Keynote: Paolo Giudici (University of Pavia), SAFE AI evaluation models

The application of Artificial Intelligence (AI) to industry services and products requires AI governance and, in particular, AI risk management models that can balance opportunities with risks. The most important risks of Artificial Intelligence applications can be resumed from the existing regulations and standards such as the European AI Act and ISO AI risk management. They can be summarised in four main risk drivers: i) "Sustainability": the resilience of AI models, and of their ecosystems, under cyber attacks; ii) "Accuracy": the truthfulness of the AI output iii): "Fairness": the inclusiveness and privacy preservation of AI systems; iv): "Explainability": the controllability of AI systems. A valid risk management model must include metrics to measure the above risks, in a consistent manner, to measure the level of risk, in comparison with a set risk threshold. This can, in turn, allow appropriate governance actions, such as risk mitigation or insurance.

To date, there exists very few models that establish operational risk measurement models for AI. We will present a mathematically consistent approach, the S.A.F.E. (Sustainable, Accurate, Fair and Explainable) AI approach, which proposes a set of commensurable statistical metrics, all derived from the same mathematical concept: the Lorenz Curve, developed to study economic inequalities, and the related Gini index. Both can be interpreted resorting to normalised Cramer distances between distributions, which makes them a natural extension of the predictive R^2 metric, applicable to all data types.

The S.A.F.E. AI framework can be applied to any AI model, and to any types of data: tabular, textual, images, multimodal. In the talk we present applications to credit lending, medical images, judicial decisions and cyber risk detection.