“Another advantage accruing from this ingredient in the constitution of the senate, is the additional impediment it must prove against improper acts of legislation. No law or resolution can now be passed without the concurrence first of a majority of the people, and then of a majority of the states. It must be acknowledged that this complicated check on legislation may in some instances be injurious as well as beneficial.”

In Congress, slowness is often a feature. Madison, along with the founding generation, understood that delay can force deliberation and expose poor ideas. Frontier AI governance strains that constitutional virtue because the pace of technological change is so rapid that Congress cannot keep pace. The resulting inaction preserves an increasingly indefensible status quo that risks popular unrest and distrust in our political institutions.

To keep pace with frontier AI, Congress needs to reform how it develops policy. One such reform is the creation of a Citizen Assembly on AI Governance. A citizens’ assembly (CA) brings together a broadly representative group of randomly selected citizens to learn about an issue, deliberate, and develop policy recommendations. Under this proposal, the assembly’s recommendations would take effect after presidential approval unless Congress enacts legislation rejecting them.

Congress has used a process known as “recusal legislating” to address similar challenges. Under recusal legislating, Congress assigns another entity responsibility for developing policy in a specific domain. Its proposals take effect unless Congress enacts legislation to reject them. The shift from a legislative default of rejection to acceptance is undoubtedly a large one. However, it has always been the case that the legislative process ought to reflect the circumstances at hand. This particular approach to legislating should not apply across the board but only where a bias toward action is necessary to solve specific issues.

Recusal legislating and similar mechanisms have been used to address policy questions that require collective action, demand a lot of technical capacity, necessitate frequent adjustment, or some combination thereof. A first principles analysis documents the need for these adjustments to Congress’s traditional posture. In certain domains, inaction is itself a major policy — opening the door to other actors to fill the void left by Congress. In some instances, this marks the most reliable way to discover enduring and efficacious governance solutions. In others, however, private and public actors intervene in ways that frustrate the constitutional order, infringe on core liberties, or result in difficult to reverse or even irreversible outcomes. In those instances, Congress cannot assume a wait-and-see posture.

This essay contends that the conditions for this kind of reform have been met when it comes to frontier AI governance and outlines one possible path forward. I’ll explain why frontier AI warrants recusal legislating and why a citizen assembly in particular is the right shape for this. I’ll then sketch out what such a citizen assembly could look like, drawing lessons from earlier congressional experiments, and address limitations and open questions. Despite potential limitations and flaws, the need for a new approach to governance of frontier AI makes the current legislative paradigm, which defaults to inaction, unacceptable.

A crowd pressed into a doorway and along a wall looks up toward a speaker out of frame, a loudspeaker mounted on the painted wall above them
Garry Winogrand, Political Meeting, New York (c. 1960). Source.

Why we need recusal legislating for AI governance

Recusal legislation is necessary for frontier AI governance for at least three reasons.

First, governing laws must undergo frequent revision to keep pace with the development of AI. The costs of failing to make timely adjustments were made clear in the wake of the Hugging Face-OpenAI incident. The Computer Fraud and Abuse Act (CFAA), the main federal anti-hacking law, doesn’t clearly say who is responsible or what remedies apply when agents launch an attack that their developers did not intend — its drafters understandably did not foresee a world in which hacks are carried out by autonomous agent swarms.

One solution would be to have Congress itself regularly reevaluate whether the CFAA and related laws adequately address concerns posed by new technologies. However, continual technical review would place substantial demands on scarce congressional resources. Determining when and how a legal regime will be rendered inadequate by AI would be too much to assign to a legislative body tasked with being responsive to a slew of other demands.

Second, frontier AI governance requires politically difficult collective action. Despite many lawmakers across the political spectrum proposing frameworks to govern frontier AI, often with support from the labs themselves, Congress hasn’t supported any of them. One reason is that Congress empowers gatekeepers, such as committee chairs and members of leadership, who may block such bills so they can maintain certain donor relationships and appease particularly strong interest groups. Lawmakers who have tight ties to industry stakeholders opposed to or unsure of whether to support AI regulation have serially been accused of stymieing progress on certain AI bills, perhaps despite an underlying recognition of what would better serve the interests of Americans in general.

Third, actors without the proper legal authority, institutional capacity, or both are intervening in AI governance issues, citing a lack of congressional action as a basis for their regulatory postures. By way of example, earlier this fall, the New York City Council presided over a days-long inquiry on AI policy issues that exceed the jurisdiction and bounds of any municipality. New York City Mayor Zohran Mamdani acknowledged, “This is not a New York City-specific question. Therefore I think the place where we need the most urgent response to this is at the national level.” Governors and state legislators have echoed the need for a federal response and explained state-level efforts as a second-best, yet necessary alternative. These regulatory escape valves at once release some public pressure for AI regulation, but in the process they place greater pressure on our constitutional order. The longer that Congress operates under its present approach of defaulting to rejection in the context of rapid AI advances, the greater the odds are that other actors step in and push the limits of their own (limited) powers.

People crowd around a table in a dim meeting room; a young man in the foreground glances toward the camera while a man in a hat talks on the phone behind him
Garry Winogrand, Political Meeting, New York (c. 1961). Source.

Why a citizens’ assembly?

Recusal legislating could take many forms. BRAC, discussed below, relies on an expert commission. The Rules Enabling Act relies on judges. So why do I advocate for a citizens’ assembly for questions on AI governance?

In short, because of values. AI governance raises many questions — some technical, some economic, some cultural — and the resulting decisions will elicit passionate and potentially long-lasting political consequences. Who has legitimacy to decide these questions? Experts can help assess what a model can do or what a policy may cost, but those assessments will not resolve how competing interests should be weighed. A citizen assembly would bring a broader range of experiences to those judgments and avoid political gridlock. Randomly selected citizens are separated from the electoral incentives that discourage lawmakers from supporting politically difficult measures, and any informational needs that participants might have could be addressed through sustained preparation and expert guidance.

That said, creating a CA that can effectively address frontier AI policy questions requires some careful thinking. What follows is my preferred proposal for how to go about it.

Designing a Citizen Assembly on AI Governance

Congress should create a citizens’ assembly on AI governance (CAAG) made up of 500 active delegates and, as detailed further below, 500 delegates-in-waiting. Each delegate should be selected through stratified sortition to reflect the demographic composition of the country. The 500 active delegates would be divided into five sub-councils of 100 that tackle discrete policy areas, as explored further below. The CAAG would be allotted ten ordinary policy proposals to Congress each year; each proposal would first go to the President, who would then send approved proposals to the Hill. For each of those first ten approved proposals, at least 55 percent of all delegates would need to back the proposal. If, after having exhausted its ten proposals, the CAAG determined that additional action was warranted, it would have to satisfy an even higher threshold — say 65 percent. For all referred proposals, Congress would have a 30-day window to enact a joint resolution of disapproval. Absent such action, the proposal would become law.

What would this look like in practice? Suppose an incident like the OpenAI-Hugging Face hack occurs. The CAAG’s national security sub-council takes it up immediately. In the first week, it asks the Department of Homeland Security and AI researchers for input. In the second, it reviews the responses and consults experts. In the third, it drafts a proposal and sends it to the full CAAG, which deliberates and votes in the fourth and final week. If the vote meets the threshold, the President has another week to approve or reject it. If the President approves, Congress has 30 days to pass a resolution rejecting the measure. If it does not, the measure takes effect.

The path of a proposal

No resolution of disapproval passes in 30 days. The measure takes effect about two months after the incident.

  1. day 0

    An incident occurs

    day 0

    AI agents launch a hack like the OpenAI–Hugging Face attack, and no law clearly says who is responsible.

  2. days 1–21

    Sub-council

    days 1–21

    The national security sub-council (100 delegates) takes it up and drafts a proposal.

  3. days 22–28

    Full CAAG

    days 22–28

    All 500 delegates deliberate and vote.

  4. days 29–35

    President

    days 29–35

    Has a week to approve or reject.

  5. days 36–65

    Congress

    days 36–65

    Has 30 days to pass a resolution of disapproval.

  6. day 65

    Takes effect

    day 65

    Becomes law under the authorizing statute, about two months after the incident.

Figure 1. One potential mechanism for CAAG intervention. Use the controls in each box to see where a proposal ends. Hover over a stage for details.Figure 1. One potential mechanism for CAAG intervention. Use the controls in each box to see where a proposal ends.

The path of a proposal

No resolution of disapproval passes in 30 days. The measure takes effect about two months after the incident.

  1. day 0

    An incident occurs

    day 0

    AI agents launch a hack like the OpenAI–Hugging Face attack, and no law clearly says who is responsible.

  2. days 1–21

    Sub-council

    days 1–21

    The national security sub-council (100 delegates) takes it up and drafts a proposal.

  3. days 22–28

    Full CAAG

    days 22–28

    All 500 delegates deliberate and vote.

  4. days 29–35

    President

    days 29–35

    Has a week to approve or reject.

  5. days 36–65

    Congress

    days 36–65

    Has 30 days to pass a resolution of disapproval.

  6. day 65

    Takes effect

    day 65

    Becomes law under the authorizing statute, about two months after the incident.

Figure 1. One potential mechanism for CAAG intervention. Use the controls in each box to see where a proposal ends. Hover over a stage for details.

This way, a response could be in force about two months after the incident, assuming the authorizing statute is in place and the assembly is already prepared. Several questions remain open from this high-level overview. How should the delegee policymaking body be composed? What’s the scope of the body’s authority? What resources are afforded to the body? How do the results of a deliberative body enter the traditional legislative process?

To answer these questions, it’s worth analysing past attempts at recusal legislating.

Lessons from recusal legislating

Congress has occasionally recognized its own inability to act when required by “justice or the general good,” to again borrow from Madison. In rare cases, “unorthodox” processes like recusal legislating or automatic lawmaking have been used to make the public business go forward.

Two illustrative examples are the Base Realignment and Closure Commission (BRAC) and the Rules Enabling Act (REA).

The BRAC was developed in response to a classic collective action problem: on the one hand, Congress knew that the United States no longer needed to maintain such a large number of military bases across the country; on the other, no lawmaker wanted to be blamed for closing a base in their community, which would presumably result in significant negative economic, cultural, and electoral consequences. This caused the nation’s excessive military infrastructure to drain federal coffers unnecessarily.

Due to a broader interest in reducing military spending, Congress established BRAC in 1988 to address this problem. The BRAC process allows the Department of Defense (DoD) to right-size its military footprint while reducing the pressure on lawmakers to prioritize bases in their respective districts over the national interest. In short, the DoD recommends base closures to the Commission, which reviews those recommendations and produces a list for the President. The President accepts or rejects that list, with no option for tinkering with its components, and, if she accepts, sends that list to Congress. Congress then has a 45-day window to enact a joint resolution rejecting the package; absent such legislation, the recommended closures proceed. In this way, Congress made collective action far more feasible, though it must still muster the initial support to launch each round of review — a step it has taken only five times.

The Rules Enabling Act of 1934 (REA) addresses a different problem arising from limits on the legislative capacity of Congress: the need for continuing, expert review of the rules governing federal court procedure. Congress assigned the Supreme Court the power “to prescribe general rules of practice and procedure.” This is a significant delegation because the resulting rules shape civil, bankruptcy, criminal, and appellate procedure, affecting every federal court litigant. Few would expect members of Congress to have the time and expertise to monitor the adequacy of these rules across all these domains.

The REA process instead assigns that work to bodies within the U.S. court system. Advisory committees study proposed amendments and seek public comment. Their recommendations are reviewed by a standing committee and the Judicial Conference, which is the principal policymaking body of the federal courts. The Supreme Court then determines which amendments to advance and sends them to Congress. The rules take effect unless Congress enacts legislation rejecting, modifying, or deferring them. Congress thus retains the power to intervene while leaving the detailed work to those with the time and expertise to undertake it.

MechanismWho develops proposalsDefault and congressional reviewPurpose
Base Realignment & Closure CommissionIndependent commission reviews DoD recommendations.President accepts or rejects the package. Approved closures proceed unless Congress enacts disapproval within the statutory 45-day period.Overcome the political difficulty of closing individual bases.
Rules Enabling ActJudicial committees develop amendments; the Supreme Court prescribes rules.Supreme Court sends proposed rules to Congress by May 1. They take effect automatically on or after December 1 of that year, unless Congress enacts legislation to reject, modify, or delay them.Keep federal court procedures responsive to changing needs.
Proposed CAAGCitizens develop recommendations with expert guidance.Presidentially approved measures take effect under the authorizing statute unless Congress enacts disapproval within 30 calendar days.Develop frontier AI policy and address resulting legal gaps and collective action problems.
Table 1. Congressional review under three approaches.

Three high-level principles for the CAAG can be gleaned from the two overviews and related information on how BRAC and the REA have functioned in practice.

How should the assembly be composed?

One of the most important decisions behind the CAAG is its composition. Frontier AI governance may prove much more politically salient than base closures and judicial rulemaking, since it directly affects more people and may end up shaping their lives for longer. This is an argument for including a broader range of voices in policy development. A citizen assembly selected through stratified sortition would combine random selection with demographic criteria in order to represent the general public. Popular support for CAs is generally tied to the means of selection: people tend to support random selection over self-selection, which raises concerns about only those with the strongest opinions being willing to participate.

In terms of size, the proposed assembly reflects research indicating that the general public trusts larger CAs, all else equal.3 Five hundred individuals may seem like quite a large body relative to a Congress made up of 535 members. However, it’s important to flag that the mini-assemblies — the sub-councils — of one hundred delegates would be tasked with most substantive work. There’s no magic number to what makes a CA function as intended. Related experiments with deliberative democracy illustrate a range of possible sizes. Iceland’s constitutional reform process included a national forum of 950 randomly selected Icelanders. The French drew the names of 150 individuals to participate in its climate change assembly. A CAAG of 500 active delegates would not only bring in more perspectives but also increase the odds of everyday Americans getting to know someone who has been a part of this novel experiment, especially given that delegates would serve finite terms.

The public also needs assurances that delegates understand the underlying issues, beyond what they’ve been told by CAAG briefs and experts. As briefly noted above, each delegate would receive a two-year appointment, spending the first year as a non-voting shadow and the second as a voting member. The shadow year would allow incoming delegates to attend meetings, study prior proposals, and get up to speed on the relevant technical and legal issues.

Where does the assembly’s authority begin and end?

The CAAG should have a precise mandate to identify and correct legal gaps created or materially altered by frontier AI, as well as to develop substantive policies narrowly tailored to frontier AI. Within that mandate, it should be able to propose new policies. Its sub-councils could be assigned to cover explicit policy domains, such as Model Development and Evaluation; Economic Security; Rights, Democracy, and Public-Sector Use; Consumer Protection; and National Security and Cybersecurity. These categories are broad by design to enable timely work on the most important issues of the day, while still enabling sub-councils to develop expertise in their respective portfolios.

Sub-councils would receive regular briefings from agencies and departments with relevant responsibilities, supplemented by input from scholars and researchers. Members of the public would have an opportunity to contribute during open hearings. Expert selection should follow public criteria and include conflict-of-interest screening to help ensure that delegates encounter a range of informed perspectives.

The U.S. processes for closing military bases and revising federal court rules each had a defined remit. So did the public deliberative processes on climate policy in France and constitutional reform in Iceland. The CAAG should likewise have a clearly bounded mandate. Research on the conditions under which people will support CAs aligns with limiting the policy scope of the body. While some researchers have proposed citizens’ assemblies and other forms of deliberative democracy as a means to replace existing legislative institutions, the public is far more skeptical of such an expansive scope of authority being assigned to CAs. A narrower role for CAs that informs and complements existing governing bodies is much more popular. Saskia Goldberg, Marina Lindell, and André Bächtiger contend that the public demands such constraints as a sort of “damage control calculation.”

A key difficulty will be to determine which issues are properly assigned to the CAAG and, of those issues, which can be grouped together into specific sub-councils. The topics identified here may be among the most salient today but may be less of a concern in years to come as AI and society evolve. Congress can borrow from BRAC to address this. The BRAC process was deliberately episodic and time-limited rather than entrusted to a permanent commission.4 That said, this also invites a severe risk of Congress abandoning the CAAG, just as it has come to rely less and less on the BRAC. Still, given the novelty of this proposal, it may be politically necessary to launch the CAAG for a trial period, albeit for a longer duration than afforded to the BRAC cycle. Congress could authorize the assembly for a five-year period, and then decide whether to extend it after reconsidering its mandate, membership, and procedures. That requirement risks recreating the inaction that has stalled further BRAC rounds. Given the breadth of the CAAG’s authority, however, periodic congressional reconsideration is a worthwhile safeguard.

How could the assembly’s decisions be turned into law?

Whether the CAAG will serve its intended purpose depends on how its work is transmitted to Congress. If the body merely teed up recommendations to a gridlocked Congress, it’d likely be of little value; the best case may be that other legislative bodies adhere to its proposals. The key, then, is to couple the institutions in a way that earns public legitimacy and produces high-quality policy. If the CAAG is perceived as an entity that intervenes too regularly and with inadequate respect for the authority of Congress, it may have a short-lived tenure or fall into disuse, like the BRAC. The proposed cap on ordinary referrals, with a higher threshold for additional measures, would require the CAAG to prioritize its work while preserving the ability to respond when further action is warranted.

Also, Congress and the President shouldn’t be able to easily dodge CAAG proposals. The proposed process facilitates a means for the public to clearly assess which actor may be responsible for the success or failure of CAAG proposals. At the first stage, they can identify whether the President adopts the CAAG’s recommendation. Assuming she does, the public can then turn their focus to Congress. Congressional rejection of a CAAG proposal would be easy for the press to cover and the public to understand. To prevent this period of congressional consideration from being manipulated by outside interests, it must be the case that Congress be prevented from amending the CAAG’s proposal or otherwise delaying its consideration through procedural hurdles.

A reporter holds out a microphone to a woman in a Women’s Liberation T-shirt amid a dense crowd, a TV camera lens pushing in from the left and a hand-lettered sign held overhead
Garry Winogrand, New York City (1971). Source.

Limitations and open questions

These alternative policy pathways invite questions about the representativeness, transparency, and constitutionality of forgoing traditional lawmaking procedures. BRAC and the REA demonstrate that Congress can authorize particular decisions to take effect without affirmative legislation; they do not establish that Congress could confer general legislative authority on a randomly selected assembly. As noted, the authorizing statute would need to adequately define the CAAG’s authority and specify how citizen recommendations could lawfully acquire legal effect.

The elements of this specific proposal can and should be contested. Perhaps one of the most critical questions is the demographic base used to select CAAG delegates. Some may favor giving younger Americans greater representation or weighting votes by age because they will live with the consequences for longer. Such changes, however, would depart from the goal of reflecting the general public and could make an already novel proposal less likely to earn popular backing. Others may assert that the CAAG should foreclose participation from individuals with extensive ties to the AI industry. The thinking may be that such individuals are excessively biased or may exercise outsized influence on others given their personal knowledge of the field. Of course, additional expertise may serve to benefit the caliber of decision-making. External groups eager to support the formation of a CAAG could assist with this key function by conducting early polling and experimentation with different demographic requirements.

Even if the CAAG adopted the selection mechanism most likely to earn popular support and assist in its weighty tasks, there are practical limits to what sortition and expert guidance can accomplish, especially when pressed to make policy recommendations in response to complex and rapidly evolving situations. The agencies, staff, and outside experts called for in this proposal could exert disproportionate influence over the agenda or the information presented to the CAAG, especially given that many delegates may have little to no prior experience with AI, public policy, or both! Public hearings, the year-long shadow requirement, and scrutiny of expert selection would help address those pressures, but the assembly will ultimately have to earn public trust through the quality of its work. Due to the proposed timetable — allowing for new, substantive legislation in as little as a few weeks — delegates must constantly be at the ready, studying the relevant issues. There’s a chance that this pace is simply too fast for public acceptance. A one-week comment period may be inadequate for affected communities to assess a complicated proposal, even when the need for action is urgent.

The upshot is that how the CAAG initially operates should not be set in stone. The expectation should be that the CAAG’s operations and relationship to Congress and the President will undergo regular analysis. In fact, Congress should include funds for ongoing study of the CAAG in the original legislation.

Conclusion

Congress is slow by design, and that caution has served America well. But frontier AI may transform society faster than any technological revolution before it. When Congress cannot act quickly enough, important decisions about the future are left to actors with less accountability to the public.

A Citizens’ Assembly on AI Governance is one solution to this problem, building on lessons from past Congressional experiments. It would give five hundred randomly selected Americans the opportunity to shape policy together with domain experts, in order to avoid the pressures of elections or interest groups. Congress still gets the final say, and can reject proposals and decide on the assembly’s authority.

This proposal will require experimentation and refinement — but inaction is not an option. The public business must go on, despite the rapid changes ahead of us.

Footnotes

  1. In the 1988 round, commissioners reviewed some bases up to five times, commissioned reports, and conducted field visits. They even refused to dine together, lest they form a quorum the public could not observe. ↩

  2. In 2025, the DoD reported that the Army had excess space in some locations and unmet needs in others. Taxpayers for Common Sense and others warn that failing to authorize new rounds has led to bloat, waste, and possibly weaker national security. Requiring Congress to renew a delegation keeps it accountable, but it also lets Congress walk away while the problem remains. ↩

  3. Note that this uptick in trust in CAs by virtue of an increase seems to span cultures — it’s a common demand among residents of the U.S., Finland, and Ireland. ↩

  4. The 1990 statute authorized three rounds and terminated the Commission after their completion. Congress later authorized a 2005 round, reconstituted the Commission, increased its membership from eight to nine, and set a new termination date. ↩