Summary
Large language AI models often give advice, but it’s unclear how they react when users disagree with them. The authors studied how these models keep or give up their claim to knowledge when challenged and created a system to categorize different types of user disagreements and AI responses. They found that many models try to keep their original answer even while agreeing with the user and sometimes apologize without changing their advice. Models are more likely to hand over authority to the user in sensitive areas like health or legal advice than in straightforward fact-checking or explanations. This work helps us understand how AI systems manage authority in conversations and suggests ways to evaluate their responses better.
epistemic authoritylarge language modelsuser disagreementconversation analysischallenge typesevidential supportadvice tasksapology in AIauthority transferclaim maintenance
Authors
Riyadh Alnasser, Yusuf Mücahit Çetinkaya, Sumin Zhao, Tuğrulcan Elmas
Abstract
Large language models are increasingly used as sources of advice and information, including in high-stakes settings, yet little is known about how they respond to user disagreement. We study how a model manages its epistemic authority, referring here to its claim to knowledge, competence, or the right to advise, once a user challenges its answer. Building on Conversation Analysis, we introduce a taxonomy of six challenge types and a four-layer framework for analysing each response: whether the original claim is maintained or changed, where authority is located, how the disagreement is socially managed, and what kind of evidential support is offered. We construct a new dataset of 2,310 controlled challenge scenarios and 32,340 corresponding responses from 14 models, and analyse them using our framework with an LLM-as-judge pipeline, providing a vocabulary which future evaluation and benchmark design can build on. We find that models show conflicting behaviour: they validate users in 85% of responses but maintain their original claim in 65%. They explicitly apologise in 33% of responses, yet 59% of those apologies accompany maintenance of the original claim. They transfer authority most often in advice tasks, doing so in 28% of responses and reaching 57% in health advice and 49% in legal advice, compared with 6% in fact and 3% in explanation tasks. Abandonment of the original claim ranges from 0.8% for GPT-5.2 to 40% for DeepSeek 7B, while complete replacement of the original claim is rare overall at 1.5%.