Leveraging artificial intelligence to identify the psychological factors associated with conspiracy theory beliefs online

The COVID-19 pandemic highlighted the significant societal impact of conspiracy theories, primarily disseminated through social media. These theories, often lacking credible evidence, can profoundly influence public health, impacting compliance with preventative measures and promoting dangerous alternative treatments. They also negatively affect intergroup relations and societal polarization. While research has explored psychological factors associated with conspiracy belief, it has largely relied on self-reported data, lacking integration with large-scale behavioral insights from social media. This study bridges this gap by combining self-reported psychological data with an AI-driven analysis of millions of social media engagements to identify risk factors for online support of conspiracy theories. The study focuses on US-based Twitter (currently X) users due to the platform's high volume of COVID-19 conspiracy theory content during the pandemic. Understanding these factors is crucial for developing targeted interventions to mitigate the spread of misinformation and harmful beliefs.

Existing research suggests several psychological factors associated with conspiracy theory beliefs. Narcissism, the need for chaos, and denialism have been consistently linked to these beliefs in self-report studies. Narcissistic individuals might find conspiracy theories appealing for attention-seeking and self-affirmation. Denialism, characterized by the rejection of expert narratives, frequently accompanies conspiracy beliefs. The need for chaos reflects a desire to provoke disorder. Political alignment also plays a significant role, with individuals at both extreme ends of the political spectrum exhibiting higher rates of conspiracy belief. The Theory of Reasoned Action suggests that attitudes toward a behavior and normative beliefs influence it; therefore, trust in information sources and confidence in identifying false information are relevant. Finally, conspiracy mentality and susceptibility to misinformation, as measured by specialized scales, are also considered key factors. However, the literature presents inconsistencies, with some studies emphasizing personality factors while others highlight distrust in institutions. The over-reliance on self-reported data limits the understanding of the actual online behavior associated with these beliefs. This study aims to address this limitation by integrating self-report data with large-scale behavioral data from social media.

This study utilized a unique dataset merging self-reported psychological data from 2506 U.S.-based Twitter (currently X) users with machine learning analysis of their 7.7 million engagements during the COVID-19 pandemic (December 1, 2019 – December 31, 2021). Participants completed an online survey including various psychometric measures (political orientation, political affiliation, need for chaos, narcissism, denialism, conspiracy mentality, misinformation susceptibility, and Theory of Reasoned Action related factors). They also granted access to their Twitter (currently X) activity. A two-stage machine learning approach was used to assess behavioral support for six overarching COVID-19 conspiracy theories. First, a sentence transformer model was used to identify engagements semantically related to each theory (similarity score > 0.25). This reduced the computational burden for the second stage. Second, an OpenAI GPT 3.5 model classified each relevant engagement as either supporting or not supporting the theory. The model's performance was validated against human annotations. Finally, the engagement data were merged with individual-level psychological data, creating a multi-level dataset. Bernoulli multi-level generalized linear mixed models were used to analyze the relationships between psychological variables and behavioral support for each conspiracy theory. Age, education, gender, political orientation, political party affiliation, and various psychological measures were included as predictors, controlling for the number of followers and accounts followed.

The analysis revealed that older age was the most consistent risk factor across all six conspiracy theories. Participants identifying as very left or very right on the political spectrum exhibited significantly higher behavioral support for three out of six theories. Belief in false information consistently predicted support for three theories. Denialism was associated with support for the theory that the public is being intentionally misled. Perceived ability to recognize misinformation was positively associated with support for the theory that governments are intentionally spreading false information. However, conspiracy mentality and narcissism did not show significant associations. The number of Twitter (currently X) engagements supporting each theory varied significantly, with the highest support for theories about government misinformation and the public being misled about the virus. The overall variance explained in the models was relatively low, ranging from 7% to 22%, suggesting other factors may be at play. The intraclass correlation coefficient indicated that about half the variance was attributable to the participant level. The study found limited evidence to support the idea that education serves as a resilience factor or that being male is a risk factor.

This study provides valuable insights into the psychological factors associated with online support for conspiracy theories. The findings highlight the importance of age and political extremism as risk factors. The consistent association between belief in false information and conspiracy support underscores the significance of this factor. The results partially support existing research while also identifying limitations of some previously proposed measures. The finding that confidence in recognizing misinformation was associated with increased support for one theory warrants further investigation, possibly related to overconfidence. The study’s relatively low explained variance suggests additional factors beyond those examined here may influence conspiracy theory endorsement. The limitations of self-report measures are highlighted by the study's findings which demonstrate that only some factors proposed in the self-report studies are also apparent in actual online behavior. The integration of large-scale behavioral data with self-report data, therefore, enhances the ecological validity of findings. The limitations of the study, which mostly involve the limitations of the data used, are discussed further in the separate limitation section.

This research demonstrates the value of combining AI-powered analysis of large behavioral datasets with self-report surveys to study online phenomena. The findings confirm some previously identified risk factors while revealing new insights and limitations of existing measures. Future research should explore the interplay of various factors influencing online conspiracy theory spread and the impact of these beliefs on real-world outcomes. Moreover, future studies could examine additional variables proposed in recent literature and account for the influence of social media platforms themselves on the spread of misinformation.

The study's reliance on Twitter (currently X) data may limit its generalizability to other platforms or contexts. The relatively low explained variance suggests that additional, unmeasured factors may contribute to conspiracy theory support. The study's cross-sectional design prevents causal inferences. Selection bias might have influenced the sample. The fact that Twitter was actively combating COVID-19 misinformation during the study period could limit the detection of overt support for certain theories. Finally, the study is limited by the relative lack of accuracy of the individual scales used, specifically the scale assessing belief in false information that only demonstrated acceptable reliability. Future research could investigate other methods of measuring support for conspiracy theories.