Research Summary: Measuring AI harms with multidimensional Lorenz Zonoids

Current AI risk management models are insufficient because they focus primarily on compliance and likelihood, rather than the real-world severity and multidimensional nature of harms. This limitation prevents effective prioritization of interventions. A new methodology proposes using multidimensional Lorenz Zonoids and Gini indices to model harm data, which is typically ordinal and multidimensional, thereby enabling a more effective risk assessment that considers harm severity.

Why it matters

The development of robust methods for measuring and assessing AI harms is critical for effective governance and responsible deployment of AI systems. Accurately quantifying the severity and multidimensional nature of AI harms will enable better strategic decisions regarding risk mitigation and intervention priorities across various sectors.

Key insights

• Existing AI risk management models are primarily compliance-driven and provider-centric, lacking focus on the severity of harms.
• These models offer limited insight into the dangerousness of harms and the prioritization of interventions.
• The inherent nature of harm data, being ordinal and multidimensional, amplifies the problem of effective risk assessment.
• The research proposes extending Lorenz Zonoids and Gini indices to a multidimensional context to model harm data.
• This extended methodology aims to provide an effective risk assessment that accounts for harm severity, moving beyond mere likelihood.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.16004