ARCHIVES

Large language models are increasingly embedded within everyday conversational environments, yet most evaluations of conversational AI continue to privilege informational accuracy over interactional structure. This paper argues that the epistemic effects of conversational AI emerge not only from what systems say, but from how conversational trajectories are sustained across interactional time. Building on interactional sociolinguistics, conversation analysis, and recent critiques of alignment optimization, the study introduces the Alignment–Epistemic Drift Model (AEDM), a minimal heuristic framework for examining the relationship between alignment, friction, and interpretive stabilization in human–AI dialogue. The analysis draws on a small annotated corpus of Bengali conversational interactions comprising two contrasting regimes: affective ChatGPT-style interactions and transactional customer-service chatbot exchanges. Each conversational turn is coded along two dimensions—alignment and friction—using a lightweight interactional annotation scheme. Drift is operationalized heuristically through the directional formulation 𝐷 ≈ 𝐴ˉ − 𝐹ˉwhere sustained alignment combined with reduced friction increases the likelihood of interpretive stabilization across conversational sequences. The findings reveal a consistent structural asymmetry. Affective interactions exhibit persistent alignment and minimal friction, producing positive drift trajectories and recursive stabilization of user framing. Transactional interactions, by contrast, display structurally embedded interruption, procedural constraint, and recurrent friction, thereby suppressing drift accumulation. Importantly, these differences emerge independently of propositional correctness, suggesting that epistemic effects in conversational AI are shaped as much by interactional continuity as by informational content. The paper further proposes the Bengali Fallacy Hypothesis, which identifies a conditional amplification effect between alignment-heavy conversational systems and sociolinguistic environments where politeness, relational continuity, and indirectness carry epistemic weight. Rather than attributing cognitive susceptibility to users, the argument locates epistemic drift within interactional structure itself. The study contributes to emerging debates on conversational AI by shifting analytical attention from isolated outputs toward the temporal organization of alignment and friction in computationally mediated interaction.