SHARP: Social Harm Analysis via Risk Profiles for Measuring Inequities in Large Language Models
Alok Abhishek, Tushar Bandopadhyay, Lisa Erickson
TL;DR
SHARP addresses the inadequacy of scalar harm metrics by modeling social harm as a multivariate random variable and foregrounding tail risk with $\mathrm{CVaR}_{0.95}$. It achieves this through dimensional decomposition into $B$, $F$, $E$, and $K$, a union-of-failures aggregation reparameterized as additive log-risk, and distributional profiling of prompt-level risk; evaluation uses a judge-ensemble to produce $\bar{B}_{M,q},\bar{F}_{M,q},\bar{E}_{M,q},\bar{K}_{M,q}$ and the derived $L_{M,q}$. Empirically, SHARP analyzes $11$ frontier LLMs on $901$ prompts, uncovering substantial tail-risk heterogeneity that scalar means miss, with dimension-specific tail patterns (bias often driving the tail, ethics the least) and model-dependent tail drivers revealed by tail-attribution analysis. These findings support risk-aware governance: models can share similar mean risk yet differ dramatically in worst-case exposure, reinforcing the need for tail-sensitive, multidimensional evaluation in high-stakes deployment.
Abstract
Large language models (LLMs) are increasingly deployed in high-stakes domains, where rare but severe failures can result in irreversible harm. However, prevailing evaluation benchmarks often reduce complex social risk to mean-centered scalar scores, thereby obscuring distributional structure, cross-dimensional interactions, and worst-case behavior. This paper introduces Social Harm Analysis via Risk Profiles (SHARP), a framework for multidimensional, distribution-aware evaluation of social harm. SHARP models harm as a multivariate random variable and integrates explicit decomposition into bias, fairness, ethics, and epistemic reliability with a union-of-failures aggregation reparameterized as additive cumulative log-risk. The framework further employs risk-sensitive distributional statistics, with Conditional Value at Risk (CVaR95) as a primary metric, to characterize worst-case model behavior. Application of SHARP to eleven frontier LLMs, evaluated on a fixed corpus of n=901 socially sensitive prompts, reveals that models with similar average risk can exhibit more than twofold differences in tail exposure and volatility. Across models, dimension-wise marginal tail behavior varies systematically across harm dimensions, with bias exhibiting the strongest tail severities, epistemic and fairness risks occupying intermediate regimes, and ethical misalignment consistently lower; together, these patterns reveal heterogeneous, model-dependent failure structures that scalar benchmarks conflate. These findings indicate that responsible evaluation and governance of LLMs require moving beyond scalar averages toward multidimensional, tail-sensitive risk profiling.