Underreporting AI Use

Today, I would like to share a recent AI SoTL article entitled, “Underreporting of AI use: The role of social desirability bias” by Ling et al. (2025) (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5464215 ).

The working paper addresses a subtle but crucial issue in research on human–AI interaction: users’ self-reported AI usage often underestimates actual use due to social desirability bias, a well-documented response bias in social science where individuals tailor responses to appear favorable to others. Social desirability bias can distort survey and self-report data when use of generative AI tools is stigmatized, policed, or subject to social evaluation, leading researchers and practitioners to underestimate how pervasively AI supports performance and decision-making across contexts.

The authors situate their analysis at the intersection of behavioral measurement and AI adoption research, drawing on classical social-psychological work on social desirability (Krumpal, 2013) as well as decision science frameworks that consider how context and social pressures shape self-presentation. They argue that conventional self-report instruments common in educational research, workplace studies, and technology adoption surveys may not accurately capture true AI use patterns because respondents tend to suppress admissions of reliance on AI tools when awareness of evaluation is present. The methodological insight is straightforward but impactful: without accounting for social desirability bias, inferences about AI adoption, impacts on learning, and the real-world integration of AI technologies may be systematically skewed.

In practical terms, this finding has implications for learning sciences and educational assessment. In educational research, self-report surveys are frequently used to gauge student strategies, study habits, and technology use. If students underreport use of AI tools like generative language models because they want to appear independent or “authentic” then educators and researchers may overestimate autonomous reasoning and underestimate AI-supported study activities. This suggests that measurement designs need to incorporate indirect or behavioral measures (usage logs, task performance patterns) alongside self-reports to more accurately capture learners’ interactions with AI. Such triangulation aligns with construct validity principles in educational measurement, where multiple indicators are combined to assess the true underlying construct rather than a potentially biased self-description.

Reference

Ling, Y., Kale, A., & Imas, A. (2025). Underreporting of AI use: The role of social desirability bias (SSRN Working Paper No. 5464215). Social Science Research Network.