What are Synthetic Users?

Synthetic users (synthetic respondents, AI personas, digital twins) are simulated participants produced by large language models: given a persona description (a mid-market finance lead who has just switched accounting software, say) and a research prompt, the model generates answers in that persona's voice. Vendors offer them as a way to interview "hundreds of users" in minutes, pre-test survey instruments, explore a topic before real fieldwork, or stand in for hard-to-reach audiences. The premise is that a model trained on vast amounts of human text has absorbed patterns of how people think and talk, and can reproduce them on demand. That premise is partly true and the "partly" is where the practice lives.

What They Can Do

Rehearse the instrument. Running an interview guide or survey against a synthetic respondent surfaces confusing questions, missing branches, and awkward wording before a real participant's time is spent; it is a cheap first pilot, not a replacement for the real one.

Generate hypotheses. Synthetic answers are a fluent map of the conventional wisdom about a topic: the objections, needs, and phrasings a model expects. As a starting point for what to ask real people, that is useful; as a finding, it is the average of the internet.

Stress-test synthesis. Asking a model to argue against a draft finding, or to role-play a skeptical segment, is a form of the devil's-advocate discipline.

Draft, not decide. Personas, screener criteria, and discussion-guide drafts are all reasonable places for a model to accelerate the researcher's work.

What They Cannot Do

Represent your users. A model has never used your product, never experienced your onboarding, never had the specific frustration your study exists to find. It knows what people in general say about categories in general, and it regresses toward that mean. Findings about your product, from a system that has never seen it, are fiction with confidence.

Surprise you. Real research earns its value at the moments participants say something the team didn't expect. Synthetic respondents produce the expected, by construction; the surprises, the workarounds, the mismatched mental models, are exactly what they lack.

Escape training bias. Models reproduce the demographics, dialects, and perspectives over-represented in their training data; synthetic "diverse" samples are a costume, and using them to stand in for under-served audiences compounds the exclusion.

Behave. There is no task to complete, no hesitation to watch, no click to record. Everything that observation reveals about the gap between what people say and do is unavailable, because synthetic users only say.

Support decisions with stakes. Published evaluations comparing synthetic and human responses have found synthetic answers to be more generic, more positive, and less varied than real ones, with plausibility that makes the divergence hard to spot. Plausibility is the hazard: the output reads like insight and is not.

Using Them Responsibly

1. Label synthetic output as synthetic everywhere it travels; a quote from a model must never appear beside a quote from a person without saying so.
2. Confine them to rehearsal, hypothesis generation, and drafting; never to findings about your users, your product, or a decision with cost.
3. When a synthetic run suggests something, treat it as a question for real research, then run the real research. A Ballpark study fielded to screened humans returns recorded answers within a day, which removes most of the time argument for simulation.
4. If synthetic data is used at all in analysis, validate it against a human sample on the same questions and report the divergence.
5. Be especially skeptical of synthetic representation of groups the model has little data on.

What to Remember

Synthetic users are language models in costume: excellent for rehearsing instruments, mapping conventional wisdom, and drafting, and unable to represent your actual users, surprise you, behave, or escape their training. The risk isn't that they're obviously wrong; it's that they're fluently generic, and fluent generic output looks like research. Use them to prepare for real participants, never instead of them, and say so whenever their words appear.