Ferrario & Hatherley expose 'update opacity' in AI—offer governance fix

Machine learning models in deployed AI systems are routinely updated to maintain performance, but these updates often introduce 'update opacity'—users cannot understand why the same input now yields a different output. Ferrario and Hatherley frame this as a diachronic failure of epistemic accessibility: relevant changes may remain invisible to users under real role- and time-specific constraints. This makes update opacity a governance problem, not just a technical one. Not every change is worth disclosing; over-disclosure can overwhelm users and undermine trust.

To address this, the authors propose a hybrid governance framework that combines the EU AI Act's system-level perimeter for normatively relevant change with Machine Learning Operations (ML Ops) operational tracking tools. They model system change through trustworthiness profiles and trustworthiness levels, then use threshold-based disclosure to selectively surface materially relevant within-envelope changes to different stakeholders over time. The approach is illustrated with a medical AI example, showing how clinicians and regulators can be alerted only to updates that impact clinical decisions.

The paper derives practical implications for lifecycle documentation, post-market monitoring, and update disclosure. By formalizing when and how to communicate model changes, the framework aims to preserve user understanding, calibrated reliance, and appropriate action. For professionals deploying AI systems, this research provides a structured way to maintain transparency and trust as models evolve—a critical need as regulatory scrutiny like the EU AI Act tightens.