1. Introduction
The Internet and artificial intelligence have become indispensable infrastructure within contemporary education systems. The instant accessibility of digital resources, the capacity to generate personalised learning pathways, and the efficiency gains offered by AI-assisted writing and research have delivered unprecedented improvements in flexibility and cost-effectiveness. Beneath this techno-optimistic narrative, however, a systemic risk is accumulating: students are abandoning printed textbooks, library-based inquiry, group study and fieldwork at an unprecedented rate, turning instead towards the ’ready-made’ knowledge supplied by search engines and generative AI.
This shift is not a mere change of method. The choice of learning medium profoundly shapes the manner and depth of cognitive processing. When students withdraw from the cognitive chain of ’searching - filtering - synthesising - critiquing’ and instead accept AI-generated summaries, answers or drafts, what is damaged is not merely the durability of knowledge retention but the structural foundation of thinking itself. Recent research has begun to document the cognitive consequences of this trend systematically, and its findings warrant serious attention from educational policymakers.
2. Screen Reading and the Attenuation of Cognitive Depth
Empirical research on the relationship between reading medium and comprehension has reached a broadly consistent conclusion: for lengthy, linear informational texts, paper-based reading systematically outperforms screen-based reading. This finding does not arise from nostalgic preference but rests on cognitive-scientific foundations. Printed texts provide stable spatial cues: readers form a ’cognitive map’ of where information is located on the page, and this map plays a vital supporting role in retrieval, integration and memory recall. Screen reading, particularly on smartphones, weakens this spatial anchoring effect while introducing additional sources of attentional distraction.
An experimental study of Argentine university students provides direct evidence: when reading the same expository text, the smartphone group’s comprehension scores were significantly lower than those of the personal computer group, with students in engineering programmes who had weaker vocabulary skills performing especially poorly under the smartphone condition. The researchers attributed this to ’higher cognitive costs, attentional distraction or insufficient reading depth’ (Delgado et al., 2018) [2]. More notably, OECD PISA data reveal a clear pattern: students who frequently read long texts (of more than ten pages) score 31 points higher in reading than those who read only short texts, and this difference remains significant after controlling for students’ and schools’ socio-economic backgrounds (OECD, 2021) [5]. Embedded links, pop-ups and multimedia elements in digital texts not only fail to enhance comprehension but correlate negatively with reading performance.
This chain of evidence points towards an uncomfortable inference: as education systems increasingly treat digital reading as the default option while printed textbooks and long-form reading are marginalised, students’ deep reading comprehension is undergoing a silent decline.
3. AI Dependence and the Erosion of Critical Thinking
If screen reading erodes the depth of cognitive processing, the widespread use of generative AI may directly dismantle the procedural conditions under which critical thinking develops. Critical thinking is not a static body of knowledge to be ’acquired’ but a cognitive habit gradually constructed through repeated cycles of ’attempt - failure - reflection - revision’. The core function of generative AI (instantly supplying seemingly complete answers, drafts or solutions) precisely removes the ’productive struggle’ indispensable to that constructive process.
A large-scale empirical study published in 2026 provides quantitative support for this judgement. Drawing on a sample of 250 higher education students and employing multiple regression analysis to examine the effects of AI use frequency and dependence on critical thinking, the study found that both AI use frequency (, ) and AI dependence level (, ) were significant negative predictors of critical thinking skills, with the model explaining 50.5% of the variance in critical thinking (Zhao et al., 2026) [12]. In other words, the deeper students’ dependence on AI tools, the lower the level of their critical thinking development.
By contrast, ’independent learning behaviour’ was a significant positive predictor of critical thinking (, ). The significance of this contrast lies in the fact that the growth of critical thinking depends on learners’ willingness and capacity to bear cognitive load actively, rather than outsourcing cognitive tasks to automated systems.
Theoretical analysis further illuminates the mechanism. A systematic study of cognitive decline notes that sustained use of generative AI is fostering habits of ’shared metacognitive offloading’, whereby learners progressively transfer responsibility for planning, monitoring and evaluating their own understanding to AI systems, thereby weakening self-regulated learning capacity (Kajiwara & Kawabata, 2024) [3]. The same study cites findings from the MIT Media Lab (2025) that students who habitually use large language models for writing show declines in both revision cycles and originality; Microsoft Research (2025) similarly describes the rise of the ’shallow knowledge worker’, proficient in prompt manipulation yet weak in abstract reasoning and ethical judgement.
More troubling still, this cognitive degradation may carry intergenerational effects. As younger cohorts habituated to automated cognitive offloading gradually assume decision-making roles, their tolerance for and capacity to handle contradiction, ambiguity and ethical complexity may be systematically lower than those of previous generations. This is no longer merely a question of educational quality but a structural risk to society’s collective cognitive resilience.
4. The Institutional Roots of Negative Externalities and the Governance Dilemma
The negative externalities described above are not inevitable products of technology itself but consequences of institutional design failure. Contemporary education systems exhibit structural deficiencies in two respects.
First, there is a mismatch between assessment systems and technological convenience. Where coursework and examinations remain predominantly ’product-oriented’, and AI can generate formally compliant outputs at near-zero cost, students rationally choose the most efficient path: using AI. The problem is not student ’moral laxity’ but the failure of assessment design to distinguish effectively between ’using a tool to complete a task’ and ’developing capability through completing a task’. Where the two cannot be distinguished, substitution will occur.
Second, digital literacy education’s ’protective’ orientation fails to reach deeper cognitive habits. Existing AI literacy education focuses largely on ethical warnings, privacy protection and academic integrity, a ’defensive - preventive’ paradigm. Such education is necessary but insufficient to address the risk of cognitive degradation. What genuinely needs cultivating is ’critical - reflective’ capacity: the cognitive habit of systematically questioning, verifying and integrating AI outputs. Yet this capacity cannot be acquired through exhortation; it can only be formed through repeated cognitive practice.
5. A Policy Framework: The Principle of Preserving Cognitive Friction
On the basis of the foregoing analysis, this article proposes the following policy framework, whose core principle is the prudent preservation of cognitive friction, that is, deliberately maintaining within educational design those cognitive challenges that digital convenience has eliminated, in order to safeguard the continued development of deep thinking.
First, dynamic adjustment at the curriculum and assessment level. Educational institutions should clearly distinguish ’AI-assisted’ tasks from ’cognitive-capability-developing’ tasks. The latter should restrict or prohibit the use of generative AI and require process evidence (such as drafts, revision records and reading notes) as the basis for assessment. Assessment design should shift from ’judging final products’ to ’evaluating cognitive processes’, rendering critical thinking, synthesis and original contribution observable and assessable.
Second, the reconstruction of reading and study norms. Given the systematic advantages of paper-based reading for long-text comprehension, curriculum design should retain a defined proportion of printed textbook reading and library-based inquiry requirements. This is not a rejection of digital tools but a respect for the principle of medium–task matching: deep comprehension of lengthy linear texts should preferentially employ paper, while information retrieval and preliminary exploration may utilise digital resources. OECD data already show that the positive association between long-text reading frequency and reading achievement is independent of socio-economic background, implying that such norms possess universal applicability (OECD, 2021) [5].
Third, the institutionalisation of AI use norms. Educational institutions should formulate explicit AI use policies, distinguishing ’permitted uses’ (such as language polishing and data organisation) from ’prohibited uses’ (such as draft generation and problem solving), and requiring students to label AI-assisted content explicitly. More critically, policies should require students to demonstrate evidence of their independent cognitive work (such as problem analysis frameworks, hypothesis-generation records and refutation-and-revision processes) to ensure that core thinking capacities are not displaced by tools.
Fourth, the repositioning of the teacher’s role. In an environment of ubiquitous AI, the teacher’s core value lies no longer primarily in information transmission but in designing learning experiences that ’cannot be automated away’: tasks requiring real-time judgement, contextualised feedback and cognitive challenge. Teachers require corresponding training and institutional support to accomplish this role transition.
6. Conclusion
The efficiency dividends that the Internet and AI bring to education are real, but so too is the tension between efficiency and depth. When ’obtaining answers faster’ becomes the default mode of learning, the capacity to ’think about problems more slowly’ is silently lost. The cost of this loss will not manifest immediately, yet it may crystallise over ten or twenty years in the form of degraded individual cognitive capacity and diminished societal capacity for innovation.
The core task of educational policy is not to reject technology but to draw prudent boundaries between technological convenience and cognitive development. This entails consciously preserving certain ’inefficient’ modes of learning (paper-based reading, handwritten notes, face-to-face discussion, fieldwork), not because they are more ’traditional’, but because they carry cognitive functions that technology cannot yet replace. An education system that abandons this boundary exchanges short-term efficiency gains for long-term cognitive capital. The cost of that exchange will ultimately be borne by society as a whole.