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

1. 1Define harmful capability uplift as a core AI safety metric
2. 2Develop human-centered evaluation methods to measure uplift
3. 3Assess marginal increase in user's ability to cause harm with frontier models
4. 4Ground evaluations in real-world scenarios and user interactions

AI researchers and safety experts on a team benefit from this framework as it provides a novel approach to evaluating AI safety, while product managers and entrepreneurs can use it to inform responsible AI development

💡 Harmful capability uplift should be a core metric in AI safety evaluations

💡 New framework for evaluating human-AI safety: measuring harmful capability uplift