AI as a cognitive partner in learning: Emerging epistemic practices of reflective thought

Synthesis: Dalsgaard and Prilop (2026) ask what students actually do with Generative AI when it is treated as a cognitive partner rather than an answer machine, and whether those uses produce reflective thought or replace it. Their evidence comes from a one-year action and design-based research project across eight Danish upper secondary schools, where 59 teachers designed their own AI activities without being told how to use the tools, and 48 students were interviewed in eight group sessions. The analysis yields four practices — explanation, creation, inspiration, improvement — unfolding into eleven more specific processes, each of which can either substitute for thinking or extend it. The authors' conclusion is pedagogical rather than technological: the answer is not to ban AI but to cultivate students' ability to use it as an instrument for examination, construction, and reflection.

Key Findings

1. Four practices organized student use: explanation (using a chatbot to get an overview of a text, clarify concepts, or learn a method), creation (generating examples, structure, adaptations, or results), inspiration (for content and for expanding an idea), and improvement (identifying shortcomings and suggesting fixes).
2. Those four practices break down into eleven distinct processes, which the authors treat as the analyzable unit — the same practice, such as asking a chatbot to summarize a reading, can be outsourcing or comprehension repair depending on how the student uses it.
3. Students used explanation to repair comprehension rather than only to avoid reading: one student pastes academic prose they cannot process and asks for it in plainer terms, which the authors read as extending epistemic practice rather than replacing it.
4. The design gave teachers no prescribed activities or AI training, only the framing concepts of cognitive partnership and epistemic practices, so what emerged reflects what teachers invented under a conceptual frame rather than a scripted intervention.
5. Reflection appeared in both theoretical and empirical registers — students reflecting on disciplinary content and on the process of working with AI itself.
6. The authors frame the outcome as a hybrid mind: cognition distributed across student and tool, which they argue requires new epistemic practices that current schooling does not teach.
7. They position the study as complementing rather than replacing substitution research, calling for attention to what AI makes newly possible alongside what it replaces.
8. The study is a two-cycle design: four group interviews at four schools after the first cycle (December 2024) and four more at the other four schools after the second (March 2025), with teacher action plans revised in between.

The practices, and the line between extending and replacing

The value of the four-practice scheme is that it does not treat AI use as a single behavior to be approved or prohibited. Explanation can be reading avoidance or comprehension scaffolding; creation can be delegation of the work or construction of a worked example to think with; inspiration can substitute for having ideas or seed them; improvement can be outsourcing judgment or sharpening it. The paper's contribution is to show students moving along both sides of that line within the same practice, and to argue that the pedagogical task is to push use toward the reflective side rather than to remove the tool. This connects the study to Distributed Cognition and to debates about Cognitive Offloading, where the same ambiguity — a tool that supports performance while weakening the capacity it substitutes for — has been documented in writing and problem solving.

What this means for practice

• Instructors. Design for the process, not the prohibition. The practices identified here were not taught as procedures; they emerged when teachers framed AI as a partner for examination, construction, and reflection rather than as a source of answers (AI Literacy).
• Faculty developers. The intervention's light touch is worth noting: teachers received concepts, not activities, and developed their own plans across two design cycles. That suggests professional learning about AI can proceed through a shared vocabulary plus room to invent.
• Instructors assessing learning. Watch which side of a practice a student is on. The same prompt — summarize this text, generate an example — appears in both reflective and substitutive use, so the artifact alone does not show what happened.
• Researchers. The study's unit of analysis, the process within a practice, is a usable frame for studying Human AI Collaboration qualitatively rather than measuring adoption or attitudes.

Limitations

• A qualitative design with a small, self-selected sample. Eight group interviews with 48 students, selected by their teachers, in one national context and one school type; the practices are interpretive themes, not a distribution of how often each occurs.
• Teachers designed the interventions. Because teachers were given concepts rather than activities, what students did depends on choices the study did not control, so the practices reflect this project's framing rather than a replicable intervention.
• Interviews report practice, they do not measure learning. Students describe how they used AI; the study does not test whether reflective use produced better or worse learning outcomes than substitution.
• The authors used an AI tool in preparing the article, disclosing that ChatGPT 5 was used to improve readability and language, with the authors taking responsibility for the content — worth noting given the paper's subject.

Citation

Dalsgaard, C., & Prilop, C. N. (2026). AI as a cognitive partner in learning: Emerging epistemic practices of reflective thought. Journal of Research on Technology in Education.