Ethics by Design for Intelligent and Sustainable Adaptive Systems

AI systems are increasingly dependent on the data and information sources they are developed with. In particular, learning machines are highly exposed to undesirable problems due to biased and incomplete coverage of training data. The autonomy exhibited by machines trained on low-quality data raises an ethical concern, as it may infringe on social rules and security constraints. In this paper, we extensively experiment with a learning framework, called Ethics by Design, which aims to ensure a supervised learning policy that can pursue both the satisfaction of ethical constraints and the optimization of task (i.e., business) accuracy. The results obtained on tasks and datasets confirm the positive impact of the method in ensuring ethical compliance. This paves the way for a large set of industrial applications, whose ethical dimension is critical to increasing the trustworthiness with respect to this technology.