The United States has a Nuclear Regulatory Commission. It has an Environmental Protection Agency. It has a Consumer Financial Protection Bureau (the most recent product of a sustained legislative and executive branch effort to design a new agency).
The United States needs an AI Regulatory Commission (AIRC), and it needs it now. It needs an AIRC both because of current risks and because of growing public concerns.
The AIRC could be accompanied by other measures from Congress; it need not and should not exhaust legislative responses. But it would be an important part of those responses.
The hard questions, of course, involve design details. The AIRC could be relatively weak; it could be extraordinarily powerful. Consider a few notes, which should be taken as preliminary, incomplete, and tentative.
The Nuclear Regulatory Commission was created as an independent agency; the EPA was created as an executive agency.
The NRC is a multimember agency with a chair. The EPA is a single-headed agency.
There is a good argument, in principle, that the AIRC should be a multimember agency with a chair. A multimember agency (five is a good number) is more deliberative, of course, and it incorporates more perspectives at the highest levels.
The Supreme Court has recently thrown the whole idea of independent agencies into the constitutional garbage heap, with an exception for the Federal Reserve Board (alongside some other possible exceptions). Putting the question of constitutional authority to one side: There is a good argument, in principle, that the AIRC should be an independent agency, to reduce political pressures in this domain. (There are counterarguments.)
It is unlikely (but not impossible) that Congress has the constitutional authority to create an independent AIRC. If the AIRC is not independent, the President might agree, in public, that he will respect a “for cause” protection of its members from removal. This agreement would not be enforceable, but still, it could be important.
The AIRC might have duties of consultation in certain cases, as, for example, with the Secretary of Defense, the Secretary of State, and the Director of OMB.
The AIRC might be required to produce a report on risks and dangers within 90 days of its creation. There mjght be a classified and a nonclassified version.
The AIRC should be given investigatory and factfinding authority.
The AIRC might be funded by AI companies, a requirement of law. (This is not clear.)
The AIRC might have licensing authority. (This is not clear.)
The AIRC should be given the authority to require periodic disclosures and reports from AI companies, with appropriate safeguards and qualifications.
The AIRC should be given the authority to impose civil penalties, subject to judicial review.
The AIRC should be given rulemaking authority, subject to ordinary standards in the Administrative Procedure Act, including judicial review.
The AIRC’s rulemaking authority might generally require consideration of costs and benefits, specified and spelled out in a sentence or two.
The AIRC should be given the authority to issue emergency rules and orders against imminent or catastrophic threats. Such authority could exist under (13), with the APA’s good cause exception, but it would be best to have explicit authority to this effect, with judicial review (probably).
The AIRC should be authorized to coordinate with and gather information from parallel or similar entities in other nations, in consultation with (for example) the Secretary of State, the Secretary of State, USTR, and the Director of OMB.
These notes are preliminary, incomplete, and tentative. But the following seems clear: The United States needs an AI Regulatory Commission (AIRC), and it needs it now.


I basically agree. Certainly with the central thesis, if not with all the precise details. But it's point # 16 that grabs me the most. Given our interconnected world, I wonder--indeed, doubt--whether a national AI regulator will suffice. I keep reading that China's AI companies are 6-12 months behind the US. Other. nations are working on AI, as well. Without agreed upon and enforced global guardrails, a domestic regulator seems more likely to simply retard US development of AI while leaving the world vulnerable to whatever comes out of foreign labs.
Not one of Cass's better ideas. Unless there is some high-profile focussing event that changes the politics, there is no realistic prospect that a new federal regulatory commission can be created by Congress and up and running in less than several years. Given the speed of AI development, the cows are likely to be long gone before the barn door is shut. Staffing such an agency with people who know what they are doing would be very difficult, given the difference between government and industry pay scales. For better or worse, the response of the federal government to AI will have to come from existing agencies like OMB, Commerce, CAISI, and the FTC. Of course, if you want to freeze AI development in the U.S., NRC-style regulation is the way to go. The NRC either saved us from another Chernobyl or blocked a fix for Global Warming. Opinions differ.