PCFBench Tests AI Reliability in Product Carbon Footprint Estimation

Summary

Researchers introduce PCFBench, a 614-item benchmark that evaluates AI systems on decomposition, retrieval, ontology matching, and numerical extraction for product carbon footprint estimation. Across eight frontier LLMs, no model dominates. While leading systems stay within twice the declared total for 77% of products, step-by-step generation reduces this to 37–58%, and only 45–75% of outputs satisfy mass conservation. The dataset and evaluation harness are released for further research.