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

Human research is underutilized despite broad agreement on its value: While 93 surveyed AI Safety & Ethics researchers acknowledged that empirical studies with human subjects are important for understanding AI risks, technical methods like model benchmarks and simulations remain dominant, creating a significant gap between what experts believe matters and what gets funded and published.

Technical researchers show resistance to human-centered methods: The study found that researchers with technical backgrounds tend to value human research less and collaborate across disciplines less frequently than their sociotechnical, governance, and normative peers, revealing an "epistemic tension" that may be limiting how comprehensively AI safety risks are evaluated.

Practical barriers, not just philosophical disagreements, block human research: Beyond methodological preferences, researchers face concrete obstacles including resource constraints, validity concerns about human studies, and lack of infrastructure support from the broader research community—suggesting that increasing human research requires systemic changes, not just persuasion.

Risk of "human-washing" without genuine integration: The authors warn against superficial adoption of human research as a checkbox for responsible AI, emphasizing that meaningful progress requires addressing the underlying epistemic and structural barriers rather than performative inclusion of human studies.

Interdisciplinary collaboration is key to progress: The research suggests that bridging the gap between technical and human-centered approaches requires not just individual researchers changing methods, but fostering genuine collaboration across different disciplinary backgrounds in AI safety and ethics work.