From the Introduction:
Generative AI tools such as GPT-4 and DeepSeek R1 have moved rapidly from experimental novelties to everyday academic companions. A recent global survey reports that 86% of university students now use AI in their studies, with more than half engaging with these tools weekly, primarily to summarise documents, check grammar, paraphrase, and generate first drafts (Digital Education Council, 2024). While often perceived as convenient (though this perception is contested; see Selwyn, 2025), these tools are not designed with educational goals in mind. Large Language Models (LLMs) generate responses using probabilistic language modeling, predicting likely word sequences from training data, rather than through conceptual understanding or reasoning. This distinction has significant pedagogical implications. Without careful integration, widespread adoption risks cognitive offloading, diminished metacognitive engagement, and weakened epistemic agency, which we define as the learner’s capacity to critically evaluate, justify, and take ownership of knowledge.
This paper advances the position that LLMs are not educationally neutral; their effects are contingent rather than fixed. While emerging research has documented both beneficial and harmful outcomes (e.g., Deng et al., 2024; Gerlich, 2025), their ultimate impact depends on how they are designed, including what they afford, obscure, or prioritise, and on the pedagogical context in which they are implemented. These factors jointly influence how they are used, what forms of learning they support or constrain, and whose epistemic values they reflect or exclude. In response to this complexity, we propose a normative, design-oriented pedagogical framework for the intentional integration of LLMs into higher education, with the specific goal of fostering critical thinking. Rather than banning or embracing these tools wholesale, we argue that educators must cultivate learning environments in which students engage critically with AI, using it to extend their reasoning rather than to replace it.
https://www.sciencedirect.com/science/article/pii/S2666920X26000342
