Computer-assisted classification of contrarian claims about climate change A growing body of scholarship investigates the role of misinformation in shaping the debate on climate change. Our research builds on and extends this literature by (1) developing and validating a comprehensive taxonomy of climate contrarianism, (2) conducting the largest content analysis to date on contrarian claims, (3) developing a computational model to accurately classify specific claims, and (4) drawing on an extensive corpus from conservative think-tank (CTTs) websites and contrarian blogs to construct a detailed history of claims over the past 20 years. Our study finds that the claims utilized by CTTs and contrarian blogs have focused on attacking the integrity of climate science and scientists and, increasingly, has challenged climate policy and renewable energy. We further demonstrate the utility of our approach by exploring the influence of corporate and foundation funding on the production and dissemination of specific contrarian claims. Organized climate change contrarianism has played a significant role in the spread of misinformation and the delay of meaningful action to mitigate climate change1. Research suggests that climate misinformation leads to a number of negative outcomes such as reduced climate literacy2, public polarization3, canceling out accurate information4, reinforcing climate silence5, and influencing how scientists engage with the public6. While experimental research offers valuable insight into effective interventions for countering misinformation3,7,8, researchers increasingly recognize that interdisciplinary approaches are required to develop practical solutions at a scale commensurate with the size of online misinformation efforts9. These solutions not only require the ability to categorize relevant contrarian claims at a level of specificity suitable for debunking, but also to achieve these objectives at a scale consistent with the realities of the modern information environment.An emerging interdisciplinary literature examines the detection and categorization of climate misinformation, with the vast majority relying on manual content analysis. Studies have focused on claims associated with challenges to mainstream positions on climate science (i.e., trend, attribution, and impact contrarianism)10,11, doubt about mitigation policies and technologies12,13, and outright attacks on the reliability of climate science and scientists14,15. Researchers, moreover, have examined the prevalence of contrarian claims in conservative think tank (CTT) communications14,16, congressional testimonies17,18, fossil fuel industry communications19, and legacy and social media20,21. Given the significant costs associated with manual approaches for content analysis, several recent studies have explored computational methods for examining climate misinformation, ranging from applications of unsupervised machine learning methods to measure climate themes in conservative think-tank articles15,22, to supervised learning of media frames such as economic costs of mitigation policy, free market ideology, and uncertainty23.Our work builds on and extends existing computational approaches by developing a model to detect specific contrarian claims, as opposed to broad topics or themes. We develop a comprehensive taxonomy of contrarian claims that is sufficiently detailed to assist in monitoring and counteracting climate contrarianism. We then conduct the largest content analysis of contrarian claims to date on CTTs and blogs—two key cogs in the so-called climate change ‘denial machine’24—and employ these data to train a state-of-the-art deep learning model to classify specific contrarian claims (Methods). Next, we construct a detailed history of climate change contrarianism over the past two decades, based on a corpus of 255,449 documents from 20 prominent CTTs and 33 central contrarian blogs. Lastly, we demonstrate the utility of our computational approach by observing the extent to which funding from ‘dark money’25, the fossil fuel industry, and other conservative donors correlates with the use of particular claims against climate science and policy by CTTs.Figure 1 displays the taxonomy used to categorize claims about climate science and policy commonly employed by contrarians. To develop this framework, we consulted the extant literature on climate misinformation to identify relevant claims, while further extending and refining this initial set by reading thousands of randomly selected English language paragraphs from prominent CTTs and contrarian blogs (see Supplementary Tables S4 and S5). This process yielded five major categories: (1) it’s not happening, (2) it’s not us, (2) it’s not bad, (4) solutions won’t work, and (5) climate science/scientists are unreliable. We describe these categories as the five key climate disbeliefs, mirroring the five key climate beliefs identified in survey research26. Nested within these top-level categories were two sub-levels (27 sub-claims, 49 sub-sub-claims), allowing a detailed delineation of different specific arguments (see Supplementary Methods for additional information on how we developed the taxonomy). This work is, to our knowledge, the first framework incorporating climate science misinformation, arguments against climate solutions, and attacks undermining climate science and scientists in a single, comprehensive taxonomy.Figure 1Taxonomy of claims made by contrarians. This figure displays the three layers of claim-making by climate change contrarian actors. The original version of this taxonomy with more detailed claim descriptions can be found in Supplementary Table S2.Yet developing a comprehensive taxonomy presents a number of conceptual challenges. Distinguishing between claims that are best described as skeptical (i.e., expressing a reasonable level of doubt based on available evidence), contrarian (i.e., contrary to mainstream views), or outright climate misinformation is a demanding task27. The claims presented in Fig. 1 generally fall into three groups. First, many of the claims—and the majority of claims in categories 1–3—have been shown to contain reasoning fallacies28 and thus may be confidently labeled as misinformation. Second, a number of claims are factual statements and are only contrarian when used as a rhetorical device to express doubts on the scientific basis for climate change and the need to take policy action. For instance, the ‘weather is cold somewhere on a certain day’ may be factually correct, while ‘the weather is cold today, therefore global warming is not happening’ is not. And while climate models are uncertain, this does not imply that all climate science is unreliable. Lastly, our taxonomy includes several claims that well-known contrarians tend to make, yet are not necessarily contrary to mainstream views. For instance, both contrarian and mainstream advocates have expressed concerns that ‘CCS is unproven’ (4.4.2), and a number of scholars argue that nuclear energy (4.5.3) can make significant contributions to climate mitigation efforts29. It is important to note, however, that claims which fall into this third group constitute only a small share of the total claims in the taxonomy; the overwhelming majority of claims directly challenge mainstream views on climate science and policy.Climate change contrarianism over the past two decadesWe developed custom software to harvest all English language textual content from 33 prominent climate contrarian blogs and the climate-related content of 20 conservative think-tanks over the period from 1998 to 2020. Supplementary Tables S4 and S5 provide a full list of the blogs and CTTs included in this study, as well as the number of documents provided by each source. In total, we collected 255,449 climate change relevant documents—which contain over 174 million words (tokens)—from these 53 sources over the studied time period. Almost all of the CTTs (95%) and the majority of blogs (64%) were from the United States. The only non-US CTT was Canadian while there were a number of non-US blogs (Australia, 12%; Iceland, 6%; New Zealand, 6%; Canada, 3%; Czech Republic, 3%, Germany, 3%; and UK, 3%). Supplementary Figure S1 illustrates the total document frequencies over time, offering the monthly counts of documents for blogs and CTTs.The 20 most prominent CTTs were identified in previous literature on organized climate contrarianism14,15. The selection criteria of the 33 contrarian blogs were based on (1) the list of central contrarian actors presented by Sharman30 and (2) the Alexa Rank for each blog. Note that the Alexa Rank score is calculated based on the number of daily visitors and pageviews over a rolling 3 month period. The score provides a rough estimate of the popularity of a particular website. While our list of blogs ((n = 33)) does not capture the entire contrarian blogosphere, it does cover a large proportion of the movement’s most prominent actors, including 139,912 blog posts over the period 1998 to 2020.With these data in hand, we adopted a supervised learning approach to classify relevant claims by (1) employing a team of climate-literate coders to categorize a sample of 87, 178 paragraphs along the three levels specified in our taxonomy (Methods and Supplementary Methods) and (2) training a model to accurately classify around 4.6M paragraphs from our corpus of contrarian blogs and CTTs.Figure 2 provides the prevalence of the five key climate disbeliefs for CTTs (Fig 2b) and blogs (Fig 2c) over time, while also providing the distribution of claim prevalence across relevant sub-claims (Fig 2a). The figure offers insights into the key similarities and differences in claims across contrarian blogs and CTTs, as well as the evolution of claims over time. In general, CTTs focus predominantly on the shortcomings of climate solutions (category 4) and attacks on climate science and scientists (category 5). While the initial years of the series were marked with approximately equal levels of emphasis on these two categories, category 4 gained prominence following 2008. This shift in the focus of the (mainly US-based) CTTs coincides with the transition of power from Republican to Democratic hands and the corresponding threat of climate legislation: in 2007, for the first time since 1993, the Democrats obtained a majority in both congressional chambers and in 2008 Senator Obama, consistently leading the presidential election opinion polls, promised comprehensive climate legislation in his presidential campaign. However, category 4 claims have dominated the CTT discourse for the remainder of the sample period, indicating a more permanent shift towards attacks on climate solutions. Blogs, on the other hand, have consistently devoted the largest share of their claims to attacking climate science and scientists. Yet, even for blogs, discussion of climate policy has risen over the last decade while challenges to the reliability of climate science and the climate movement have been on a downward trend, indicating that future contrarian claims are likely to increasingly focus on climate solutions.For both CTTs and blogs, claims which outright deny the existence and severity of anthropogenic climate change (categories 1–3) have been stable or have declined in relative terms in recent years. Claims for categories 1–3 are much more likely to be present in blogs than in CTT materials, although the pre-2010 period exhibited non-trivial levels of these claims even among CTTs. These results suggest that the blogs seem to be acting as the pseudo-scientific arm of the climate change counter-movement, with authors from this corpus being more likely to offer alternative explanations for scientific observations and predictions found within the climate science literature. This result is consistent with social network analysis finding the most central networked contrarian blogs are focused on science rather than policy30.Figure 2Prevalence of super- and sub-claims by CTTs and contrarian blogs. (a) illustrates the share of claim-making paragraphs related to the sub-claims of our taxonomy by CTTs (circle) and blogs (hollow square). (b) and (c) Display the share of 515,005 claim-making paragraphs devoted to the following super-claim categories: 1. Global warming is not happening (green hollow circle), 2. Humans are not causing global warming (yellow diamond), 3. Climate impacts are not bad (blue filled square), 4. Climate solutions won’t work (black circle), and 5. Climate movement/science is unreliable (orange hollow square). Note that estimates prior to 2007 in (c) are derived from a relatively small number of blogs.A significant advantage of our model is that it can detect claims at a more granular level, which allows us to determine which lower-level claims are driving the macro disbelief trends described above. Figure 2a visualizes the prevalence of selected sub-claims over the entire time period in CTTs (circles) and blogs (boxes), with the list sorted by CTT sub-claim prevalence. Here, we see how the driver of the category 4 arguments made by CTTs has been the claim that mitigation and adaptation measures will be harmful to the economy, environment, and society more generally. Category 5 claims were also prominent in both corpora; however attacks on the science and the climate movement were roughly equally frequent among the blogs, whereas CTTs were more likely to focus on attacking the movement by accusing climate scientists and activists of being alarmist and biased. Note that due to the thematic overlap between sub-claims 5.2 (Movement is unreliable) and 5.3 (Climate is a conspiracy), we collapsed these claims into a single measure both when training our model and presenting results. Further, our results show how the most common sub-claim for both CTTs and blogs not covered by categories 4 or 5 is that observed climate change is simply due to natural cycles.A closer look at conservative think tank climate messagingNext, given the considerable attention paid to CTT discourse in the literature on organized climate contrarianism14,15,22,24,31,32, we offer a more detailed examination of the specific claims of these organizations over two decades. Figure 3a examines the dynamics of two prominent policy-related sub-claims—’Climate policies are harmful’ and ‘Clean energy won’t work’—while also overlaying major US climate policy events, from the 2003 Climate Stewardship Act to the Obama administration’s Clean Power Plan. The highlighted sections of Fig. 3a indicate the relevant beginning and ending dates for these efforts, with the most common being the introduction of and voting on a Congressional bill. The figure demonstrates that claims on the harmful effects of climate policy, particularly for the economy, closely align with changes in the US policy environment: CTTs tend to first ramp up discussion following the announcement of a bill, and then again prior to a bill reaching the floor for a vote. Particularly salient is the spike in policy claims in late 2009, which not only coincided with intense debate on the American Clean Energy and Security Act (ACES), but also with the COP15 climate summit in Copenhagen. This summit was billed as an especially consequential meeting for progressing mitigation policies. Claims that challenge the efficacy of clean energy, however, appear less sensitive to policy events and yet have increased considerably over time, with the second quarter of 2020 representing the highest share of these claims to date. Notably, this trend runs counter to the plummeting cost of renewable energy production33.Figure 3Prevalence of selected contrarian sub-claims in CTT communication. This figure illustrates the temporal variation (quarterly) in the proportion of sub-claims found in CTT documents related to (a) ‘Climate policies are harmful’, ‘Clean energy won’t work’, and (b) ‘Climate movement is unreliable’, ‘Climate science is unreliable’. Highlighted periods in the time series include: (A) 2003 Climate Stewardship Act; (B
https://www.nature.com/articles/s41598-021-01714-4
Computer-assisted classification of contrarian claims about climate change
