AI personas are cheap. That doesn't make them your sample

Why everyone's tempted

Tell a model who to pretend to be and it answers instantly, for pennies, at any sample size. For early-stage idea sorting, that's genuinely useful. As a replacement for human respondents, it's a trap — a confident, articulate, average-of-the-internet trap.

Legitimate uses

• Breaking your questionnaire: Catch routing bugs and confusing wording before paying for field.
• Sorting hypotheses: Decide which concepts deserve real-sample testing.
• Planning hard-to-reach screens: Rehearse screening logic for rare audiences before fielding.
• Training QC models: Generate edge cases to sharpen fraud and coding classifiers.

Where it goes wrong

Models echo their training data: overconfident on majority-culture views, thin on low-income, low-digital-literacy, and non-Western perspectives — often exactly the people you need to hear. And synthetic answers can't surprise you the way real markets do. No model tells you the thing nobody has said yet.

Our rule

We use synthetic respondents as a rehearsal stage, never as the audience. When the decision is worth real money, we talk to real people — and we say so in the methodology section.