Measuring justice in machine learning

Rawlsian fairness frameworks in ML may inadvertently encode bias against people with disabilities. The paper argues that capability theorists' long-standing philosophical critique of John Rawls's theory of distributive justice—that it uses the wrong measure of justice—applies directly to machine learning systems that have adopted Rawlsian principles for fairness.

Current fair ML approaches may be measuring the wrong thing. The paper suggests that by operationalizing Rawls's theory in machine learning without addressing capability theorists' objections, the field risks systematically encoding the same biases that philosophers have identified in Rawls's original framework.

There is an alternative philosophical framework available. Capability theorists offer a different measure of justice beyond Rawls's approach, which the paper suggests fair ML should consider adopting to build more genuinely just systems.

Defining justice in AI requires both philosophical rigor and practical measurement. The paper emphasizes that building more just ML systems requires not only choosing the right definition of justice but also ensuring that definition can be properly measured and implemented in technical systems.

Fairness in ML is not a purely technical problem. The paper demonstrates that choices about how to measure fairness in machine learning systems are fundamentally philosophical choices with real consequences for which groups benefit and which groups are harmed.