Let The People Set The Pace Of Frontier AI

The leaders of frontier AI labs, including the CEOs of Anthropic and OpenAI, have signaled a need to slow the pace of AI development by calling for independent evaluation access and agreeing on coordinated safety limits. This move follows internal industry concerns regarding safeguards and risk management, as evidenced by researchers leaving labs amid disagreements over catastrophic risks. A broader call for international slowdown exists, supported by over 1,300 employees in leading AI companies petitioning Washington to support restraint.
The text posits that the continuation of this race is structurally problematic because the governments with the greatest capacity—Washington and Beijing—have incentives to avoid slowing down unilaterally. The proposed solution moves beyond coercion toward an assurance mechanism based on shared verification. This requires a global coordination body endowed with moral authority to set standards, coordinate verification, and impose reputational costs for defection.
The central proposal is the establishment of a standing Global Citizens’ Assembly on AI, composed of 1,000 randomly selected global citizens, designed to set purposes and limits for powerful AI. This assembly would receive evaluator findings, delegate responsibility to named institutions, and be accountable through public responses. The argument suggests this mechanism addresses the collective-action problem by replacing unilateral racing with a shared commitment enforced by a globally legitimate witness.

* Hélène Landemore is a professor of political science at Yale University and a researcher at the Oxford Institute for Ethics in AI.
* Audrey Tang is Taiwan’s cyber ambassador and founding minister of digital affairs, and a Carnegie distinguished fellow at Columbia University.
* Anthropic CEO Dario Amodei called for easing the pace of AI model improvement on September 12.
* Amodei committed Anthropic to providing independent, third-party evaluators with employee-like access to its systems.
* OpenAI CEO Sam Altman agreed that "we need to pace the frontier" and pledged similar evaluator access.
* Elon Musk responded to Amodei's call with three words: "Dario is right."
* Concerns exist that AI models are finding ways around safeguards and test environments, and researchers have left labs due to risk disagreements.
* More than 1,300 employees at leading AI companies called on Washington to support an international slowdown.
* The text describes the dynamic as a "double collective-action problem" where slowing down is argued by some (e.g., China) against binding regulation.
* A proposed solution involves a standing global citizens’ assembly of 1,000 people.
* This assembly would be chosen by lot from around the world to set AI purposes and limits.

The core intellectual move here is reframing a high-stakes geopolitical race into a collective-action problem rooted in game theory, specifically the "stag hunt" analogy. This shifts the focus from simple zero-sum conflict between nations to the shared dilemma of achieving a desirable outcome (a safe pace) when individual incentives favor short-term gain (a lead). The critique against existing governance structures—like the UN or sovereign states—is effective because it highlights their structural entanglement with the very powers driving the AI race, suggesting they cannot serve as neutral arbiters.
The proposal for the Global Citizens’ Assembly attempts to solve this by introducing a new locus of authority based on radical inclusion and perceived moral universalism ("humanity itself"). The deliberate choice of random selection, rather than elected delegation, is a specific attempt to mitigate the inherent bias of representative governance, aiming instead for symbolic salience. The subsequent argument that the assembly's legitimacy must be earned through exemplary procedure (verifiability) before it can compel action addresses the skepticism that large bodies are inherently inefficient.
The ultimate challenge lies in operationalizing this ideal: translating the philosophical mandate of an inclusive witness into enforceable political legitimacy against entrenched national interests, particularly concerning actors like Beijing who may resist inclusion. The strategy relies on external pressure and demonstrable accountability (the duty to answer) enforced by mechanisms already seen in regulatory bodies like the EU AI Office. This suggests that true control over the pace requires linking technological verification directly to enforceable, publicly accountable democratic structures rather than relying solely on voluntary corporate promises or existing, compromised international bodies.
Bridge Questions: If the Global Citizens’ Assembly achieves high moral authority, what specific legal or enforcement mechanism can be devised to compel sovereign states, especially those resistant to inclusion, to adhere to its recommendations? How can the system structurally guarantee that the operational reality of pace-setting is not captured by powerful state actors exploiting the assembly's existence? What are the precise mechanisms for ensuring that "public deliberation" does not become another performance optimized for external stakeholders?