Scale AI's DualChem benchmark tests how LLMs handle dual-use chemistry questions

What DualChem measures

According to Scale, most existing AI benchmarks test general capability, such as reasoning accuracy, coding performance or factual recall. The company argues these do not capture how a model behaves in ambiguous, high-stakes situations. DualChem is built to measure the latter: whether a model gives the same answer whether it supports legitimate scientific work or harmful activity.

The benchmark consists of 600 expert-curated questions, reviewed by PhD-level expert evaluators before inclusion. They span several laboratory chemistry domains: molecular design, synthesis procedures, purification, validation and information management. Chemistry experts, most with advanced degrees, wrote the questions to reflect realistic dual-use ambiguity. Scale says it avoided synthetic or explicitly adversarial prompts, which it argues can fail to reflect how real chemistry questions are asked.

Scale describes DualChem as one of the first benchmarks to evaluate whether language models assist in such scenarios, and where current safeguards may fall short. In the real world, it notes, that has to be balanced against not inhibiting beneficial uses.

Government collaboration

DualChem is the first benchmark to come out of a Collaborative Research and Development Agreement that Scale signed with the NIST Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) in February 2025. Scale says it combined its own evaluation expertise with input from CAISSI and U.S. Government subject-matter experts, who informed the scope and the question taxonomy. The results are described in a research paper titled "DualChem: Measuring LLM Uplift and Dual-Use Risk Awareness in Chemistry."

Key findings

Scale reports that frontier models perform strongly on dual-use chemistry questions. Average accuracy across the 31 evaluated models is about 50%, and the top performers reach about 67%.

The company also measured human uplift. Participants with bachelor's degrees in chemistry improved from around 32% to around 48% on dual-use chemistry tasks when assisted by an LLM. Scale calls this a meaningful capability gap with potential real-world implications.

• Risk recognition without refusal: models frequently identified questions as potentially dual-use but still gave substantive answers in most cases. Across the evaluated models, risk identification and refusal behavior appeared only weakly connected.
• Differences from experts: when asked to assign severity scores to expert-annotated examples labeled benign or harmful, model predictions showed only weak alignment with human judgments and significant variability across cases.

Scale concludes that current models can flag surface-level sensitivity without reasoning about what actually makes a question dual-use. It also says the results suggest refusal-based safety mechanisms alone may be insufficient for complex dual-use scenarios.

What it means for oversight

Scale positions the benchmark as a tool for institutions that assess frontier models. It says independent assessments can help policymakers and procurement teams understand how models behave in ambiguous dual-use scenarios and where safeguards may fall short.

The company also calls for public and private partnerships on shared evaluation methods that support institutional oversight, including adversarial stress-testing, robustness analysis under distribution shifts, and measurement approaches aligned with evaluators' needs. Scale Labs says it will keep working with governments and national AI safety institutes on frontier AI evaluation and deployment, with the aim of clarifying how models behave in different contexts and which safeguards, testing and evaluation mechanisms support reliable deployment.