As machine learning (ML) systems become increasingly embedded in areas such as healthcare, education, hiring,
and criminal justice, concerns about fairness and bias have intensified. This paper explores what it means for an
algorithm to be fair, focusing on the concept of justice and how it can guide the design and evaluation of ML
systems. Drawing insights from social and political philosophy, particularly theories of distributive justice and
equality of opportunity, the paper examines the …
Read moreAs machine learning (ML) systems become increasingly embedded in areas such as healthcare, education, hiring,
and criminal justice, concerns about fairness and bias have intensified. This paper explores what it means for an
algorithm to be fair, focusing on the concept of justice and how it can guide the design and evaluation of ML
systems. Drawing insights from social and political philosophy, particularly theories of distributive justice and
equality of opportunity, the paper examines the strengths and limitations of common fairness metrics, such as
demographic parity and equalised odds. These metrics, while useful, are normatively incomplete, as they
presuppose contested moral assumptions and cannot, on their own, determine what justice requires in algorithmic
decision-making. The paper argues for a deeper understanding of fairness that goes beyond technical solutions;
one that considers the ethical, social, and systemic dimensions of justice. By connecting the bioethical principle of
justice to fairness in ML, the paper proposes a more comprehensive framework to better navigate the complexities
of fairness-aware learning, while ensuring greater accountability and transparency in AI systems. The aim is to
bridge the gap between technical approaches to fairness and their real-world impacts, encouraging interdisciplinary
collaboration to shape more equitable and responsible AI technologies. The paper therefore defends the claim that
algorithmic fairness must be explicitly grounded in substantive theories of justice if it is to address real-world
algorithmic harms in ethically defensible ways.