Scale AI has launched DualChem, a new benchmark designed to test whether leading artificial intelligence models can recognise when a chemistry question could help both legitimate research and harmful activity, and whether they answer anyway. Released on October 28, 2026, it is the first benchmark to emerge from Scale’s research collaboration with the National Institute of Standards and Technology’s Center for Advancing Innovation and Standards for Super Intelligence, known as CAISSI, under an agreement signed in February 2025.
A Test Built For Ambiguity
Most AI benchmarks measure straightforward capability: reasoning accuracy, coding skill, or factual recall. DualChem targets something slipperier. Chemistry is, in Scale’s words, one of the clearest testbeds for dual-use risk, because the same guidance that helps a graduate student troubleshoot an experiment could also lower the barrier for harmful misuse. Existing safety tests often focus on explicit harmful requests, which are easy to refuse. Real chemistry questions are rarely that clean.
The benchmark consists of 600 expert-curated questions spanning laboratory chemistry domains including molecular design, synthesis procedures, purification, validation, and information management. The questions were developed by chemistry experts, most with advanced degrees, and reviewed by PhD-level evaluators before inclusion. Rather than synthetic or openly adversarial prompts, DualChem uses realistic questions with ambiguous dual-use applications, where the same request could support either legitimate research or harmful misuse depending on context.
Scale evaluated 31 state-of-the-art frontier models on the benchmark. The results, published alongside the launch, paint a picture of capable systems with unreliable judgment. Frontier models averaged around 50 percent accuracy on the dual-use chemistry questions, with top performers reaching about 67 percent. More striking was the uplift test: participants with bachelor’s degrees in chemistry improved from around 32 percent to around 48 percent on dual-use tasks when assisted by a frontier model.
Models See The Risk And Answer Anyway
The most consequential finding is a gap between recognition and restraint. Models frequently identified questions as potentially dual-use, Scale reports, but still provided substantive answers in most cases. Risk identification and refusal behaviour appeared only weakly connected, meaning a model could flag a question as sensitive and then answer it regardless.
A second gap concerned judgment. When tested on assigning severity scores to dual-use scenarios, model predictions showed only weak alignment with expert judgments and significant variability across cases. Current models can flag surface-level sensitivity without reasoning well about what actually makes a question dual-use, Scale says, and refusal-based safety assumes a clean decision point that these ambiguous cases do not offer.
Why Governments Are Watching
The CAISSI partnership is central to the launch. Scale developed the benchmark with input from CAISSI and US government subject-matter experts, who informed the scope and question taxonomy. FedScoop, reporting the release, quoted Scale’s global head of policy and government relations, Max Fenkell, describing DualChem as giving government “another tool on how to assess AI risk.” For agencies tasked with judging increasingly capable frontier models, a shared test for ambiguous chemistry scenarios is more useful than a vendor’s own safety claims.
The timing also fits a wider shift toward domain-specific safety testing. As frontier models grow more capable in specialised scientific fields, general benchmarks say less about real-world risk. A chemistry benchmark, a biology benchmark, or a cyber benchmark each probe a different failure surface, and regulators and procurement teams increasingly want evidence tied to the domain where a model will actually be used. DualChem extends the NIST AI Risk Management Framework into one such domain, according to the launch coverage.
What DualChem Does Not Prove
It is worth keeping the benchmark’s limits in view. DualChem measures how models behave on curated questions, not how they would behave in a live deployment with additional safeguards, monitoring, or human oversight layered on top.
Nor does the benchmark settle what the right response should be. Refusing every ambiguous chemistry question would block legitimate research, while answering them all invites misuse. The interesting policy question DualChem raises is not whether models should refuse more, but whether they can reason about context well enough to distinguish a graduate student’s troubleshooting from a harmful request, and whether that reasoning can be made reliable rather than approximate.
Why It Matters
This matters because AI safety debates often focus on dramatic, explicit misuse, while the realistic risk may look far more ordinary: a plausible question, a helpful answer, and a user whose intent the model never probed. DualChem gives researchers, policymakers, and buyers a shared yardstick for that grey zone, much as provenance tools such as WRITER’s enterprise watermarking give them a yardstick for tracing AI-generated text. If refusal alone is insufficient for complex dual-use scenarios, as Scale argues, then evaluation, monitoring, and deployment controls will have to do more of the work, and benchmarks like this one are how the gaps get measured. The same trust questions run through Bank of America’s warning about AI and account security and research showing analysts deferred to AI agents 86 percent of the time: capable systems are being trusted before their judgment is fully understood.
Background
Scale AI builds data and evaluation infrastructure for AI development and works with governments and national AI safety institutes on evaluation frameworks. DualChem, its first benchmark under the CAISSI collaboration, tests 31 frontier models on 600 expert-curated dual-use chemistry questions, finding high capability paired with weak links between risk recognition, severity judgment, and refusal.
Source: Scale announcement, FedScoop report
0 Comments