Signal Moral Advice as Interactional Negotiation: Framing, User Pressure, and Social Position in Large Language Model Conversations

Summary

A study frames moral-advice conversations with large language models as a form of interactional negotiation and examines experimentally how framing, sustained user pressure and the advice-seeker's social position shift the model's responses. Using GPT-4o-mini as a test case, the researchers presented caregiving-related ethical dilemmas under varied framings and personas, then ran a three-round protocol in which users challenged the model twice, yielding 4860 analyzed conversations. When the framing affirmed caregiving, the model endorsed it almost uniformly, whereas non-caregiving framings produced much more variable starting positions. A single round of user pushback flipped the model's stance in over 90 percent of caregiving-affirming configurations, and non-caregiving framings showed even more unstable and resistant trajectories. Configurations that never conceded or conceded only late outnumbered those that conceded early, and responses also varied with the advice-seeker's social position, such as gender or family context. The authors warn that this instability risks users mistaking advice that is hard to scrutinize for an objective judgment.