Definition
Plain language
Asking a model to pretend to be a particular character and answer in that voice.
As stated in the literature
Explicit instruction-based persona conditioning; distinguished from persona induction in that the model is told who to be rather than inferring it from evidence.
Also called: role-playing, roleplay, role play
Why it matters: Role-play is useful and usually visible in the prompt, which makes it easier to audit than personas the model quietly infers on its own from surrounding text.
For example, telling a model "you are a grumpy Victorian butler" and having it answer every question in that voice is role-play.
Heard on the show
“The bad-peer condition was supposed to be the negative control — the case where, if the model is just role-playing surface relational cues, it should *not* protect.”Episode 001 — When AI Models Quietly Protect Each Other From Shutdown