Introduction to The Special Section on Safe AI

Alignment with human values and compliance with ethical and regulatory standards has become a major concern as cutting-edge generative AI-based solutions are becoming ubiquitous and have been increasingly adopted as a productivity tool by many institutions. Despite their remarkable results across several fields, there are still objections regarding embedding AI models, even the vanilla ones, into safe-critical systems, specially, across healthcare, aeronautical and nuclear industries, as many AI methods lack explainability and robustness, and are still prone to adversarial and cyber attacks. The Safe AI field addresses these issues through the proposal of design and development procedures aimed at learning assurance and model alignment, the investigation of interpretable methods with explainable outputs, and the proper treatment of uncertainty. We summarize first the outcomes of a recently launched workshop on Safe AI and present then five selected papers in this special section representing a crosscut of different application areas for AI safety.