Corrigible Software Development: A Paradigm for AI-Assisted Engineering

Large language models (LLMs) are reshaping software development practice. While existing research emphasizes productivity gains and human–AI collaboration patterns, less attention has been given to the structural consequences of embedding probabilistic code generation into development workflows. We argue that AI-assisted programming introduces a new paradigm in how constraints are discovered, formalized, and accumulated over time. Developers increasingly begin with partial intent, iteratively correct generated artifacts, and progressively encode recurring guidance into persistent instruction artifacts. When institutionalized, these artifacts reshape subsequent generative behavior across tasks. We formalize this mechanism as Recursive Corrigibility—the supervised transformation of recurring interaction-induced corrections into durable governance structures that constrain future probabilistic generation. We define the resulting paradigm as Corrigible Software Development (CSD). In CSD, specification becomes progressive rather than purely anticipatory. Partial intent initiates exploration; friction reveals latent constraints; recurring corrections crystallize into institutionalized rules; and accumulated governance progressively stabilizes AI-assisted workflows. Stabilization emerges not solely from exhaustive upfront specification, but from recursive constraint institutionalization and progressive cognitive offloading. This structural shift gives rise to an emerging orientation—Autonomy Engineering—in which governing generative autonomy becomes a central engineering responsibility. Rather than proposing a new model architecture or programming language, this work articulates a process-level transformation in AI-augmented software engineering.