An Invitation to the AI Safety Research Community

SIRDA is seeking criticism of a specific hypothesis: that a narrowly bounded, distributed transnational government can supply legitimate decision procedures for catastrophic AI risk while constraining its own power. The inquiry builds on AI 2040: Plan A, the debate it prompted on LessWrong, and the governance, assurance, and rights work mapped in the Ecosystem.

Here is the case we want help examining. A consequential AI development is visible, its risks remain disputed, and the parties have incentives to disagree. What procedure determines whether it proceeds? Who authorizes that procedure? What prevents its administrators from converting temporary safety responsibility into lasting control?

A useful contribution can address one question, a source, an assumption, or a proposed mechanism. Describe the actor, the incentive, the point of failure, and the consequence. Alternative arrangements will face the same stated tests as the founding proposal.

Researchers inside labs are welcome in a personal capacity. Organizations can submit authorized responses. We also seek contributions from technical safety groups, independent evaluators, governance and legal scholars, public institutions, infrastructure providers, economists, security researchers, rights organizations, and researchers across countries. Stage One maps who already holds consequential powers, so we especially want people who know how licensing, compute access, procurement, insurance, investment, standards, and certification work in practice. Public communicators and podcasters can help identify questions that specialist discussions leave unclear to people affected by the technology.

Authors whose work the inquiry discusses are invited to correct our account of it. Factual corrections will be made with a visible revision record. With permission, substantive responses will appear alongside the proposal, including arguments that the project should stop. Participation expresses engagement with a question; endorsement is recorded separately.

Challenge the Argument