Drafting Regulations with Generative AI and the Rationale for Administration: Reflections on the U.S. Department of Transportation Case

0 Introduction

In January 2026, the U.S. investigative news organization ProPublica reported that the U.S. Department of Transportation (DOT) was advancing a plan to use Google's generative AI, "Gemini," to accelerate the drafting of regulations related to traffic safety (Note 1). According to the report, Gregory Zerzan, the DOT's General Counsel, stated in a meeting with staff that "we don't need a perfect rule. We don't even need a very good rule. Good enough is fine," signaling an intent to mass-produce draft regulations in a short period (Note 1). Similar content has also been reported by Ars Technica (Note 2).

This story is often understood from a technical perspective as "streamlining document creation" or "automating routine tasks." However, from the perspective of administrative law, it is somewhat problematic. Regulations are not merely documents; the process of drafting them is a mechanism for the administration to demonstrate to the outside world what the basis for its actions is, what it has considered, and why it reached its conclusions (Notes 3 and 4). Furthermore, generative AI is extremely adept at creating "plausible-sounding text." The greatest danger in administrative law is not the wrong conclusion itself, but rather that a wrong conclusion is protected by plausible-sounding reasoning and embedded into the system in a way that cannot be verified.

This article is a collection of reflections that uses the DOT case as material to examine where the law stumbles when generative AI enters administrative rulemaking, and what kind of design would allow the choice to "use" it to be institutionally viable.

1 The Location of the Problem: What Does Generative AI Speed Up?

What is symbolic in the ProPublica report is the vocabulary of speed-first. Ideas emerge such as creating draft regulations in 30 days, Gemini being able to spit out a draft in 20 minutes, and the notion that much of the preamble is just "word salad" and therefore suitable for AI (Notes 1 and 2). The point to pause at here is the distinction between whether the target of "speed" is the body of the regulation (the normative part that enters the Code of Federal Regulations) or the explanatory part, including the preamble.

The U.S. Administrative Procedure Act (APA) establishes a typical notice-and-comment rulemaking procedure, setting a framework for public notice, submission of comments, and publication of the final rule (Note 3). What is important is that the Act does not merely say to create the "body of the regulation," but requires that the adopted rule incorporate a "concise general statement of their basis and purpose" (Note 3). In other words, the preamble is not a decoration, but the core that supports whether the administration's judgment can withstand external scrutiny as a "reasonable administrative judgment."

On the other hand, in Japan, the amendment of the Administrative Procedure Act has also legalized the public comment procedure, adopting a structure where opinions are widely solicited during the process of establishing orders, etc., considered, and the results are publicly announced (Notes 5 and 6). Here, too, the point of the system is not just "creating a draft," but externalizing "what procedures and reasons were used to form the draft."

Therefore, if generative AI only speeds up the generation of text, the administrative law problem is relatively simple. The problem arises when the acceleration of text generation also strips away the institutional friction of the process of forming reasons, checking consideration factors, and accumulating records for accountability.

2 The Temptation of "Word Salad" and the Legal Function of the Preamble

The idea that "the preamble is word salad" is understandable from a practical standpoint. The preamble of a draft regulation has a certain form, such as a list of legal authorities, background circumstances, cost-benefit analysis, consideration of alternatives, and responses to comments. Since there is a form, it is an area where generative AI excels. In fact, the area where AI is most proficient is the reproduction of forms.

However, there is a double danger here.

First, there is the danger that by regarding the preamble as "routine anyway," the legal function it performs will be underestimated. The preamble is not just an explanation, but a structure for justifying that the administration reached its judgment. In judicial review under the APA, it is questioned whether administrative action is "arbitrary and capricious" (Note 4), and this question is directly linked to the quality of the reasoning in the preamble. If the preamble is thin, it gets attacked. If the preamble is thick, it provides peace of mind. However, generative AI generates "thick preambles" at a low cost. Here, a form of text inflation that the system did not anticipate can occur.

Second, the moment one says "AI is fine for word salad," there is a danger that the role of the preamble will be inverted from "external persuasion" to "internal procedural management." Rulemaking only becomes possible through documentation, allowing for coordination within the organization's legal, policy, and technical departments, and even with oversight agencies. Text is a common language for bearing the costs of internal coordination. However, if text can be created cheaply with generative AI, the common language for coordination appears first as a premise for coordination rather than as a result of it. Text becomes a tool for skipping discussion rather than the fruit of discussion.

This inversion is incompatible with the design philosophy of administrative law. The system does not slow down administration through friction. It makes administration accountable through friction.

3 Three Externalizations Created by Generative AI: Judgment, Responsibility, and Record

What the DOT case shows is the possibility that generative AI will induce not the outsourcing of document creation, but the externalization of administrative judgment. There are at least three types of externalization.

First is the externalization of judgment. Generative AI appears to be reaching a conclusion. However, in reality, it is merely a reconstruction of past text patterns. What is needed in administrative rulemaking is consistency with existing laws, technical feasibility, enforceability, interest adjustment, and normative judgment on which values to prioritize. The faster text generation becomes, the stronger the temptation to absorb judgment into the quality of the text. When the reasoning appears well-ordered, the verification of whether the conclusion is appropriate recedes into the background.

Second is the externalization of responsibility. In the ProPublica report, the image is described as humans doing the work of proofreading what AI has created (Note 1). One DOT employee described the future vision painted by the presenter as "our job will be to proofread the machine's output" (Note 1). What happens here is a fragmentation of responsibility. AI wrote it. Humans corrected it. But ultimately, who can say, "This regulation is based on these findings of fact and value judgments"? The act of proofreading, while appearing to be an act of taking responsibility, can actually obscure the location of responsibility. This is because proofreading can fix the surface of the text, but it does not reconstruct the basis of the judgment.

Third is the externalization of records. Administrative law emphasizes procedures and records to enable ex-post verification. In the U.S., the review of administrative action is fundamentally based on the "record," and it is questioned whether a reasonable explanation is included in the record (Note 4). In Japan, too, through the publication of the results of the public comment procedure, how opinions were handled is externalized (Note 6). When using generative AI, which prompt, which model, and how the output was adopted or modified is essentially the path of judgment. However, this path is not designed as an object that current administrative procedures naturally preserve. As a result, the worst composition is established where only the path of judgment becomes a black box, and only a splendid preamble is visible from the outside.

4 Hallucinations and "Wildly Irresponsible" Behavior

The problem of so-called hallucinations is an unavoidable risk factor for generative AI. This is a phenomenon where AI confidently generates information that is not based on facts as if it were true. Even in 2025, there have been successive cases where courts have been confused and lawyers have been sanctioned due to AI hallucinations (Note 2).

The fact that the subject of AI drafting in this instance involves safety regulations for aircraft, gas pipelines, and freight trains carrying hazardous chemicals further amplifies ethical concerns. As one official described it as "wildly irresponsible," deficiencies in regulations in these areas lead directly to disasters such as physical explosions, crashes, and derailments (Note 2).

Mike Horton, a former AI lead at the DOT, criticized the plan to use Gemini to write regulations, calling it "like having a high school intern do the rulemaking" (Note 1). Pointing out that traffic safety regulations are directly linked to human life, he stated, "They want to move fast and break things, but moving fast and breaking things means people get hurt" (Note 1).

According to reports, in a demonstration of drafting by Gemini, the generated document lacked the actual text of the provisions that should be included in the Code of Federal Regulations, resulting in a situation where staff had to fill in the gaps (Note 1). This indicates that large language models are not drafting laws by understanding their "meaning," but are merely outputting "law-like text" probabilistically.

5 Disregard for Expert Knowledge and Transformation into Monitors

Another point suggested by this case is the transformation of the role of administrative officials and the disregard for expert knowledge.

The DOT's vision assumes that AI will handle 80% to 90% of rule creation, with human officials retreating to the role of "monitoring AI-to-AI interactions" (Note 1). However, this may have to be called an optimistic design based on the premise that "humans can appropriately correct AI errors."

As insights from automation bias in cognitive science teach us, humans tend to neglect verification and overestimate the legitimacy of information presented by automated systems. Faced with a 90% complete, fluent draft created by AI, spotting the subtle but fatal logical contradictions or factual errors lurking in the remaining 10% is a task that requires higher concentration and expertise than drafting from scratch.

As officials pointed out in reports, DOT rulemaking is a complex task that requires "decades of subject matter expertise and familiarity with existing statutes, regulations, and case law" (Note 1). By delegating drafting to AI, there is a risk that the tacit knowledge and legal sense cultivated by officials as they work through the process of refining legal text will be lost.

6 The Friction of Administrative Procedure: "Reasoning" is Not a Byproduct of Slowness

What I want to emphasize here is that the slowness of administrative procedure is not mere inefficiency, but a structure for accountability. The U.S. APA requires the form of notice and comment (Note 3), and subsequently requires an explanation of the basis and purpose (Note 3). This structure institutionally forces the administration to put its judgments into writing and place them under the possibility of criticism from others. Japan's public comment procedure also visualizes the administrative formation process to a certain extent by soliciting opinions, considering them, and announcing the results (Note 6). In other words, procedures exist not to create documents, but so that documents can be subject to criticism.

If generative AI enters rulemaking, this relationship may collapse. Generative AI does not create documents that can be criticized; it creates documents that are difficult to criticize. Documents that are difficult to criticize are those with a large vocabulary, smooth conjunctions, and preemptive mentions of counterarguments. This is equally dangerous in academic papers, but when it happens in administrative preambles, the level of danger jumps. This is because the document functions as a substitute for accountability to courts, legislatures, oversight bodies, and society.

In short, what generative AI excels at is the "camouflage of accountability" that administrative law dislikes. Of course, it does not mean that the administration intentionally disguises it. The problem lies in the transformation of the system into a form where it is difficult to detect the disguise.

7 What Should Be Designed If Using Generative AI?

So, should the administration not use generative AI at all? It is easy to conclude so, but it is not realistic. Rather, the point is "in which processes" and "under what conditions" it should be used.

In this regard, there are policy documents in the U.S. that promote AI utilization; for example, the OMB asks each agency to accelerate AI utilization while maintaining the protection of civil liberties and privacy (Note 7). On the other hand, another OMB memo on AI procurement explicitly states that "regulatory activities governing AI use by non-governmental entities" are out of scope (Note 8). There is a gap here in that AI promotion as a policy and the introduction of AI into rulemaking are not necessarily organized in the same context.

To bridge this gap, there are three points that should be designed at a minimum.

First is the limitation of use. Processes where the output is unlikely to become the "basis of judgment itself," such as the classification and summarization of opinions or assistance in searching existing materials, are relatively more acceptable. On the other hand, the generation of core parts, including draft provisions of the regulation itself, legal authority, fact-finding, and cost-benefit evaluations, is in principle high-risk. In particular, the part of the preamble that corresponds to the "basis and purpose" (Note 3) is not just a template but the focus of legal review, so there is a great danger in letting AI write it to sound plausible.

Second is the design of records. When using generative AI, it is necessary to save prompts, reference materials, model versions, outputs, and the history of adoption/revision in a way that can be verified later. This is not a technical log issue, but an issue of preserving the "path of reasoning" under administrative law. The future image of a "proofreader" shown by the ProPublica report (Note 1) is a future where, without logs, the validity of the proofreading cannot be verified either.

Third, there is the articulation of accountability. When the administration uses generative AI, it should clearly state in the preamble or public notice of results which parts were used and which parts were judged by humans under their responsibility. Administrative procedures are, ultimately, a mechanism for presenting human judgment to society. If the use of generative AI is concealed, society will misidentify the 'subject of judgment.' This is a matter of trust, and if trust is lost, the costs will eventually bounce back in the form of litigation, political backlash, and institutional changes.

8 Reflections: Good writing is not a substitute for good rules

The fear in the story of generative AI writing administrative preambles is not that the AI will make mistakes. What is scary is that the AI will generate text that sounds correct while hiding the errors, and the institution will accept it as an explanation.

Administrative law places great importance on writing. However, this is not because the writing is beautiful. It is because writing fixes the basis of judgment and enables criticism and verification. When writing is fixed, the administration cannot move the goalposts after the fact. Therefore, writing is a constraint for the administration and a clue for society. When generative AI becomes able to create text freely, this constraint loosens. When the constraint loosens, the possibility of criticism decreases. When the possibility of criticism decreases, accountability becomes hollow. Once it reaches this point, the foundation of administrative law is shaken.

Technology is fast. Law is slow. However, this 'slowness' is not merely being outdated, but is also a safety device. The DOT case shows that the entry of generative AI into administrative writing is not just about efficiency, but requires a redesign of the structure of reasons and records upon which administrative law relies. The point of contention in the future should not be whether to prohibit or permit generative AI, but should emerge as a problem of institutional design: which frictions to leave and which frictions to reduce.

We cannot view this as a fire on the other side of the river. In Japan, too, digital administrative and fiscal reform and the efficiency of administration are being called for, and the verification of the use of generative AI is progressing. For Japanese administrative sites suffering from labor shortages, the slogan 'months of work in 20 minutes' has a sweet ring to it. However, we need to stop and think. Are laws and regulations merely a collection of text data? Are they not the crystallization of wisdom accumulated by members of society to agree upon and protect safety and order? When 'good enough' regulations mass-produced by AI threaten the safety of the skies or the safety of railways, it is not the AI or its developer that pays the price, but flesh-and-blood human beings.

(Reference)

(Magazine) 'AI and Law - Reflections'

* Please refer to the following for the table of contents

note General Terms of Service Article 3, Paragraph 2, First Sentence
3.2 The copyright of digital content created by the creator belongs to the creator.