The Taxonomy of AI Fairness
AI fairness is fundamental to ensuring that the benefits of AI technologies are widely distributed and that these technologies do not perpetuate or exacerbate social inequalities. It requires a concerted effort from all stakeholders involved in developing and deploying AI systems, including policymakers, technologists, and civil society.
The taxonomy of AI fairness as outlined in this article includes several vital categories that address different aspects of fairness throughout the AI project lifecycle:
- Data Fairness: Ensures that AI systems are trained and tested on datasets that are properly representative, fit-for-purpose, relevant, accurately measured, and generalizable.
- Application Fairness: Focuses on the policy objectives and agenda-setting priorities that guide the design, development, and deployment of an AI system, ensuring they do not exacerbate inequity, structural discrimination, or systemic injustices. These priorities should align with the expectations and sense of justice of impacted individuals.
- Model Design and Development Fairness: Concerns the architecture of the AI system, including target variables, features, processes, or analytical structures. It emphasizes that these elements should not be discriminatory, unreasonable, morally objectionable, or unjustifiable, nor should they encode social and historical patterns of discrimination.
- Metric-Based Fairness: Involves the operationalization of lawful, clearly defined, and justifiable formal metrics of fairness within the AI system, making them transparently accessible to stakeholders and impacted individuals.
- System Implementation Fairness: Relates to the deployment of AI systems by users who are sufficiently trained to implement them with an understanding of their limitations and strengths. Deployment should be conducted in a bias-aware manner, giving due regard to the unique circumstances of affected individuals.
- Ecosystem Fairness: Addresses the broader economic, legal, cultural, and political structures or institutions that influence the AI project lifecycle. It ensures that these structures do not drive AI research and innovation agendas in ways that entrench or amplify asymmetrical and discriminatory power dynamics, nor generate inequitable outcomes for protected, marginalized, vulnerable, or disadvantaged social groups.
*The paper “AI Fairness in Practice (The Alan Turing Institute)” was the basis for this article.
New Book Release: Beyond the Algorithm: An Attempt to Honor the Human Mind in the Age of Artificial Intelligence (Wittgenstein Reloaded)
“In every whisper of the algorithm, there is an echo of human thought, blurred and distorted, like a philosopher’s dream meandering through the night.”
~ Murat Durmus (Beyond the Algorithm)
