Introduction The dominant paradigm of generative artificial intelligence (AI) in education positions it as an omniscient oracle, a model that risks hindering genuine learning by fostering cognitive offloading. Objective This study proposes a fundamental shift from “AI as Oracle” model to a “Cognitive Mirror” paradigm, which reconceptualizes AI as a teachable novice engineered to reflect the quality of a learner’s explanation. The core innovation is the repurposing of AI safety guardrails as didactic mechanisms to deliberately sculpt AI’s ignorance, creating a “pedagogically useful deficit.” This conceptual shift enables a detailed implementation of the “learning by teaching” principle. Method Within this paradigm, a framework driven by a Teaching Quality Index is introduced. This metric assesses the learner’s explanation and activates an instructional guidance level to modulate the AI’s responses, from feigning confusion to asking clarifying questions. Results Grounded in learning science principles, such as the Protégé Effect and Reflective Practice, this approach positions the AI as a metacognitive partner. It may support a shift from knowledge transfer to knowledge construction, and a re-orientation from answer correctness to explanation quality in the contexts we describe. Conclusion By re-centering human agency, the “Cognitive Mirror” externalizes the learner’s thought processes, making their misconceptions objects of repair. This study discusses the implications on assessment, addresses critical risks, including algorithmic bias, and outlines a research agenda for a symbiotic human-AI coexistence that promotes effortful work at the heart of deep learning.
Tomisu et al. (2025) studied this question.