Definition
Silicon sampling is the practice of prompting large language models to role-play as specific human demographics or personas and treating their outputs as stand-ins for real survey or poll responses. Researchers use it to cheaply approximate public opinion or behavioral data at scale, but its validity is contested since model outputs can reflect training-data artifacts and persona stereotypes rather than how those groups would actually respond.