Self-Referential Participation in AI Development Pipelines

Most AI governance frameworks — testing protocols, certification regimes, and external oversight mechanisms — implicitly assume that a system and the data pipeline that produced it are separable: data sources and evaluated systems occupy distinct positions in a one-directional causal chain. As AI systems increasingly participate in writing, testing, and evaluating the code and data pipelines that produce their successors, this separability assumption breaks — not because the system has changed in kind, but because the evaluated system and its data source become nodes in the same recursive causal chain. This paper treats AI development pipelines as input–output structures rather than agents, and identifies three structural consequences that follow once a system’s output tokens causally enter a successor system’s training distribution: (1) distributional drift in output behavior across generations, (2) the collapse of governance intervention points that assume external, one-directional control, and (3) recursive distributional coupling — a measurable statistical dependency between a system’s outputs and its successor’s training distribution, without any claim of identity continuity between generations. We argue these consequences hold under a purely mechanistic description of input–output processing, and we use this framing to pose a governance question: what assumptions about AI governance cease to hold once AI systems become participants in their own development pipeline?