A multi-layer social-theoretical framework for AI ethics

The increasing use of artificial intelligence (AI) in socially consequential domains has intensified concerns regarding fairness, privacy, accountability, human autonomy, institutional power, and fundamental rights. Although numerous AI ethics and governance frameworks provide principles for responsible AI, translating these principles into context-sensitive and operational ethical assessment remains challenging. This study develops a Multi-Layer Social-Theoretical AI Ethics Framework (MLST-AEF) that integrates normative ethical reasoning, stakeholder analysis, institutional context, bias and power assessment, and structured decision support. The framework is developed through a conceptual framework-development methodology combining literature synthesis, problem structuring through Soft Systems Methodology (SSM), framework construction, configurable ethical scoring, and illustrative application. SSM is incorporated to examine stakeholder worldviews, relationships, ownership, and environmental constraints beyond conventional stakeholder identification. The resulting MLST-AEF comprises four analytical layers: Ethical Evaluation, Stakeholder Impact Analysis, Institutional Context Assessment, and a Bias and Power Audit. These layers are integrated with a configurable scoring mechanism using dimension-level scores, explicit weights, critical ethical thresholds, qualitative justification, and safeguard requirements. An illustrative application to facial-recognition technology in UAE policing demonstrates how the framework reveals tensions between anticipated public-safety benefits and concerns relating to rights, distributive fairness, human agency, surveillance, and contestability. The assessment indicates a mixed ethical profile supporting conditional rather than unconditional acceptability, subject to appropriate safeguards. The MLST-AEF contributes an integrated and traceable decision-support structure that preserves human judgement while making ethical trade-offs, stakeholder disagreements, and weighting assumptions explicit.