A Seven-Layer Model for Standardising AI Fairness Assessment

AI bias is a systemic problem across the entire lifecycle: The paper identifies that AI systems are "prone to biases at every stage of its lifecycle, from inception to its usage," meaning fairness cannot be addressed through a single intervention but requires attention throughout the development and deployment process.

Current research has significant blind spots: The authors found that peer review shows "excessive focus on biases in the datasets, fairness metrics, and algorithmic bias," while "other aspects affecting AI fairness get ignored"—suggesting that many organizations may be overlooking critical sources of bias outside these commonly-studied areas.

A standardized seven-layer framework is needed: The paper proposes organizing AI fairness assessment into seven distinct layers (inspired by networking standards), with each layer corresponding to "a well-defined AI model-building or usage stage," providing a structured approach similar to how other technical standards are implemented across industries.

Practical implementation tools are provided: The model includes "checklists for each layer" and identifies "potential sources of bias in each layer and their mitigation methodologies," offering organizations concrete guidance rather than abstract principles for assessing fairness at each stage.

Context matters for fairness standards: The authors emphasize that "AI fairness and other ethical requirements depend on multiple factors such as context, use case, type of the AI system," indicating that fairness cannot be defined universally but must be tailored to specific applications.