Anthropic CEO Amodei proposes three-stage AI development coordination plan
In brief
- External evaluators embedded in Anthropic with access comparable to internal risk teams
- Government-mediated coordination across U.S. frontier AI developers via regulation
- Verifiable international agreements among democracies to preserve strategic pacing relative to China
Building Scrutiny from Within
Amodei's first stage proposes inviting external evaluators into Anthropic with access comparable to internal risk teams. The proposed review team would receive company equipment, workspace access and opportunities to speak with employees, with contracts permitting publication of key findings without Anthropic controlling conclusions.
This approach leverages Anthropic's corporate structure. The company operates as a Public Benefit Corporation under Delaware law, requiring directors to balance stockholder interests with interests of affected people and specified public benefit. Its Long-Term Benefit Trust holds board-selection powers intended to support the company's mission—a governance layer that creates room for safety-aligned decisions internally.
But internal governance alone has limits. A public-benefit charter can authorize safety-minded decisions inside one company. It cannot bind a competitor that rejects the same trade-off.
Scaling to Regulation and International Coordination
Amodei's second and third stages address that gap. Amodei's second stage calls for regulation and government-mediated coordination across a critical mass of U.S. frontier AI developers. His third seeks verifiable agreements among states, with democracies preserving enough strategic room relative to China to pace development.
The proposal steers between two opposing risks. Full nationalization—taking government ownership stakes in frontier labs—creates its own concentration risk by placing model development and the decision to stop it in the same institution. Yet purely voluntary coordination fails when competitive pressure incentivizes races to the bottom.
The Narrow Brake Alternative
Recent legal scholarship points toward a middle path. An August 2026 legal paper by Yonathan Arbel, Simon Goldstein and Peter Salib proposes a narrow, discretionary and temporary government power to halt frontier AI training or deployment when catastrophic risk or corporate power is involved. A halt order reaches pacing decisions more directly than public equity. The state would not need to own every model or operate every laboratory before suspending covered training or deployment.
The authors favor conventional regulation or taxation for addressing monopoly, inequality and other harms—keeping the brake narrow and emergency-use only.
The Decentralization Counter
Open-weight models press in the opposite direction by widening access. Researchers can inspect and adapt systems without relying on a handful of corporate gatekeepers. That distribution of capability creates its own form of scrutiny and reduces single points of control.
Amodei's proposal doesn't resolve which path dominates. It maps the terrain: nationalization, narrow government brakes, and decentralized access each carry tradeoffs. The choice between them will shape not just AI governance, but the structure of computational power for years ahead.
