Programming Cognition with Language: Toward a Dialogic Architecture for AI-Mediated Learning

Abstract

Generative artificial intelligence is becoming a cognitive infrastructure in education, reshaping how problems are framed, explored, and evaluated across disciplines. While these developments create learning opportunities, they also raise concerns about epistemic automation and cognitive offloading in AI-mediated environments. Current educational responses often focus on tool adoption or prompt-engineering skills, which remain insufficient to address regulatory challenges of large language models.This paper reconceptualizes Natural Language Programming as cognitive and metacognitive programming: the design and regulation of reasoning through language-mediated interaction with generative systems. To operationalize this perspective, we propose the Minimal Dialogic Cognitive Programming Architecture (MDCPA), a domain-independent framework composed of five dialogic cognitive primitives organized as a recursive regulatory cycle.Prompt-Based Learning is introduced as the pedagogical instantiation of this architecture, reframing prompts as programs of thought and providing a structured approach to teaching and assessing AI-mediated reasoning. The framework applies across analytical and interpretative knowledge practices, supporting relevance in multidisciplinary contexts.Illustrative observations from a Master’s course suggest that students engage in co-reasoning with generative systems, while their ability to regulate reasoning varies. These observations motivate the need for pedagogical and assessment approaches capable of capturing AI-mediated reasoning and support interpreting digital competence as regulatory capacity.By positioning natural language as a medium for programming cognition rather than querying systems, this work contributes a framework for sustaining epistemic control in AI-mediated education.