Artificial intelligence has entered education faster than most institutions could prepare for it. A 2025 survey of 1,041 UK undergraduates found that 92% used AI in some form, up from 66% a year earlier. Its capability has kept pace: in a blind test at the University of Reading, 94% of AI-written exam submissions went undetected and, on average, earned higher grades than real students’ work. This raises a legitimate concern. If a machine can do the intellectual work, will students still develop the ability to do it themselves? The concern is well founded. The answer lies in understanding what learning requires, and in using AI in ways that support those requirements rather than replace them.



The nature of the problemLearning is not producing correct answers, but the process by which a person builds understanding, memory, and skill through effortful engagement with material. Cognitive scientists describe this as ‘desirable difficulty’ as the struggle of recalling information, wrestling with an unfamiliar problem, or organising scattered ideas into a coherent argument. In experiments by Roediger and Karpicke, restudying led to better recall after five minutes, but testing produced better retention after one week, even though restudying left students more confident.




When a student delegates that struggle to an AI system, learning is diminished. In a field experiment with nearly 1,000 Turkish high-school mathematics students, unrestricted GPT-4 access raised practice performance by 48%, yet students who later lost access scored 17% worse than those who never had it. The OECD (2026) similarly concludes that general-purpose tools can improve task performance without producing learning gains. However, the evidence is not uniformly negative: outcomes depend on how students use the tools, according to a study published in 2025.




The central question for any student is therefore not ‘Is this permitted?’ but ‘Who is doing the thinking?’ If AI performs the cognitive work the assignment was designed to develop, the student has compromised their own education, regardless of whether they broke a rule.



Using AI as a tutor , not a substituteThe most productive use of AI is as an explanatory aid. A student who encounters an unfamiliar concept, such as the causes of the French Revolution or the mechanics of mitosis, can ask an AI system to explain it simply, offer analogies, or approach the topic differently. In a Harvard randomised trial, students learned significantly more, in less time, from a purpose-built AI tutor than from an active-learning class. However, the study covered only 194 students and two physics topics. Usage matters: students who used language models as tutors benefited, whereas those who had them solve practice exercises learned less. Students must therefore remain active participants, asking follow-up questions, challenging explanations, and restating ideas in their own words.




A useful practice is to try a problem independently before seeking assistance, then use AI to check the reasoning or locate an error. Seeking a hint rather than a complete solution preserves productive struggle: in the Turkish experiment, a tutor configured to give incremental hints largely avoided the learning loss. Sequence also matters; in a small MIT preprint, participants who wrote first and then used an LLM showed higher recall than those who began with the LLM. Students may also ask the AI to question them rather than answer, turning a passive tool into an active learning partner.



Preserve the core intellectual tasksCertain tasks should remain the student’s own because they are the skills education aims to build: formulating an argument, structuring an essay, interpreting a text, and solving an original problem. Writing often discovers thought, and doesn’t just record it. A meta-analysis of 48 studies found that writing-to-learn activities modestly improve achievement. Participants who wrote with an LLM reported the lowest sense of ownership over their essays and struggled to quote them accurately.




This does not mean AI has no role in writing. It can suggest alternative phrasings, identify unclear passages, or flag grammatical errors in a completed draft. The distinction is between refining work the student has produced and generating work they have not. The former strengthens the writer; the latter replaces the writer.



Verify and think criticallyAI systems are fallible. They can produce confident but inaccurate statements, cite non-existent sources, and reflect biases in their training data. In a study of 636 ChatGPT-generated citations, 55% from GPT-3.5 and 18% from GPT-4 were fabricated. A UNESCO-commissioned study found clear bias against women in every large language model it examined (UNESCO & IRCAI, 2024). Every factual claim obtained from AI should therefore be checked against textbooks, peer-reviewed publications, and established reference works.




This requirement is itself educational. Among knowledge workers, greater confidence in AI is associated with less critical thinking , whereas greater confidence in one’s own ability is associated with more. Students who interrogate AI outputs develop critical judgment that no tool can supply on their behalf.



Maintaining academic integrity Institutions and instructors have different policies, and students are expected to follow the rules for each assignment. Clarity is often lacking: in a 2024 survey of more than 3,800 students in 16 countries, 86% were not fully aware of their university’s AI guidelines. Where AI assistance is permitted, it should be acknowledged transparently; where it is prohibited, circumvention undermines students’ learning and the trust on which educational communities depend.




Beyond compliance, submitted work should honestly represent what the student understands. Grades and credentials are meaningful only insofar as they reflect real competence. A student who presents AI-generated work as their own may earn a favourable mark, yet is likely to encounter examinations, interviews, or professional demands that expose the gap, as the unassisted exam results in the Turkish experiment suggest.




Regular self-assessment is a practical safeguard, because AI can distort self-judgment: AI users scored higher on reasoning tests yet consistently overestimated their performance. After studying with AI, a student can explain the material aloud without notes, write a summary from memory, or attempt practice problems unaided; difficulty signals that the AI did too much of the work. Practice testing and spaced retrieval were rated the most effective of ten common learning techniques , and they cannot be outsourced.




Students should also set deliberate boundaries, such as completing first drafts without assistance or limiting AI to concept clarification and exam preparation. Defining boundaries consciously distinguishes purposeful use from habitual dependence.



Is AI truly a threat to education?AI is neither a threat to education nor a guaranteed benefit; its effect depends on how it is used. The OECD Digital Education Outlook 2026 finds that using it with pedagogical intent can improve learning, whereas completing a task with AI does not automatically produce learning. Students who treat it as a shortcut risk eroding their skills, while those who treat it as a tutor, a sounding board, and a source to be scrutinised can deepen their understanding. The guiding principle is simple: the thinking belongs to the learner. AI may illuminate the path, but the student must still walk it. By protecting core intellectual tasks, verifying information, upholding honesty, and monitoring their own progress, students can enjoy the advantages of this technology while keeping their education genuinely their own.

Contact to : xlf550402@gmail.com


Privacy Agreement

Copyright © boyuanhulian 2020 - 2023. All Right Reserved.