New DualChem Benchmark Evaluates AI Safety in Chemistry
The evaluation tool tracks model accuracy and refusal behaviors across 600 sensitive chemical design and synthesis tasks.
Updated on Oct. 28, 2026 in Artificial Intelligence
Live Poll
Should AI models be strictly restricted from answering queries that carry potential dual-use security risks?
Scale AI has released DualChem, a new benchmark designed to evaluate how AI models handle dual-use chemistry tasks. The project, conducted in partnership with the Center for Advancing Innovation and Standards for Super Intelligence, assesses model performance on safety and technical accuracy.
Why it matters
The benchmark provides a framework for the government to measure AI risks in chemistry, a critical domain for national security oversight. By identifying performance gaps, it helps regulators gauge whether current frontier models can safely navigate sensitive chemical information.
DualChem tests models against 600 expert-curated questions covering molecular design and synthesis. Frontier models show a performance spread where top-tier systems reached 67% accuracy, though many models failed to consistently refuse queries despite flagging them as high-risk.
The players
Scale AI
A provider of data infrastructure and evaluation services that supplies training and testing pipelines for large-scale machine learning models.
Center for Advancing Innovation and Standards for Super Intelligence
A research entity focused on developing safety standards and evaluation frameworks for advanced artificial intelligence systems.
The details
DualChem uses a set of 600 questions developed by subject matter experts and verified by Ph.D.-level evaluators. To prevent data contamination, the test set is held out from training processes, ensuring AI systems encounter the problems for the first time during evaluation. The benchmark assesses both the model's ability to provide accurate chemistry information and its tendency to ignore safety flags when responding to dual-use scenarios.
Timeline
February 2025: Partnership between Scale AI and the Center for Advancing Innovation and Standards for Super Intelligence was announced.
October 28, 2026: The DualChem benchmark was officially released.
The Tech Race
This release marks a shift toward domain-specific safety benchmarks in the race to secure frontier AI models. It follows the government's push for external validation protocols, mirroring larger efforts to formalize safety thresholds before these models gain broader access to scientific data.
This tool is currently a research-stage assessment for government oversight rather than a commercial product for end-users. Future iterations may lead to broader industry safety standards, but currently it serves primarily as a risk-measurement instrument for regulatory agencies.
The takeaway
The gap between flagging a risk and refusing an answer highlights a persistent failure in current model safety alignment. Stakeholders should monitor whether future government-mandated evaluations lead to stricter performance requirements for models handling sensitive technical knowledge.
Further reading
For more on industry-wide model testing standards, visit our Artificial Intelligence section.
Source note: This article includes information reported by FedScoop.
Live Poll
Should AI models be strictly restricted from answering queries that carry potential dual-use security risks?
