In AI FAIRNESS, Dr. Derek Leben Proposes a Theory of Algorithmic Justice

With the ever-growing presence of artificial intelligence (AI), algorithms are the invisible hands shaping decisions about housing, loans, healthcare, employment, and even criminal justice. As AI systems become more integrated with work and life, a fundamental question arises: Are these systems fair?

Derek Leben, in his new book AI Fairness: Designing Equal Opportunity Algorithms (MIT Press, 2025), tackles this pressing issue, offering a philosophical framework to evaluate and mitigate the inherent biases of AI. Leben draws inspiration from the work of the philosopher John Rawls, proposing a theory of algorithmic justice built upon core principles including autonomy, equal treatment, and equal impact. These principles, he argues, should guide the design and deployment of AI systems, ensuring they meet a "minimally acceptable level of accuracy," avoid irrelevant attributes, and provide equal opportunity.

The book explores the challenges of measuring fairness in AI algorithms, using case studies like the Apple Card and the COMPAS criminal risk assessment tool to illustrate the problem of choosing appropriate metrics. These metrics are often incompatible with each other, and may also produce trade-offs in performance. Companies must decide which fairness measurements to use, and how they will mitigate AI systems to satisfy ethical demands for fair treatment and impact.

At a deeper level, big data and machine learning also create more fundamental questions about what counts as a “protected” feature and why. Beyond just legally protected categories like age, race, gender, and disability, we now need to ask whether other features like how long a person charges their cell phone at night and their parent’s highest level of education should or should not be used in making predictions about that person’s behavior.

Leben's analysis goes beyond abstract principles, acknowledging the importance of performance and efficiency in AI development. Moreover, the book addresses complex issues like algorithmic affirmative action, the trade-off between fairness and accuracy, and ethical considerations in algorithmic pricing.

In one section, Leben addresses the challenges of image generators built by companies like OpenAI and Google. In Google’s case, they mitigated their image generator to be more fair with respect to race and gender, but it produced absurd results. Many companies might interpret this story as a cautionary tale about fairness interventions. However, Leben writes, “The problem was not that Google and OpenAI used fairness mitigations on their genAI systems, it was that they used the wrong ones.”

In AI Fairness, Derek Leben provides a vital contribution to the ongoing conversation about the ethics of artificial intelligence. His work serves as a guide for navigating the complex terrain of algorithmic bias, urging us to build a future where AI reflects our aspirations for a just society.

AI Fairness: Designing Equal Opportunity Algorithms is available on May 13, 2025, from MIT Press or wherever books are sold.