The source offers a model, not classroom evidence
The first Edmetry Weekend Long Read points to Luke Rowe's From Gradual Release of Responsibility to Gradual Release of Technology: A Case for the Staged Use of AI in Formal Education, published in Education Sciences on 13 August 2026. The date is confirmed in the journal's version record. The full text is freely available, and the source takes approximately 30–35 minutes to read.
It is a conceptual article, not an intervention study. That distinction matters. Rowe offers a useful vocabulary for making better decisions about AI in learning, while leaving the model itself open to empirical testing.
The useful move is to separate performance from learning
Education debates often ask whether AI is good or bad for students. That is an understandable question, but it is too blunt for the classroom. An AI system can help a learner produce a stronger answer today while making it harder to tell what the learner can do tomorrow without that help.
Rowe's argument begins with this difference between AI-assisted performance and durable learning. The question is not simply whether a tool improves an output. It is whether the learner is building the knowledge, skill, judgement, and understanding needed to use the tool without surrendering the important thinking.
That reframes AI access as an instructional decision. The right level of support depends on the task, the learning goal, and what the learner has demonstrated—not only on what the product makes possible.
Three stages for a less binary policy
The proposed model, called the Gradual Release of AI Technology, extends the familiar gradual release of responsibility. The original model asks who is doing the cognitive work. This extension asks when and how AI should be allowed to share, support, or take on part of that work.
- Independent cognition. Foundational knowledge and skills are developed with AI substitution deliberately restricted. The point is not technological purity. It is to make sure the learner has an opportunity to form the underlying understanding.
- Shared cognition. AI enters through constrained, teacher-guided interaction. Learners can question, compare, critique, and revise with the tool, while the teacher sets the boundaries and keeps the learning goal in view.
- Offloaded and reallocated cognition. Strategic delegation becomes more appropriate once learners can evaluate outputs, recognize limits, and retain responsibility for the judgement that matters. AI can take on selected work, but the learner remains accountable for the result.
The stages are overlapping rather than age bands. Rowe proposes progression based on demonstrated knowledge, skill proficiency, epistemic understanding, and AI literacy. That is a more useful design signal than age, elapsed time, or enthusiasm for the technology, although the proposal still needs testing.
What this asks of an EdTech product
Most AI products expose a permission switch: available or unavailable. A staged model suggests a richer contract.
The product should be able to make the learning condition visible. Was the learner working independently? Was AI available only for hints? Did the teacher approve a particular use? Which parts of the task were delegated, and which judgement remained with the learner?
This does not mean turning classroom work into surveillance. It means recording only the context needed to interpret evidence, with a clear audience and a clear retention boundary. A result produced with extensive assistance should not be presented as if it were evidence of independent mastery.
The practical consequence is a separation between the learner's work and the conditions under which that work was produced. In a teacher-led system, that could support three better decisions:
- keep foundational or diagnostic tasks AI-free when independent evidence is the goal;
- allow bounded scaffolding when the purpose is to help a learner reach the next step; and
- permit selective offloading when the learner can explain, evaluate, and take responsibility for the result.
The system should support those decisions without pretending that a usage log is a measure of learning. Evidence of assistance is context. It is not, by itself, evidence of understanding.
Use the model for a bounded pilot, not policy
GRAIT is a strong design hypothesis and a useful vocabulary, but it is not settled evidence. The article is conceptual and the framework awaits empirical validation. It should therefore inform product questions and classroom experiments, not become a universal policy or an automated gate that decides what every learner is ready to do.
That restraint is what makes the piece worth reading. It replaces the exhausted allow-or-ban argument with a more demanding question: what kind of thinking is this task meant to develop, what support is appropriate now, and what would demonstrate readiness for a different kind of support later?
For now, school leaders should treat GRAIT as a prompt for a bounded, teacher-led pilot rather than a policy template. Choose one learning goal, define one allowed AI role, preserve one independent check of understanding, and record what would justify expanding, changing, or stopping the approach.
For EdTech, that is a better starting point than adding another AI feature. The product decision comes first. The model follows.