PluriHarms: Benchmarking the Full Spectrum of Human Judgments on AI Harm

Human disagreement on AI harms is systematic and predictable: Rather than treating disagreement as noise, the research shows that disagreement patterns follow identifiable rules—for example, people with different educational backgrounds or prior experiences with toxicity consistently judge the same AI outputs differently, suggesting that pluralistic AI systems should account for these patterns rather than ignore them.

Context matters more than binary categories: The benchmark demonstrates that whether an AI output is considered harmful depends heavily on specific details like whether risks are immediate versus theoretical, and whether harms are tangible versus abstract. This suggests current "yes/no" safety frameworks are too simplistic for real-world deployment.

Current AI safety models fall short on representing human diversity: When tested on PluriHarms, existing AI safety approaches showed that while personalization helped, there remains "considerable room for future progress"—indicating that today's systems still struggle to reflect the legitimate range of human values and judgments about what constitutes harm.

Demographic and psychological traits systematically shape harm judgments: Annotator characteristics like toxicity experience, education level, and psychological traits weren't random noise but explained meaningful disagreement patterns, suggesting that responsible AI systems need to acknowledge these differences rather than assume universal agreement on safety standards.

Targeting disagreement is a practical path forward: By deliberately studying 150 cases where humans disagree rather than focusing only on consensus cases, the research provides a concrete methodology for building AI systems that respect value pluralism rather than imposing a single definition of safety.