Update Opacity: Epistemic Accessibility and Governance Under AI System Change

Researchers propose a governance framework addressing 'update opacity'—the problem that AI system updates can change outputs without users understanding why. The framework combines EU AI Act requirements with Machine Learning Operations tools to enable threshold-based disclosure of materially relevant changes to stakeholders, using trustworthiness profiles to determine what information different parties need.

Update opacity represents a critical governance challenge in deployed AI systems where routine model updates obscure changes in behavior from end users and operators. The core issue extends beyond simple transparency: users cannot calibrate reliance on systems when they lack accessible explanations for output differences, creating safety and trust failures in high-stakes domains like healthcare. This research reframes update opacity as a diachronic epistemic accessibility problem—the challenge of maintaining human understanding across system changes over time rather than at a single moment.

The governance landscape has evolved significantly with regulatory frameworks like the EU AI Act establishing baseline requirements for system documentation and monitoring. Simultaneously, MLOps methodologies provide technical infrastructure for version control and change tracking in machine learning pipelines. The authors leverage both approaches, recognizing that regulatory perimeters define what constitutes normatively relevant change while operational tools enable practical implementation of monitoring and disclosure mechanisms.

For AI developers and healthcare providers deploying clinical decision support systems, this framework offers concrete implementation guidance through trustworthiness profiles that specify stakeholder-specific disclosure thresholds. Rather than overwhelming users with every update, the approach identifies which changes materially affect system behavior and routes information appropriately. Medical AI exemplifies the stakes: clinicians must understand when diagnostic recommendations shift to adjust their reliance calibration and practice accordingly.

Future regulatory compliance will likely incorporate these principles, making lifecycle documentation and post-market monitoring central to deployment governance. Organizations developing AI systems should begin mapping their update processes against these frameworks to anticipate compliance requirements and build systems supporting differentiated stakeholder disclosure.

• →Update opacity occurs when AI system changes produce unexplained output differences, preventing users from understanding why the same input yields different results.
• →The framework combines EU AI Act governance requirements with MLOps technical tools to create threshold-based disclosure mechanisms for materially relevant changes.
• →Trustworthiness profiles enable different stakeholders to receive appropriate information about system changes without overwhelming users with every minor update.
• →Post-market monitoring and lifecycle documentation become critical governance practices for maintaining epistemic accessibility in deployed AI systems.
• →Healthcare AI systems exemplify the necessity of this approach, where clinician understanding of recommendation changes directly impacts patient safety.