Persona simulation is conditioning a language model on a detailed persona and letting it answer as that person: a profile, a history, the material it knows, and a state that changes as the conversation goes. The research literature has called the same idea silicon samples, Turing experiments, homo silicus, generative agents and LLM personas. What separates a simulation worth using from a chat with a system prompt is what the persona is built from, whether it can show its reasoning, and whether anyone has checked it against real people.
The market sells synthetic users. The papers describe persona simulation. The words point at the same mechanism from different rooms, and the mechanism is worth understanding before you buy anything that claims to do it.
The definition
A language model predicts what text comes next. Give it a person, in enough detail, and what comes next is what that person would say. Persona simulation is the deliberate version of that: a persona is specified, the model is conditioned on it, and its answers are read as evidence about the people the persona stands for. The persona can be a sentence or a dossier. That difference decides everything downstream, which is why the definition of synthetic users spends its time on what is inside one.
The research lineage
The idea has a short, dense history, and the vocabulary in the market comes straight from it.
| Term | Who, when | What they showed |
|---|---|---|
| Silicon samples, algorithmic fidelity | Argyle and colleagues, 2022 | GPT-3, conditioned on thousands of real respondents’ backstories, reproduced the response distributions of human sub-populations. They named the property algorithmic fidelity. |
| Turing experiments | Aher, Arriaga and Kalai, 2022 | A test of whether a model can simulate a representative sample of participants well enough to replicate classic human-subject findings, and reveal its consistent distortions. |
| Homo silicus | Horton, Filippas and Manning, 2023 | Language models as implicit computational models of humans, given endowments, information and preferences and run through economics experiments. |
| Generative agents | Park and colleagues, 2023 and 2024 | Agents with memory, reflection and planning that behave believably; then agents built from two-hour interviews with 1,052 people that reproduced their survey answers 85% as well as they reproduced their own. |
| LLM personas | Jiang and colleagues, 2023 | Personas assigned Big Five profiles, tested on the personality inventory and a writing task, to see whether the assigned traits actually show in the output. |
Argyle’s group framed the promise first: language models could be “effective proxies for specific human sub-populations in social science research”, and the “silicon samples” they built were conditioned on real people’s backstories, not invented ones. Park’s generative agents supplied the architecture the market now imitates, “computational software agents that simulate believable human behavior” with a stored record of experience, reflection and planning. The 2024 follow-up grounded each agent in a real person’s two-hour interview, and that grounding is why its accuracy result exists. The evidence note walks through what happened when the grounding was missing.
What a persona is made of
- A profile. Traits on named frameworks, so the persona’s dispositions are stated rather than implied by a job title. Jiang’s LLM personas started here, with the Big Five.
- A history. Formative experiences, beliefs with a strength, wounds with triggers. Argyle’s silicon samples worked because they carried real backstories.
- Material. What the persona knows, attached with a declared level of fluency, and a rule for what happens beyond it. This is the grounding that PersonaCite formalised in 2026 and that separates a grounded persona from a prompted one.
- State. Memory that persists between sessions, emotion that moves and settles, trust that is earned. Park’s agents introduced memory and reflection; a persona without state answers every question as a stranger.
- A record. Why it answered as it did. Without a trace, a surprising answer cannot be told from an artefact.
A persona is only the attributes the system can act on. Everything else is a name.
From prompt to instrument
Every paper above was written by people who could inspect their own setup. A buyer cannot, so the question is what makes a simulation inspectable from the outside. Three things. The persona is built from material you can see, so its answers have a source. Each answer shows what it drew on and why, so a claim can be checked. And someone has compared its output with real people on the kind of question you ask, so the number it produces has a meaning. PersonaCite, in January 2026, put it as reframing personas as “evidence-bounded research instruments”. That phrase is the whole difference between an instrument and a chat.
What it is not
- Not a prediction of individuals. The individual-level record is poor even where aggregates hold. A 2026 cross-domain benchmark found that at the individual level “no LLM beats even the strongest baseline”.
- Not a replacement for fieldwork. Nielsen Norman Group’s verdict on synthetic users stands for persona simulation too: they “cannot replace the depth and empathy gained from studying and speaking with real people.”
- Not a model of a named person. Personas are fictional or composite. Simulating a real, identifiable individual is a different act with different ethics, and Perplica’s terms forbid it.
Why Perplica uses the term
Because it describes what the system does rather than what it replaces. A synthetic user is a claim about output. A simulated persona is a mechanism you can ask questions of: what was it built from, what does it know, why did it answer that way, what did it withhold. Perplica’s personas carry a profile on named frameworks, a narrative identity, six kinds of memory and your own material at a declared fluency, and every reply stores its reasoning. Whether that produces answers that track your real customers is the concordance question, and it is answered by measurement, not by the name.
What is persona simulation?
Conditioning a language model on a detailed persona, a profile, a history, material it knows, and letting it answer questions as that person would. In market research it stands in for a participant during the rehearsal phase, before real fieldwork.
Where does the idea come from?
From a run of research papers between 2022 and 2024: silicon samples and algorithmic fidelity (Argyle and colleagues), Turing experiments (Aher, Arriaga and Kalai), homo silicus (Horton), generative agents (Park and colleagues) and LLM personas (Jiang and colleagues). Persona simulation is the plain name for what those papers do.
Is persona simulation the same as synthetic users?
The market words overlap. Synthetic users, synthetic respondents and digital twins describe the output; persona simulation describes the mechanism. A synthetic user can be a one-line prompt. A simulated persona has a profile, a memory and material behind it, or it is not one.
Does persona simulation predict real people?
Sometimes, at the aggregate level, when the persona is grounded in real material about real people. The strongest result, 85% of human test-retest accuracy, came from agents built on two-hour interviews. Prompt-only personas match averages and miss variation. Perplica makes no accuracy claim until a public concordance study exists.