Plan A authors defend AI governance proposal against mischaracterizations

In a detailed rebuttal on LessWrong, MKodama and Thomas Larsen address what they see as serious misrepresentations of their AI governance proposal, Plan A, by critic Séb Krier. They argue that Krier’s characterization—that Plan A “bakes in too much” and leaves little space for trial-and-error—is the exact opposite of the truth. Plan A is designed to be extremely iterative, offering more time for AI companies to generate evidence and for governments to respond reasonably. Total research transparency (TRT) ensures that all evidence informing decisions is public and accessible to academics, independent researchers, and the public, not locked away in corporate labs. This, they claim, maximizes learning and room for experimentation in a way the current market-driven system does not.

The authors also address claims about economic assumptions, global coordination, and elite control. They reject the notion that Plan A assumes all profits accrue to labs or that it creates a central planner—instead, it coordinates US and Chinese regulators to ban only the most egregiously dangerous research directions, with all decisions backed by public evidence. They distinguish their proposal from older safety plans that relied on solving alignment and handing it to the UN. In Plan A, private companies iteratively solve alignment while regulators set a low floor. The response underscores that Plan A is not a top-down diktat but a framework for transparency, adaptation, and shared responsibility in AI development.