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

Current AI safety tests miss a critical question: Most evaluations focus on whether AI systems themselves cause harm, but this paper argues we should instead measure whether AI makes it easier for people to cause harm—a distinction the authors call "harmful capability uplift" that existing benchmarks largely ignore.

The metric matters more than the model's internal safeguards: A frontier AI model might refuse to help with harmful tasks directly, but still enable users to accomplish those tasks more effectively than before. The paper suggests this practical capability increase is what safety evaluations should prioritize measuring.

Real-world context is essential: The framework emphasizes "human-centered evaluations" grounded in social science research rather than abstract benchmarks, meaning safety assessments should consider actual user scenarios and compare AI assistance against "conventional tools already available"—not just whether harm is theoretically possible.

Measurement requires systematic methodology: The paper provides concrete guidance for developers and researchers to standardize how harmful capability uplift is evaluated, suggesting this should become standard practice across the industry rather than ad-hoc or inconsistent approaches.

Multiple stakeholders need to act: The authors call on AI developers, researchers, funding organizations, and regulators to collectively adopt harmful capability uplift evaluation, indicating this shift requires coordination across the entire AI ecosystem rather than isolated efforts.