Persona Engineering: A Field Guide to AI Synthetic Personas — Ishan Anand, InsightSciences.ai
A research team compared responses from over a thousand real people answering market surveys with AI agents playing the same role.
The agents matched people well on average but lacked human variation — and small changes in the prompt (instructions) caused purchasing decisions to swing dramatically, because the model guessed at hidden patterns that were never mentioned. The problem is that a model can look correct on average while simultaneously distorting minority groups and concealing the variation that actually matters. The solution is to first compare people against people to know what agreement is even possible, then judge the synthetic personas against that standard — and most importantly: treat them as economic actors that must be validated against real results before influencing any decision.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
