Title: Update Opacity: Epistemic Accessibility and Governance Under AI System Change

Abstract:

Machine learning models integrated into active AI systems are frequently modified to ensure their continued operational accuracy. However, these modifications can create a phenomenon known as "update opacity," where users struggle to comprehend why a previously identical input now produces a divergent result. We posit that update opacity should be interpreted as a diachronic breakdown in epistemic accessibility. The core issue lies in the fact that changes with material significance often fail to remain visible to human users in ways that facilitate understanding, enable calibrated reliance, and support appropriate action, particularly given the constraints specific to certain roles and timeframes. Consequently, this phenomenon constitutes a significant governance challenge. It is important to note that not all system changes hold equal relevance; furthermore, the mandatory disclosure of every minor update could overwhelm users and hinder system utility. To resolve this dilemma, we integrate two complementary governance strategies: the EU AI Act, which delineates the system-level boundaries of normatively significant change, and Machine Learning Operations (MLOps), which offers the operational mechanisms necessary to track and compare changes across time. Building on this integration, we introduce a framework that characterizes system evolution through trustworthiness profiles and levels, employing threshold-based disclosure to highlight materially relevant changes that fall within established parameters to various stakeholders over time. We demonstrate the efficacy of this approach using a case study in medical AI and outline practical consequences for lifecycle documentation, post-market surveillance, and update transparency.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC