Toward Personalized Explainable AI

点击翻译按钮获取中文摘要

The AI community is increasingly interested in understanding how to build artifacts that, in addition to performing useful tasks, are well accepted and trusted by their users. It is undeniable that the explainability of an AI system can be an important factor for acceptance and trust. However, there is still limited understanding of the actual relationship between explainability, acceptance, and trust and which factors might impact this relationship. In particular, although existing research on Explainable AI (XAI) suggests that having AI systems explain their inner workings to their end users can help foster transparency, interpretability, and trust. there are also results suggesting that such explanations are not always wanted by or beneficial for all users. These results indicate that research in XAI needs to go beyond one-size-fits-all ex-planations and investigate AI systems that can personalize explanations of their behaviors to the user's specific needs. There is general agreement that such needs may depend on context, e.g., the type of AI application and criticality of the targeted tasks, but there is also evidence that, given the same context, user differences play a role in defining when and how explanations may be useful and effective. These results call for the need to investigate personalized XAI, namely how to create AI systems that understand to whom, when, and how to deliver effective explanations of their actions and decisions. This is the objective of this proposal. AI-driven personalization has been an active field of research for several decades, spanning fields such as recommender systems, intelligent-tutoring systems, conversational agents, and affect-aware systems. To provide personalization, an AI system needs to have an adaptive loop in which it acquires a model of its user by inferring relevant user properties from available observations and decides how to personalize its behavior accordingly, to favor at best the goal of the interaction In this proposal, we frame explanations as yet another element of personalization in the adaptive loop, where the system ascertains if and how to explain its behavior to the user based on its best understanding of user properties specifically relevant to evaluate the need for explanation. What these relevant properties are, how an AI system can assess them and how it can respond with adequate personalization of explanations is all still largely unknown. The proposed research aims to contribute to filling these gaps