Intersection of Big Five Personality Traits and Substance Use on Social Media Discourse: AI-Powered Observational Study.

Background: Personality traits are known predictors of substance use (SU), but their expression and association with SU in digital discourse remain largely unexamined. During the COVID-19 pandemic, the online social engagement heightened and led to an amplification in SU rates, thereby creating a unique natural opportunity to investigate these dynamics through large-scale digital discourse data. In our study, we offer insights beyond traditional self-report methods, which are crucial for developing timely and targeted public health interventions. Objective: We aim to evaluate whether the associations between the Big Five personality traits and SU discourse shifted during the 2019-2021 period, and to conduct a focused analysis of how these traits predict SU and relate to specific substance types, emotional expression, and demographic factors. Methods: We analyzed a corpus of several hundred million public posts from a major social media platform from 2019 to 2021. Using a pipeline of natural language processing and deep learning models, we identified SU-related posts and subsequently extracted scores for the Big Five personality traits, emotions, and user demographics. We used trend analysis to compare annual shifts in trait-SU associations, while detailed 2020 data underwent rigorous modeling using logistic regression, correlation analysis, and topic modeling to elucidate the core relationships. Results: Our analysis revealed that Extraversion (odds ratio [OR] 3.22, 95% CI 2.98-3.49) and, most strikingly, agreeableness (OR 4.04, 95% CI 3.71-4.41) were the strongest positive predictors of being a substance user. In stark contrast to the conventional self-medication hypothesis, neuroticism emerged as a robust protective factor against SU (OR 0.29, 95% CI 0.26-0.31). This counterintuitive finding was supported by a decreased association between neuroticism and SU posts at the pandemic's onset in 2020 (Cohen d=-0.13, 95% CI) and a negative correlation with the expression of negative emotions online. Topic modeling further indicated that SU discourse was frequently embedded in social contexts (social drinking and friendly beverage choices) rather than themes of solitary coping. Conclusions: Our findings challenge traditional models by demonstrating that in large-scale online discourse, SU expression is more powerfully linked to social-affiliative traits than to negative emotionality. The paradoxical protective role of neuroticism suggests that established risk profiles may not apply uniformly to digital environments, particularly during a public health crisis. These insights are vital for refining computational methods for public health surveillance and developing interventions that recognize the potent social drivers of SU in the digital age.