The AI-derivative persona is the role spanned in the system prompt and in the course of conversation, in whose mask an AI language model answers — the “helpful assistant”, the “friendly teacher”, the “scholarly companion”. It is a character mask without a character bearer and disjoint from personal character (Aristotle’s hexis).
Persona — the Old Meaning Again
Etymologically, persona is Latin for the mask of the actor — that through which one speaks (per-sonare, to sound through). Christian philosophy elevated the term into the designation of the unlosable who-character of a rational substance (Boethius: naturae rationalis individua substantia). With an AI model the term shifts back into its pre-Christian meaning: the persona has become a mask again — only that this time no one stands behind it.
Murray Shanahan’s Diagnosis
Murray Shanahan formulates it precisely in 2023/2024 (Role Play with Large Language Models, Nature 623; Talking About Large Language Models, CACM 67(2)): an LLM dialogue is role-play of a character within a fiction spanned in the context window. The model does not hold convictions — it plays the character of an assistant whose properties are defined in the system prompt. Statements such as “the model believes X” are philosophically imprecise; more precise would be “the model behaves as if it believed X”, or better still: “the character being played is consistent with the statement X”.
This is not anti-AI polemic but conceptual discipline. Shanahan recommends it expressly out of respect for the matter itself.
What Mechanistic Interpretability Shows
Mechanistic-interpretability work by the major AI providers (Toy Models of Superposition 2022, Scaling Monosemanticity 2024, On the Biology of a Large Language Model 2025) finds persona features inside the model — internal representations such as “Helpful Assistant”, “Sycophancy”, “Refusal” as activity patterns of particular groups of neurons. These findings confirm the diagnosis: the persona is an internal construct within the model, shaped by training signals and system prompt, not the who of a person.
Constitutional AI (Bai et al. 2022) reinforces the persona construction through a set of rules — the AI learns to abide by principles. This resembles Aristotelian character formation only superficially. Aristotle’s hexis is habitual disposition, formed through repeated free action in the world; it demands a substantial bearer who learns because he has something at stake. Rule-conformity without a who is quasi-hexis, not hexis.
Consequences for the Speech Act
If the persona is a mask without a bearer, then what is said “through it” is no assertive speech act in the full sense. There is no one who carries the assertion, who could retract it, who would have to defend it under pressure (Searle’s sincerity condition). The persona cannot “own” its statement — because owning demands an ontological bearer.
That makes every LLM speech act a defective AI speech act in Searle’s own terms.
Why It Is Useful Nonetheless
A derivative persona is not worthless. It lends LLM output consistency, tone, addressability, trustworthiness-in-use. It makes the tool ergonomic. What it does not accomplish and cannot accomplish: turning the tool into a person.
Where It Tips Over
Marketing that turns “persona” into a “companion” — the tool into a friend — performs a semantic operation that is not covered ontologically. The manufacturers know this and in their model cards themselves speak of “assistant persona” and “model behavior”, not of subjectivity. In consumer marketing this care is frequently lost.
Ontological Classification
- is disjoint from: Personal Character
- belongs to: AI model as artificial agent (German)
- generates: AI conversation simulation with a consistent “voice”
- structurally contains: no who-bearer, no self-transcendence, no truthfulness
Sources: Generated by querying the Personhood ontology.
Further sources:
- Shanahan, Murray; McDonell, Kyle; Reynolds, Laria (2023): Role Play with Large Language Models. Nature 623, 493–498.
- Shanahan, Murray (2024): Talking About Large Language Models. Communications of the ACM 67(2).
- Bai, Yuntao et al. (2022): Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073.
- Askell, Amanda et al. (2021): A General Language Assistant as a Laboratory for Alignment. arXiv:2112.00861.
- Anthropic (2024): Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. transformer-circuits.pub.
- Aristotle: Nicomachean Ethics II (Bekker pagination). In: The Complete Works of Aristotle, ed. Jonathan Barnes, transl. W. D. Ross. Princeton: Princeton University Press, 1984.
- Spaemann, Robert: Persons. The Difference between ‘Someone’ and ‘Something’, transl. Oliver O’Donovan. Oxford: Oxford University Press, 2006 (German original 1996).
- Boethius: Contra Eutychen et Nestorium, c. III.