The proliferation of climate change misinformation on social media platforms poses a significant challenge to efforts to mitigate climate change and foster public understanding of the scientific consensus. Misinformation undermines public support for climate action, hinders effective communication of accurate information, and creates a self-silencing effect among individuals who hold scientifically accurate views but fear expressing them due to the perceived prevalence of contrarian viewpoints. The decentralized nature of the internet and algorithms that prioritize engagement exacerbate the spread of misinformation. Social media platforms act as conduits, mainstreaming contrarian claims, which then may be adopted by news outlets and political actors. The issue is further compounded by advancements in generative AI, which can dramatically increase the scale and frequency of false narrative creation. Given this escalating problem, automated detection of climate change misinformation presents a crucial need. Previous research has shown that contrarian tweets and polarization have grown substantially, with hoax-themed tweets prevalent in certain geographical regions. Studies have also revealed the use of aggressive language and incivility in climate change denial tweets. This necessitates the development and deployment of effective interventions to counter these negative impacts, making automated detection and response mechanisms particularly relevant.

Existing research highlights the negative consequences of climate change misinformation, including reduced public support for mitigation policies and the self-silencing effect of perceived contrarian prevalence. The role of social media in amplifying misinformation has been widely studied, with concerns raised about its impact on trust in scientific and journalistic expertise. Studies have analyzed climate change discussions on Twitter, identifying growing polarization and the geographical distribution of hoax-themed tweets. The use of incivility and aggressive language in contrarian tweets has also been documented. Prior work has explored automated approaches to misinformation detection across various domains, but these efforts are often domain-specific. The CARDS (Computer Assisted Recognition of Denial and Skepticism) model, a supervised machine learning approach, represents a notable previous attempt to categorize contrarian climate claims but primarily focused on blogs and conservative think-tank websites, and its applicability to social media data has not been fully evaluated. This study builds upon that foundation by enhancing the model's performance on Twitter data.

This study enhances the existing CARDS model by creating an 'Augmented CARDS' model. The Augmented CARDS model uses a two-stage hierarchical architecture. The first stage is a binary classifier that distinguishes between convinced and contrarian tweets using a DeBERTa language model. This addition allows the model to first filter out tweets aligned with the scientific consensus before classifying the remaining tweets into specific categories of misinformation. This initial binary classification step improves the model's ability to function effectively in the noisy environment of Twitter, where a majority of tweets agree with the scientific consensus. The second stage employs a multi-label classifier, also based on the DeBERTa language model, to classify contrarian tweets based on the CARDS taxonomy (Fig. 1). This taxonomy categorizes contrarian claims into five main categories and further subcategories. The model is trained on a dataset that includes both tweets and text from previous sources used to train the original CARDS model, which helps mitigate class imbalances inherent in Twitter datasets where convinced claims heavily outweigh contrarian claims. The inclusion of Twitter data in the training data enhances the model’s ability to capture the linguistic nuances specific to the platform. The models were fine-tuned for 3 epochs using a learning rate of 1e-5, a batch size of 6, and input sequences constrained to 260 tokens. The performance of the Augmented CARDS model was assessed using an expert-annotated test set of tweets, comparing its performance to the original CARDS model. Additionally, a large dataset (over 5 million tweets) of climate-related tweets collected between July and December 2022 from the Online Media Monitor (OMM) at the University of Hamburg was used to examine temporal trends in climate contrarianism and to identify the events that correlate with peaks in contrarian claims. Word frequency analysis was used to identify shifts in word usage during periods of high contrarian activity. The analysis included comparisons of the log-fold change and p-values for word frequencies. These methods were used to determine and characterize the correlation between specific events and the emergence of peaks in contrarian tweets.

The Augmented CARDS model demonstrated significantly improved performance compared to the original CARDS model in classifying contrarian claims from Twitter data. In the binary detection task (distinguishing between convinced and contrarian claims), the Augmented CARDS model showed better performance on Twitter data, while maintaining similar accuracy on the original data set consisting of contrarian blogs and conservative think-tank articles. In taxonomy detection (identifying specific types of contrarian claims from the CARDS taxonomy), Augmented CARDS performed substantially better on Twitter data than the original CARDS model. This improvement is attributed to the inclusion of Twitter data in the training set and the hierarchical model architecture, which effectively addresses class imbalances. Analysis of over five million climate-related tweets from 2022 revealed that over half of contrarian claims involved attacks on climate actors (category 5.2) or conspiracy theories (category 5.3). Four distinct categories of events were identified as stimuli for increased climate contrarianism: (1) Natural Events (e.g., Hurricane Ian); (2) Political Events (e.g., Biden's climate emergency declaration, COP27); (3) Contrarian Influencers; and (4) Convinced Influencers. Natural events led to a surge in claims that extreme weather events are unrelated to climate change. Political events were associated with increased criticism of climate policies. Influencer activity, regardless of stance, led to an increase in conspiracy theories. The analysis also showed the existence of accounts spreading misinformation, with many posting numerous repetitive tweets with minor variations in wording, indicating possible automated generation.

The Augmented CARDS model addresses a limitation of the original CARDS model, improving its ability to classify contrarian claims within the context of diverse social media content. The findings highlight the significant role of ad hominem attacks and conspiracy theories in climate change misinformation on Twitter, which contrasts with the emphasis on policy claims found in other sources like conservative think-tanks and contrarian blogs. The identification of four distinct stimuli—natural events, political events, and contrarian/convinced influencer activity—provides valuable insights into the dynamics of climate misinformation spread and offers targets for more effective interventions. The identification of accounts exhibiting high volumes of repetitive tweets raises concerns about the use of automation in disseminating misinformation. These findings demonstrate the effectiveness of this new model in detecting misinformation on Twitter, opening up a new line of inquiry into the specific factors that drive misinformation spread in the context of climate change.

This study demonstrates a significant advance in the automated detection of climate change misinformation on Twitter. The Augmented CARDS model outperforms the original CARDS model in classifying contrarian claims in the complex context of Twitter. The findings highlight the prevalence of ad hominem attacks and conspiracy theories and identify four key stimuli triggering misinformation spikes. This research provides valuable insights for developing more effective interventions and strategies to combat climate change misinformation. Future research should focus on expanding the model's capabilities to other social media platforms and languages, analyzing long-term trends, and integrating logical fallacy detection to enhance the accuracy and effectiveness of automated debunking.

The study focuses solely on Twitter data, limiting the generalizability of findings to other social media platforms. The model is currently trained only on English language text, necessitating further development for multilingual applications. The time frame of the data analysis is relatively short, preventing conclusive assessment of long-term trends and the impact of recent platform changes. The accuracy of the model's classification might be affected by the use of sophisticated AI-generated text, which could evade detection.