Environmental impact statements can take federal agencies months or years to draft. A new tool from Pacific Northwest National Laboratory and OpenAI is testing whether AI can speed that process up – without cutting corners on rigor.

The benchmark project, called DraftNEPABench, paired the U.S. Department of Energy's Pacific Northwest National Laboratory with corporate partner OpenAI. 

PNNL’s PermitAI research team worked with OpenAI over the past year to evaluate whether AI agents originally designed for coding tasks could help draft complex sections of environmental impact statements. The reports are often required under the National Environmental Policy Act when agencies consider major federal projects, such as siting new data centers or electricity infrastructure.

DraftNEPABench can evaluate systems that write specific subsections following detailed instructions, pull facts from multiple references, and ground their output in technical and authoritative sources that can be cited, with pointers to the original documentation for expert review. 

Unlike a simple AI chat system, a coding agent can break a task into steps, search and extract information from documents, and revise its draft along the way – much like a human analyst.

The benchmark includes 102 real-world test cases curated from published environmental impact statements spanning 19 government agencies and a range of action types, from energy development to restoration and waste-related projects. 

Nineteen subject matter experts with drafting experience created the challenge tasks, bringing expertise across disciplines such as biology, geology, engineering, anthropology and law.

“Our evaluation showed that AI coding agents can generate structured and domain-specific draft sections for environmental impact statements with promising results,” said PNNL data scientist and DraftNEPABench research lead Anurag Acharya in a release. “While the systems still require human oversight, the benchmark highlights both the potential and current limitations of these approaches.”

“The combination of data science expertise and private-sector AI prowess has produced a system that empowers the federal professionals responsible for scientifically credible and publicly accountable decision-making,” said Sameera Horawalavithana, a principal investigator of the PermitAI project.

PermitAI began as a pilot project sponsored by DOE’s Office of Policy to centralize NEPA decision data. Driven by a federal priority to accelerate and improve environmental reviews, the team has steadily expanded its focus. 

Rigorous benchmarks for drafting documents in regulatory settings have been largely absent until now, in part because most environmental permitting documentation remained siloed and inaccessible before the PermitAI team developed data standards and metadata coding. 

The team later released a machine-readable dataset, known as the NEPA Text Corpus (NEPATEC), making historical NEPA data and decisions far easier to search. DraftNEPABench adds to that suite of tools. 

A previous release, SearchNEPA, is an interactive AI-driven toolkit with a plain-language interface designed for federal NEPA reviewers, drawing on NEPATEC’s more than 140,000 NEPA documents and decisions spanning more than 50 years of data collection. More documents are coming online regularly as agencies comply with a federal mandate to adopt uniform standards.

Looking ahead, the PermitAI team is expanding its outreach beyond federal agency tools to include search functions that could help the public and corporations more rapidly prepare and submit documents required for federal permitting.

The research was presented as peer-reviewed work at the first annual Association of Computing Machinery Conference on AI and Agentic Systems, held May 27-29 in San Jose, California.