Ethical Data Governance Models for AI Training Pipelines

Abstract

Data governance for AI training pipelines -- the policies, processes, and technical mechanisms governing the collection, curation, annotation, and use of data to train machine learning models -- is a foundational requirement of responsible AI that has received substantial regulatory attention but limited operational specification. The EU AI Act (2024) Article 10 mandates data governance practices for high-risk AI training data covering quality criteria, representativeness, bias examination, and personal data handling, while the GDPR (2016) imposes independent data protection obligations across the full data lifecycle. This paper proposes the Ethical Data Governance Framework for AI Pipelines (EDGAP), a structured governance model operationalising these regulatory obligations as concrete data governance practices across five pipeline stages: data acquisition, data curation, data annotation, data preprocessing, and data versioning and provenance. EDGAP comprises twenty-two governance practices, each specified with a formal requirement, verification method, regulatory compliance mapping, and risk severity classification. The framework is evaluated through a governance audit of 26 AI training pipelines across six industry sectors and a controlled experiment assessing the data quality improvement achieved by EDGAP adoption in four organisational case studies. EDGAP audits identify a mean of 5.8 governance gaps per pipeline (SD = 1.7), of which 43.1% are classified Critical or High severity. EDGAP adoption in case studies achieves a 52.4% reduction in data quality defects over 90 days (SD = 6.8%) and full Article 10 compliance documentation for all four adopting organisations. The study contributes the EDGAP specification, a Data Governance Maturity Index (DGMI), and an empirical benchmark of data governance gaps across AI sectors.