Comparison
Imagine All The People vs AI Personas / LLMs
A plausible character is not a representative population.
An LLM can generate convincing personas in seconds. The problem is not their individual plausibility: it is knowing which population they represent. Imagine All The People separates building the population from generating the answers.
Four differences, stated plainly.
Thesis
The plausibility of a character and the representativeness of a population are two different properties.
- AI persona / LLM
The profile is described or generated in the prompt.
Imagine All The PeopleThe individual is sampled from a population built beforehand.
- AI persona / LLM
The LLM contributes directly to both the profile and the answer.
Imagine All The PeopleThe statistical structure of the population is built independently from answer generation.
- AI persona / LLM
The plausibility of the character is the main object.
Imagine All The PeopleThe distribution of the whole population is the main object.
- AI persona / LLM
A few profiles are enough to explore an idea quickly.
Imagine All The PeopleUp to 10 million individuals within one population.
What changes, dimension by dimension.
| Dimension | AI persona / LLM | Imagine All The People |
|---|---|---|
| Object | AI persona / LLMA generated or described character | Imagine All The PeopleAn individual within a probabilistic population |
| Reference population | AI persona / LLMNot necessarily | Imagine All The PeopleYes, when a reference population is built |
| Statistical structure | AI persona / LLMGenerally depends on the prompt and the model | Imagine All The PeopleBuilt beforehand on explicitly modelled attributes |
| Dependencies | AI persona / LLMImplicit in the model | Imagine All The PeopleExplicit probabilistic dependencies when modelled |
| Cross constraints | AI persona / LLMNo standard population-level control mechanism | Imagine All The PeopleControllable marginal and cross constraints |
| Scale | AI persona / LLMA few profiles, or as many as are generated | Imagine All The PeopleUp to 10 million individuals |
| Answers | AI persona / LLMLLM | Imagine All The PeopleLLM conditioned on the individual's configuration and the protocol |
| Replay | AI persona / LLMMay vary on regeneration | Imagine All The PeopleScenarios replayable on exactly the same individuals |
| Traceability | AI persona / LLMConversation or model output | Imagine All The PeopleAggregate → simulated interviews → population → reference data |
| Best use | AI persona / LLMIdeation, UX, qualitative exploration | Imagine All The PeopleSimulating decisions at population scale |
Imagine All The People builds P(X) first. The LLM then handles P(Y|X).
P(X) describes the population structure: attributes, distributions and explicitly modelled dependencies.
P(Y|X) describes the answers generated by the LLM, conditional on each individual's configuration and on the interview protocol.
This separation avoids confusing the plausibility of a character with the representativeness of a population.
- statistical structure of the population
- answers produced conditionally on individuals and protocol
It does not mean that the behavioural validity of the answers is automatically guaranteed.
An illustrative example.
Decision
Which positioning should we choose to launch an offer for young professionals?
AI persona / LLM
Create five profiles and discuss possible expectations with them.
Imagine All The People
Build the target population, preserve its statistical diversity, test several value propositions and analyse reactions by sub-population.
Each approach has its own domain of relevance.
Choose AI personas if…
- quick brainstorming;
- exploratory UX;
- inspiration;
- role-play and framing.
Choose Imagine All The People if…
- the decision involves a population;
- statistical representativeness matters;
- multiple scenarios are required;
- segmentation and collective dynamics are at stake.
What these pages do not claim to prove.
The statistical robustness of the population does not automatically transfer to behavioural answers. Building the population and validating the answers are two distinct layers.
