AI Will Not Remake School.

Learning Concepts Will.

A learning concept explains what a learner needs to do for learning to occur. Here, the term covers ideas such as retrieval practice, formative assessment, and project-based learning. These concepts deal with processes including recall, feedback, modelling, transfer, reflection, and independent performance. Stanford’s 2026 review of learning research uses the same underlying concerns, especially metacognition, transfer, and ‘productive struggle’, when examining whether educational activities produce durable knowledge. [1]

Schools often encounter generative AI as a question-writing problem. Teachers are shown how to ask for a lesson plan, an explanation, a rubric, or a student activity. Students are taught to improve prompts so that the system returns a more useful answer. The OECD’s 2026 Digital Education Outlook draws a line between better task performance and actual learning: generative systems can support learning when clear teaching principles guide their use, while outsourcing a task may improve the completed work without producing a corresponding learning gain. [2]

A question tells the system what to produce. A learning concept determines what the learner must practise. The two can appear together, but they do different jobs.

Stanford’s SCALE Initiative examined 818 papers related to AI and K–12 education that were available in its repository by October 2025. Only 20 met its standard for strong causal evidence. Across those studies, students often performed better while support was available, but the gains weakened or disappeared in some assessments completed without it. The review also found more promising results when systems included pedagogical guardrails, such as hints and step-by-step guidance, rather than immediately supplying complete answers. [1]

Consider a Year 9 science standard about relationships within marine ecosystems. A teacher could ask a chatbot to explain food webs, produce a worksheet, and generate ten comprehension questions. The material might be accurate and visually clear. That still leaves the learning process unspecified.

Retrieval practice would require students to recall ecological relationships before receiving help, then compare their thinking with feedback. Cognitive apprenticeship would begin with an expert modelling how a marine biologist interprets field observations, followed by coached practice and reduced support. Career-connected learning would place the standard inside the decisions made by someone monitoring reef health. Project-based learning would organise the work around a sustained ecological problem, constraints, investigation, revision, and a public product.

A 2025 randomised controlled trial at Harvard University provides one example of that distinction. Researchers compared an AI tutor with an in-class active-learning lesson in an undergraduate physics course involving 194 students. The median learning gains in the tutored condition were more than twice those of the classroom condition, while students using the tutor spent a median of 49 minutes on the lesson. [3]

The intervention was more than access to a chatbot. Its designers built active engagement and cognitive-load management into the system prompt. The platform controlled the sequence of multi-part problems, included detailed step-by-step solutions to reduce inaccurate explanations, allowed students to work at their own pace, and delivered targeted feedback. The researchers reported that prompt instructions alone could not reliably provide enough structure, so they designed a platform around the learning process. [3]

The result also has boundaries. The study covered two lessons in undergraduate physics, with students meeting particular material for the first time. The authors did not claim that the same approach would always outperform classroom instruction in work requiring complex synthesis or higher-order critical thinking. They suggested using structured tutoring for introductory material while preserving class time for advanced problem-solving, projects, and group work. [3]

The wider evidence is positive, although it is not consistent enough to support a general rule. A 2025 meta-analysis examined 68 peer-reviewed experimental and quasi-experimental studies and found a moderate positive effect on learning outcomes. It also reported very high variation between the studies, with an I² value of 95 per cent. The effect changed across educational levels, subjects, intervention lengths, and study designs. [4]

This makes “using AI” a weak description of an educational intervention. A general-purpose chatbot providing completed answers, a structured tutor providing graduated hints, a teacher-guided investigation, and a system used to generate practice material expose students to different kinds of intellectual work. They should not be expected to produce the same results.

Learning concepts give schools a more exact way to describe those differences. Under this frame, an AI-supported activity would identify the learning concept, the cognitive work expected from the student, the boundaries placed around assistance, and the evidence that remains when support is removed.

Which learning concepts should schools build around AI, and which ones lose their educational value when the tool is present?

Phil

References

[1] Fesler, L., Martinez Claeys, J. P., Agnew, C., and Loeb, S. The Evidence Base on AI in K–12: A 2026 Review. Stanford University SCALE Initiative, 2026.

[2] OECD. OECD Digital Education Outlook 2026. OECD Publishing, 2026.

[3] Kestin, G., et al. “AI Tutoring Outperforms In-Class Active Learning: An RCT Introducing a Novel Research-Based Design in an Authentic Educational Setting.” Scientific Reports, 2025.

[4] Han, X., Peng, H., and Liu, M. “The Impact of GenAI on Learning Outcomes: A Systematic Review and Meta-Analysis of Experimental Studies.” Educational Research Review, 2025.

Let’s get work.

- Phil