What is a synthetic persona?

A synthetic persona is an AI-generated individual unit: one computational agent built to think, respond, and reason as a single specific person or person-type would. It is the person-level building block of a synthetic study, the unit that sits below a cohort and below a synthetic audience. The credible ones are grounded in real human data, not invented from a demographic prompt, and they can be queried one-to-one or assembled into a population.

The term is used both narrowly (one agent representing one real person) and slightly more loosely (one agent representing a defined demographic or attitudinal segment). Both senses appear in the literature and in vendor language. The post that follows treats the narrow sense as the load-bearing definition.

A working definition

Synthetic persona. An LLM-based individual-level simulation of a single person or person-type, conditioned on real-world data sufficient to make the agent answer questions the way that person, or a person like them, plausibly would. A persona is one unit. A cohort is a grouping of personas defined by shared characteristics. A synthetic audience is the full simulated population for a study.

The persona is the engine. A synthetic respondent is the record it produces when it answers a questionnaire. A synthetic audience is the aggregate population those records belong to. In practice, the terms get collapsed, and many practitioners use “persona” and “respondent” interchangeably. They are not the same level of object.

Where the word comes from

“Persona” is older than the LLM era. Alan Cooper introduced it as a practitioner tool for interaction design in The Inmates Are Running the Asylum, published in 1998. By Cooper's own account, the method evolved out of a 1983 project where he interviewed a real woman named Kathy at Carlick Advertising and role-played her perspective while designing a project management program. The first formally named Goal-Directed personas (Chuck, Cynthia, and Rob, built for the founders of Sagent Technologies) followed in 1995. The 1998 book is the accepted public origin point. Cooper would later note that personas, like all powerful tools, can be grasped in an instant but take years to master.

That lineage matters for two reasons.

First, “persona” entered the design vocabulary already pointing at the individual level. A Cooper persona is one named person with goals, habits, and tolerances, not an aggregate. The connotation carried forward: when LLM researchers reach for “persona,” they typically mean an individual unit with attitudes and behavior, not a market segment.

Second, the original method was grounded in field research. Cooper's personas were extrapolated from interviews with real people; the rigor of the method was always tied to the quality of that grounding. The same standard now distinguishes credible synthetic personas from generic ones.

No current synthetic-persona vendor explicitly invokes Cooper. The lineage is implicit. It is worth surfacing because it explains why the word carries the weight it does.

Synthetic persona, synthetic respondent, synthetic audience

The category vocabulary is unsettled. The literature does not establish a single fixed semantic distinction between “synthetic persona” and “synthetic respondent,” and vendor pages use the terms loosely. Academic work uses “generative agents” (Park and colleagues, 2023 and 2024) or “silicon samples” (Argyle and colleagues, 2023). The American Association for Public Opinion Research, in its May 2026 Task Force report on Responsible AI Integration in Survey Research, prefers “synthetic responses” over “synthetic samples” because the method estimates what a specified set of respondents would say rather than designing a sample. None of these are wrong. They emphasize different layers of the same object.

The working distinction Replism uses, and the one this post recommends as the cleanest carving:

Synthetic persona. The person-level unit. One agent representing one specific person or person-type. The thing you can interview, query one-to-one, or instantiate at scale.

Synthetic respondent. One answering instance produced by a persona inside a specific study. The row of the dataset. A persona that completes a questionnaire produces a respondent record.

Cohort. A defined grouping above the persona level. Likely voters in a state, buyers of a category, executives in a sector. Cohorts are how studies are sampled and how findings are reported.

Synthetic audience. The full simulated population for a study, assembled from personas sampled to a frame.

This distinction is a convention, not a settled industry standard. It is offered to make internal memos and vendor briefs cleaner, not to claim the field has agreed on usage. Where this post writes “persona” without qualification, it means the person-level unit.

Synthetic persona is also distinct from synthetic data. Synthetic data is artificially generated information designed to mimic real datasets, often for privacy preservation or model training. A synthetic persona uses real human data as grounding and makes that grounding queryable as a person.

Grounded versus prompted personas

The single most consequential property of a synthetic persona is what it is conditioned on. The literature now distinguishes cleanly between two approaches.

Grounded. The persona is built from real human data: interview transcripts, survey responses, psychometric instruments, behavioral records. The agent is conditioned on individual-level evidence drawn from a specific human or a sampled population.

Prompted. The persona is described in natural language to a general-purpose LLM, with no grounding in observed human responses. The agent is asked to roleplay a persona defined by demographic descriptors alone.

The clearest empirical comparison is Park and colleagues (2024). Using a national sample of 1,052 Americans, they built agents from two-hour semi-structured interviews, from structured surveys, or from both. On held-out General Social Survey items, agent accuracy reached 83% (interview only), 82% (surveys only), and 86% (combined) of participants' own two-week test-retest consistency. Agents prompted only with demographic descriptions reached 74%. The grounding gap is roughly nine to twelve percentage points, and the gap widens further on personality measures and on accuracy disparities across racial and ideological groups, where grounded agents reduced bias relative to demographics-only prompting.

The framing matters. The 83% to 86% figures are not absolute accuracy against a ground truth. They are accuracy relative to the human self-replication ceiling, the rate at which a real respondent matches their own prior answer when retested. A synthetic persona cannot be more reliable than the human it is trying to replicate. Claiming 100% would mean overfitting on noise.

Prompted personas fail in characteristic ways. Cheng, Durmus, and Jurafsky (ACL 2023) found that when GPT-3.5 and GPT-4 are asked to generate persona descriptions for marked demographic groups, the portrayals “contain higher rates of racial stereotypes than human-written portrayals using the same prompts.” The failure is systematic, not incidental. Gupta and colleagues at the Allen Institute (ICLR 2024) showed that persona assignment surfaces latent bias even in models that overtly reject stereotypes when asked directly: 80% of the personas they tested across five socio-demographic groups exhibited bias under task conditions, with some groups showing performance drops above 70% on reasoning benchmarks.

The Pew Research Center, in a May 2026 institutional position, summarized the published findings on AI-as-respondent more bluntly: AI estimates “tend to stereotype groups of people, have a harder time representing Republican viewpoints than Democratic ones, and understate the level of disagreement in public opinion.” The variance-compression problem is real. A prompt-only persona will frequently produce the average opinion of its demographic descriptor and very little of the disagreement that actually defines that demographic.

None of this means LLMs cannot be used as personas. It means the grounding layer is what separates a defendable persona from a stereotyped one.

What synthetic personas are used for

Use cases sort by the level at which the persona is queried.

Concept and message testing. Run a single concept or message past a persona, or a small set of personas, to get fast directional reaction before committing budget to a fielded study. Cross-vertical use case. Common across campaigns, brand, and policy.

Pricing and willingness-to-pay exploration. Probe a persona on price sensitivity, alternative formats, and tradeoff scenarios. Useful in the early-iteration phase, before a conjoint or Van Westendorp study with humans.

Qualitative-style interviews. One-to-one dialogue with a single persona, treating the agent the way a research interviewer would treat a real participant. Park and colleagues (2024) describe the architecture as the language model role-playing as the person it just interviewed, drawing on the interview transcript plus the social and psychological knowledge embedded in the model.

Agent-based simulation. Instantiating hundreds or thousands of personas to study emergent dynamics. Park and colleagues (2023) demonstrated the genre in a 25-agent sandbox where agents formed opinions, made plans, and coordinated events with no top-down direction. At population scale, this is the lineage that connects synthetic personas to social simulation.

Hard-to-reach populations. Specialist audiences (rare patient populations, single-district voters, C-suite buyers) that are slow or expensive to recruit through traditional panels. The persona-level unit is what makes one-to-one engagement possible with these groups at all.

None of these uses substitute for human research. They expand what can be tested cheaply and iteratively before the high-stakes round goes to people.

How to tell a good synthetic persona from a bad one

Four criteria separate a defendable persona from a convincing one.

Grounding source. Real interview transcripts, survey responses, or behavioral records produce a persona that reflects the idiosyncrasies of a real individual. A demographic description alone produces a persona that reflects the model's stereotype of the demographic. Ask what the persona is conditioned on. If the answer is “the demographic descriptors in the prompt,” you have a prompted persona, not a grounded one.

Evaluation methodology. A credible persona is benchmarked against known answers. The honest benchmark is the human self-replication ceiling, not a hand-picked accuracy claim. A vendor that cannot report performance relative to that ceiling, on held-out items, in writing, is selling assertion rather than evidence.

Response variance. Real populations disagree with themselves. A good persona, asked the same question across slightly different framings, produces variation that resembles real human variation. A bad persona collapses toward the median answer for its demographic, which is exactly the failure mode Pew, Cheng, and Gupta describe.

Coherence across questions. A persona that answers each question plausibly in isolation but contradicts itself across a battery of related items is producing distribution-level statistics without individual-level structure. Persona coherence is what enables qualitative interviews, multi-question survey designs, and any analysis that depends on within-respondent consistency.

These criteria are not exotic. They are the same criteria traditional survey methodology applies to its instruments, adapted for an agent rather than a paper questionnaire. The bar should not be lower for a synthetic persona than for a real one.

References