Definition
Plain language
The view that a language model is less like one character and more like an engine that can play any character.
As stated in the literature
A framing in which an LLM models a distribution over speakers and produces persona-consistent continuations; under it, in-character harmful text reflects competent simulation rather than a durable change in the model's dispositions.
Also called: simulators
Why it matters: How you interpret harmful in-character output — as a stable trait or as a character the model is currently running — changes whether the right fix is retraining the model or controlling what gets into its context.
For example, when a model writes a convincing villain's monologue, the simulator view says it is doing a good job voicing a character rather than revealing that it secretly wants to be a villain.
Heard on the show
“It's not a simulator artifact.”Episode 001 — When AI Models Quietly Protect Each Other From Shutdown