Dualchem Evaluating Ai Models Dual Use Chemistry Risks

Gives a concrete benchmark for how frontier models handle ambiguous dual-use requests. Relevant if you build on or evaluate models where safeguards affect permitted use.

• Scale's DualChem has 600 PhD-reviewed chemistry questions where the same request could serve legitimate research or harm. It is the first benchmark from Scale's research agreement with NIST's CAISSI, with US government experts shaping its scope and taxonomy.
• Across 30+ frontier models, average accuracy was about 50% and the best reached about 67%. Model help lifted chemistry bachelor's-degree holders from about 32% to about 48% on dual-use tasks, which Scale calls a meaningful capability gap.
• Models often recognized a question as potentially dual-use and still gave a substantive answer; Scale found risk identification and refusal only weakly connected across models.
• When asked to rate how severe dual-use scenarios were, models aligned only weakly with expert labels, which Scale reads as flagging surface sensitivity without reasoning about what makes a question dual-use. It concludes refusal-based safety alone may be insufficient.

• benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

• articleSwe Atlas Complete
• articleLlm Novice Uplift
• articleHil
• articleAspi

• articleStrategic Warning For Ai Risk Progress And Insights From Our Frontier Red TeamAnthropic research
• articleNuclear Safeguards For AiAnthropic research
• articleAi2 uses psychometrics to audit what LLM safety benchmarks measureAi2

Checking sign-in…

Loading comments…