Global Citizens' Assembly Proposed to Govern Frontier AI Development

A New Governance Model for AI

Leaders of major AI companies increasingly acknowledge their technology race may be moving dangerously fast, but structural incentives prevent them from slowing down unilaterally. Now two prominent governance experts have proposed a novel solution: a standing global citizens' assembly of 1,000 people chosen by lottery to set boundaries for frontier AI development.

Yale political scientist Hélène Landemore and Taiwan's cyber ambassador Audrey Tang detailed the proposal in Noema Magazine. Their core argument is that neither the United Nations nor any single nation possesses sufficient legitimacy to coordinate a slowdown between competing powers, particularly the United States and China.

The Double Race Problem

The challenge facing AI governance is what the authors characterize as a nested collective-action problem. Domestically, competing labs cannot afford to slow down if rivals continue racing. Internationally, Washington and Beijing each fear that restraint will hand strategic advantage to the other.

This dynamic resembles a "stag hunt" rather than a prisoner's dilemma, the authors argue. In Rousseau's parable, hunters collectively pursuing a stag would all benefit from cooperation, but each risks breaking away to chase an individual hare if they lose faith in the group. Similarly, AI labs and governments would prefer coordinated pacing over a dangerous race, but lack assurance that others will cooperate.

Recent statements from industry leaders underscore the tension. In September, Anthropic CEO Dario Amodei called for easing the pace of AI model improvement, with OpenAI's Sam Altman quickly agreeing that "we need to pace the frontier." Yet without structural coordination mechanisms, these commitments remain voluntary and fragile.

How the Assembly Would Function

The proposed Global Citizens' Assembly on AI would comprise 1,000 members selected by lot from every continent, serving fixed part-time terms with one-third of seats regularly redrawn. Members would receive financial compensation, translation services, secure communication tools, and protection from government retaliation.

The assembly would operate continuously rather than producing a single report. It would set risk thresholds for AI systems, receive findings from independent evaluators with employee-like access to frontier labs, and issue recommendations to named institutions with deadlines for public response.

Crucially, the body would not require unanimous global participation to function. The European Union's AI Office already possesses enforcement powers over major labs operating in its market. A revised code of practice could obligate signatories to answer the assembly's recommendations publicly, giving the duty to respond its first legal anchor.

Why it matters

This proposal addresses a fundamental legitimacy gap in AI governance. Current frameworks rely on labs self-regulating, governments protecting national interests, or expert panels lacking democratic mandate. A lottery-selected global body would represent humanity's universal stake in AI outcomes without the structural conflicts that compromise existing institutions. If even partially implemented, it could transform voluntary industry commitments into accountable public processes.

Precedents and Path Forward

The authors note that more than 800 similar deliberative bodies have been convened globally, according to the OECD. The Global Assembly for COP26 drew 100 people worldwide to deliberate on climate change. In 2023, Anthropic and the Collective Intelligence Project engaged about 1,000 Americans to draft principles for an AI model.

Landemore and Tang propose starting with willing participants: a coalition of governments and foundations could fund the assembly through an independent secretariat, while Anthropic and OpenAI could open their labs to evaluators vetted under assembly-set rules.

The details were first reported by Noema Magazine.