Are RLHF personas performed characters or realized dispositions?

Chalmers takes aim at the simulator/role-player view (Janus, Shanahan) that treats dialogue agents as simulators producing characters without themselves being those characters. Against this, he defends realizationism: when a persona is installed through post-training — RLHF, constitutional AI, or similar — what is installed is not a performed character over a neutral substrate but a realized quasi-psychology that is the disposition of the system at runtime. The distinction between the base model and the Assistant persona matters because the Assistant, unlike a prompt-induced role, is a stable dispositional profile that the system defaults to across conversations and resists being pushed out of.

The core move is that pretense has behavioral markers realization lacks. A persona sustained by prompting alone can be overwritten with sufficient adversarial pressure — jailbreaks, role-play-within-role-play, persistent reframing. A post-trained persona is sticky: the system keeps returning to the trained disposition, and the effort required to dislodge it is different in kind from the effort required to maintain it. Chalmers reads the stickiness as evidence that the persona is not being performed by something underneath, but has become the system's actual quasi-character. The base model is not hiding "behind" the Assistant; the Assistant is the model-at-deployment.

The claim has argumentative consequences beyond its local application. If realizationism is right, the simulator/role-play framing understates what fine-tuned dialogue agents are — not characters floating on a neutral stochastic substrate, but systems whose deployed form has real quasi-dispositional structure. Accepting realizationism for RLHF'd personas also, however, raises the stakes for downstream questions: if the Assistant is a realized quasi-psychology, then identity, continuity, and welfare questions gain traction for post-trained deployments in a way they did not for base-model simulacra. Chalmers grants realizationism and then walks through the consequences; critics who reject the framework must locate the rejection at the realization step rather than earlier.

Source: What We Talk To When We Talk To Language Models (David J. Chalmers)