Introducing DualChem

Scale AI has introduced DualChem, a benchmark designed to evaluate AI safety in dual-use chemistry scenarios. Developed in collaboration with NIST, it measures whether large language models provide assistance to requests that could be used for either legitimate research or harmful activities. The benchmark uses 600 expert-curated questions to test models in ambiguous, high-stakes environments. Findings indicate that frontier models often provide substantive answers even when they identify potential risks, and their risk assessments do not align well with human expert judgments. This research suggests that simple refusal-based safety mechanisms are insufficient for handling complex dual-use risks.