Our study of LLM hallucination on environmental regulations is published in ES&T

Our study, Patterns and Mitigation of Large Language Models’ Hallucinations on Environmental Regulatory Facts, is now online in Environmental Science & Technology.

Language models can be confidently wrong, and using them responsibly means understanding how and why. On environmental regulatory facts, we found their errors follow a deep geographical and linguistic bias. Looking inside the models, we traced where these facts are recalled, to early-to-middle layers, and found that editing one limit there often shifts others in the same standard. Grounding answers in the regulation text proved the safer fix. Read the study and let us know what you think!