Mirrors and Their Limits: A Review of Mamun's "In Silico Society: A Comprehensive Analysis of Generative Agent Simulations at Population Scale"

This review examines Dr. Syed Muntasir Mamun’s paper “In Silico Society: A Comprehensive Analysis of Generative Agent Simulations at Population Scale” (KSCCN, 2026), an extended analysis of the study “Generative Agent Simulations of 1,000 People” by Park and colleagues (2024). That study built generative agents for 1,052 real individuals by placing each person’s full interview transcript into a large language model’s prompt, and reported that the agents reproduced participants’ survey answers with a normalised accuracy of 85% and replicated the same four of five experiments that the human participants replicated. Mamun’s paper explains the study’s data pipeline, architecture, validation and ethical framework, and discusses emergent social dynamics, methodological tensions and future applications. This review identifies three strengths: a clear exposition of a complex study, including a careful explanation of what “normalised” accuracy means; a useful distinction between believability and predictive fidelity; and a balanced treatment of interpretability, cost, bias and ethics. It then raises four points for further discussion: separating the findings of the reviewed study from those of other studies; the presentation of the replication results; the reliance on secondary sources; and the strength of some concluding claims. It closes with broader implications for researchers and readers.