Assessing and Adjusting Social Desirability Bias in Self-Reported Surveys Using Large Language Models

Abstract

Abstract

Much Information Systems (IS) research relies on self-reported surveys to examine individuals’ beliefs, attitudes, and behaviors. However, self-reported surveys are susceptible to social desirability (SD) bias, whereby respondents tend to overreport socially desirable behaviors and underreport socially undesirable ones. Prior research has proposed several approaches to assess and control SD bias. Among them, covariance techniques that incorporate SD scales—such as impression management (IM) and self-deceptive enhancement (SDE)—have been widely used (Kim et al., 2026). Although these approaches help mitigate SD bias, they have notable limitations. First, including SD scales increases survey length and respondent burden. Second, because these techniques rely on statistical control, they may not fully eliminate response distortion, particularly when both moralistic and egotistic biases—captured by the IM and SDE scales—simultaneously influence responses. To address these limitations, this study proposes a novel approach that leverages large language models (LLMs) to assess and adjust for SD bias in survey responses. Specifically, we utilize LLMs to estimate the susceptibility of survey responses to SD bias and to generate bias-adjusted estimates of survey responses. We then compare the proposed LLM-based approach with conventional SD control techniques to evaluate its effectiveness in diagnosing and mitigating response distortion. This study introduces an AI-assisted approach that complements traditional SD control techniques and provides a new methodological tool for improving the validity of self-reported survey research in IS.