Studying People to Study AI: Expert Perspectives on the Epistemic Fit and Barriers of Human Research in AI Safety & Ethics

A recent study, based on expert surveys and interviews, highlights a significant disconnect in Artificial Intelligence Safety & Ethics (AISE) research. While there is broad consensus among experts regarding the value of human-centric research for assessing AI risks, its practical integration is hampered by perceived validity issues, resource constraints, and epistemological challenges. This suggests a systemic undervaluing of empirical human research in favor of more technical evaluation methods, despite the increasing evidence of AI risks in human interactions.

Why it matters

This analysis reveals a critical gap in how AI risks are assessed, potentially leading to incomplete or inaccurate understanding of real-world impacts. For senior executives, addressing this requires re-evaluating current risk assessment methodologies to ensure a comprehensive, human-centered approach, thereby safeguarding organizational reputation and operational resilience in an AI-driven landscape.

Key insights

• AI safety risks are increasingly manifest in human interactions with AI technologies.
• Current predominant approaches to AI risk evaluation favor technical methods, such as model benchmarks and LLM simulations.
• Empirical research involving human subjects is often sidelined in AISE.
• There is expert consensus that human research is valuable for generating evidence in AISE.
• Adoption and acceptance of human research in AISE are constrained by perceived validity issues.
• Tangible resource barriers hinder the implementation of human research in AISE.
• Epistemic barriers further limit the integration of human research findings.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.05656