This paper asks how generative AI tools, which students now use for writing, coding, reasoning and analysis, can be designed to support learning instead of only producing better outputs. The authors call this the learning-performance paradox, since systems that improve immediate task performance may weaken the cognitive and metacognitive processes behind lasting learning. The paper reports no empirical results. It builds a design framework by combining current LLM-based study support with earlier work in artificial intelligence in education, then applies that framework to 5 case studies across different educational contexts, levels and tool designs.
Its main claims are these:
- AI learning companions should be pedagogically grounded, adaptive, trustworthy, embedded in authentic learning settings and aimed at durable learning over short-term performance.
- A pedagogical foundation should address how students learn with AI.
- An adaptive foundation should address how the AI learns about the students it works with.
- A responsible design foundation should cover transparency, accountability, privacy, inclusion and trust.
- Applying the framework to the case studies shows both the promise of companion-like systems and their current limitations.
The authors conclude that the central design task is making LLMs educationally accountable. In their account, such systems would strengthen learner agency, support metacognitive growth and adapt as learners develop. They'd also be judged by what students can understand and apply on their own once AI support is withdrawn. The abstract doesn't say which limitations the case studies revealed.
