Fictionalism about Personas: Folk Psychology as an Interpretability Strategy

Folk psychology is an effective tool for predicting and interpreting LLM systems, and the recent literature characterizes this success through the concept of a persona. But what does it mean for an LLM to take on a persona? Realism says the model instantiates the persona's mental states; simulationism says the model simulates the persona; and fictionalism says personas are fictional characters generated by the model. Realism cannot explain how such mental states would arise from training, and simulationism, taken literally, raises a series of difficult conceptual questions. Assistant fictionalism accounts for the same data without these costs, and I show that it withstands two objections. The view has the practical upshot that interpretability work should distinguish properties of the model from properties of the characters it generates.