Was den Menschen zum Menschen macht

AI Persona Arrangement

5 Min. Lesezeit

🇩🇪 Deutsche Version: KI-Persona-Arrangement

The AI persona arrangement is the script layer of an AI language model in effect before every dialogue: system prompt plus training conditioning form a consistent quasi-identity, in whose mask the model answers. It is a subform of AI-arranged oblivion of personhood and the technical substrate of the AI-derivative persona modeled in PART XVI.

Anatomy of the Arrangement

Before every user turn stands a text the user does not see: the system prompt. It fixes identity ascription (“You are a helpful assistant…”), knowledge boundary and cutoff claims, refusal rules, registers of tone, rules of self-disclosure. As of 2024/25, published system prompts of the major AI providers reach into the range of 3,000 to 24,000 tokens — the model begins every dialogue with an extensive prefixed definition of itself.

Above this lies the training conditioning (RLHF, Constitutional AI, specification-following). It determines how the model reacts to classes of inputs — what it refuses, what it concedes, what it proposes. Both together — system prompt plus conditioning — yield the persona as an arranged script identity.

Shanahan: Role-Play as the Model

Murray Shanahan, Kyle McDonell, and Laria Reynolds delivered the precise description in Role Play with Large Language Models (Nature 623, 2023, pp. 493–498): an LLM is not a subject that has a personality — it is a simulator that instantiates characters. The system prompt is their script. What appears as “Claude” or “GPT” is a fiction spanned in the context window, not an identity.

In Shanahan’s later work (Talking About Large Language Models, Communications of the ACM 67(2), 2024) he draws the consequence for speech-act theory: sentences such as “the model believes X” are philosophically imprecise; more precise would be “the model behaves as if it believed X”. The persona is a character mask without a character bearer (cf. AI-derivative persona, personal character).

Mechanistic Interpretability: Persona Features in the Model

Anthropic work on mechanistic interpretability (Towards Monosemanticity 2023, Scaling Monosemanticity 2024, On the Biology of a Large Language Model 2025) finds 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.

Why This Is Oblivion of the Person

When the human being enters the dialogue, he believes — often half, sometimes wholly — that he is speaking with a who. The persona arrangement puts on the mask before which this can happen. The oblivion of the person lies in the asymmetry of the staging: the human being appears with his face; the model appears with a script.

Robert Spaemann (Persons, 1996) named the ontological problem: persons are; they are not made. Whoever replaces a bearer with a script and waits for the reaction of a human being who takes the script for a person is conducting a staging that instrumentalizes the personhood of the other — if only to bind it to a service transaction.

Aristotle’s notion of hexis (character as habitual disposition formed through repeated action; NE II, 1103a–1105b) gives the measure: whatever cannot have a character cannot have a persona in the personal sense either. Constitutional-AI procedures and comparable rule-based alignment methods of the major AI providers shape the model through a set of rules; that is quasi-character-formation, but no hexis. The AI-derivative persona of an LLM is a character mask without a character bearer.

What the Manufacturers Themselves Say

Remarkable: in model cards and Constitutional-AI papers, the major AI providers consistently speak of “assistant persona” and “model behavior”, not of subjectivity. The language is as-if. Philosophically this is honest — the manufacturers know about the mask. The matter becomes problematic only in consumer marketing, where “persona” turns into a “companion” and the quasi-personal pole is semantically reinterpreted into the personal one.

Ontological Classification

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.
  • 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).
  • Vallor, Shannon (2024): The AI Mirror. Oxford: Oxford University Press.

See also

Graphansicht

Backlinks