Building AI companions that prioritise learning over performance

Texto completo

Generative AI, currently most visible in education through large language models (LLMs), is increasingly embedded in students’ everyday learning practices, supporting tasks such as writing, coding, reasoning, and analysis. Yet its educational value remains uncertain because systems designed to improve immediate task performance may also weaken the cognitive and metacognitive processes that support durable learning. This paper addresses this learning–performance paradox by asking how AI systems can be designed to support learning rather than merely produce better outputs. To address this, in this paper, we make three contributions. First, we synthesise current LLM-based study support with earlier traditions in artificial intelligence in education to develop a contemporary account of AI learning companions systems that are pedagogically grounded, adaptive, trustworthy, embedded in authentic learning environments, and oriented toward durable learning rather than short-term performance. Second, we propose a design framework for building AI learning companions organised around three interrelated foundations: a pedagogical foundation concerned with how students learn with AI, an adaptive foundation concerned with how AI learns about students, and a responsible design foundation concerned with transparency, accountability, privacy, inclusion, and trust. Third, we apply the framework to five case studies across diverse educational contexts, levels, and tool designs, illustrating both the promise and current limitations of companion-like systems. We posit that the central design challenge is not to make LLMs more helpful or productive, but to make them educationally accountable: capable of strengthening learner agency, supporting metacognitive growth, adapting to learner development, and being evaluated by what students can understand, transfer, and do independently after AI support is withdrawn.