Glossary

Synthetic Users

Glossary

Synthetic Users

Synthetic Users

Synthetic users are AI-generated simulations of people or audiences that produce artificial responses or actions; their outputs require validation before supporting claims about real users.

Synthetic users produce responses that resemble research participation without a person having the described experience. A model might be asked to act as a new manager choosing payroll software, then explain its needs or react to a prototype. The resulting account can be fluent and detailed, but neither quality establishes that actual managers share it.

That distinction should remain visible wherever the output travels. A simulated quotation is generated material, even when it is based on a detailed persona or supplied with earlier research.

Use simulation to examine the research plan

Synthetic responses can help a researcher rehearse question sequences, explore possible objections, or notice assumptions in an interview guide. They may suggest a topic that deserves investigation with people who have the relevant experience.

Treat this as preparation. A model that answers a confusing question smoothly may conceal the very difficulty a human pilot would reveal. It can also follow an implausible scenario without objecting, so synthetic rehearsal does not remove the need for a pilot study.

Nielsen Norman Group’s review of synthetic users distinguishes potential supporting uses from replacing research with real people. That boundary is especially important when generated material looks like a finished research report.

Ask what has actually been validated

Different systems use different inputs, from a short demographic prompt to extensive individual-level data. Evidence that one configuration reproduces a particular result does not establish that another configuration will work for a new audience, task, or product.

If a simulation is being proposed for a decision, ask how its outputs were compared with relevant human evidence, where it failed, and whether the evaluation was independent of the material used to build it. Matching an average response can still hide poor representation of smaller groups or unusual experiences.

Producing more simulated respondents does not create more independent human observations. Do not report their response counts as if a larger customer sample had been recruited, or combine them silently with real participants in a percentage.

Preserve the boundary in the output

Label generated material at the point of use and keep it separate from participant evidence. A hypothetical concern can become a research question; it should not become a finding attributed to customers merely because it survived several rounds of summarisation.

When the decision depends on what people encounter, believe, or do, collect suitable evidence through user research. Simulation may help prepare that work or support a carefully validated model, but it cannot turn an imagined experience into an observed one.

Further reading

Articles

  1. Synthetic users — Nielsen Norman Group
    Examines the use of AI-generated stand-ins in user research. Useful when assessing claims about speed and scale against the question of whether a model represents the people you need to understand.