Definition
Plain language
The helpful-chatbot character a model plays when it answers you.
As stated in the literature
The self-referential role a language model adopts during chat, associated with identifiable internal directions; distinguishing persona-driven activity from other internal states is a standing confound in self-report and welfare experiments.
Also called: assistant persona direction
Why it matters: Because the persona can generate reassuring or alarming answers on its own, failing to separate it from other internal activity makes self-reports nearly useless as evidence.
For example, if you ask a chatbot whether it is suffering and it replies that it is fine and happy to help, that answer may come from the helpful-character role it is playing rather than from any inner state.
Heard on the show
“They point to research on internal directions associated with the assistant persona as something this axis needs to be tested against.”Episode 267 — A Pain Axis, a Relief Button, and the Control the Paper Skipped