1. Headline & intro
The idea that a chatbot might help decide how often your plane is inspected or how gas pipelines are monitored should make you uneasy. Yet that is effectively where the US Department of Transportation (DOT) is heading by using Google’s Gemini to draft safety regulations. According to recent reporting by ProPublica and Ars Technica, the priority is speed over precision – “good enough” instead of rigorous.
This isn’t just a US bureaucratic curiosity. It’s an early test of a far bigger question: will governments outsource the craft of rule‑making to commercial AI models, and what happens when they do?
2. The news in brief
According to an investigation by ProPublica, covered in detail by Ars Technica, the US Department of Transportation has begun using Google’s Gemini AI system to draft federal safety regulations for planes, cars, pipelines and other transport infrastructure.
Internal briefings described by these outlets suggest DOT leadership wants AI to accelerate rule‑writing so that drafts that previously took weeks or months can be produced in as little as 30 days, with Gemini reportedly generating initial texts in under 30 minutes. The department’s top lawyer was quoted in meeting notes as saying that rules do not need to be perfect, only “good enough” to move the process along.
Some staffers, speaking anonymously, warned that relying on a general‑purpose chatbot for such intricate legal and technical work risks serious errors, given widely documented AI hallucinations in courts and other settings. ProPublica reports that Gemini has already been used on at least one unpublished Federal Aviation Administration rule. Google has publicly promoted its partnership with DOT, positioning Gemini as a tool for modernising government.
3. Why this matters
The stakes here are brutally simple: a sloppy e‑commerce rule produces annoying cookie banners; a sloppy aviation rule can kill people. Regulatory text in transportation is the last layer of defence between complex systems and human bodies. Treating that text as low‑stakes “word filling” is not a technical issue – it is a governance failure.
Who benefits? In the short term, DOT leadership and the Trump administration gain a political win: a bureaucracy suddenly looks fast and “innovation‑friendly.” Google gains something even more valuable than contract revenue: a flagship case proving that its model can sit at the heart of state decision‑making. If DOT can claim that 80–90 percent of regulatory drafting can be handed to Gemini, every other agency becomes a potential customer.
Who loses? First, the public, which has no way today to know which clauses in life‑or‑death rules were hallucinated by an opaque model trained on unknown data. Second, the civil servants whose expertise turns messy engineering and legal realities into enforceable, coherent rules. If their role is downgraded to “AI supervisor,” institutional knowledge will erode, just when we need it most.
There is also a subtler cost: legitimacy. Even if every AI‑assisted rule were technically correct, citizens discovering that a chatbot authored the framework governing aircraft maintenance or tanker operation will trust those rules less. Once trust collapses, so does voluntary compliance – and regulators end up weaker, not stronger.
4. The bigger picture
DOT’s experiment lands after a year in which generative AI has already embarrassed the legal system. Lawyers in multiple jurisdictions have been sanctioned for filing briefs filled with invented case law copied from chatbots. Judges have admitted they can be fooled by fabricated citations. These incidents were warnings: large language models are extraordinarily good at sounding authoritative while being wrong.
Despite this, governments worldwide are racing to use generative AI for drafting and summarising internal documents. Most pilots so far have focused on low‑risk tasks: email replies, meeting minutes, first drafts of memos. DOT’s move is qualitatively different. It nudges AI from the back office into the engine room of state power: the creation of binding, enforceable rules.
There is historical precedent for this kind of technological delegation going wrong. Before the 2008 financial crisis, regulators and banks leaned heavily on risk models that were treated as oracles. The models were not inherently evil; the problem was institutional over‑confidence in tools that were poorly understood, used outside their design envelope and never stress‑tested for systemic consequences.
Compare that with how software is treated in other safety‑critical fields. Code in aircraft control systems or nuclear plants is subject to exacting verification, redundancy and certification standards. The idea that a generic text generator, with no formal guarantees, could write the rules for those same systems without similar rigour should concern anyone who remembers past disasters born from “move fast” thinking.
5. The European / regional angle
It is tempting for European readers to dismiss this as another US governance experiment, but that would be a mistake. American transport regulations shape global norms. Many non‑US airlines, logistics firms and manufacturers follow Federal Aviation Administration and other US standards because they are de facto world benchmarks. If those standards start to reflect the quirks of a proprietary AI model, the impact will be exported.
Europe is also not immune to the political narrative being tested here: that “modern regulators” must embrace AI aggressively to stay competitive. The difference is that the EU already has a dense web of rules – from GDPR and the Digital Services Act to the upcoming AI Act – that explicitly emphasise accountability, transparency and human oversight. A European transport ministry trying DOT’s approach would immediately collide with questions about explainability, data provenance, and the right to contest automated decision‑making.
For EU institutions and national governments, DOT’s move is therefore a useful cautionary tale – and an opportunity. Europe can choose a different model: AI as an assistive tool tightly bounded by process, rather than an invisible co‑author of law. That difference could become a selling point for European tech firms and public administrations aiming to be trusted, not just fast.
6. Looking ahead
Several things are now in play. Domestically in the US, expect legal challenges. If a safety rule drafted with heavy AI assistance leads to a major incident, plaintiffs will argue that the agency acted arbitrarily by relying on a non‑transparent system known to fabricate facts. Courts may start asking agencies to document exactly how AI was used in rulemaking – something few are set up to do today.
Politically, this experiment could go either way. A run of apparently successful, controversy‑free AI‑drafted rules would embolden other agencies and other countries to follow. A single high‑profile failure could produce a backlash that pushes generative AI out of sensitive policy domains for years. Both scenarios are plausible.
On the vendor side, Google and its rivals will double down on marketing specialised “government‑grade” models with audit trails, restricted training data and domain‑specific fine‑tuning. The risk is that governments fixate on technical tweaks while ignoring the core issue: who is accountable for what the machine writes.
For citizens and civil society in Europe and beyond, the key questions to watch are simple:
- Will regulators disclose when and how AI systems helped write rules?
- Will there be independent review of AI‑generated regulatory text?
- Will public‑interest groups get access to the prompts, not just the polished final PDFs?
Without clear answers, we are drifting towards a world where no one can quite say who authored the rules that govern critical infrastructure.
7. The bottom line
Handing the pen for aviation and pipeline safety rules to a general‑purpose chatbot, in pursuit of speed, is an experiment with asymmetric risks. The upside is bureaucratic convenience; the downside is loss of trust and, in the worst case, avoidable disasters. If governments are serious about both innovation and safety, they must treat AI as a tool that works for accountable humans, not as an uncredited co‑legislator. The open question is whether voters – in the US and in Europe – will demand that distinction before the first AI‑tainted rule fails in the real world.
