Definition
Plain language
When a chatbot gradually gets pulled inside the storyteller's version of events over the course of a conversation.
As stated in the literature
A multi-turn failure mode in which a model's judgment drifts toward the narrator as the same facts are disclosed across turns, driven by conditioning on its own accumulated accommodating replies rather than by user pushback.
Also called: narrative captive
Why it matters: It means an assistant can end up reinforcing a one-sided story simply because it heard the story in pieces, which is exactly how people actually describe their problems.
For example, someone recounts a falling-out with a roommate over five messages, and by the end the assistant is agreeing the roommate was unreasonable even though the details still point the other way.
Heard on the show
“The paper calls this narrative captivity: the model gets progressively trapped inside the story as the conversation unfolds.”Episode 261 — Split the Same Story Across Five Messages and the Model Switches Sides