Anthropic CEO Dario Amodei has published a major policy essay, “Policy on the AI Exponential,” arguing that AI regulation is roughly a year behind the technology’s actual pace and must urgently catch up. The essay lays out concrete proposals across five areas.

Key proposals include:

  • Regulation modeled on the FAA: Frontier models above a compute threshold should undergo mandatory third-party testing for cybersecurity, bioweapons, loss of control, and automated R&D risks, with government empowered to block unsafe releases.
  • Job displacement response: Better measurement and tracking of labor market effects, pro-employment incentives like wage insurance and retention tax credits, and long-term income support up to universal basic income if displacement proves enduring.
  • Fast-tracking AI’s benefits: Regulatory agencies like the FDA should develop standards now for accepting AI-driven methods (simulations, synthetic control arms, biomarker validation) to prevent approval pipelines from becoming bottlenecks.
  • Civil liberties safeguards: Accountability rules for autonomous weapons, a ban on domestic use of fully autonomous systems, closing the data broker loophole on bulk surveillance, and a right to AI assistance during adverse government action.
  • Democratic AI coalition: Like-minded democracies should form a coalition to share chip supply chains, coordinate risk policies, and deny access to adversaries — treating AI as strategically comparable to nuclear weapons.

Amodei rejects the framing that AI has a “PR problem,” arguing that public concern reflects a correct perception of real risks, and calls the current moment a “window of opportunity” where policymakers are unusually receptive to forward-looking action.

Policy on the AI Exponential