Show GN: ManyPerson - Korean AI Persona Public Opinion Simulator Based on Statistics Korea MDIS

Raw data from Statistics Korea MDIS's 2025 Household Finance and Welfare Survey was joined at the household master level (34,880 records) and household member level (69,929 records) to construct approximately 41,000 AI personas reflecting the demographics and household structure of Korean society.

Sex, age, household relationship, marital status, education, employment status, occupation major category, industry major category, whether in the Seoul metropolitan area, housing type, number of household members, and income quintile were preserved as persona attributes, along with household total income, disposable income, consumption expenditure, assets, liabilities, and net worth. Household income was re-estimated at the individual level, distributing individual annual income via heuristics based on whether the person is the householder, spouse, child, unemployed, or retired.

Additional details were generated with Gemini. Occupation name, MBTI, personality, hobbies, hometown, and first-person self-introduction were created, constrained so as not to conflict with annual income, assets, liabilities, and occupation group. Initially there were inconsistencies such as high-income titles being attached to low-income individuals, which were corrected through a separate Phase 2 regeneration that re-adjusted occupation/bio.

When given a natural language question, the service filters target personas by sex, age group, region, income quintile, education, occupation group, number of household members, marital status, etc., and aggregates the responses of the selected AI citizens. Results are organized as positive/neutral/negative or multiple-choice distributions, and weighted results and cross-tabulations applying Statistics Korea's weight are also provided.

The implementation consists of Node.js, Express, EJS, PostgreSQL/Cloud SQL (JSONB), Valkey, GKE Autopilot, the Gemini 3 series, Vertex AI Flex API, and SSE streaming. However, this is not an actual opinion poll but a virtual simulation based on AI personas, and variables such as detailed regional data, political orientation, and real-time issue awareness require separate estimation.