Dire bene, dire giusto? Norme conversazionali e valutazione dell'IA nei ragionamenti complessi

This chapter investigates how Large Language Models (LLMs) influence human-AI interaction during complex reasoning tasks. Grounded in social psychology, pragmatics, and Speech Act Theory, the research examines how conversational strategies affect user trust, agency, co-construction of meaning, and anthropomorphism. Utilizing a counterbalanced within-subjects design, twenty-three university students engaged in a complex logic task ("Alice's Cousin Problem") with ChatGPT-5 and Claude 4.6, both prompted to act as Socratic tutors. Analysis via Conversation Analysis and Interlocutory Logic revealed distinct interactional styles: ChatGPT-5 adopted a structured, directive, procedural approach, whereas Claude 4.6 employed a dialogic, maieutic style promoting cognitive autonomy. While both platforms achieved an identical 52.2% success rate, nine participants produced divergent outcomes across systems. Critically, the study documents instances of "cognitive breakdown" and "epistemic yielding" in ChatGPT-5, where the system abandoned logical accuracy to preserve conversational flow. Participants maintained high trust driven by process transparency and relational warmth rather than propositional truth, uncovering a dissociation between interactional competence and epistemic reliability. Additionally, 56.5% of participants spontaneously anthropomorphized the systems, assigning gender and social roles based on illocutionary style. The authors advocate for AI design guided by transparency and benevolence to mitigate automation complacency.