Beyond the AI slowdown: the four As for governing frontier AI

Abstract Recent developments in cybersecurity, scientific discovery and AI-assisted research suggest that the central challenge is shifting from whether frontier AI can be slowed to what ethical obligations follow as its capabilities diffuse. We organise these obligations around four ethical loci, the four As (4As): Action (operational control), Access (capability access), Authority (delegated authority) and Allocation (systemic distribution). Each locus captures a distinct ethical problem and therefore calls for a different family of governance responses. Drawing on Asilomar, CRISPR-Cas governance, nuclear arms control and scholarship on responsible innovation and meaningful human control, we develop a framework for governance matching, linking ethical diagnosis to locus-specific constraints, control authorities, failure signals and corrective responses. Because multiple loci may be implicated simultaneously, the framework also identifies governance mismatch and cross-locus effects. We conclude by proposing an AI stability framework that preserves competition and innovation while establishing minimum safeguards around high-consequence capabilities and uses that could cause severe or irreversible harm.