Deontological Ethics By Monotonicity Shape Constraints

AI systems easily violate basic ethical principles: Modern machine learning models can readily violate common deontological ethical rules like "favor the less fortunate" and "do not penalize good attributes," demonstrating that ethical considerations are not automatically built into AI systems.

Shape constraints offer a practical solution: The paper proposes adding "shape constraints" to machine-learned models that force them to respond only positively to relevant inputs, providing a concrete technical method to embed ethical principles directly into AI systems.

Ethical and fairness goals can work together: The research shows that deontological constraints (rules about individual rights and duties) can be compatible with consequentialist fairness goals like equal opportunity and statistical parity, suggesting these different ethical frameworks don't necessarily conflict.

Works with real-world attributes: The approach is practical for sensitive attributes commonly used in high-stakes decisions, including both Boolean attributes (yes/no categories) and continuous ones like income and age, making it applicable to real AI systems used in lending, hiring, and other domains.

Contributes to trustworthy AI development: By demonstrating how to systematically incorporate ethical principles into machine learning models, this work supports the broader goal of building AI systems that are more responsible, transparent, and aligned with human values.