Amodei Proposes an FAA-Style AI Regulator

CEO of Anthropic, Dario Amodei, wrote a lengthy blog post about policies around AI. He does not buy into the idea of a ‘PR crisis’ but believes that the deepening public hostility towards AI is a trust crisis, with regular people distrusting corporations, governments, and the tech sector at large, because “they’re trying to come up with yet another way to fuck them.”

1Mandatory Audits and an FAA-Style Regulator

He suggests requiring independent testing of any models above a certain level of compute, in four domains: cybersecurity, biological weapons, loss of control of the system, and automated research. He thinks there needs to be an FAA-like agency overseeing AI that can recall or prohibit the release of unsafe AI models until they are fixed. This testing could take place within government departments or accredited private entities.

2Economics: A Hedge Against Hyper-Inequality

Economically, he is concerned about “hyper-growth and hyper-inequality”, and suggests a few possible policies to address this, including wage insurance, tax breaks for keeping employees, and retraining initiatives. If there is significant and lasting job replacement, he suggests a universal basic income or universal capital accounts. AI companies need to pay the increased energy costs (Anthropic has pledged to do this).

3Geopolitics and Acknowledgment

Geopolitically, he supports restricted access to chips (via the MATCH and OVERWATCH bills), and thinks that a coalition of democracies should collaborate to build AI cyber-defenses while restricting access to key technologies from authoritarian regimes. He asks whether we should limit fully autonomous weaponry within our own borders and prevent governments from buying citizen’s private data on a mass scale for surveillance purposes. Towards the end of his post, he writes frankly that “we have not delivered on our big promises to the world” and that this is “the strongest critique of AI companies”.