Algorithmic Persuasion in Human–LLM Interactions: How Conversational AI shapes Political Opinion Formation

The rapid integration of conversational artificial intelligence (AI) into citizens’ information environments has transformed how public opinion is formed and expressed. While previous scholarship has focused on broadcast media, interpersonal influence, and algorithmically curated social media, large language models (LLMs) introduce a new form of interactive, personalized political communication. Existing research has primarily examined LLMs as sources of misinformation or synthetic political content but has paid limited attention to how routine human–AI interactions reshape political beliefs, influence affective responses, and alter the broader distribution of public opinion. This study addresses that gap by conceptualizing conversational AI as an active site of opinion formation rather than merely a channel of information exposure. The study investigates four questions: (1) whether conversational AI systematically shifts, rather than reinforces, users' expressed opinions; (2) whether the emotional tone of AI responses moderates opinion change; (3) whether prolonged interactions produce stable or oscillatory belief trajectories; and (4) whether individual-level opinion change contributes to aggregate polarization with implications for democratic resilience. The analysis integrates Social Impact Theory, the Elaboration Likelihood Model, Zaller's Receive–Accept–Sample model, affective polarization research, and the Computers-as-Social-Actors paradigm to theorize LLMs as scalable belief-shaping agents capable of influencing both individual attitudes and the broader opinion landscape. Empirically, the study analyzes approximately 1.8 million human–AI exchanges drawn from four large open conversational datasets (OASST1, OASST2, LMSYS-CHAT-1M, and WildChat-1M), comprising around 400,000 users interacting with 35 LLMs. Using sentence-transformer embeddings, Bayesian-inspired opinion-update estimation, sentiment and emotion analysis, lexical entrainment measures, clustering techniques, and conversation-dynamics modelling, the study quantifies opinion change across successive user interactions. A politically relevant subsample focusing on electoral politics, immigration, climate change, and gender equality is used for confirmatory analyses. Five principal findings emerge. First, conversational AI consistently shifts users' expressed opinions rather than merely reinforcing prior beliefs. Second, emotionally positive AI responses generate significantly larger opinion changes than neutral responses, suggesting an affective pathway of algorithmic persuasion. Third, despite substantial conceptual movement, users exhibit minimal lexical imitation of AI language, indicating that conceptual influence occurs without linguistic convergence. Fourth, over half of extended conversations display oscillatory rather than linear belief trajectories, highlighting the dynamic nature of AI-mediated persuasion. Finally, although individual opinions become more diverse, the overall opinion space converges into a smaller number of stable clusters. We describe this phenomenon as the “entropy paradox,” whereby conversational AI simultaneously broadens individual exposure while reducing aggregate viewpoint diversity. The study contributes to public opinion research by operationalizing algorithmic persuasion at scale, extending dual-process theories of persuasion to human–AI interaction, and identifying a novel entropy-reduction mechanism distinct from selective exposure and homophilous network effects. These findings have important implications for democratic resilience, suggesting the need for greater transparency in AI persuasive design, sentiment-aware auditing frameworks, AI literacy initiatives, and computational indicators capable of monitoring opinion shifts in increasingly AI-mediated public spheres.