Better Grades, Weaker Skills? New Framework Rethinks AI in Education

A new education framework argues that the most important question about artificial intelligence (AI) in schools is no longer whether students should use it, but when they are ready to do so without weakening the learning process. In "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, Luke Rowe of the Institute for Learning Sciences & Teacher Education at Australian Catholic University proposes the Gradual Release of AI Technology, or GRAIT, model.

The framework is designed primarily for K–12 education and separates AI-supported performance from durable learning. Rather than treating AI as inherently beneficial or harmful, the author argues that its educational value depends on when it enters the learning process, what cognitive work it performs, and whether students are sufficiently prepared to evaluate its output.

Better Work Is Not Always Better Learning

The paper insists that educational success cannot be measured only by the quality of work produced while AI is available. Learning involves relatively durable changes in knowledge and skills that support later reasoning, retention and transfer, while performance describes what a learner can produce during or immediately after instruction. Generative AI can improve speed, fluency and apparent competence while potentially bypassing some of the cognitive effort through which durable understanding develops.

A large field experiment involving nearly 1,000 secondary mathematics students found that standard GPT-4 increased performance during assisted practice by 48 percent, yet was followed by a 17 percent decline from baseline when students later completed an unassisted assessment. A guarded GPT tutor delivered an even larger 127 percent practice improvement, but that performance advantage disappeared when the AI was removed.

The concern is not that cognitive offloading is always harmful. People routinely rely on calculators, search engines and other technologies to reduce unnecessary mental effort, and such tools can be valuable when they free cognitive capacity for more sophisticated work. The educational risk arises when AI performs the very reasoning, retrieval, generation or evaluation that the student is supposed to be learning.

This can create what the paper describes through established learning science as an illusion of competence. Students may interpret the ease and fluency created by technological assistance as evidence that they understand the underlying material, even when their performance depends heavily on support that will later disappear. In an AI-rich classroom, the polished product can therefore become a poor proxy for the learner's actual capability.

Build the Skill Before You Automate It

GRAIT responds by organising AI use into three overlapping stages, beginning with Independent Cognition. At this point, AI should be restricted where it would substitute for the foundational processes needed to develop knowledge, procedural fluency and reliable judgement. Students should retrieve, explain, apply and solve without AI so educators can determine whether competence belongs to the learner rather than the technology.

This first stage is not a blanket prohibition on technology. Rowe explicitly distinguishes AI that performs targeted cognitive work from accessibility technologies, such as screen readers or voice interfaces, that may support equitable participation without doing the learning task on a student's behalf. The principle is narrower: before intellectual work is delegated, students should demonstrate that they can perform the relevant activity independently.

It carries potentially significant consequences for assessment. If an institution cannot establish whether an answer reflects a student's knowledge, AI assistance or an inseparable combination of both, it becomes harder to validate the learning represented by a grade or qualification. GRAIT therefore supports AI-free assessment at points where educators need to establish underlying knowledge and skills independently of technological support.

The readiness principle also challenges age-based or institution-wide AI rules. Under the framework, progression is determined by mastery of specific knowledge and skills, epistemic understanding and, where relevant, sufficient AI literacy rather than by age, technological access or student preference.

Turn AI From an Oracle Into a Coach

The second stage, Shared Cognition, introduces AI under more structured and teacher-regulated conditions. Instead of functioning as an answer machine, AI can operate as a coach, critic or creative partner by offering hints, challenging arguments, suggesting alternative explanations or providing feedback on work that students have already generated. The learner remains responsible for questioning, refining and ultimately judging the machine's contribution.

Knowing how to write effective prompts is insufficient if students cannot recognise uncertainty, verify claims against disciplinary knowledge, identify plausible errors or decide when human judgement should override an automated answer. Under GRAIT, technological fluency without evaluative competence does not qualify a learner for unrestricted AI use.

The third stage, Offloaded and Reallocated Cognition, permits greater delegation once the relevant knowledge or skill has been mastered. Students might use AI to organise disparate information, generate a visual or handle routine components while reallocating their attention toward synthesis, evaluation and refinement. The offloaded activity should not itself be the central learning objective, and students must remain capable of critiquing and overriding AI output.

This is where the model becomes more than a case for restricting AI. It recognises that intelligent systems can genuinely expand productivity and learning opportunities when students possess enough expertise to supervise them. The goal is not to preserve every manual cognitive task indefinitely, but to ensure automation follows competence rather than replacing the process through which competence is acquired.

AI Policy Now Needs a Theory of Timing

The broader significance of GRAIT lies in moving education beyond the crude choice between banning AI and embracing it everywhere. It proposes that schools think about sequencing: which capabilities students should first establish independently, where guided human-AI collaboration adds value, and which mastered activities can eventually be automated. The approach shifts the policy conversation from access alone toward learning design.

For governments and education systems, such an approach could demand more granular curriculum and assessment policies. Instead of declaring AI either permitted or prohibited across entire institutions, educators could distinguish between tasks where independent competence must be demonstrated, tasks where AI can provide structured scaffolding and tasks where automation is educationally appropriate. Such differentiation may become increasingly important as generative AI becomes embedded in mainstream digital tools.

The implications may be particularly relevant to developing countries, where governments face pressure to prepare students for AI-enabled economies while working with uneven infrastructure, large class sizes and varying levels of teacher capacity. GRAIT offers a useful principle, technology adoption should be judged by whether it strengthens learning, but the paper does not test how the model would function in low-resource education systems. Its practical feasibility across those settings remains an open research question.

The author also makes clear that GRAIT itself remains provisional. It is a conceptual framework rather than an experimentally validated intervention, and it requires empirical testing across educational levels, subjects, tasks and learner populations. The model is primarily intended for teacher-led formal learning environments, particularly K–12 classrooms, while implementation in large university courses or other settings may be more difficult.

The limitation points directly to the next research agenda. Researchers need longitudinal and randomized studies comparing unrestricted AI use, staged AI use and AI-free instruction while measuring not only immediate task performance but retention, transfer and independent capability after AI disappears. Evidence across different subjects, socioeconomic contexts and education systems will also be essential before staged AI access can become a robust policy prescription.