Expert Survey Reveals Barriers to Human Research in AI Safety and Ethics

A new study published on arXiv (ID: 2608.05656) investigates the underutilization of empirical human research in the field of AI Safety and Ethics (AISE). The research, conducted via an expert survey of 93 participants and in-depth interviews with 17 experts, explores why technical methods like model benchmarks and LLM simulations are favored over studies involving human subjects. The findings indicate that while there is broad consensus on the value of human research for generating evidence, its adoption is hindered by perceived validity issues, resource constraints, epistemic and personal methodological preferences, and infrastructural limitations within the research community. Notably, researchers with technical backgrounds tend to undervalue human research and collaborate less across disciplinary boundaries. The study includes perspectives from experts in Technical, Sociotechnical, Governance, and Normative fields, highlighting a significant gap in the acceptance of human-centric approaches in AI safety evaluation.