An interpretable and adaptive AI governance framework

Abstract

Abstract In recent years, the rapid development of artificial intelligence has raised a series of ethical issues. How to balance AI innovation and AI governance has attracted wide attention from society. Some principle-based and practical approaches have already emerged. However, these strategies mainly face problems of time, speed, and cost. They fail to fundamentally balance dynamic factors and the need for explainability. This study proposes an explainable and adaptive AI governance approach from the perspective of time to balance AI governance and AI innovation. Causal models have causal explanation capabilities. They meet the requirements of responsible innovation. They can improve governance speed and reduce governance and innovation costs. Taking Bayesian networks as an example, they have explainability and dynamic adaptation ability. Bayesian networks can be used as a governance step before a regulatory sandbox. Bayesian networks can implement a human-assisted regulatory approach. This approach can combine the EU Artificial Intelligence Act with scholars’ recommendations on quantifiable risk levels. The explainable and adaptive AI governance approach has three important meanings for AI governance. First, it can provide an inclusive governance path. Second, it can supplement static governance strategies. Third, it can automatically generate explainable value alignment schemes.