⚡ Quick Summary
Published by the Institute for AI Policy and Strategy, this report conducts a systematic search for non-AI regulatory precedents that could inform the regulation of advanced AI in the United States. Its purpose is to identify promising agencies and individual regulations for later case-study research, rather than to endorse regulatory approaches or present definitive rankings. It applies Gerring’s approach to selecting “extreme” cases: agencies or regulations with unusually high values on variables selected as relevant to advanced AI regulation.
The study initially defines six variables: intensiveness, expertise, enforcement against powerful companies, risk-assessment, uncertainty-assessment and focus on tail-risks. It discards the quantitative tail-risk measure after finding it unreliable, and aggregates the other five at agency level. Regulatory intensity is measured through RegData counts of restrictive terms in the Code of Federal Regulations (CFR); expertise through the proportion of employees with doctorates or post-doctorates; and enforcement through fines involving current S&P 500 companies. Risk- and uncertainty-assessment are measured through CFR text mining, supplemented by five expert interviews. The report identifies the EPA, financial regulators, FAA/AST, DOE, DOD, NRC and FAA among the candidate precedents.
🧩 What’s Covered
The report proceeds from case-selection rationale to methods, rankings and agency profiles.
- Background and variable selection: Explains why non-AI regulation may provide lessons for advanced AI regulation and argues that past case studies have largely been chosen through researchers’ intuitions. It uses an “extreme” case-selection approach and focuses on US federal agencies because of their relevance to the regulation of frontier AI companies.
- Intensity, expertise and enforcement methods: Defines intensity as the number of restrictions in the CFR, using RegData counts of binding terms such as “shall” and “must.” Expertise is the proportion of agency employees with a doctorate or higher, while enforcement combines the value and number of major fines imposed on current S&P 500 companies since 2000.
- Text-mining methods: Details the creation, expert review and refinement of keyword lists for risk-assessment, uncertainty and tail-risks. The researchers search individual CFR Parts using Python, treating the CFR as a census rather than a sample.
- Reliability and aggregation: Reports interrater reliability measures for keyword lists and uses five anonymous interviews with regulatory specialists to triangulate results. Agency scores combine five variables after min-max normalisation; regulation scores combine intensity, risk-assessment and uncertainty-assessment.
- Variable-level results: Lists high-scoring agencies and regulations for each measure. EPA leads the intensity, enforcement and uncertainty results; the Federal Reserve, OCC and FDIC lead risk-assessment counts; and the SEC leads the enforcement composite.
- Tail-risk findings: Finds the tail-risk keyword measure unreliable and excludes it from aggregate results. Qualitative interview responses are retained but are described as inconclusive.
- Discussion and conclusion: Profiles selected agencies and regulations, including EPA emissions standards, financial capital-adequacy rules, FAA/AST launch safety, NRC licensing, DOE nuclear safety, FAA airworthiness, and DOD rules on classified information and radiation-dose reporting. It concludes that these are leads for further research rather than settled models for AI regulation.
💡 Why it matters?
The report gives AI governance researchers and policymakers a reproducible way to move from broad analogies to a structured shortlist of regulatory case studies. It distinguishes between agency design and individual regulations, and makes the operational choices behind its rankings visible. This can help users investigate mechanisms such as licensing, capital requirements, safety analysis, inspection, reporting, enforcement and risk management in domains outside AI.
Its cautions are also practically important: high scores are not evidence that a regulatory model should be transferred to AI. The findings are intended to guide exploratory case selection, not to establish which regulatory approaches are most effective.
❓ What’s Missing
The report does not provide full case studies of the agencies or regulations it identifies, nor does it determine whether their approaches would be effective for advanced AI. Its empirical scope is confined to US federal agencies and the CFR; it explicitly notes that an ideal study would also cover the UK and EU, but that practical constraints prevented this. Several measures are partial proxies: doctorate counts omit external expertise, enforcement data omits non-monetary interventions, and CFR keyword frequencies may not represent day-to-day agency practice. Missing data also affects some agencies. The uncertainty measure has weaker reliability than the risk-assessment measure, while the tail-risk measure is discarded as invalid. The authors also describe the percentile thresholds as arbitrary.
👥 Best For
Best suited to AI policy researchers, public-sector teams and regulatory-design analysts selecting non-AI precedents for deeper study. It is especially useful for users comparing licensing, safety assurance, risk assessment, disclosure, capital adequacy or enforcement arrangements across US environmental, financial, aviation, nuclear, transport and defence regulation.
📄 Source Details
AI-relevant regulatory precedents: A systematic search across all federal agencies was published by the Institute for AI Policy and Strategy on 3rd April 2024. The 77-page English-language report lists Bill Anderson-Samways, Oscar Delaney, Ashwin Acharya and Zoe Williams as authors. No URL to this report itself is printed in the document.
