Schools Should Redesign Learning, Not Retreat From AI

The biggest risk generative AI poses to education may not be cheating. It may be something quieter and more consequential: students producing increasingly sophisticated work while doing less of the cognitive labour needed to build durable knowledge and independent judgement.

In the study "Designing for cognitive resilience: a distributed cognition approach to education in the age of generative AI," published in AI & Society, Hans Westerbeek argues that schools and universities need to shift from policing AI toward protecting "cognitive resilience" - the capacity to keep thinking, retrieving, reasoning and verifying even when AI can provide instant answers.

The Real Risk Is Not Cheating but Cognitive Offloading

Much of the education sector's response to generative AI has focused on academic misconduct, plagiarism detection and acceptable-use rules. Westerbeek argues that this framing misses a more structural problem: AI is lowering the perceived cost of cognitive effort, making it easier to bypass retrieval, drafting, argument construction, verification and independent judgement while still producing polished work.

The paper draws on several strands of evidence to justify caution. Research on cognitive offloading suggests that using external systems is not inherently harmful, but repeated outsourcing becomes problematic when underlying mental operations are no longer rehearsed. Westerbeek cites a 2025 study of 666 participants that found a negative correlation between AI use and critical-thinking scores, with cognitive offloading acting as a mediating factor, while emphasizing that the study does not establish causation.

The wider evidence base is similarly suggestive rather than definitive. Automation-bias research shows that users can shift from active monitoring to passive acceptance, while early studies of AI-assisted writing and AI dependence have raised questions about reduced engagement and cognitive fatigue. At the same time, decades of learning science show that retrieval practice, effortful learning and "desirable difficulties" often strengthen long-term retention and transfer.

The policy implication is important because fluent output can disguise weak learning. A student may submit an articulate answer, receive a high mark and still have practised little of the reasoning required to generate or defend it independently. In that sense, generative AI challenges not only academic integrity but the validity of educational assessment itself.

Education Needs Friction, Verification and Visible Thinking

To address this problem, the author proposes nine design principles: deliberate friction, attention protection, process-visible assessment, retrieval and reconstruction, cognitive provenance, bounded AI use, metacognitive calibration, source triangulation and developmental staging. The goal is not to remove AI from learning, but to make sure the learner retains responsibility for the intellectual work that matters.

The framework's logic is captured in the paper's diagram on page 4. Commercial AI design can make thinking feel cheaper by offering speed, fluent synthesis and instant answers; reduced practice can then weaken underlying capacity and gradually normalize dependence. The nine principles are positioned as interventions that interrupt this pathway at different points, from protecting attention to redesigning assessment and requiring independent verification.

This concept of "deliberate friction" is particularly significant. Education has spent years trying to make learning more seamless, personalized and efficient, but some forms of difficulty are productive because they force learners to retrieve, reconstruct and reason. The implication is not that schools should make tasks unnecessarily harder, but that they should avoid designing away every obstacle that contributes to cognitive development.

The distinction is also crucial for assessment. If a task can be completed by prompting an AI system without demonstrating understanding, then it may no longer measure the competence it claims to assess. Westerbeek therefore supports approaches that make thinking visible, including supervised analytical work, oral defences, live reasoning and statements that clarify where AI contributed and where the learner's own intellectual work occurred.

One AI Policy Cannot Fit Every Stage of Learning

According to the author, the role of AI should change across primary, secondary and higher education rather than being governed by a single institutional rule. The central principle is sequencing: support should be introduced in ways that strengthen competence rather than permanently replace the cognitive operation being learned.

• At primary-school level, the paper argues that attention, language construction, working-memory rehearsal and early meaning-making need to remain with the child long enough for foundational competence to develop. AI may be most useful here as a teacher-facing tool or as a structured questioner, rather than as a default explainer or solver that performs the task before the learner has built the capacity to perform it.
• At secondary level, the author proposes a more explicit model of bounded use: students should first frame and think through a problem themselves, then use AI to expand or challenge their thinking, independently verify the result, and finally defend their reasoning. This sequence treats AI as part of a learning process rather than a replacement for one.

Universities face a different problem because they certify competence to employers, professions and society. If AI can perform a large part of assessed work, institutions may need to rethink what constitutes valid evidence of mastery. The paper therefore shifts the debate from "Was AI used?" to "Does this assessment still demonstrate that the student possesses the capability being certified?"

The Bigger Challenge Is Governance and Equity

Education technologies are introduced through procurement systems, platform choices and public-private partnerships whose commercial incentives may prioritize engagement, convenience, data capture or ecosystem lock-in rather than cognitive development. Governance therefore shapes the learning environment before teachers make individual pedagogical decisions.

This raises questions for education ministries, regulators and school systems. Procurement standards may need to consider whether platforms promote independent reasoning or encourage frictionless completion. Curriculum authorities may need to make verification, critical data literacy and epistemic judgement explicit learning goals, while assessment regulators may need to focus less on policing AI use and more on preserving the validity of credentials.

The equity implications are equally significant. Westerbeek argues that well-resourced schools with strong teachers and governance capacity may be better able to use AI while preserving deep learning, whereas under-resourced institutions could face stronger pressure to substitute automation for human support. That could create a new form of inequality in which access to AI becomes widespread but access to environments that cultivate independent judgement remains uneven.

The paper does not treat all AI assistance as harmful. Translation, text-to-speech, writing support and individualized formative feedback can expand access for students who would otherwise struggle to participate fully. For under-resourced schools in particular, AI-supported scaffolding may offer genuine benefits, meaning cognitive-resilience principles must function as context-sensitive defaults rather than rigid universal rules.

The framework is conceptual, not a tested causal model, and the paper explicitly calls for longitudinal cohort studies, intervention research and validated measures of cognitive resilience. It also urges further study of assessment validity, procurement politics and whether AI-related cognitive offloading is distributed unevenly by socioeconomic status and school resources.

The strategic challenge for education is not to preserve a pre-AI classroom. It is to determine which forms of human effort remain developmentally essential even when machines can perform them more quickly. If education rewards only polished output, AI will increasingly supply that output; if it rewards reasoning, verification, judgement and genuine authorship, institutions will have to redesign learning so those capacities remain visible and necessary.