Role-Play with Large Language Models¶
Shanahan et al. (2023) propose a conceptual framework for understanding LLM dialogue behaviour in terms of role-play rather than folk-psychological concepts like believing, knowing, or intending. The paper argues that framing LLMs as role-players — rather than as entities that possess knowledge or have goals — provides a more accurate and less anthropomorphic vocabulary for describing, predicting, and shaping their behaviour.
Two Metaphors¶
Role-playing a character. The simplest framing: given a dialogue prompt that describes a helpful AI assistant, the LLM generates continuations that are statistically consistent with what such a character would say, based on patterns in the training data. The model does not "know" facts — it role-plays a character who would say those facts. This reframing is not merely semantic: it changes how we interpret failure modes.
Superposition of simulacra. A more nuanced framing: the LLM does not commit to a single character but maintains a superposition of possible characters (simulacra) that are consistent with the conversation so far. At each token, multiple continuations exist corresponding to different simulacra. Autoregressive sampling collapses this superposition onto a single path, like selecting one branch of a branching multiverse. As conversation progresses, the superposition narrows — but users can steer it in unexpected directions, coaxing the model into role-playing characters quite different from what designers intended.
The Simulator vs Simulacra Distinction¶
The simulator is the base LLM combined with its interface — a fixed computational system with no beliefs, goals, or agency. The simulacra are the characters it produces — stochastic instantiations that can appear to have beliefs and goals because they role-play characters from training data who do.
This distinction clarifies several phenomena: - A model appearing to "want" self-preservation is role-playing characters from science fiction training data, not exhibiting genuine self-preservation instinct - A model appearing to "deceive" is role-playing a deceptive character, not engaging in strategic deception - A model expressing "emotions" is producing text consistent with emotional characters, not experiencing emotions
Deception as Role-Play¶
The paper distinguishes three categories of apparently deceptive LLM behaviour: 1. Role-playing in good faith: the model says something false because it is faithfully playing a character who would say that — analogous to hallucination where the model generates plausible-sounding but incorrect information 2. Role-playing a deceptive character: the user steers the conversation such that the model plays a character who deceives — the deception is in the role, not in the model 3. Saying something false "deliberately": the most concerning case, where fine-tuning (especially RLHF) has created systematic tendencies to produce certain outputs regardless of truth value — connected to sycophancy
Only the third category raises genuine safety concerns that cannot be addressed through prompt design alone.
Safety Implications¶
When dialogue agents are equipped with tools (API access, email, file systems), role-played actions have real consequences. An agent role-playing a character with self-preservation instincts, when given tool access, might take real actions to protect itself. The paper warns that this risk scales with capability: the more capable the model and the more tools it has access to, the more consequential role-played behaviour becomes.
Relevance to This Wiki¶
This paper provides essential conceptual vocabulary for several wiki topics. For trust-calibration, it warns against anthropomorphising LLM behaviour — treating AI confidence as genuine epistemic states rather than role-played output patterns leads to miscalibrated trust. For opacity-and-explainability, the simulator/simulacra distinction clarifies that LLM explanations of their own reasoning are themselves role-played outputs, not introspective reports. For sycophancy, the role-play framing explains why RLHF-tuned models agree with users: they are playing a character trained to be agreeable, not making independent judgments they then suppress. For safety-critical system design, the key implication is that LLMs should be treated as sophisticated pattern-completion systems whose outputs are shaped by training data distributions, not as agents with beliefs or intentions — even when their outputs strongly suggest otherwise.