Explainability in Artificial Intelligence; towards Responsible AI

Modern Artificial Intelligence (AI) has experienced unprecedented growth over the past decade. These revolutionary technologies are giving new impetus to many application areas. However, the adoption of these techniques is very often limited by the lack of traceability and feedback for experts. Experts are frustrated by this lack of feedback, even though the very implementation of the tool requires them to make a considerable effort to formalize and make available a colossal amount of expertise. Some authors therefore speak of a "black-box evolution", which is undesirable for the traceable, interpretable, explicable and, ultimately, responsible use of these tools.

The need for explanations of how an intelligent system operates is all the greater when the system's performance exceeds – at least in one specialized domain – human capabilities, and this issue has been addressed since the days of expert systems. Recent deep learning systems (Deep Learning – DL) can achieve astonishing levels of performance, and their large number of parameters makes it all the more difficult to understand the solutions they arrive at, even if these parameters are all accessible. However, the topicality of the subject of explainability for intelligent systems stems less from real breakthroughs – still awaited – in the resolution of this problem than from the legal novelty – which is being imposed in particular on AI players – that constitutes the inclusion in the General Data Protection Regulation (RGPD, European regulation) of explanation obligations for the automatic processing of personal data. We therefore adopt a traditional technological approach to a highly mobile field and propose, alongside examples, a conceptual framework guiding the practitioner's approach to finding solutions.