Researchers have identified a critical flaw in evaluating large language models used as social simulators, including those employed as synthetic survey respondents. The primary concern is that these models can produce human-like outcomes without relying on the same rationale or reasoning patterns as humans. To address this issue, a study involving a 94-person sunscreen concept test was conducted to investigate the discrepancies between simulated and human-derived responses. The findings highlight the need for more robust evaluation methods that assess not only the accuracy of outcomes but also the underlying reasoning processes. This is crucial because a simulator that matches human outcomes but uses incorrect reasoning patterns can lead to flawed decision-making. The development of reason-mediated behavioral models for auditing large language models is essential to ensure the validity and reliability of social simulators, so what matters most to practitioners is the ability to distinguish between simulators that truly mimic human behavior and those that simply produce similar outcomes1.