Across the federal government, agencies are beginning to integrate artificial intelligence (A.I.) tools into their operations. One area where A.I. may have particular benefits is in rulemaking, where lengthy processes and review procedures can slow down regulatory policymaking. Using A.I. tools may enable agencies to make the regulatory process significantly more efficient, enabling the government to govern more effectively and responsively.
The Administrative Procedure Act (APA), which governs federal rulemaking processes, does not bar agencies from using A.I. tools to draft components of rules, review public comments, or perform underlying analyses. But the use of A.I. in rulemaking may sacrifice caution for the sake of increased efficiency, raising the risk that agencies finalize rules that are fundamentally flawed. To avoid this risk, agencies should thoroughly review a model’s training data, parameters, and methodology; ensure that data submitted to the model is clean and structured and that prompts provided to the model are clear and accurate; and carefully check the model’s outputs for signs of hallucination or bias. Otherwise, final rules may exhibit logical or factual flaws that could undermine their efficacy and make them vulnerable to legal challenge. In sum, the promise of A.I.-powered rulemaking will only be realized if agencies treat A.I. as a complement to rigorous reasoning, not a substitute for it.
This report discusses how the structure of A.I. models shapes their outputs and explores the opportunities and challenges of A.I. use for rulemaking.
