Definition
Plain language
Using an AI as a stand-in for human survey respondents by giving it a fake persona and tallying its answers as if it were polling real people.
As stated in the literature
The practice of prompting an LLM with demographic personas and treating repeated per-call outputs as independent draws from a simulated population; relies on the flawed assumption that each call is an i.i.d. respondent sample.
Also called: silicon sampling method
Why it matters: This practice matters because it promises cheap, fast stand-ins for expensive human surveys, but it breaks down if each answer isn't really an independent, realistic response.
For example, a researcher might tell an AI to 'answer as a 40-year-old teacher from Ohio' hundreds of times and treat the replies as if they were a real poll of such people.
Heard on the show
“Over the last few years there's been a fast-growing shortcut called silicon sampling.”Episode 230 — Why AI Survey Panels Break Before the Dice Ever Roll