Building AI Companions that Prioritize Learning over Performance

Synthesis: The argument of this position paper is that the problem with generative AI in education is not weak prompting but a design objective: general-purpose models are built and tuned for task completion, so the better they get at helping, the more reliably they absorb the cognitive work through which capability develops. The paper responds not with another set of Guardrails but with a definition and a design framework for AI learning companions, deliberately distinguishing companionship from assistance. It identifies three limits of prompt-level scaffolds that explain why they produce near-null learning effects, derives eight design commitments from them, organizes those commitments into pedagogical, adaptive and responsible design foundations, then reads five real deployments against the framework and finds that the adaptive loop, durable-learning evidence and inclusivity are where current systems fall short.

Key Findings

1. The paper frames the central tension as learning versus performance: systems optimized for immediate task completion can improve output while weakening the cognitive, metacognitive and self-regulatory processes that produce durable capability, a dynamic the authors describe as the learning-performance paradox.
2. Prompt-level remedies are characterized as insufficient for three structural reasons: they are reactive rather than proactive, reaching only learners who already recognize that they need help; they are largely stateless with respect to prior knowledge, goals, confidence and Misconceptions about AI; and they operate at the level of individual tasks rather than across a curriculum or over time.
3. The authors distinguish assistance from companionship: assistance helps with the task at hand, while companionship participates in the learner's development over time. A friendly tone, occasional questions and refusal to give direct answers do not by themselves produce companionship, since such a system can remain reactive, prompt-bound and weakly persistent.
4. The stakes are extended beyond formal education to professional formation: if generative AI absorbs the routine but effortful work through which novices once built expertise, tasks get completed while the developmental experience is bypassed, with recent evidence that workers delegate routine tasks to AI and in some cases replace peer or subordinate interaction with AI assistance.
5. An AI learning companion is defined as a pedagogically grounded, adaptive and responsibly designed AI-mediated learning partner embedded in authentic learning environments and designed to support durable learning across diverse learner populations, with the term offered as a design and relational construct rather than a claim about machine intention or personhood.
6. Eight design commitments follow from that definition: pedagogically grounded, adaptive, responsibly designed, character-based presence, agency-preserving, embedded in authentic learning environments, learning-centered and inclusive.
7. The commitments are operationalized through three foundations, each with four components: a pedagogical foundation (deep and interactive learning, guided Scaffolding, learning to learn and higher-order learning, contextual learning), an adaptive foundation framed as capture, model, adapt and evolve, and a responsible design foundation covering security, transparency, accountability and inclusion.
8. Five case studies written by researchers who led or contributed to the tools are analyzed against the framework: Khanmigo, the RiPPLE student co-creation and peer learning platform, CodeHelp for scaffolded programming support, the JeepyTA teaching assistant for course discussion, and Recast, a university platform for educator-designed dialogic assistants.
9. The cross-case synthesis finds the pedagogical foundation relatively mature but realized through different mechanisms, the adaptive foundation the most uneven with RiPPLE closest to persistent longitudinal learner modeling and CodeHelp and JeepyTA adapting only within sessions, and responsible design shaped by context rather than following one technical pattern.
10. Agency preservation emerges as the strongest shared commitment and inclusivity the least realized: several tools provide scale, privacy, availability or multilingual interaction, but explicit design for diverse learner needs remains limited. Recognisable AI character is fully evident only in Khanmigo and RiPPLE.
11. The research agenda names five priorities: evidence of durable learning rather than proximal indicators such as engagement, helpfulness or task completion; richer persistent learner models including metacognitive development; interaction strategies that preserve agency and productive struggle; responsible design treated as a research object rather than a compliance layer; and inclusion and equity placed at the center of companion design.

Why the definition carries the argument

The paper's analytic move is to refuse the shortcut of calling any educationally prompted model a companion. Its definition is assembled from two prior traditions that it argues were each incomplete: general-purpose generative models, which bring conversational flexibility, scale and cross-domain reach but are organized around assistance, and earlier educational agents and tutoring systems, which contribute companion roles, social presence, dialogue, learner modeling, feedback, adaptivity, and the transparency tradition of open learner models and explainable AI in education, together with algorithmic fairness work in education, but were limited by bounded domains, scripted interaction, authoring cost and system-led models of support that constrained learner agency.

Generative AI, on this account, does not resolve those limitations so much as raise the stakes: because a companion draws on largely opaque training corpora and captures rich open-ended dialogue, what earlier narrow systems could treat as refinements become preconditions for trustworthy use. The eight commitments are assembled as the union of what each tradition lacked, which is why the adaptive foundation is defined as four operations over the learner rather than as a feature list, and why the responsible design foundation appears as a design foundation rather than as an appendix about compliance.

The authors also spend care on what they are not claiming. Companion is a design and relational construct, not a claim that AI possesses care, understanding or the capacity to accompany a person; continuity resides in the design, governance and educational purpose of the system rather than in the model. Nor is the companion meant to substitute for educators and peers, since the same technical system can be deployed either to strengthen educational relationships or to reduce investment in them.

What the case studies expose

Reading five deployments against the framework turns an abstract agenda into a gap analysis. The pedagogical foundation is where tools already converge, and the recurring design decision they identify is which forms of cognitive work remain with the learner: Khanmigo and CodeHelp withhold the direct solution, RiPPLE moves students into authoring and peer review, JeepyTA embeds support inside course discussion and formative feedback, and Recast supports educator-designed assistants for reflection and role-play.

The adaptive foundation is where they diverge most, and persistent longitudinal adaptivity is named as the clearest gap between current implementations and the fuller vision. That diagnosis connects directly to work on Learner Modeling and Adaptive Instruction and Knowledge Tracing, which supply the machinery for exactly the cross-session learner state the case studies lack.

Responsible design appears everywhere but in different sociotechnical forms: conversation visibility and parent or teacher oversight in a K-12 platform, an analytics dashboard reflecting a university position that accountability means knowing how its AI tools are used, instructor configurability and human review in course-bound assistants, and explainable learner modeling with human-gated decisions in the co-creation platform. The paper's point is that these are context-sensitive arrangements, not interchangeable templates.

What this means for practice

• Judge an educational AI by what cognitive work it leaves with the learner, and by whether it can act before the learner asks for help rather than only after.
• Treat persistent learner state as the design target, since in-session and task-local adaptation is the norm and the framework's clearest gap.
• Evaluate against delayed retention, transfer, conceptual change and metacognitive growth instead of engagement, helpfulness or completion metrics.
• Decide the institutional conditions of use, including oversight, data governance and whether the tool strengthens or substitutes for teaching relationships, before deployment rather than after.
• Plan for learners who do not ask for help, and for those who over-rely on it, since proactive initiation and monitoring support are separate capabilities.

Limitations

• The framework is a conceptual synthesis supported by illustrative rather than comparative evaluation; the cross-case ratings are authored judgments by people involved in the tools, not blind or independent assessments.
• Five case studies, concentrated in university and platform deployments led by the authors' networks, cannot represent the diversity of educational AI, particularly non-English and low-resource settings, and the framework's applicability to early childhood, vocational and special education settings is not examined.
• The paper proposes what companions should be measured against but reports no new outcome data on durable learning, so the learning-performance paradox remains supported by other studies rather than tested here.
• The eight commitments are normative and largely unfalsifiable in their current form, since the paper specifies no thresholds at which a system would count as meeting each one, and implementation cost, institutional incentive structures and teacher workloads are named as context but not analyzed in depth, leaving the adoption side of responsible design underdeveloped.

Connected Concepts

• Cognitive Offloading — the mechanism by which performance gains mask capability atrophy
• Agentic AI — companions as persistent, proactive systems rather than answer services
• Framing AI Use for Students — how institutions frame what AI is for shapes what tools get built
• Pedagogical Agent — the earlier tradition the paper absorbs and extends
• Adaptive Learning — the adaptive foundation, and the gap in persistent adaptation
• Learner Modeling and Adaptive Instruction — the learner state a companion must maintain across sessions
• Productive Failure — the effort the design must protect rather than optimize away
• Self-Regulated Learning — planning, monitoring and evaluation as companion design targets
• Metacognition — the paper's proposal for metacognitive onloading
• Transfer of Learning — the durable-learning outcome current systems rarely measure
• Explainable AI — transparency from open learner models, now a precondition
• Equity and Inclusive Learning — the least realized commitment
• Trust Calibration — learner trust as a design responsibility, not a by-product
• AI Governance — institutional purposes and incentives as part of responsible design

Connected Articles

• TeachLM: Post-Training LLMs for Education Using Authentic Learning Data — post-training approaches that make a system persist in a pedagogical role
• EduClaw-Bench: A Long-Horizon Benchmark for Pedagogical LLM Agents with Simulated Learners — what pedagogical agents actually know about pedagogy when tested
• Confirming Correct, Missing the Rest: LLM Tutoring Agents Struggle Where Feedback Matters Most — tutoring agents evaluated against adaptive support criteria
• AI Advice Suppresses People's Willingness to Say "I Don't Know", Even When the Advice Is Wrong and Accuracy Is Incentivized — learners who stop recognising what they do not know when assisted
• Agentic AI in educational environments and its association with social and motivational pathways to learning performance — how agentic AI use relates to motivational pathways

Citation

Khosravi, H., Gašević, D., Sadiq, S., Yan, L., Lodge, J. M., Tangen, J. M., Denny, P., DiCerbo, K., Buckingham Shum, S., & Baker, R. S. (2026). Building AI Companions that Prioritize Learning over Performance. Computers and Education: Artificial Intelligence.