Evaluating Human-AI Safety: A Framework for Measuring Harmful Capability Uplift

In a position paper, the researchers argue that AI safety research should shift toward human-centered evaluation, taking harmful capability uplift as a core safety metric, i.e., the marginal increase in a user's capability to cause harm using frontier models relative to the existing level of conventional tools. The authors argue that current frontier AI safety evaluations overemphasize static benchmarks, third-party annotation, and red teaming; the paper grounds this metric in existing social science research, provides concrete methodological guidance for systematic measurement, and lists action steps for developers, researchers, funders, and regulators to make it standard practice.