Option B: Error Rates

Introduction to Rawls

John Rawls (1921–2002) was a highly influential American political philosopher, best known for his seminal work, A Theory of Justice (1971). In this book, Rawls sought to address fundamental questions about fairness and the organisation of a just society. His aim was to develop a systematic theory of justice that could serve as a foundation for political institutions and social arrangements, balancing the competing claims of liberty and equality.

Rawls proposed a thought experiment known as the "original position," where rational individuals choose principles of justice from behind a "veil of ignorance." This veil ensures that decision-makers are unaware of their own personal characteristics, such as race, class, or gender, forcing them to consider justice from an impartial perspective. This approach aims to create a framework for fairness, as no one would design rules that unfairly disadvantage a particular group without knowing whether they themselves might belong to that group.

So, let's explore what this theory could tell us about our COMPAS case.

Rawls's Principles of Justice

Applying Rawls to COMPAS

Rawls's approach to justice, particularly through the veil of ignorance and the difference principle, may arguably prioritise equal error rate balance in the COMPAS case, even if this comes at the cost of maximising overall accuracy.

From the perspective of the veil of ignorance, decision-makers evaluating the algorithm would not know their race, gender, or likelihood of being classified as high or low risk. In this position, they would prioritise fairness in how errors are distributed across groups, ensuring that no one group is disproportionately burdened by false positives (being labelled high risk when not dangerous) or false negatives (being labelled low risk when actually dangerous). A system that produces more false positives for one group would likely be rejected as unjust under Rawls’s framework, as individuals behind the veil would seek protection against unequal treatment that could harm them unfairly.

The difference principle reinforces this focus. It demands that inequalities or disadvantages within a system must work to the benefit of the least advantaged. In the context of COMPAS, a higher rate of false positives for certain groups (e.g., Black defendants) disproportionately harms individuals who may already face systemic disadvantages. Rawls’s theory would oppose such an imbalance, as it exacerbates the burdens on the least advantaged rather than alleviating them.

By contrast, maximising overall accuracy may prioritise efficiency at the expense of fairness. If achieving high accuracy involves tolerating disparities in error rates, this would be incompatible with Rawls's principles, as it allows one group to bear a greater share of harm. Rawls’s approach would instead demand that the algorithm ensures equal error rate balance, even if this slightly reduces overall predictive accuracy, as it aligns with the commitment to fairness and the protection of those in vulnerable positions.

Potential Criticisms

1. Possible Trade-Offs with Overall Accuracy: Balancing error rates may reduce the system’s overall predictive accuracy. Critics might argue that prioritising fairness could make the algorithm less reliable overall, potentially compromising public safety if certain high-risk individuals are incorrectly flagged as low-risk. This trade-off could be seen as an ethical tension between individual fairness and collective security.
2. Challenges in Implementation: Adjusting for error rate balance can be technically complex, requiring sophisticated adjustments to the algorithm. Additionally, the process of balancing error rates may inadvertently introduce other forms of bias, complicating efforts to achieve a truly fair system. This raises the question of whether it is feasible to achieve a perfectly balanced system in a way that doesn’t introduce unintended consequences.
3. Tension with Efficiency Goals: A Rawlsian framework’s commitment to fairness could be seen as inefficient by those who prioritise maximised accuracy. Balancing error rates may require that the system occasionally over- or under-flag certain groups to maintain equality, a choice that could be perceived as suboptimal for purely operational or security-focused goals.

Let's quickly recap: