Mapping Data Governance Requirements Between the European Union’s AI Act and ISO/IEC 5259: A Semantic Analysis
Keywords: Artificial intelligence, Data governance, Data quality management, Semantic analysis, Ontological mappings, Responsible AI governance
Abstract: The rapidly evolving landscape of artificial intelligence (AI) has made regulatory frameworks essential to guide the development, deployment, and usage of AI technologies responsibly. Recently, the European Union (EU) has approved the AI Act to address these needs, laying out comprehensive requirements for AI systems. Similarly, the EU published a request for harmonized standards that would support implementation of the AI Act across a number of topics related to trustworthy AI and AI quality and management. One source for such European Harmonised Standards are International Standards, and ISO/IEC JTC1 SC42 has a number of standards published and in development that may be appropriate, but official analysis shows some gaps that require additional features for existing standards. It is not clear that near term modifications to existing standards will satisfy all the requirement of the AI Act given the complexity and lack of state of the art in many areas, especially in novel area such as protection of fundamental rights.We propose therefore the use of semantic web vocabularies to track the mappings of AI Act requirements that will enable the progressive tracking to third party guidelines, standards and specification. In particular, we demonstrate this approach by producing an requirement analysis of Article 10 of the AI Act on Data Governance and map it to the relevant provisions of the SC42 standards 5259 on Data Quality for Machine Learning. This study conducts a semantic analysis of the EU’s AI Act and ISO/IEC 5259 requirements, utilizing the Simple Knowledge Organization System (SKOS) ontology to map concepts between these two frameworks. We identify areas of alignment, partial alignment, and disparities between these regulatory requirements. Our analysis covers various dimensions, including completeness of satisfaction, partial satisfaction, normative language differences, definition disparities, and associated costs for compliance. Our findings reveal instances of direct alignment, partial alignments, variations in normative language and disparities in concept definitions, highlighting nuanced differences in terminology and scope.
Submission Number: 5
Loading