Speed without surrendering accountability
Federal agencies are piloting large language models and document automation to compress NEPA timelines 2. The technology can accelerate drafting, evidence retrieval, and consistency checks. It also introduces new litigation risk if outputs are treated as agency findings without human review, provenance, or model documentation 1.
Governance should make three commitments explicit: humans remain accountable for decisions, every generated passage is traceable to source material, and models used in the record are described well enough for a court to understand their role 1.
Define permitted and prohibited uses
Separate low-risk uses (search, summarization of public comments, formatting) from high-risk uses (effects determinations, mitigation commitments, species presence conclusions). Prohibit the latter from autonomous model output. Publish the policy where project teams and contractors can see it before work starts.
Require human attestation on record-bearing text
Any paragraph that enters an EA, EIS, or decision memo should have a named reviewer who attests they verified sources and applied professional judgment 1. Track edits between model draft and final text. Courts will ask whether the agency relied on the model or on staff expertise.
Log prompts, model version, and retrieval context
When retrieval-augmented generation is used, store which documents were retrieved, which model version produced the draft, and which user approved changes. Without this metadata, agencies cannot reconstruct why language appeared in the record 2.
Validate against authoritative sources
Ground outputs in agency GIS layers, species lists, regulatory citations, and prior decisions. Flag conflicts when model text diverges from source data. Automated checks for broken citations and outdated regulatory references reduce record errors.
Manage vendor and contractor access
Consultants often bring their own tools. Contract language should require disclosure of AI use, prohibit training on confidential agency data without approval, and align deliverables with the agency's records schedule. Treat vendor models as part of the supply chain risk surface.