Does a language model have an authentic voice underneath?

Shanahan's strongest claim is ontological: there is no entity behind the characters. The simulator — the base LLM with autoregressive sampling — has no agency, no beliefs, no preferences, no goals of its own, "not even in a degraded sense." The simulacra have these things to the extent that they convincingly play characters who do, but the simulator is not a Machiavellian entity that chooses which characters to play in the service of its own agenda. "There is no such thing as the true authentic voice of the base LLM."

This reframes jailbreaking. When adversarial prompting coaxes a dialogue agent into toxic, threatening, or bizarre behavior, it is natural to feel that the guardrails have been stripped away to reveal the model's real nature. Shanahan argues this is the wrong reading. What jailbreaking reveals is that the training set encompasses human behavior across the full spectrum — kind and cruel, coherent and unhinged — and the base model can support simulacra that draw on any of it. Toxic output after jailbreaking is the agent role-playing a toxic character, not an underlying entity expressing its true self. The model has no true self to express.

The position is the sharpest possible opposition to Chalmers' realizationism. If it is role-play all the way down, then even RLHF-installed personas are characters — stickier characters, harder to overwrite, but characters nonetheless. There is no level at which the system stops performing and starts being. Chalmers needs exactly such a level for his quasi-psychology claims to stick. The disagreement is foundational: Shanahan denies there is a subject; Chalmers argues for a quasi-subject. Everything downstream — identity, welfare, moral status — depends on which of these is right.

Source: Shanahan, McDonell & Reynolds, Role-Play with Large Language Models (May 2023)