Yale Journal on Regulation: Navigating the Web of Agency Authority with AI (pacificlegal.org)

Yale’s Journal on Regulation highlights the Pacific Legal Foundation’s new Nondelegation Project, an AI-driven database that links every part of the Code of Federal Regulations (CFR) to its cited statutory authority and codifies each congressional delegation as specific, general, or none. The system uses large language models (Gemini, GPT-3.5/4, Claude, Grok) to extract authorities from the electronic CFR, classify the delegation type, and rate the statute–regulation relationship (directly mandated, authorized, related, or unrelated). The team ran comparative evaluations for accuracy and cost; Google’s Gemini 2.0-flash achieved a best balance (94% accuracy with low cost). The beta release (excludes some IRS tax regs) also computes restrictiveness by counting modal terms (“shall,” “must,” “prohibited”), and exposes the AI’s coding decisions so users can inspect results. This matters because it turns an otherwise intractable 190,000+ page regulatory corpus into a searchable, auditable tool that can support litigation, rulemaking review, and deregulatory initiatives. Key findings: the database captures 56,000+ delegations with 37% coded as general; 26 U.S.C. §7805 is the most-cited general delegation, §42 the most-cited specific; the EPA and FERC have the most general delegations, and the EPA is by far the most “restrictive” agency under the project’s metric. By combining transparent LLM evaluation with reproducible coding, the project demonstrates how AI can accelerate regulatory analysis, surface potential nondelegation vulnerabilities, and materially influence policy and judicial scrutiny of agency power.