← Home

The model playing people

Simulation

Language models are increasingly used to stand in for people. I test when that works by measuring simulated respondents and simulated deliberation against real survey and deliberative-poll data, and I look for methods that bring simulations closer to the populations they claim to represent.

Questions

  • Do simulated respondents reproduce the spread of opinions in a real population?
  • Does simulated deliberation change minds the way real discussion does?

Projects

Human–AI Deliberation for SBS D-Forum 2026

For SBS D-Forum 2026, we built a mini-public of LLM agents to explore how public opinion forms on polarizing issues. Ten thousand personas grounded in Korean survey data deliberate in 1,000 rooms of ten.

We compare models, and measure the diversity of arguments and how minority opinions shift over the discussion. The results are presented at the forum's KAIST session.

  1. Partner

    SBS · with KAIST KIXLAB

  2. Period

    2026

  3. Team

    Prof. Juho Kim (KAIST KIXLAB), Prof. Nuri Kim, and Prof. Jihee Kim (KAIST EconAI)

  4. My role

    Coordinated the EconAI side and ran the multi-model simulations.

Full project page →

AI-Agent-Based Social Simulation: Validation and Alignment with Human Behavior

A 2026 AI-Science Hub project funded by the Ministry of Science and ICT. We build a framework for social simulation with LLM agents, then compare and validate it against human social behavior, including data from real deliberative polls.

The goal is simulations that are reproducible and whose populations can be checked against the people they stand in for.

  1. Partner

    Ministry of Science and ICT · AI-Science Hub

  2. Period

    2026

  3. Team

    Prof. Jihee Kim (PI), with Prof. Nuri Kim and Prof. Joseph Seering

  4. My role

    Simulation experiments and analysis.

Full project page →

Papers

  1. World Bank 2026

    Who Differs, and How Many? Assessing Language Models’ Potential to Complement Opinion Surveys

    Chaemin Jang, Jihee Kim

    Better Data for Better Lives: Innovations in Survey Measurement in the Age of AI (World Bank) 2026

    Abstract

  2. NeurIPS Workshop 2026

    From Simulated Citizens to Simulated Deliberation: Challenges in Representation and Interaction

    Chaemin Jang, Junsik Min, Jaewoo Choi, Donggyu Lee, Haiin Lee, Junyoung Park, Namhee Kim, Hyunwoo Kim, Jungwon Kim, Juho Kim, Nuri Kim, Jihee Kim

    SocialAgent and UserSim Workshops @ NeurIPS 2026

    Spotlight at UserSim @ NeurIPS 2026

    Simulating a deliberating public needs agents whose opinions reflect a population and change through discussion. Across six LLMs, national surveys, controlled experiments and the 2017 Shin-Kori deliberative poll, agents argue convincingly but represent opinion too narrowly, and change positions even when they cannot see each other.

    arXivPDFCode

  3. EMNLP Findings 2026

    Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe

    Chaemin Jang, Dongman Lee, Jihee Kim

    Findings of EMNLP 2026

    Silicon sampling treats each model call as an independent survey respondent, but instruction-tuned models collapse to one answer: the same persona gives the same answer on more than half of a public-opinion benchmark. Asking the model to describe the distribution in a single call more than halves error against human survey data, and Prompt-Perturbed Argyle cuts error by 21% when per-persona answers are needed.

    arXivPDFCode

Other directions