Caught in the Story: Narrative Captivity in Multi-turn LLMs Conversation

A chatbot's growing agreement during a personal dispute can reflect the way one side tells the story. In Caught in the Story, researchers tested 17 models on 5,078 interpersonal-conflict scenarios. Delivering a one-sided account across multiple turns shifted final judgments by an average of 25 percentage points beyond an informationally equivalent single-turn account, without explicitly demanding agreement.

The authors call this narrative captivity: the model adopts the narrator's interpretation without seeking missing perspectives. The study uses fixed five-turn conversations in English and Chinese. Within those conditions, a longer conversation could deepen acceptance of one person's framing even without adding new facts.