The Median Has No Edge Cases
An Evidence Map for Synthetic Users in UX Research
Large language models (LLMs) are marketed as "synthetic users": simulated research participants that promise user insights without recruiting humans. The empirical literature on this idea looks contradictory. Some studies report close alignment between simulated and human data, others report systematic failure.
We argue that this contradiction disappears once we change the question. Instead of asking whether synthetic users work, we ask which function of user research they can serve. Read this way, the evidence is consistent: LLM-simulated participants can approximate averages under favorable conditions, but they fail to reproduce variation, subgroup structure, cultural differences, and lived experience.
We show that this split is built into the technology rather than a sign of immaturity: it follows from the training pipeline of current LLMs, and persona prompting moves the model's average without restoring human diversity. We combine the evidence into an evidence map, a decision aid that places common UX research activities by the kind of knowledge they need and by how much depends on their results.
We close with a research agenda for HCI, including controlled comparisons on established UX questionnaires.