Dr Hassan Khosravi is the lead author of the new paper Building AI companions that prioritise learning over performance, which explores how educational AI can be designed to strengthen long-term learning rather than simply improve task completion.
The research addresses a growing challenge as large language models and other generative AI tools become increasingly embedded in students’ everyday study practices.
At the centre of the paper is what the authors call the learning–performance paradox.
General-purpose AI systems are largely designed around productivity. They can reduce effort, speed up tasks and help users produce higher-quality outputs. Those qualities can be highly valuable in workplaces, but education presents a different challenge.
Some of the effort that AI is designed to remove — thinking through a difficult problem, retrieving information, testing ideas, reflecting on mistakes and monitoring understanding — is also part of the process through which people learn.
As the paper explains, students may therefore perform better while using AI without necessarily developing the knowledge and skills needed to perform independently later. AI-supported performance can mask over-reliance, reduced self-regulation and what has been described as “metacognitive laziness”.
The key question is therefore not simply whether AI makes a student’s work better, but whether the student is learning more because of the interaction.
To address this problem, the researchers propose a different way of thinking about educational AI: the AI learning companion.
The paper defines an AI learning companion as a pedagogically grounded, adaptive and responsibly designed AI-mediated learning partner that is embedded within authentic learning environments and designed to support durable learning across different groups of learners.
This distinction is important.
A conventional AI assistant is typically designed to respond to a request as efficiently as possible. A learning companion would instead be designed to support the student’s development over time.
Rather than immediately providing an answer, for example, an effective learning companion might ask a student to explain their reasoning, attempt the problem first or retrieve something they have previously learned. It could then provide a hint when needed, gradually reduce that support as the student’s ability develops, and encourage the student to reflect on how confident they are in their own understanding.
The goal is not to make learning unnecessarily difficult. It is to preserve the productive cognitive effort that helps students build knowledge they can later use without AI.
The researchers propose a framework built around three interconnected foundations.
The first is the pedagogical foundation, which focuses on how students learn with AI. AI learning companions should be grounded in established principles of learning, including deep and interactive learning, appropriate scaffolding, metacognition and learning within meaningful contexts. Instead of simply producing information, an AI companion should encourage students to retrieve knowledge, explain concepts, justify their reasoning and apply what they have learned.
The second is the adaptive foundation, which examines how AI learns about students. An effective companion would develop an understanding of a student’s existing knowledge, misconceptions, goals, confidence, learning strategies and progress, allowing support to become more personalised over time. Importantly, adaptivity should not mean the AI simply taking control of a student’s learning. Students should remain involved in decisions about goals, strategies and support so they continue to develop the ability to regulate their own learning.
The third is the responsible design foundation, which focuses on how AI acts with integrity. The framework emphasises transparency, accountability, privacy, security, inclusion and human oversight.
The paper also identifies eight design commitments for AI learning companions. These include being pedagogically grounded, adaptive, responsibly designed, agency-preserving, learning-centred, inclusive and embedded in authentic learning environments, while also providing a recognisable pedagogical presence for the learner.
A particularly important principle is preserving learner agency.
The researchers do not propose AI as a replacement for educators or peers. Instead, AI should extend the support available within human-led learning environments by providing additional opportunities for practice, reflection and guidance while keeping the learner responsible for thinking and decision-making.
The paper applies its framework to five existing educational AI systems to examine how elements of the learning-companion approach are already emerging in practice.
The analysis also highlights areas where further work is needed, including more persistent and meaningful adaptivity, inclusive design and stronger evidence that AI-supported learning continues after the technology is removed.
Ultimately, the paper proposes a different benchmark for evaluating AI in education.
The challenge is not simply to create AI that is more helpful, faster or better at producing answers. Instead, educational AI should be judged by whether it strengthens learner agency, metacognitive development and durable understanding — including what students are able to understand, transfer and accomplish independently after AI support is withdrawn. As AI becomes an increasingly common part of education, the research shifts the focus from asking “What can AI do for the student?” to asking “What can the student learn to do because of their interaction with AI?
