Toward Psychological Digital Twins: Evaluating Large Language Models for Simulating Assessment Responses From Interview Data

Large language models (LLMs) have demonstrated remarkable capabilities in simulating human behavior, raising the prospect of creating “digital twins” that can predict individual responses to psychological assessments. However, systematic validation of LLM-based respondent simulation remains limited, particularly for non-English populations. We present a comprehensive evaluation of five state-of-the-art LLMs—GPT-5, Gemini 2.5 Pro, Kimi-k2-0905, Kimi-K2-Thinking, and Qwen3-235B—in their ability to simulate responses to psychological questionnaires based on semi-structured interview transcripts. Using data from 645 Russian-speaking participants who completed interviews covering seven life domains and seven validated psychological instruments (38 scales total), we assessed simulation accuracy via correlations between actual and predicted scale scores, complemented by mean absolute error, root mean squared error, bias, intraclass correlation, and calibration analyses. Results revealed substantial variation across constructs and models: correlations ranged from non-significant (hedonism, proportionality) to strong ( <tex-math notation="LaTeX">$r = 0.67$ </tex-math> for religiosity, <tex-math notation="LaTeX">$r = 0.63$ </tex-math> for extraversion). After Benjamini–Hochberg FDR correction across 190 scale–model tests, 184 remained significant ( <tex-math notation="LaTeX">padj< .05 </tex-math>), confirming the robustness of the rank-order findings, while error and calibration analyses revealed systematic underestimation of dark triad and material-values scales. GPT-5 achieved the highest overall accuracy, while Kimi-K2-Thinking performed best among open-source alternatives. We further demonstrate that context reduction improves accuracy for smaller models but not for larger ones (observed for the Material Values Scale and two models, and therefore offered as a hypothesis for broader testing rather than a general rule), and that adding LLM-generated psychological portraits to prompts yields construct-specific benefits. Our results indicate that current LLMs can approximate psychological assessment responses with moderate accuracy for verbally transparent constructs, but they do not constitute reliable digital twins of individuals; the predicted scores should be interpreted as probabilistic approximations rather than as substitutes for direct human assessment. Our findings establish the first large-scale validation of LLM-based respondent simulation on a Russian sample, provide practical guidelines for model selection and prompt engineering, and highlight which psychological constructs are most amenable to, and most resistant to, psychological digital twin approaches.