NCEAS Publishes 10 Rules for Using AI in Environmental Science

Researchers codify lessons from the trenches

A team at UC Santa Barbara's National Center for Ecological Analysis and Synthesis has published formal guidelines for using generative AI in environmental data science, converting internal struggles into community standards now available to the entire field.

The work, published in PLOS Computational Biology, emerged from practical challenges the center's researchers faced while building the Wildfire Resilience Index — an open-access tool measuring community and landscape preparedness across two countries and 13 jurisdictions. According to ecologist and data scientist Rachel King, teams kept having identical conversations about when to trust AI suggestions, how much autonomy to grant AI agents, and what to do when chat sessions lost context.

"We kept repeating the same conversation, project after project," King said. "At some point it made more sense to work it out together, once, than have every team rediscover it independently."

Why it matters

Environmental science teams work with messy, multi-source datasets and widely varying technical expertise — conditions that existing AI guidance, written primarily for software engineers or abstract scientific contexts, doesn't address. Without field-specific standards, researchers risk embedding errors in work that informs policy decisions on wildfire, biodiversity, and climate change. The guidelines also surface equity concerns: male researchers report larger productivity gains than female counterparts, and while roughly two-thirds of people in high-income countries use GenAI tools, usage in many low-income countries sits near 5 percent, according to UN data cited by the researchers.

From crisis to curriculum

The Wildfire Resilience Index project began in 2023 as AI coding tools were emerging. By the project's completion, an entirely new generation of capabilities had arrived, rendering six-month-old lessons obsolete. That velocity wasn't unique — across NCEAS, junior researchers leaned on AI they didn't fully understand while veteran scientists questioned whether it could be trusted at all.

Rather than let each project reinvent protocols, NCEAS convened 22 researchers, developers, and data analysts to synthesize guidance through literature review and collaborative writing. Senior author Cat Fong noted that some participants entered the process openly skeptical of GenAI, which kept the final recommendations honest about tradeoffs rather than purely promotional.

The resulting 10 rules organize around three phases: preparation (including AI tool selection), active coding (promoting verification practices), and post-coding (documentation requirements).

Costs beyond code quality

The researchers document environmental and economic externalities. Data centers supporting GenAI are projected to consume 4 to 12 percent of U.S. electricity by 2030 and up to 32 billion gallons of water annually by 2028, though uncertainty remains high because much relevant data is privately held. The shift toward paid tiers for the most capable tools threatens to exclude researchers at underfunded institutions.

Meanwhile, GenAI's rise has coincided with rising unemployment among recent computer science graduates and steep declines in software development job postings, raising questions about disruption to a talent pipeline built with public investment.

The researchers deliberately avoid prescribing whether GenAI should be used for specific tasks, arguing that decision requires dedicated ethical scrutiny. Their focus is narrower: once a researcher decides to use it, doing so competently is a learnable skill.

The Wildfire Resilience Index, used throughout the paper as a real-world example, is available at wildfireindex.org. Details of the guidelines were first reported by UC Santa Barbara.