LLM 媒介仮想実験法の妥当性ランドスケープ：4 軸タクソノミーと経験的状態の体系的レビュー （2022-2026）

Abstract

Virtual experimentation—the use of large language models (LLMs) as synthetic respondents, populations, or scenario generators in psychological and social science research—expanded explosively after Argyle et al.'s (2023) "silicon sample" proposal. Subsequent empirical work (Bisbee et al. 2024; Boelaert et al. 2025; Hullman et al. 2026) has identified systematic methodological limitations: low variance, topicspecific machine bias, prompt sensitivity, WEIRD population skew, and algorithmic monoculture. Yet integrative reviews that map which validity claims hold under which simulation strategy for which epistemic purpose remain absent, leaving practitioners and reviewers without a shared evaluative framework. We propose a four-axis taxonomy spanning (1) Simulation Layer (LLMas- Respondent / LLM-as-Population / LLM-as-Scenario), (2) Validity Strategy (Face / Benchmark / Mechanistic-Process / Stress-Adversarial / Participatory-Ecological), (3) Epistemic Goal (Prediction-Substitution / Theory Probing / Design Exploration / Normative Representation), and (4) Empirical Status (Established / Contested / Frontier / Empirical-Void). We apply this taxonomy systematically to 18 representative studies, producing an empirical-status matrix that identifies three regions: (a) a saturated region (LLM-as-Respondent for survey-style aggregate prediction, contested but well-charted), (b) a frontier region (LLM-as- Scenario for design exploration and mechanism probing, promising but under-tested), and (c) an empirical-void region (normative representation of silent majorities, future generations, and marginalized groups). We derive four falsifiable hypotheses with pre-registered prediction thresholds and rejection criteria, and articulate eight honestdesign principles, each operationalized as a 3-5 point reporting checklist with concrete thresholds. We address known reviewer critiques (recursive limitation, selection bias, search transparency) by preemptive disclosure and conclude with implications for participatory design, research ethics, and the representation of voices that cannot speak for themselves.

Keywords