Why AI’s Exponential Rise Demands Faster Policy Action, Says Dario Amodei

On June 10, Anthropic CEO Dario Amodei published “Policy on the AI Exponential,” asserting that AI is moving exponentially while policy mechanisms remain designed for a slow world. He likens the situation to hobbits urging Treebeard to act quickly, yet Treebeard takes a day just to greet them.

Anthropic has long supported transparency requirements such as safety tests, system cards, and risk reports, but Amodei says this is no longer sufficient because risks are now concrete. He cites Claude Mythos Preview’s ability to discover many high‑severity software vulnerabilities that affect operating systems, browsers, and critical‑infrastructure, and argues that frontier models have become tools of strategic national importance, with cybersecurity, bio‑risk, and AI‑autonomy risks likely to follow.

Anthropic released two policy frameworks: one for frontier‑model safety and another for AI’s impact on employment and the economy.

Frontier models should be accepted like aircraft

Amodei proposes mandatory third‑party testing before releasing models that meet specific thresholds—training compute over 10²⁵ FLOPs and either AI‑related revenue above $500 million or AI R&D spend above $1 billion. Testing focuses on four catastrophic risks: cybersecurity, bioweapon, AI loss‑of‑control, and risk amplification from automated R&D.

The framework requires protecting model weights and training infrastructure, regular red‑team and penetration testing, reporting major security incidents, and granting governments legal authority to block or deter dangerous deployments, with penalties tied to global annual revenue and escalating for repeat violations.

He notes the controversy: giving governments gate‑keeping power sounds reasonable but risks regulatory capture. Large firms can absorb compliance costs, while small teams and open‑source communities may be squeezed. Over‑reliance on safety can become a moat, and the industry lacks enforceable rules rather than moral statements.

Work issues placed on the table

Amodei does not claim AI will cause permanent mass unemployment, but argues past technology‑revolution analogies are insufficient because AI replaces broader cognitive abilities. He proposes three policy categories: (1) better labor‑market monitoring with real‑time AI impact visibility; (2) employment incentives such as wage insurance, tax breaks for retained jobs, training subsidies, and job‑matching infrastructure; (3) long‑term income support like universal basic income or universal capital accounts funded by relevant corporate or capital‑gains taxes.

He observes that data‑center electricity‑price anxiety often masks larger economic anxiety; Anthropic has pledged to absorb electricity‑cost impacts from its data centers, suggesting public discontent may stem from perceived private profit and externalized costs.

AI accelerates science, regulation may bottleneck

Amodei is optimistic about AI in biomedicine, believing it could flood drug pipelines with candidates and improve safety and efficacy. However, FDA and EMA processes remain slow—typically 7–8 years per candidate—so accelerated R&D could make existing regulatory pipelines a bottleneck.

He recommends agencies pre‑define standards to accept AI‑generated simulations and analyses, including pharmacokinetic modeling, toxicity prediction, dose selection, synthetic control arms, and surrogate endpoints. This reverses intuition: stricter standards for AI itself, faster acceptance for AI‑derived downstream results.

States must not wield AI as unlimited power

Amodei warns AI could let state power bypass traditional checks, enabling fully autonomous weapons, massive surveillance, and government use of AI tools that outmatch citizens in litigation and regulation.

His policy suggestions include: (1) fully autonomous weapons must be subject to court, legislative, and lawful command‑chain constraints; (2) ban domestic law‑enforcement use of fully autonomous weapons; (3) shut down data brokers and bulk‑data‑collection loopholes; (4) ensure that when governments use AI against individuals or organizations, the affected parties can obtain AI assistance of comparable capability. He notes future procedural justice may involve AI parity, not just lawyers.

National alliances and chip supply chain

Amodei views AI as a core military and economic power, comparable to or surpassing nuclear weapons. He recommends nations form AI alliances to share chips and semiconductor equipment while restricting adversaries from accessing critical supply chains. Alliances would coordinate safety assessments, export controls, AI defense, AI‑driven intelligence and manufacturing, and reject AI‑driven high‑pressure domination.

This reflects a realistic, US‑centric view; once AI safety enters national‑security frameworks, it moves beyond laboratory ethics committees.

The real contradictions remain unresolved

Critics ask how the rules can constrain firms that both demand mandatory testing and hand frontier capabilities to national‑security agencies, and who will prevent bureaucrats, interest groups, and large firms from defining safety. These doubts cannot be dismissed with vague future‑humanity arguments.

Amodei’s article’s value lies in moving the conversation from abstract AI risk to concrete policies: testing, auditing, reporting, deployment blocks, employment compensation, drug‑regulation reform, civil‑rights protections, and chip‑supply‑chain governance. The more specific the rules, the more they expose power distribution. According to Anthropic, the transparency era has ended, and the next question is who will place their models, customers, deployment processes, and commercial interests under the same rule set.

AI policy mechanisms have finally awakened; now we must decide whose trees get cut.