We study how water systems can use less carbon, recover more resources and put AI to responsible use—all while fulfilling their essential purpose: providing clean and safe water for all.
Patterns and mitigation of large language models’ hallucination in environmental domain
LLMs can produce confidently wrong results, known as hallucination. Responsible integration requires our understanding of how and why it happens, and the measures to mitigate. We found in the environmental domain, LLM hallucination exhibits a deep geographical and lingual bias. Looking inside the models, we traced recall of these limits to early-to-middle layers—but editing one limit there often shifted others in the same standard. Grounding answers in the regulation text proved the safer fix.
Putting language models into environmental context
What does AI need to work reliably on environmental problems? Our perspective connects environmental data, traceable evidence and tools—with physical constraints and human oversight throughout.
Hong Kong PolyU
