community links). Figure 4 shows the network of communities detected in the (a) COP21 and (b) COP26 retweet networks using the Louvain algorithm [36] . In the visualisation, each node represents a community of users, with edges representing inter-community links. For both COP21 and COP26, we find multiple majority communities which dominate the retweet networks. In contrast, only one minority community is found for COP21. In COP26 there are multiple minority communities, but these are rare relative to majority communities. Around 3% of users fall in the minority group for COP21, with this rising to 12% for COP26. Edges are observed between minority and majority communities during COP21 and COP26. However, the heatmaps of user ideology against average neighbour ideology in Figure 4 panels (c) and (d) show that cross-ideology communication is rare. The clear separation between minority and majority groups reveals the presence of echo chambers, defined as an environment in which users reinforce their opinion about a topic by repeated interactions with content showing similar attitudes [18] . Such echo chambers have been observed previously in the context of the online climate discussion [14] . Finally, heatmaps of the latent ideology against independent third party measures of news media credibility and transparency (panels (e) and (f)) confirm that the ideology distributions shown in Figure 2 reflect real-world interpretations of ideology. Taking all the tweets of users with an assigned ideology score, we identify URLs and cross-reference the extracted domains with a database of "trust scores", provided by NewsGuard (see Methods). This reveals that the ideological majority preferentially reference news domains with high trust scores, whereas the minority often reference domains with low scores. We now analyse the distinct topics referenced by each ideological group during COP21 and COP26. Figure 5 shows the top hashtags used during COP21 and COP26 in panel (a). We then quantify whether hashtags are used FIG. 4 . Twitter communities, ideological echo-chambers and correlation with distinct groups of media outlets in COP21 and COP26. Top: The communities formed in the (a) COP21 and (b) COP26 retweet networks, coloured according to their average ideological score. Each node represents a community of users, with edges between nodes representing inter-community interactions. The figure reveals a large number of communities who are part of the ideological majority, alongside two major communities which are part of an ideological minority for COP26 and only one community for COP21. Communities in the ideological minority correspond to 3% of users in COP21 and 12% of users in COP26. Middle: Heatmaps showing the density of individual ideology scores vs average neighbour ideology for (c) COP21 and (d) COP26. The isolated clusters observed illustrate the presence of echo-chambers, see [18] . Bottom: Heatmaps showing the density of news media trust scores provided by NewsGuard, against the average ideological score of each media outlet's Twitter audience for (e) COP21 and (f) COP26. disproportionately by the minority or majority groups in panels (b) and (c) using the Shiftiterator approach [37] (see Methods). The top hashtags used during COP21 and COP26. (a) The most popular hashtags during COP21 and COP26; hashtags in green are common to both conferences. In panels (b) and (c) we rank hashtags based on whether they are used disproportionately by the minority or majority groups for COP21 and COP26 respectively. In each panel, hashtags used predominantly by the ideological majority are shown on the right in yellow, and those used by the ideological minority and shown on the left in blue. Hashtags are ranked from top to bottom based on the absolute difference in their frequency in majority versus minority tweets. Note that due to minimal engagement with the ideological minority during COP21, hashtags in this group are dominated by the themes from a small number of tweets and users, leading to the presence of rare hashtags in the ranking. Similarly, a small number of climate related tweets are likely responsible for #globalwarming, #climate and #climatechange appearing at the top of the COP21 minority ranking. Conference specific. Several hashtags are specific to the COP process. COP21 includes treaty references (#parisagreement, #climatetreaty), terminology (#indc), and references to particular conference groupings (#lpaa, #adp2). For COP26, majority terms are more general, but minority terms refer to the G20 preceding COP26 (#g20romesummit), and the prominence of the #netzero theme. There are also references to specific blogs critical of the COP process (#flop26). Climate urgency. Between COP21 and COP26, there is a shift in the urgency of climate language, with terms like #climatechange and #globalwarming becoming less prominent, and new terms such as #climatecrisis emerging. Climate activism. Hashtags associated with climate activism are prominent in both COP21 (e.g., #climateac-tion, #actonclimate, #climatejustice, #climatemarch) and COP26 (e.g., #togetherforourplanet, #climateactioninyourarea). For COP26, there are also hashtags associated with the Korean K-Pop group BLACKPINK who acted as climate ambassadors (e.g., #blackpink, #blinks). Local politics. Specific regional hashtags illustrate local engagement with COP. In the minority group, COP21 has references to Canadian politics (e.g., #polcan, #cdnpoli, #onpoli, #ableg), while in COP26 we observe the prominence of the host country with critical references to Scottish First Minister Nicola Sturgeon (e.g., #resignsturgeon, #greatpretender, #elsiemcselfie, #fmqs). Scottish politics is also referenced by the majority group in COP26 with the pro-independence hashtag #yesscots. The majority reference #auspol, highlighting criticism of the Australian Government's climate policy. Covid-19. During COP26, the ideological minority reference #covid19 and #novaccinepassports. This is emphasised if we filter minority tweets according to their use of the term "covid" (see SI), where we find a core set of tweets with a strong climate focus, and a wider secondary group who discuss issues of broader relevance to the political right. We have investigated ideological polarisation around climate change, analysing how the online discussion around COP has evolved between 2014 and 2021. Our results show that content creation and engagement increased on Twitter, Youtube and Reddit, and that all platforms include prominent accounts, channels, or subreddits, which are outside the climate mainstream. On Twitter, where the COP discussion is most active, we find that ideological polarisation was low and largely flat between COP20 and COP25, before a significant increase in COP26 driven by growing right-wing engagement, as well as a degree of overlap with the online discussion around Covid-19. The analysed time period is important, spanning the start and end of the Trump presidency, and the rise of the "Fridays For Future" activist movement. On Reddit, the US election in 2016 was found to be a major polarising event [38] . However, our data shows almost no change in climate polarisation at the time. Rather, our increase in polarisation coincides with the Covid-19 pandemic; how the pandemic is affecting online polarisation remains a hotly debated topic. Our assessment of structural polarisation reveals majority and minority groups in the COP discussion. Within the minority group, the key COP21 influencers have a strong focus on environmental issues, self-identifying as outside the climate mainstream, e.g., @JunkScience. While several such accounts reappear in the COP26 data, minority influencers are largely right-wing media organisations and politicians. Most of these accounts do not appear to have been engaged with COP prior to COP26; user engagement with the ideological minority rose from 3% in COP21 to 12% in COP26. This shows that, while the political right on social media largely ignored COP in the past, there is a growing online community who oppose the climate mainstream. However, we note that although there is evidence of an overlap between right-wing political views and opposition to climate action, surveys show that right-wing views on climate are more subject to change than left-wing views [39] . Consequently, there is reason to believe that growing right-wing opposition to climate action may be reversible. Although we aimed to comprehensively study three social media platforms, the bulk of our analysis focuses on Twitter, where the COP discussion is sufficiently active to allow a robust assessment of ideological polarisation. Future work should acquire even larger datasets, and could consider a wider range of platforms, including Gab, Parler, and Gettr which may better reflect uncensored minority views [40] . Our data suggests that the COP discussion is not particularly active on Youtube or Reddit, although this may not be the case for other climate events. Focusing on technical aspects, we note that how to best measure polarisation quantitatively is an open problem. The latent ideology is one promising measure. However, a one dimensional ideology score cannot fully capture the nuance of ideological views on complex social and political issues. Future work may wish to consider how best to quantify polarisation in these more complex scenarios. Given the increase in polarisation observed during COP26, future work should continue to monitor the evolving nature of polarisation during COP27 and onwards. Possible questions include whether there are signs of interactions between polarised groups [14] , whether ideological minorities are growing or declining in influence, and whether social media polarisation is having broader influence on public debates. Finally, it is a value judgement as to what constitutes a healthy plurality of views on social media, or unhealthy polarisation. Consensus should not be expected [41, 42] . However, as we have demonstrated, tracking trends in polarisation over time provides an important insight into the political context for accelerated climate action. Twitter data including tweets and user information was collected using the official Twitter API for academic research (https://developer.twitter.com/en/docs/twitter-api), using the search query "cop2x", x ∈ {0, . . . , 6}. For each COP, data was collected from July 1st in the year of the conference, to May 31st the following year, with the exception of COP26 where data was collected up to and including November 14th, 2021. Statistics for each COP are provided in Table I . Each dataset was downloaded between October and November 2021. For Youtube, we collected videos using the official YouTube Data API (https://developers.google.com/ youtube/v3), searching for videos matching the keywords "cop2x", where x ∈ {0, . . . , 6}. Then, an in-depth search was performed by crawling the network of related videos as provided by the YouTube algorithm. Next, we filtered the videos that contained the keywords in the title or description from the gathered collection. Finally, we collected the comments of each video from our collection. Note, however, that unlike the Twitter API, the Youtube API only returns a sample of videos and comments for any given search query, typically around 10% of the total. Reddit data was collected using the pushshift.io API (https://pushshift.io/) and the official Reddit API (https: //www.reddit.com/dev/api/). For each COP, we downloaded all the submissions containing "cop2x" from the pushshift repository, and then downloaded the corresponding engagement statistics, such as the number of comments and cross-posts, through the official Reddit API. This choice guarantees that submission statistics are up to date since pushshift.io mirrors Reddit in near real-time, and thus statistics on pushshift are often unreliable. The Twitter interaction network is constructed by taking the full corpus of tweets for each COP and focusing exclusively on retweets. Such an approach is typical in the Twitter analysis literature, where retweets are considered evidence of a user endorsing the message of the original poster [43] . This is in contrast to quote tweets or comments which are less likely to represent a clear endorsement of a tweet. After selecting all the retweets from the full Twitter dataset, we filter by language using the Twitter API language metadata, selecting only those retweets written in English. From this set of English language retweets, a network is constructed by defining a node for each unique user in the dataset. This includes any user who authored an original English language tweet, or retweeted an English language tweet, containing the keyword "cop2x", x ∈ {0, . . . , 6}. A directed edge is formed from node A to node B if user A retweeted a post authored by user B. Edges are weighted according to the number of unique retweets between those two users. The latent ideology estimation was developed in [31, 32] and adapted for exploiting retweet interactions in [20] . Following [20] , we infer ideological scores for Twitter users using correspondence analysis [44] (CA) and retweet interactions. First, we build a matrix A such that each element a ij is the number of times user i retweeted influencer j. To select only users that are interested in the COP26 debate, we prune out users that retweeted fewer than two influencers. We then execute the CA method according to the following steps. Given the adjacency matrix normalised by the total number of retweets as P = A( ij a ij ) −1 , the vector of row and column sums respectively as r = P1 and c = 1 T P, and considering the matrices D r = diag(r) and D c = diag(c), we can compute the matrix of standardised residuals of the adjacency matrix as S = D Hartigan's diptest is a nonparametric test to measure the multimodality of a distribution from a sample [33] . It calculates the maximum difference over all sample points between the unimodal distribution function that minimises that maximum difference and the empirical distribution function. The test produces a statistic D which quantifies the magnitude of multimodality, and a statistical significance p. If p < 0.01, we say that the ideology distribution shows statistically significant multimodality. Conversely, if p ≥ 0.01, we cannot reject the unimodality of the distribution. Applying the latent ideology to a set of influential accounts on Twitter does not guarantee that those accounts will arrange themselves in the latent space based on political or climate ideology. In a number of cases, the dominant factor which determines the principal ideological axis is geography. By focusing exclusively on English language twitter, the effect of these geographic factors is reduced. However, some additional filtering is required to avoid the latent ideology partitioning accounts based on geography. Factors which may conflate ideological scores include (1) language (e.g., English vs. non-English), (2) geography (e.g., accounts focused on Indian politics), and (3) prominent topics outside the core discussion (e.g., discussions in the blockchain community), see SI for details. These factors are mitigated by selecting English language tweets, and by performing some minor filtering of the influencer set. For each COP, less than 2% of accounts are removed from the set of influencers as part of the filtering process (see SI). After separating Twitter users into an ideological minority and majority, we selected all the English language hashtags related to the users in each ideological group, analysing the COP21 and COP26 tweets separately. We then pre-processed the data by normalising the text (removing capitalisation, Unicode characters, punctuation, and English stop words). After cleaning, we used the Shiftiterator package [37] to analyse each set of aggregated hashtags, comparing the different ideologies for a given COP. This package allows us to quantify which words contribute to a pairwise difference between the two texts and how they contribute to understanding their differences. The comparison method is word shift analysis, based on the frequency counts of how often words appear in each of the two texts. For each word w in the text, the method evaluates the probabilities p 1 (w) and p 2 (w), that the word w appears in texts 1 and 2 respectively. It then calculates the proportion shift as their difference: If the difference is positive (δp w > 0), then the word is more common in the second text. If it is negative (δp w < 0), then it is relatively more common in the first text. The method ranks words based on the absolute difference. To ensure that hashtags are informative, we filter out some frequent hashtags related to COP, such as the location (e.g., 'paris', 'glasgow', 'france', 'scotland') and the year (e.g., 'cop21','cop26', '2015', '2021', 'paris2015', 'glasgow2021'). To highlight the different news sources used by the ideological minority and majority, we exploited data retrieved from NewsGuard https://www.NewsGuardtech.com/. NewsGuard is a tool that provides trust ratings for news and information websites. NewsGuard assesses the credibility and transparency of news and information websites based on nine journalistic criteria. These criteria are individually assessed and then combined to produce a single "trust score" from 0 to 100 for a given news media outlet. Scores are assigned by a team of journalists, not algorithmically. Scores are not given to platforms (e.g., Twitter, Facebook), individuals, or satire content. More detail regarding the rating process is available at https://www.NewsGuardtech.com/ratings/rating-process-criteria/. To complement news media trust scores, NewsGuard also provide a political leaning for news outlets (far left, slightly left, slightly right, far right) which allows us to gauge the ideological leaning of the news sources referenced in the COP Twitter discussion. Note that NewsGuard classifies a far larger set of news sources as slightly left of right, than far left or right. Using the database of news media trust scores, we cross reference the domains found in individual tweets with the corresponding trust score from the NewsGuard database. For COP21, 16% of all tweet URLs correspond to a domain with a trust score in the NewsGuard database, while for COP26 17% of URLs correspond to a domain with a score. The supplementary results include an expanded discussion of engagement with COP, including more detail on summits other than COP21 and COP26. We also provide a brief discussion of ideological polarisation on Youtube and Reddit to support the findings on Twitter. The bulk of the supplementary results expand on the latent ideology introduced in the main manuscript. Finally, we provide a brief analysis of French language Twitter to show that changes in polarisation are not specific to English language Twitter. In the supplementary methods, we outline the preliminary filtering required for the latent ideology and discuss the issue of deleted material and bots for the validity of our results. Supplementary Figure 1 shows the total number of daily posts referring to COP since Twitter was founded in March 2006. The first three COPs during this period (COP12 -COP14) were not discussed on Twitter. Between COP15 and COP19, some discussion around COP was active on Twitter, although the total number of posts are generally small compared to all COPs since COP20. Given the small datasets up to and including COP19, we exclude these conferences from our analysis in the main paper. there is some engagement, but overall content creation is far below the values observed during any subsequent conference. The limited size of the dataset prior to COP20 warrants this paper restricting our dataset to COP20 and onwards. Supplementary Figure 2 shows the engagement with COP between COP20 and COP26 on (a) Twitter, (b) Youtube, and (c) Reddit. Each subfigure shows the distribution of retweet counts or comments on each post or video that makes reference to that specific COP. In all cases, the data shows that COP engagement on Twitter is significantly higher than on either of the other two platforms. This is most clearly seen during COP20 where Twitter saw a total of 3.1 × 10 5 posts making reference to "cop20", while only 10 3 Youtube comments were on COP related videos, and 30 Reddit comments. This lack of data on Youtube and Reddit prevents a robust analysis of ideological polarisation on these platforms. However, we offer some brief qualitative observations in the following section. Reddit. The total number of retweets/ comments is shown in the inset. Engagement on Twitter is consistently at least one order of magnitude larger than on Youtube or Reddit. The lack of data for Youtube and Reddit prior to COP26 prevents a robust analysis of the evolving nature of ideological polarisation on these platforms. As seen in the engagement section, many users take part in the online debate around COP on Twitter, Reddit and Youtube. For Reddit and Youtube, we can qualitatively confirm the presence of fringes belonging to the ideological minority, although their number in terms of users is too small to allow us to deepen their study by applying the latent ideology method. In fact, for Twitter we observe in Supplementary Figure 4 that the latent ideology only gives consistent results if the number of influencers exceeds 200 accounts. On Youtube and Reddit, we are unable to extend the influencer set beyond ca. 60 and ca. 30 accounts respectively in a principled manner, largely because the bipartite network constructed to compute the latent ideology is no longer fully connected if we extend the influencer sets beyond these numbers. A lack of data and other limiting factors therefore prevent a robust, quantitative analysis of ideological polarisation on Youtube and Reddit. It is important to highlight that while it is possible to collect all posts on Reddit about a certain topic, the YouTube API provides only a sample of videos. First of all, the majority of Youtube channels in our dataset are not about climate, but their contents cover a wide range of topics (e.g. music, religion, news channels, tech). To get an idea of this effect, among the 100 channels with the most comments in our dataset, only 5 are climate specific. In our dataset, videos from those channels collect about 3% of the total comments. Therefore, ideology results for them are poorly related to the climate debate. Another issue for YouTube comes from the fact that, after selecting only English-language channels, many have a strong geographical focus presenting news specific to a given region (e.g. Al Jazeera English, South China Morning Post, WION). In this case, the geography of the channel is an important limiting factor that conflates the latent ideology. Another problem concerns the different number of comments in climate action channels vs. climate sceptic channels. In our Youtube dataset, the channel with the most interactions is Blackpink, a Korean k-pop group. They publish climate activism videos on their channel and the comments on these videos correspond to approximately 8% of all the comments in our dataset. Among the 100 channels with the most comments, channels which appear to be part of the ideological mainstream are UN Climate Change, Just Have a Think, Nick Breeze, ClimateGenn and Facing Future. They collect about 2% of the total comments. Conversely, Friends of Science is the most prominent channel which appears to oppose climate action, receiving around 0.1% of all comments. This difference between engagement with channels with contrasting ideological views limits the efficacy of the latent ideology. Finally, it is important to note that several important channels have comments turned off (e.g. the official COP26 Youtube channel). On Reddit, although we found subreddits dedicated to the climate discussion from both mainstream and minority points of view, only a small number of users refer to COP explicitly. Indeed, the share of submissions related to COP for the most debated conference (COP26), is below 4% for subreddits focused on climate topics endorsing the mainstream perspective such as "climate" (3.8%), "climatechange" (2.9%), and "environment" (1.8%), as well as for subreddits adhering to alternative views, such as "cimatedisalarm" (2%) and "climateskeptic" (1.8%). Moreover, Reddit data show a high separation in terms of geography with the presence of country-oriented subreddits. For example, in COP21 the third most used subreddit is "france", while in cop26 we find subreddits such as "Scotland", "NewsOfTheUK", "unitedkingdom" and other lesser referenced geographical subreddits. Here, we report the distribution of the political leaning of the news media outlets referenced by the majority and minority ideological groups on Twitter. We rely on the NewsGuard database to classify domains and assign them a political orientation. Results are shown in Supplementary Figure 3 . This shows that the ideological majority are dominated by left-leaning news sources, whereas the ideological minority are dominated by right-leaning sources. Note that the figure shows the fraction of news sources from each leaning category. The absolute number of links in the minority group is much larger in COP26 than in COP21. Supplementary Figure 4 shows how the polarisation statistic, D, changes with the number of influencers used to calculate the latent ideology in COP21 and COP26. We evaluate the D statistic for the users and the influencers starting with 20 influencers and increasing the number of influencers one at a time, recomputing the ideology each time until we reach 500 influencers. Influencers are ordered according to the number of retweets of each influencer (the influencer with the largest number of retweets first). The bold lines represent the mean D values at each influencer number. The shaded area indicates the 95% confidence interval. The figure shows that the latent ideology produces consistent polarisation measures if the number of influencers is larger than around 200 influencers. Larger influencer sets produce more consistent results, but for COP20, it is difficult to extend the influencer set beyond 300 accounts. Hence, we use 300 influencers for each COP. 1. Growing polarisation is driven by increased engagement, not by changing individual ideologies There are two plausible explanations for the increase in ideological polarisation seen between COP21 and COP26: 1. Users who were previously part of the ideological majority have now become a part of the ideological minority. 2. The ideological views of users already engaged with COP haven't changed. However, engagement from new users, who were previously not part of the COP discussion, has strengthened the ideological minority. Assessing which of these explanations is most likely is difficult using the unbalanced influencer sets in Figure 2 of the main paper. Therefore, we recompute the ideology using a balanced set of influencers. This requires (1) that an influencer takes part in both the COP21 and COP26 discussion, and (2) that the influencers selected are split equally between the ideological majority and minority groups. Given that the ideological minority is much larger for COP26 than COP21, we choose influencers based on their COP26 ideological score. Note, this does not mean that these influencers will, by necessity, fall into the same ideological group for COP21. Supplementary Figure 10 recomputes the latent ideology with this balanced set of influencers, merging the top 15 minority and majority influencers who appear in both the COP21 and COP26 discussions. Supplementary Figure 10 . The ideological spectrum calculated using a balanced set of influencers who appear in both COP21 and COP26. Influencer polarisation is similar between COP21 and COP26, but user polarisation increases significantly. This reflects a large increase in user engagement with the ideological minority during COP26. Hartigan's diptest shows that influencer polarisation is similar between COP21 and COP26 using the balanced influencer set (COP21: D = 0.20, 95% CI: [0.14, 0.23], p < 2.2 × 10 −16 ; COP26: D = 0.21, 95% CI: [0.14, 0.23], p < 2.2 × 10 −16 ). This indicates that minority influencers from COP26, who also engaged with COP21, were already part of the ideological minority (and likewise for majority influencers). However, user polarisation increases significantly between COP21 and COP26 ( It is interesting to note that if we focus on minority influencers that appear in both the COP21 and COP26 discussions, 7 of 15 have a specific climate focus. This exceeds the number of climate focused influencers in the minority if we simply use the 300 most retweeted accounts. Retweet networks are constructed for COP21 and COP26 where tweets are filtered for French language only. Note that due to a significant overlap between some prominent French accounts and the wider francophone discussion on Twitter, the COP26 influencer set requires filtering prior to computing the latent ideology. This is because interactions between these accounts and non-French francophone accounts lead to geographical conflation of the latent ideology. Supplementary Figure 11 . The French language ideological spectrum for COP21 (left) and COP26 (right). Top: A histogram of the influencer and user ideology scores for COP21 and COP26. The ideological minority map to −1, whereas the majority group map to +1. Bottom: The 30 most retweeted influencers and accompanying user ideology distributions. Influencers who are primarily retweeted by the ideological minority are written on the left, and influencers primarily retweeted by the ideological majority are on the right. The distributions shown alongside each influencer correspond to the distribution of user ideologies who retweeted that influencer. Expanded figure with all 300 influencers is available here. . This shows a clear increase in ideological polarisation on French language Twitter, despite the effect of local enhancement during COP21. The increase in polarisation is smaller than for English language Twitter, although there is evidence that, as in English, the change in polarisation is reflected in an increase in far right engagement with COP. For the COP21 influencer set, although none of the influencers self-identify as being climate focused, a majority of them self-identify as being right-wing partisans in France. @ntwolfmother and @Sisf94 mention being supporters of the former French president Nicolas Sarkozy and @f philippot is the president of the political movement Les Patriotes, which notably aims at cancelling the climate emergency France declared in 2019. Moreover, a large number of these influencers have been actively voicing their opposition to the sanitary pass, which was implemented in France from 9th June 2021, by using popular hashtags from the French opposition such as #AntiPass. One of these accounts, namely @EugenieBastie, works as a journalist for Le Figaro, which is one of the most consulted newspapers in France. She has also used her Twitter account to voice scepticism regarding Covid-19 vaccination and her opposition to left-wing social movements. Regarding COP26, more influencers fall within the ideological minority. One of them, @philippeherlin, published a book titled "Cancel economy : Why the energy transition is an economic catastrophe". This influencer set also contains some of the most influential French political actors from the far-right, such as Marine Le Pen (@MLP officiel) and Eric Zemmour (@ZemmourEric). During the last opinion polls for the 2022 French presidential election, 16% and 15% of the voters expressed their support respectively for Marine Le Pen and Eric Zemmour, which means that they are more likely to end up in the second round of elections than their left-wing counterparts. Two of these influencers are also the official pages of media outlets, namely Russia Today (@RTenfrancais) and Sud Radio (@SudRadio). The latter has hosted several French media personalities who expressed their hesitancy regarding the Covid-19 vaccination campaign. Other influencers, such as @AnonymeCitoyen and @Carene1984, also voice their disagreement with the French sanitary pass policy and Covid-19 vaccines. Overall, we notice that the contemporary debates about the sanitary crisis has become viral in the ideological minority, leading a set of notable political actors, media sources and common citizens to rally around right-wing social causes. F. Extended discussion of tweet content In Supplementary Figure 12 we show the results of the Shiftiterator method applied to aggregated Twitter hashtags from COP21 and COP26 editions. Panel (a) shows hashtags used by the ideological majority from COP21 and COP26, with those appearing disproportionately during COP21 on the left and those appearing disproportionately during COP26 on the right. Panel (b) shows the equivalent for the ideological minority. Supplementary Figure 12 . Pairwise comparisons between hashtags from different COP editions. Panel (a) presents the comparison of hashtags tweeted by users from the ideological majority and Panel (b) presents the comparison of hashtags used by the ideological minority. In both panels, hashtags on the left appeared disproportionately during COP21, and those on the right appeared disproportionately during COP26. The COP edition with the second most engagement is COP21. For a fair comparison between COP26 and COP21, it is necessary to explore what happens on French language Twitter as well as on English language Twitter. In Supplementary Figure 13 we show the results of the Shiftiterator applied to aggregated French Twitter hashtags within the same COP edition. For COP21 (Supplementary Figure 13 (a) ), hashtags are in general related to climate discussion themes. On the right side hashtags relate to local politics. For COP26 (Supplementary Figure 13 (b) ), climate themes are still present, alongside references to politicians and climate activists. Once again, Covid-19 remains a key focus. Supplementary Figure 13 . Pairwise comparisons between French hashtags within the same COP edition. For Figures (a) and (b), the results show the comparison between French Twitter hashtags used by the ideological majority (right side) and by the minority (left side). Supplementary Figure 14 further dissects the ideological minority by separating tweets which make explicit reference to "Covid-19" from those which do not, both for English and French hashtags. As we observe, there is a clear separation between climate themes on one side, and general right-wing themes on the other. The latent ideology can be conflated by geographical factors, as well as niche topics outside the main COP discussion. Therefore, in order to ensure that the latent ideology splits influencers based on climate ideology, rather than, for example, geography, we must remove some accounts from the top 300 most retweeted accounts used as our set of influencers. In some cases, it is also necessary to filter out accounts whose set of unique retweeters have no overlaps with any of the unique retweeters for the remaining influencer set. This is typically only an issue for small datasets. The accounts filtered out of the influencer set for each COP are listed below: • COP20: @GreenLifeStory due to disjoint set of retweeters. (Total removed: 1/300 influencers) • COP21: @narendramodi due to prominent cluster of accounts associated with Indian politics which have minimal overlap with the wider English language discussion. (Total removed: 1/300 influencers) • COP22: @BoardshortsBen, @Anggun Cipta, @lekimastores and @jpbussmann. All of these accounts are non-US/UK/Australia/Canada, which between them dominate the COP discussion. This conflation occurs in part due to very few accounts in the minority climate ideology. (Total removed: 4/300 influencers) • COP23: @earthtokens, an account which primarily interacts with the blockchain community. (Total removed: 1/300 influencers) • COP24: No filtering required. (Total removed: 0/300 influencers) • COP25: No filtering required. (Total removed: 0/300 influencers) • COP26: @narendramodi and @naftalibennett due to association with Indian and Israeli politics respectively. @cop26token and @Earthtoken io who primarily interact with the blockchain community. Note, a significant Korean cluster is observed in the latent ideology, but this does not obscure the divide between the minority and majority ideological positions on climate. Hence, following a principle of minimal interference, we do not remove any of the accounts in this Korean cluster. (Total removed: 4/300 influencers) The Twitter API does not allow access to tweets which have been deleted, or to accounts which have been suspended or removed. If a user retweets a tweet which is later deleted, individually, or because the account of the original author is suspended, then the retweet will also be removed. As a result, our datasets, particularly for earlier COPs, are likely to be missing some tweets. However, it is not possible to measure exactly how many tweets/ users have been deleted, and how this missing data will affect our results. To roughly approximate how missing data may affect our results, we note that if a tweet mentions a user, e.g., by using the tag @realDonaldTrump, that user tag remains even after the tagged users account is removed or suspended. Conveniently, these tags appear in the tweet metadata that is downloadable via the Twitter API, including those corresponding to suspended or removed accounts. Consequently, we can gauge the propensity of missing information by checking whether the accounts mentioned in our datasets were active on Twitter at the time our data was downloaded. The fraction of mentions corresponding to accounts which are suspended or removed are given below: • COP20. 1.71% • COP21. 2.57% • COP22. 1.90%. • COP23. 1.56%. • COP24. 1.06%. • COP25. 0.45%. • COP26. 0.15%. Note that these percentages are not the percentage of unique accounts which have been suspended, but rather the percentage of mentions. Therefore, if an account is mentioned multiple times (e.g.,@realDonaldTrump), it will appear multiple times in the percentage calculation. Given that these percentages fall far below the 12% of users found to be part of the ideological minority for COP26, deleted users cannot account for the significant increase in polarisation observed between COP21 and COP26. For the COP21 dataset, approximately 2.5% of mentions refer to deleted accounts for both the majority and minority groups. Supplementary Figure 15 shows the number of tweets from the ideological minority and majority which come from official Twitter sources, e.g., "Twitter for Android". Tweets from these official sources cannot be used to run bots which artificially enhance content engagement. The figure shows a steady increase in the overall fraction of tweets from official sources between COP20 and COP26, with minimal differences between the majority and minority groups, with the exception of COP21 where the minority group only sends around 55% of tweets from official sources. In COP26, more tweets in the ideological minority are from official sources than for the majority. Hence, the increase in engagement with the ideological minority between COP21 and COP26 cannot be explained by an increase in bot activity. In fact the reverse appears to be true; the COP26 ideological minority are the least dependent on bot activity relative to previous COPs. Supplementary Figure 15 . Percentage of tweets from the ideological majority (yellow bars) and the ideological minority (blue bars) coming from official Twitter clients. The percentage is calculated over the total count of tweets from both official and non official sources. The social media life of climate change: Platforms, publics, and future imaginaries Are social media making constructive climate policymaking harder? Concerning values: what underlies public polarisation about climate change? The politicization of climate change and polarization in the american public's views of global warming Individuals with greater science literacy and education have more polarized beliefs on controversial science topics Corporate funding and ideological polarization about climate change Politicization and polarization in climate change news content Scientific networks on twitter: Analyzing scientists' interactions in the climate change debate Climate change sentiment on twitter: An unsolicited public opinion poll Polarization of climate politics results from partisan sorting: Evidence from finnish twittersphere Twitter and climate change Climate change on twitter: Topics, communities and conversations about the 2013 IPCC working group 1 report Visualizing the paris climate talks on twitter: Media and climate stakeholder visual social media during cop21 Network analysis reveals open forums and echo chambers in social media discussions of climate change Social media, political polarization, and political disinformation: A review of the scientific literature The spreading of misinformation online The covid-19 social media infodemic The echo chamber effect on social media Science vs conspiracy: Collective narratives in the age of misinformation Szymanski, Shifting polarization and twitter news influencers between two us presidential elections Assessing the risks of 'infodemics' in response to covid-19 epidemics How to fight an infodemic Exposure to opposing views on social media can increase political polarization Ideological biases in social sharing of online information about climate change Media and climate change observatory special issue: A review of media coverage of climate change and global warming in 2020 Media attention for climate change around the world: A comparative analysis of newspaper coverage in 27 countries Climate journalism in a changing media ecosystem: Assessing the production of climate change-related news around the world Reliability of google trends: Analysis of the limits and potential of web infoveillance during covid-19 pandemic and for future research The utility of google trends data to examine interest in cancer screening Understanding polarization: Meanings, measures, and model evaluation Tweeting from left to right: Is online political communication more than an echo chamber? Birds of the same feather tweet together: Bayesian ideal point estimation using twitter data The dip test of unimodality Evidence-based strategies to combat scientific misinformation The who-to-follow system at twitter: strategy, algorithms, and revenue impact Fast unfolding of communities in large networks Generalized word shift graphs: a method for visualizing and explaining pairwise comparisons between texts Quantifying social organization and political polarization in online platforms Partisan asymmetry in temporal stability of climate change beliefs Comparing the language of qanon-related content on parler, gab, and twitter Negotiating climate change: Radical democracy and the illusion of consensus This is despite many Twitter users stating in their biography that retweets should not be understood as endorsements L'analyse des données Supplementary Figure 6. The latent ideology for COP22. Top: user and influencer ideology distributions using the top 300 influencers. Below: the top 30 (of 300 total) influencers with distributions showing their retweeters on the ideology scale Below: the top 30 (of 300 total) influencers with distributions showing their retweeters on the ideology scale Supplementary Figure 8. The latent ideology for COP24. Top: user and influencer ideology distributions using the top 300 influencers. Below: the top 30 (of 300 total) influencers with distributions showing their retweeters on the ideology scale All tweets referring Covid-19 are aggregated and compared to tweets that do not reference Covid-19. Hashtags which disproportionately coexist with references to Covid-19 are shown on the left side of each panel, whereas themes that do not coexist with Covid-19 are shown on the right side of each panel. Panel (a) shows the aggregated tweets for English language Twitter M.F., A.G., M.T., F.Z., W.Q. and A.B. acknowledge the 100683EPID Project "Global Health Security Academic Research Coalition" SCH-00001-3391. F.Z. acknowledges financial support from the European Union's Rights, Equality and Citizenship project EUMEPLAT grant no. 101004488. Expanded figures for the latent ideology are available for each COP at https://osf.io/nu75j/?view_only= 53e03939cd824bc680e83b7c64c80b27.Twitter and Youtube data is made available in accordance with Twitter and Youtube's terms of service. Tweet and Youtube video IDs for each COP are available at https://osf.io/nu75j/?view_only=53e03939cd824bc680e83b7c64c80b27. The corresponding tweets can be downloaded using the official Twitter API (https://developer.twitter.com/ en/docs/twitter-api). Youtube video metadata can be downloaded using Youtube's official API (https:// developers.google.com/youtube/v3). Reddit data was downloaded using the https://pushshift.io/ API, and is freely available to the public.