Can large language models replicate human decision-making in water system management?

Growing evidence across diverse disciplines demonstrates that Large Language Models (LLMs) exhibit human-like decision making in various economic, psychological, and social experiments. This capability suggests an exciting new paradigm for modeling complex human-environmental systems. Traditional agent-based models require prescribed representations of human behavior, often muddling the line between imposed and emergent system dynamics. LLMs directly address this limitation by enabling flexible, context-rich, and state-aware agent profiles that innately infer how humans shape and are shaped by their environment. As a first step in this expansive, multidisciplinary direction of research, we embed an LLM as a reservoir operator tasked with making monthly decisions for downstream water supply allocation. We benchmark the LLM’s performance, quantified by the volume of water supply shortage, against traditional control methods including deterministic and stochastic dynamic programming. Our results demonstrate that even without specific training examples, LLMs demonstrate risk-averse hedging policies and, in some cases, outperform traditional methods, particularly in adapting to changing conditions. Furthermore, we analyze the chain-of-thought (CoT) reasoning behind each of the LLM's decisions to interpret their operational logic, finding that decisions are justified using sophisticated quantitative reasoning and an understanding of risk. This research highlights the potential for LLMs to serve as a simulation engine for complex multi-actor systems, as well as to bridge the gap between quantitative models and the qualitative, value-laden decisions that confront policymakers and water managers.