Towards the Formalization of a Trustworthy AI for Mining Interpretable Models explOiting Sophisticated Algorithms

Interpretability is foundational to trustworthy AI: The paper argues that models designed to be interpretable from the start—rather than as an afterthought—are essential for building trust and enabling safe real-world deployment of automated decision-making systems, particularly in high-stakes applications.

Three types of interpretable models offer different benefits: The MIMOSA framework identifies feature importance models, rule-based models, and instance-based models as the major families of interpretable approaches, each with distinct strengths for explaining how AI systems reach their decisions.

Ethical properties must be formally embedded, not added later: Rather than treating fairness, privacy, and causal reasoning as separate concerns, the framework proposes that these ethical properties should be formally defined and built directly into the model development process from the beginning.

Trade-offs between interpretability and ethics require deliberate choices: The paper acknowledges that balancing interpretability with performance, fairness, privacy, and causal validity involves inherent tensions that developers must navigate thoughtfully rather than assuming all goals can be simultaneously maximized.

Trustworthy AI requires systematic evaluation across multiple dimensions: The framework establishes formal metrics and verification procedures for assessing not just accuracy, but also interpretability, fairness, privacy protection, and causal validity—suggesting that responsible AI development demands comprehensive, multi-faceted evaluation rather than single-metric optimization.