Monotonicity for AI ethics and society: An empirical study of the monotonic neural additive model in criminology, education, health care, and finance

Monotonicity is a critical but overlooked fairness principle: The paper demonstrates that AI systems should follow logical monotonic relationships (e.g., more education should not decrease job prospects), and violations of this principle across criminology, education, health care, and finance can lead to "disastrous consequences" for individuals and society.

Standard fairness approaches are incomplete: While the machine learning community has focused heavily on statistical parity and discrimination issues, this research shows that monotonicity represents an essential but separate facet of fairness that requires explicit attention in algorithm design.

Monotonic Neural Additive Models (MNAMs) offer a practical solution: The study provides empirical evidence that MNAMs—a fairness-aware machine learning algorithm—can effectively enforce monotonicity constraints while maintaining model performance across multiple high-stakes domains.

Real-world harms span multiple sectors: The research documents specific fairness violations across criminology (criminal risk assessment), education, health care, and finance, suggesting that monotonicity failures are a widespread problem affecting vulnerable populations in critical decision-making systems.

Explainability and fairness are interconnected: By combining explainable AI (XAI) principles with fairness requirements, the paper demonstrates that more transparent AI systems can be designed to follow logical, monotonic relationships that align with human values and domain expertise.