What anchors a stable identity beneath an LLM's persona?

Shanahan introduces the role play framing to navigate between anthropomorphism and naive dismissal. An LLM playing a helpful assistant can be described using familiar folk-psychological terms — it "believes" its answers, "wants" to be helpful — without committing to the claim that these are genuine mental states. The role play framing permits the vocabulary while marking its qualified status.

But the Simulacra paper reaches a deeper claim: with LLMs, "it's role play all the way down." This is different from saying LLMs engage in role play. It means there is no stable substrate beneath the role play that would make "the person behind the mask" intelligible.

Humans are social chameleons. Goffman documented the way humans adopt different personas across social situations — front stage vs. back stage, different registers, different self-presentations. But even for the most extreme social chameleon, there is a stable biological self underneath: needs, drives, a developmental history, a body that persists across situations. We can always meaningfully speak of the person whose mask this is.

LLMs lack even the biological needs common to all animals. They are not embodied entities with hunger, fear, comfort, desire. They are "simultaneously role-playing a set of possible characters consistent with the conversation so far" — a superposition of simulacra, generated stochastically. The "character" produced by any given conversation is not the expression of a stable underlying self; it is a sample from a distribution of possible characters.

This makes LLM identity categorically different from human identity — not just quantitatively less stable, but structurally lacking the substrate that would make stability possible. If consciousness requires co-presence (Can disembodied language models ever qualify as conscious?), the absence of stable biological selfhood makes it even clearer why the consciousness vocabulary struggles to find purchase.

The geometric evidence for "role play all the way down" comes from the Assistant Axis: since How stable is the trained Assistant personality in language models?, post-training positions models in a low-dimensional persona space where the dominant axis measures distance from the default Assistant persona. Drift along this axis in response to emotional or meta-reflective conversations demonstrates that the Assistant persona is loosely tethered, not anchored — consistent with there being no stable self beneath the role play, only a trained default position with no inherent restoring force.

The upshot: useful for thinking with but not for talking about. The intentional stance (treating LLMs as rational agents) is valid as a predictive heuristic. But it should not suggest there is something it is like to be this character, or that the character persists beyond the context window.