DGEval leaves dangerous-goods AI at 14.2%

National Cargo Bureau’s DGEval benchmark tested 13 AI models from six providers across 1,678 dangerous-goods questions, and the models averaged only 14.2% when asked to identify the regulatory section supporting an answer. The study announcement reports that Gemini 3.1 Pro reached 95.8% on multiple-choice recall versus an 83.8% corrected human baseline, but performance fell by an average of 35 points on stowage and segregation; the preprint supplies the methodology. Operators building compliance copilots should treat retrieval as an assistant and require human verification against the IMDG Code, a control boundary AISI’s agent-evaluation incident also makes hard to ignore.