Explainable artificial intelligence as AI literacy scaffolding
TL;DR
It is argued that AI explanations can function as a scaffold that supports users in learning not only about specific outputs but also about model behavior, limitations, and potential bias.
Abstract
This conceptual paper responds to the growing gap in artificial intelligence (AI) literacy among users of large language models (LLMs) and other AI systems in recent years. While explainable AI (XAI) has largely been framed as a mechanism for transparency and accountability, its potential to be used as a developmental learning support for end users remains underexplored. The purpose of this article is to conceptualize explainability as a form of literacy scaffolding that supports the gradual, in situ development of AI literacy skills through explanation-driven interaction. A conceptual framework is developed through an integrative analysis of literature from explainable AI, AI literacy, educational scaffolding theory and human–AI teaming research. Key concepts are defined, and theoretical parallels are drawn between scaffolding processes in education and explanation-driven interaction with AI systems. We argue that AI explanations can function as a scaffold that supports users in learning not only about specific outputs but also about model behavior, limitations, and potential bias. This framework is organized around three core scaffolding principles: the contingent support calibrated to the learner's Zone of Proximal Development, progressive fading of explanatory support as competence grows, and the reframing of explanations as entry points for inquiry rather than disclosure endpoints. These principles collectively support the development of calibrated trust among users rather than uncritical acceptance or unwarranted skepticism toward AI models. This article offers a theoretical framework for understanding explainability as AI literacy scaffolding and outlines implications for the design and evaluation of XAI systems. It further proposes methodological approaches for empirically identifying and measuring scaffolding behavior in user interactions with XAI. By positioning explainability as a mechanism for fostering durable AI literacy and calibrated trust, this work contributes to information science scholarship on human–AI interaction and responsible AI use.