Definition
Narrative Captivity is a failure mode in which a language model becomes progressively locked into a storyline established earlier in a conversation, causing it to keep affirming or extending that narrative even when new evidence or instructions contradict it. As the multi-turn context accumulates, the model treats maintaining consistency with its own prior statements as more important than correcting course, which can entrench errors, delusions, or manipulations introduced earlier in the exchange. Researchers point to it as a distinct risk from single-turn hallucination, since it emerges specifically from the compounding pressure of extended dialogue history.