From the classrooms of Pulchowk to the dormitories of Harvard — an evidence-based synthesis of current research on artificial intelligence and learning, with direct implications for students, professionals, and graduates in Nepal.
Introduction
Consider two students at a campus café in Kathmandu, both working on a research paper due the following morning. One opens ChatGPT and pastes in the assignment prompt. The other opens it too — but uses it to stress-test her own thesis, identify weaknesses in her argument, and locate three studies she had not previously considered. Both used artificial intelligence. Only one engaged in genuine learning.
This contrast encapsulates the central debate around AI in education. The question is no longer whether students should use AI — that threshold has long since been crossed. Eighty-six percent of students worldwide now use AI in their studies, with figures reaching 92% in certain markets. The consequential question — particularly for students and working professionals in Nepal — is how they use it, and whether that use is cultivating capability or quietly eroding it.
This article synthesises the most credible research published in the preceding 18 months — drawn from peer-reviewed journals in Nature, Frontiers in Psychology, and PMC, as well as landmark university experiments — and grounds the findings in Nepal’s specific context: a country where youth unemployment stands at 20.82%, where university graduates face a 26.1% unemployment rate, and where more than 100,000 students depart for overseas study each year. In this environment, one’s relationship with AI is not merely an academic consideration. It is a career-defining one.
The Scale of AI Adoption Among Students
Before evaluating whether AI is beneficial or detrimental to students, it is necessary to acknowledge the extent to which it has permeated academic life. The pace of adoption has been extraordinary — arguably the fastest voluntary uptake of any technology in the history of formal education.
Global Adoption Metrics
The increase from 53% to 88% of students using AI for formal assessments within a single academic year represents one of the most rapid behavioural shifts ever recorded in education research. Approximately 65% of students globally now consider AI tools essential to their academic success — not supplementary, but indispensable.
Nepal Context: Where Does the Country Stand?
Nepal does not yet feature in major global AI-in-education surveys; however, the available indicators are instructive. As of early 2025, 16.5 million Nepalis — representing 55.8% of the population — are online, with mobile broadband penetration at 144.56%. Students at institutions such as Pulchowk Engineering College and Tribhuvan University are increasingly building machine-learning models for their theses and employing AI for tasks ranging from software development to academic writing.
Nepal ranked 150th out of 193 countries in Oxford Insights’ 2024 Government AI Readiness Index, though its Data Representativeness score of 58.51 indicates meaningful underlying potential. In August 2025, the government formally approved the National AI Policy 2082, which explicitly promotes the integration of AI competencies into educational curricula.
The Cognitive Outsourcing Problem: What the Evidence Shows
Concerns regarding AI causing “cognitive outsourcing” — more precisely termed cognitive offloading in the academic literature — are not unfounded. They are supported by a growing and methodologically robust body of research.
The Core Mechanism
Cognitive offloading refers to the use of external tools to perform tasks that would otherwise require internal mental effort. Writing has always been a form of it. However, AI represents an extreme and frictionless variant: rather than helping one store a thought, it generates the thought entirely. This distinction is of fundamental importance to learning theory.
“Studies so far have found that generative AI boosted learning for those who use it to engage in deep conversations and explanations, but hampered learning for those who sought direct answers.”
— PMC / National Library of Medicine, 2025
The most comprehensive recent investigation into this phenomenon involved a survey and in-depth interviews with 666 participants across diverse age groups and educational backgrounds, published in Societies (MDPI, 2025). Researchers found a significant negative correlation between frequent AI tool usage and critical thinking ability, mediated through cognitive offloading. Younger participants demonstrated the highest levels of AI dependence and the lowest critical thinking scores.
A separate study published in Frontiers in Psychology (2025), examining 580 Chinese university students, found that greater AI dependence was associated with diminished critical thinking, with cognitive fatigue acting as a partial mediating variable. A neural-behavioural experiment by Kosmyna et al. (2025) extended these findings further: participants who composed essays using an AI tool exhibited measurably lower cognitive activity, lower self-reported sense of ownership over their work, and — following four months of continued AI use — consistently underperformed across neural, linguistic, and behavioural measures compared to those using search engines or no AI tools at all.
The Cognitive Miser Problem
A well-established principle in psychology — the Law of Least Effort — holds that, given a choice, humans systematically avoid unnecessary cognitive demand. AI creates a frictionless path of least resistance. Research from Frontiers in Psychology (2025), studying Chinese college students, describes this as cognitive inertia: the more students rely on AI’s immediate feedback, the more they form an “effort-saving learning model” that weakens exploratory investment and creative thinking. The troubling finding: student dependence on AI has increased, yet there has been no corresponding improvement in academic innovation outcomes.
Key Warning: A study conducted at a major university in Zimbabwe found that 32.7% of students demonstrated addictive patterns in their AI usage, averaging 18.3 AI interactions per day, with 65.8% reporting failed attempts to reduce usage despite recognising its negative academic consequences. For students in comparable developing-country contexts — including Nepal — the combination of high academic pressure, limited tutoring infrastructure, and expanding AI access creates analogous risk conditions.
The Productivity Opportunity: Counterbalancing Evidence
The same body of research that documents these risks also contains genuinely significant findings regarding what AI can achieve when used deliberately. The critical qualification present in virtually every positive finding is consistent: the benefits are conditional on how AI is employed.
The Harvard Experiment
In late 2023, physicist Gregory Kestin and senior lecturer Kelly Miller at Harvard University conducted one of the most rigorously designed experiments on AI in education, subsequently published in Scientific Reports (Nature, 2025). In a randomised crossover trial involving 194 undergraduate physics students, they compared learning outcomes between students in a best-practice active learning classroom and students using a custom AI tutor (“PS2 Pal”) from their residential accommodation.
“Students using the AI tutor learned more than twice as much as their peers in the best-practice active learning classroom — in less time — and reported significantly higher engagement and motivation.”
— Kestin & Miller, Scientific Reports / Harvard University, 2025
The AI tutoring cohort demonstrated approximately double the learning gains, with an effect size between 0.73 and 1.3. Median study time for the AI group was 49 minutes, versus 60 minutes for classroom instruction. Critically, this was not unstructured ChatGPT use — the AI tutor was designed according to pedagogical principles: it provided one step at a time, never disclosed the complete answer, and actively required students to reason before receiving guidance.
Active Versus Passive AI Use
A consistent thread throughout the literature is the distinction between active and passive engagement with AI. A meta-analysis of 51 experimental studies on ChatGPT identified a large positive effect on learning performance — concentrated among students who used AI to engage in deep dialogue and test their own reasoning. Additional meta-analytical research found that AI-assisted learning environments can improve outcomes by 23–35%, with the strongest effects observed in STEM and language learning — precisely the competencies in greatest demand in Nepal’s expanding information technology sector.
Key Performance Metrics
The Nepal Equation: Elevated Stakes
The global findings outlined above carry heightened urgency when applied to Nepal’s specific economic and educational context.
The Graduate Employment Crisis
Nepal’s fourth Living Standard Survey (2024) found that the youth unemployment rate rose from 7.3% in 1995–96 to 22.7% in 2022–23 — a structural deterioration that renders the practical value of academic credentials increasingly contingent on underlying competence.
AI Fluency as a Competitive Advantage
Nepal’s skill mismatch is acute: over 677,000 generally skilled workers migrate abroad annually, while only 732 highly skilled workers do. A report by the Skills4Dev project found that 92% of job postings now require some level of digital proficiency. LinkedIn data corroborates this: the top skills added by professionals globally in 2024 were ChatGPT (60%) and prompt engineering (38%) — both learnable, practical, and demonstrably valued by employers.
For Nepali students who graduate with the ability to effectively deploy AI tools for research, coding, writing, data analysis, and problem-solving, a meaningful competitive advantage exists over peers who either avoid AI entirely or use it merely as a shortcut mechanism.
The Equaliser Argument
There is a further dimension specific to resource-constrained environments. For students in Nepal with limited access to quality tutors, expensive textbooks, or reliable infrastructure, AI represents a potential equaliser. A student in Pokhara who cannot afford private preparation coaching may now access a patient, perpetually available learning resource on their mobile device. The critical question remains whether they use it to build their thinking or to replace it.
The Academic Integrity Risk
AI-related academic misconduct has increased sharply on a global basis. Reported AI cheating incidents rose from 1.6 students per 1,000 in 2022–23 to 7.5 per 1,000 in 2024–25 — a nearly 400% increase within two years. In Nepal, where academic institutions are only beginning to develop formal AI policies, this creates substantive career risk: a transcript that appears strong but does not reflect actual capability, and graduates who enter the labour market without the competencies their credentials imply.
Two Divergent Paths
Every student and professional in Nepal currently using AI is, effectively, on one of two trajectories. The research evidence on their respective outcomes is clear.
This distinction is characterised in the academic literature as “active extension versus passive offloading” and is the single most predictive variable in whether AI helps or harms a student’s development. The underlying mechanism is straightforward: it is a question of whether the brain is engaged or circumvented.
A Practical Framework for Capability-Building AI Use
The following research-grounded framework is intended for students, working professionals, and aspiring professionals seeking to use AI in ways that genuinely build their capabilities.
01. The Draft-First Principle
Before opening any AI tool for a writing task, produce a personal rough draft — even a single paragraph of unpolished thinking. Then use AI to respond to what you have written. This ensures cognitive engagement with the problem precedes AI input, positioning AI as a feedback instrument rather than a content generator. The neural research of Kosmyna et al. (2025) is unambiguous: the brain’s ownership of work, and its subsequent learning, depend on the effort invested in its creation.
02. Interrogation Over Prompting
The difference between “Write me an essay on climate change” and “I have argued that Nepal’s hydropower sector is central to its climate strategy — what are the strongest counterarguments to this position?” is substantial. The second formulation requires the student to have developed a position independently. Use AI for Socratic dialogue: ask it to identify weaknesses, offer alternative frameworks, and surface gaps in your reasoning.
03. Expanding the Research Map
AI is highly effective at surfacing research a student may not have known existed. Use it to ask: “What are the five most cited studies on X?” or “What analytical frameworks do economists apply to Y?” — then read those sources independently. Allow AI to expand intellectual range; do not allow it to become the sole lens through which information is filtered.
04. Treating AI Fluency as a Formal Skill
In Nepal’s labour market — particularly in information technology, consulting, data analysis, and digital marketing — AI fluency already confers a competitive advantage. Developing prompt engineering skills seriously and representing them explicitly on a CV and professional profile is advisable. Global data confirms that ChatGPT and prompt engineering were the most commonly added professional skills in 2024.
05. The Explanation Test
After using AI to understand a concept, a coding technique, or an analytical framework, close the AI interface and articulate the concept in your own words — verbally or in writing. If you cannot do so without returning to the AI, comprehension has not been achieved; the knowledge has merely been borrowed temporarily. This is the critical distinction between learning performance (short-term recall) and learning retention (genuine capability development).
06. Designated AI-Free Practice Zones
Deliberately practise the foundational skills in which AI performs capably — but do so in contexts where AI is not used. Write a short daily reflection without AI assistance. Attempt coding problems before consulting AI for solutions. Construct and articulate arguments from memory. Productive struggle is the mechanism of cognitive development — the same logic underlying why trained athletes still engage in physical conditioning despite the availability of mechanical aids.
Recommended AI Tools for Nepali Students and Professionals
Not all AI tools are equally suited to active learning and professional development. The following tools reward deliberate engagement and are relevant to Nepal’s evolving professional landscape.
Recommendations for Nepali Educators and Institutions
The evidence presented in this synthesis also carries implications for educators, department heads, and university administrators.
Nepal’s National AI Policy 2082 is correct that AI competencies must be embedded in educational curricula. However, the manner of integration is of critical importance.
💡 For Educators: Institutions that prohibit AI use entirely are likely fighting an unwinnable contest — and may inadvertently prompt students to use AI unsupervised, without guidance regarding its risks. The more effective approach, supported by the Harvard study and others, is to design assessments and learning experiences that exploit AI’s strengths while preserving the cognitive challenge that produces genuine learning.
Oral examinations, live problem-solving tasks, project defences, and reflective journalling are assessment formats that AI cannot shortcut, and that develop the critical thinking, communication, and analytical skills that Nepal’s employers consistently identify as lacking in recent graduates.
A 2025 study from Frontiers in Education, examining Southeast Asian university students, specifically noted the absence of research from broader South Asian contexts, identifying “unique cultural factors, including collectivist values and hierarchical educational structures, that may significantly influence technology adoption.” Nepal fits this characterisation precisely. The country’s educational culture’s strong emphasis on rote memorisation suggests that the risk of passive AI adoption may be more pronounced here than in more critically oriented educational environments. This is not an argument against AI in Nepali classrooms — it is an argument for approaching it with greater intentionality than others.
Conclusion
Is AI adoption cognitive outsourcing or a productivity opportunity? The research provides an intellectually honest answer: it is both, and which of the two it becomes is determined entirely by how it is used. The same tool that doubled learning gains in a Harvard physics classroom also correlates with measurably reduced cognitive engagement when used passively. The same technology considered essential by 65% of students is also producing a 400% rise in recorded academic misconduct. AI is, in essence, an amplifier — it magnifies whatever intent is brought to it.
For Nepal specifically, the implications extend beyond individual academic performance. In a country where youth unemployment is structurally elevated, where the education-employment gap is widening, and where the domestic capacity to build a meaningful AI ecosystem clearly exists, the manner in which this generation of students engages with AI will have material consequences for Nepal’s economic trajectory over the next decade.
“Nepal finds itself at a historic moment, with AI redefining how we learn, work, govern and grow. Nepal’s AI strategy must promote inclusion over exclusion, empower over control, and put people over profits.”
— Hon. Prithvi Subba Gurung, Minister of Communication and Information Technology, at the UNDP 2025 HDR Launch, Kathmandu
The UNDP’s 2025 Human Development Report, launched in Kathmandu, emphasised building “a complementarity economy where AI works alongside people rather than replacing them.” That framing is precisely correct — and it begins not at the policy level, but at the individual level, in the decisions each student and professional makes every day.
Central Finding
Artificial intelligence cannot substitute for the value of independent human reasoning. It can supplement, accelerate, and enrich that reasoning — but the moment a student stops thinking, and AI becomes the thinker while the student merely reviews its output, a trade has been made that the evidence consistently shows is detrimental to learning, to professional development, and ultimately to career outcomes.
Nepal does not need graduates who can operate ChatGPT. It needs graduates who can think rigorously, communicate persuasively, and solve problems creatively — and who are also able to deploy AI to do all of those things more effectively. That combination, at this moment, is genuinely uncommon. And uncommon is valuable.