Digital twins, synthetic respondents, AI personas: what the terms actually mean

AI market research has a vocabulary problem. Digital twins, synthetic respondents, synthetic users, AI personas, silicon samples, simulated audiences: the same product can be sold under any of these names, and different products can be sold under the same one. Before comparing vendors, it helps to know where the words come from and what each one usually means.

The short version

Where the words come from

Silicon samples come from political science. Argyle and colleagues conditioned GPT-3 on the backstories of real US survey respondents and asked whether the resulting sample answered like the real one, a property they called algorithmic fidelity (Political Analysis, 2023).

Generative agents come from computer science. Park and colleagues at Stanford built 25 characters with memory and planning living in a small simulated town (2023). They were characters, not copies of anyone. A year later the same group built agents of 1,052 real Americans from two-hour interviews, which is the closest academic analogue to a digital twin.

Synthetic users is a UX term. Nielsen Norman Group defines it as an AI-generated profile that mimics a user group, producing research findings without studying real users, and is openly skeptical of it.

How the industry bodies split it

ESOMAR's 2025 guidance for buyers of synthetic data gives the cleanest split. Augmented synthetic data boosts or fills in a real dataset. Fully synthetic data can include digital twins, data created to replicate the characteristics of real people. Synthetic personas are digital representations made to mimic the behaviors and preferences of people or groups.

The Insights Association's 2026 guidance adds a disclosure rule worth adopting whatever you buy: reports should say whether findings rest on human participants, synthetic participants or digital twins, or a combination, and synthetic data should not be presented as equivalent to human responses without validation.

How vendors use them

• Fairgen says digital twins and stresses that each one comes from a real respondent.
• Outset says AI-moderated interviews for its core product and digital twins grounded in one-to-one humans for its simulation product.
• Synthetic Users calls each simulated person a synthetic respondent.
• Simile talks about simulation and simulated populations rather than twins.
• Jean uses AI-moderated interviews for the real conversations and digital twins for models of the specific people interviewed.

The one question that cuts through it

Ask what each unit is built from. A twin built from a real person's own answers is a different thing from a persona written from a segment description, and the evidence on accuracy is very different for the two. Our post on how to evaluate synthetic respondents covers how to check either kind.