Dario Amodei, CEO of Anthropic, just published a comprehensive AI policy essay with a bold proposal: regulate frontier AI models like aircraft. His timing is notable—he's simultaneously calling for government oversight while building exactly the kind of powerful models he argues need regulation.
Amodei frames the challenge through a Lord of the Rings analogy: AI advances rapidly from "an amusing toy to the full country of geniuses," while policy and legislation move painfully slowly. His solution? An FAA-style framework that treats powerful AI models like aircraft or pharmaceuticals rather than consumer apps.
The proposal is detailed and specific. Mandatory third-party testing for models above a compute threshold, focusing on four risk areas: cybersecurity, biological weapons, loss of AI system control, and automated R&D acceleration. Government authority to block unsafe deployments, with safeguards against political manipulation. Third-party evaluation through government agencies or authorized private organizations. Security standards requiring model protection, red-teaming, and collaboration on threat actors.
Anthropic also released an economic policy framework addressing labor disruption. Amodei argues that "the key challenge won't be incentivizing growth, but finding a way for everyone to share in the benefits." He proposes data collection tracking AI's displacement effects, wage insurance, retention tax incentives, workforce training grants, and long-term income support mechanisms including universal basic income or capital accounts.
Here's the tension: Anthropic is in a peculiar position. They're calling for regulation while actively pushing frontier models that they argue require regulation. Build the most powerful AI possible, then advocate for rules governing it. It's not hypocritical necessarily—you can genuinely believe in both—but it's certainly a balancing act.
Critics at Mashable note that Anthropic's growth claims conflict with independent research findings. And policy experts point out that the devil is in the implementation details. An FAA-style framework could be transformative if executed properly, or it could become bureaucratic theater that slows innovation without improving safety.
The most substantive part might be the economic framework. While everyone argues about AGI timelines and existential risk, is actually grappling with the practical question of what happens when AI starts displacing workers at scale. Data collection, wage insurance, training programs—these are concrete policies that could help regardless of whether AI becomes superintelligent or just very economically disruptive.
