<p>Despite the rapid expansion of artificial intelligence (AI) and the proliferation of ethical frameworks, failures in ethical AI practice persist and, in some cases, intensify. This paradox highlights a fundamental limitation in current approaches: while normative principles are increasingly well defined, their translation into technical practice remains insufficient. Existing frameworks are largely principle-driven and governance-oriented, offering limited insight into how ethical risks arise within data pipelines, model development processes, and system architectures. This study reframes ethical risk in AI as a challenge of implementation rather than principle. It adopts a mixed-method approach, integrating a systematic literature review with PRISMA-based screening, structured comparative case analysis, and simulation-based evaluation to examine how variations in data quality and selection criteria influence downstream model behaviour and fairness outcomes. The findings demonstrate that ethical vulnerabilities are structurally embedded within processes of data generation, curation, and system design. This challenges prevailing assumption that fairness can be achieved through representational balance alone, instead emphasising the importance of data quality, conceptual validity, and contextual fitness for purpose. To address these gaps, the study introduces two complementary frameworks: ethics-based multi-criteria decision making (EbMCDM) and ethics-based system design (EbSD), which embed ethical evaluation within dataset selection and system architecture respectively, advancing a proactive, design-oriented approach to responsible AI.</p>