U.S. DOT to Pilot Google Gemini for Rulewriting, Raising Safety and Legal Concerns

A federal push to use Google’s Gemini to produce regulatory drafts at the Department of Transportation marks a sharp operational shift that prioritizes speed over traditional deliberation. Agency leaders are presenting the system as a way to accelerate the production of rule text, with internal advocates describing the tool as capable of producing draft language in measured minutes rather than months. That promise of rapid turnaround is explicitly tied to a desire to transform how rules are written across agencies, starting with transportation, which regulates aircraft safety, hazardous materials movement and commercial driver standards. Skeptics inside and outside government caution that generative models frequently hallucinate, misapply context, or produce superficially plausible but legally fragile text—problems that are especially consequential when regulatory language touches public safety. The move coincides with substantial staff departures at DOT, including hundreds of lawyers, a dynamic that officials say factors into the decision to lean on automated drafting. Past agency uses of AI focused on translation, data sorting and comment analysis; this program represents one of the first attempts to shift the tool into the core work of creating binding regulation. Legal scholars and former agency technologists argue that hurried adoption without robust human review, transparent provenance of model outputs, and strict validation processes invites litigation and potential harm. Proponents counter that compressed timelines could improve responsiveness and allow agencies to iterate, but they must still reconcile speed with the procedural safeguards required by administrative law. The claim that a draft can be produced in roughly twenty minutes signals an operational ambition but also highlights the thin margin for scrutiny before publication. If adopted widely, this model would change staffing needs, oversight workflows, and the evidentiary record that courts rely on to evaluate agency intent and reasoned decision-making. Implementing controls—explicit human signoff, versioned audit trails, and domain-tuned model evaluation—would be necessary to reduce both technical and legal exposure. The initiative will test whether agencies can harness generative systems to augment capacity without sacrificing the quality and defensibility of rules that affect public safety. Its outcomes will influence statutory challenges, internal hiring strategies, and vendor relationships across the federal government. Observers should watch for pilot documentation, red-team findings, and the first instances where Gemini-generated text becomes part of an enforceable regulation.