Conditions for the Establishment of Accountability in Generative AI — From the Perspective of Reproducibility, Persistence, and Verifiability of the Pre Judgment Structure —

This paper examines the conditions for the establishment of accountability in generative AI from the perspective of the structural conditions underlying judgment processes. Existing discussions on AI accountability often focus on the distribution of responsibility among developers, operators, users, or institutions. However, prior to the allocation of responsibility, the structural conditions under which accountability itself becomes possible have not been sufficiently clarified. This paper argues that accountability cannot be established merely by the existence of a result. For accountability based on a judgment process to be established, it must be possible to identify, reconstruct, preserve, and verify the process through which the result was formed. The paper defines the set of prior conditions that enable judgment formation as the Pre-Judgment Structure, consisting of information selection, criteria application, and scope delimitation. Based on this framework, the paper demonstrates that when the Pre-Judgment Structure is not fixed, multiple judgment processes may arise even under identical input conditions, making it impossible to uniquely identify which conditions led to a given result. Consequently, accountability based on the judgment process becomes structurally incomplete. The paper then proposes three necessary conditions for the establishment of accountability: reproducibility, persistence, and verifiability. It argues that the Pre-Judgment Structure and the judgment process must be reconstructible under identical conditions, maintained over time, and verifiable by third parties. Applying this framework to current generative AI systems, the paper argues that many contemporary AI environments do not sufficiently externalize or preserve the Pre-Judgment Structure, making post-hoc reconstruction and verification of judgment processes difficult. Therefore, the instability of accountability in generative AI should be understood not merely as a problem of ethics, institutional design, or probabilistic output behavior, but as a structural problem concerning the fixation, maintenance, and verification of the Pre-Judgment Structure.