Aligning machines with human values: opacity and ethical sustainability of AI

The increasingly rapid development of deep learning algorithms is radically transforming society. The potential of neural models is paving the way for the automation of areas previously unrelated to AI, from medical diagnostics to therapeutics, from mobility to justice and distribution of resources. However, the widespread diffusion of new AI tools carries risks. Indeed, neural models, which are increasingly entrusted with autonomous decision-making, are often opaque. Opacity entails major ethical issues: how can we trust the machine's decisions if we do not know how it works? How can we guarantee the fairness of the algorithm, given its increasingly widespread implementation in areas such as justice, social welfare and credit, if we do not understand its ‘reasoning’? To answer these questions, strategies of representational alignment will be proposed. Through such strategies, researchers intend to adapt the representational content emerging from neural models to human values and beliefs. In this perspective article, we will argue that representational alignment is a strategic tool, and more effective than XAI approaches, to ensure the ethical sustainability of the machine, avoiding the risk of bias towards social or cultural minorities, thus optimising the architecture in an ethically oriented manner.