SHARP: Social Harm Analysis via Risk Profiles for Measuring Inequities in Large Language Models

Average scores mask dangerous worst-case failures: Models with similar average performance scores can differ by more than twofold in their worst-case harm exposure, meaning that relying on mean-centered benchmarks alone fails to identify which models are safest for high-stakes applications.

Different types of harm require separate evaluation: SHARP breaks down social harm into four distinct dimensions—bias, fairness, ethics, and epistemic reliability—revealing that models fail differently across these categories, with bias showing the most severe tail risks and ethical misalignment being comparatively lower.

Tail risk metrics matter more than averages: Using Conditional Value at Risk (CVaR95) to measure worst-case scenarios rather than average performance provides a more realistic picture of how LLMs will behave in rare but critical situations where harm is most severe.

Current benchmarks obscure critical inequities: Traditional evaluation methods reduce complex social risks to single numbers, hiding important patterns about which groups or scenarios are most vulnerable to model failures and how different failure types interact.

Frontier models show heterogeneous failure patterns: Testing eleven advanced LLMs on 901 socially sensitive prompts revealed that each model has a unique "failure structure"—meaning no single model is uniformly safer across all harm dimensions, requiring more nuanced deployment decisions.