Using large-language models to analyse social media engagement with the public sector communication on circular economy: Evidence from X

This study examines how public sector communication on circular economy (CE) initiatives influences social media engagement by analysing message elaboration, emotional tone (positive, negative, neutral), and message purpose (educational, advocacy, and policy-oriented). Grounded in the Elaboration Likelihood Model (ELM), we collected over 20,000 X (formerly Twitter) posts from 2015 to 2024 using Apify and classified them through zero-shot Large Language Model (GPT-4o-mini) analysis, validated against human coders. To establish causal effects rather than correlations, we applied a two-step approach: DirectLiNGAM for causal discovery and DoWhy for causal estimation with robustness tests. Results show that high-elaboration messages foster deeper engagement, negative tones evoke urgency; neutral tones support trust and advocacy-driven responses, while positive tones yield limited effects. Educational messages drive long-term behavioural change through systematic information processing; advocacy messages generate emotional resonance for rapid engagement, and policy-oriented messages enhance transparency but lack significant emotional impact. The study advances ELM by integrating emotional tone and message purpose into sustainability communication. Methodologically, it demonstrates how LLMs combined with causal inference methods can systematically analyse large-scale social media data. Practically, the findings provide actionable guidance for designing effective public campaigns to strengthen citizen engagement with CE goals. • Integrates emotional tone and message intent into the Elaboration Likelihood Model for CE engagement. • Proposes a framework combining GPT-4o-mini, DirectLiNGAM, and DoWhy for causal inference on X data. • Finds fact-based, high-elaboration messages drive deep cognitive and sustained behavioural engagement. • Negative tones trigger short bursts of engagement; neutral tones foster trust and long-term advocacy. • Educational posts shape lasting change; advocacy sparks quick reactions; policy messages yield highest impact.