Researchers at UC Santa Barbara's National Center for Ecological Analysis and Synthesis (NCEAS) have published a set of ten practical rules for using generative AI in environmental science research. The framework, released in PLOS Computational Biology, was developed by 22 researchers, developers, and data analysts who wanted to standardize how their field handles AI-assisted coding without sacrificing scientific rigor.
The guidance emerged from a recurring problem at NCEAS, where dozens of teams work on projects spanning wildfire, biodiversity, and climate change. Junior researchers were using AI tools they didn't fully understand, veteran scientists distrusted them entirely, and everyone in between was inventing their own standards for responsible use. Those conversations became a position paper that NCEAS now offers to the broader field.
Ten rules, three phases
The rules are organized around three stages of working with AI: preparation, implementation, and verification. The first phase covers choosing appropriate AI tools before a project begins. The second promotes best practices during coding. The third focuses on verifying AI-generated code and documenting the process so reviewers can evaluate it.
The framework grew out of a real project: the Wildfire Resilience Index, which measures how prepared communities are for wildfires using satellite imagery, land-cover data, and socioeconomic variables. The project ran from 2023 through release, a period when AI tools evolved quickly enough that lessons learned six months in no longer applied by launch.
" "It started as something we needed for ourselves," said ecologist and data scientist Rachel King. "We kept repeating the same conversation, project after project: whether to trust a suggestion, how much access to give an AI agent, what to do when a long chat session forgets a decision it made an hour earlier. At some point it made more sense to work it out together, once, than have every team rediscover it independently."
Uneven benefits across the field
The authors are direct about the equity problems tied to GenAI adoption. According to King, "GenAI's benefits are not landing evenly. Male researchers report larger productivity gains than their female counterparts, and a recent UN report found that while roughly two-thirds of people in some high-income countries use GenAI tools, usage in many low-income countries hovers near just 5%."
The group also flags the fee structures now appearing in the AI industry. As companies shift toward paid tiers, the most capable tools could become out of reach for researchers at underfunded institutions, turning what was marketed as an equalizer into a new barrier to entry.
The infrastructure has an environmental footprint, too. Data centers supporting AI are projected to use 4 percent to 12 percent of all U.S. electricity by 2030 and up to 32 billion gallons of water annually by 2028, though the authors note these figures carry uncertainty because much of the underlying data is privately held.
A skill worth building - if you choose to use it
The authors don't tell scientists whether to use GenAI. They argue that decision deserves its own ethical review. Their rules apply only after a researcher has opted in. They describe effective AI use as a skill that needs explicit training and practice, not hope.
For scientists and researchers, the NCEAS framework is a starting point for codifying a repeatable workflow. The rules force explicit decisions about which AI tools serve the project, what data they can access, and how to document the process for peer review. The full paper is available through PLOS Computational Biology for those wanting the exact wording contact. The Wildfire Resilience Index, used throughout the paper as a working example, is available at wildfireindex.org.
That means you don't need to invent standards from scratch for your next project. You can adapt tested guidelines to your context, flattered only by the science you're doing. For researchers deciding whether or how to use AI in their workflows, this framework passes a baseline: it was written by people who tried scenarios and failed, not in the abstract.
Your membership also unlocks:
