research methods
Synthetic personas
What are synthetic personas, and what's the risk of using them?
Synthetic personas are user personas generated by AI from a prompt rather than built from research. Ask a model for "three personas for a banking app" and you get instant, polished, plausible-looking profiles. The trouble is they're assembled from training-data averages and stereotypes, not from real people - so they look like research output while being, in effect, confident fiction.
Also known as: synthetic personas, ai personas, generated personas
Prefer to watch?
Watch the recap 0:55
The demo
Two personas, same tidy template. One was conjured from an AI prompt; one was built from real interviews. Flip between them - and notice how alike they look, and how differently they hold up.
What this demo shows (text version)
Two personas are shown in the same professional template. The AI-generated one is fluent but generic and lightly stereotyped - "Sarah, 34, busy professional, wants things to be quick and easy, frustrated by complexity" - with goals and frustrations that could describe anyone and come from nothing. The research-based one is specific and slightly surprising - a real behaviour, an unexpected workaround, a direct quote - and every claim traces back to actual interviews.
The point is that the polished format hides whether there's any substance underneath. A persona's whole value is being a memorable summary of real research; a synthetic one keeps the format and discards the evidence, baking in stereotypes and the team's assumptions while looking authoritative. Treat synthetic personas as, at most, disposable placeholders - any persona that drives real decisions has to come from real people.
A persona's entire value is that it's a memorable summary of real research; generate one from a prompt and you keep the format while throwing away the substance. Synthetic personas are fast and convincing but built from stereotypes and your own assumptions, so they launder guesses into an official-looking artefact and quietly bake bias in. Use them at most as throwaway placeholders to pressure-test thinking - never as a stand-in for talking to users.
Side by side, the AI persona looks just as professional - a name, a photo, neat goals and frustrations. But read closely: it's generic, a little stereotyped, and grounded in nothing. The research-based one is specific, slightly surprising, and traceable to real quotes. Same template; one is evidence, the other is a plausible-sounding average dressed up as a user.
A persona is only as good as the research behind it. Its job is to make real, messy findings memorable and shareable so a team designs for actual users instead of themselves. A synthetic persona inverts that: it provides the memorable summary with no findings underneath, so the team rallies around a character that represents no one - and can't tell, from the polished output, that anything is missing.
The specific harms are stereotyping and false confidence. Generated from training-data averages, synthetic personas tend to reach for clichés (the "busy mum", the "tech-savvy millennial"), smoothing over real variation and sometimes encoding bias. And because they look authoritative, they invite teams to skip real research and to over-trust a profile that has no evidential basis - amplifying the team's own assumptions rather than challenging them.
There may be narrow uses - a disposable placeholder to structure a workshop, or a prompt to surface your own hidden assumptions so you can go and test them. But the line is the same as for synthetic users: they can support preparation, never substitute for evidence. A persona that affects real decisions must come from real people, or it's just your guesses with a stock photo.