By Sudhir Tiku
Sudhir Tiku is a Singapore-based AI Author and Technology leader with over two decades of work experience across leading Fortune 500 companies. Throughout his career, he has led initiatives in automation and the practical application of Artificial Intelligence for the Global South. In this space, he has a deep interest in Small Language Models and their localized use cases. He is the author of the book, AI in Global South: Power, Policy and Progress, published by World Scientific, Singapore.
Abstract: Artificial intelligence or AI has become one of the most consequential technological influences on contemporary education, but its relationship with learning did not begin with Generative AI. It extends through a longer history of teaching machines, computer-assisted instruction, intelligent tutoring systems, adaptive learning and educational data mining. The public emergence of generative AI after 2022 accelerated this trajectory by placing conversational and content-generating systems directly in the hands of students and educators, often outside formal institutional planning.
This paper traces that historical development, examines current educational applications and risks, and considers likely future trajectories. It argues that AI can improve personalisation, accessibility, formative feedback, teacher productivity and lifelong learning, while also creating serious concerns around academic integrity, cognitive dependence, algorithmic bias, educational inequality and the credibility of assessment. The paper probes whether AI is most useful when it expands the capabilities of teachers and learners without taking away their responsibility for judgement and independent thought. The paper approaches this historically, institutionally and prospectively.
Introduction
Education has always developed alongside changes in technology. Writing allowed knowledge to survive beyond oral memory. The printing press expanded the availability of books and formal learning. Radio and television extended instruction beyond the classroom, computers introduced interactive learning, and the internet transformed access to information. Artificial intelligence belongs to this historical sequence, but its educational implications are different in kind as well as degree.
Earlier educational technologies primarily helped people store, transmit or retrieve knowledge. Artificial intelligence can increasingly participate in activities that look much closer to intellectual work itself. Generative AI can participate in the production, reformulation and interpretation of knowledge, which places it inside processes that education has traditionally used both to teach and to assess human understanding. The public release of ChatGPT in the year 2022 brought this shift into mainstream education. Students no longer needed specialist software or technical knowledge to work with an advanced language model. A learner could ask AI for a theory to be simplified, request a different analogy, generate revision notes, receive help on a mathematical problem, improve an essay or produce an entire assignment. The same accessibility that created opportunities for individualised support also unsettled assumptions about authorship, effort, assessment and academic integrity.
Students have adopted AI rapidly. The Digital Education Council’s Global AI Student Survey 2024, based on responses from students across 16 countries, found extensive use of AI within higher education.¹ The Organisation for Economic Co-operation and Development has similarly noted that generative AI is already affecting teaching, learning and assessment while governance frameworks are still developing.² Students began using general-purpose AI before many schools and universities had decided what responsible educational use should mean. Hence, AI is already a part of our education system. The difficult question is what education should preserve as machines become capable of producing explanations, answers and polished language almost instantly.
Historical Foundations
The ambition to personalise education through technology predates modern Artificial intelligence systems. During the 1920s, psychologist Sidney Pressey developed mechanical teaching devices capable of presenting questions and providing immediate feedback.³ These machines were primitive, but they introduced an enduring idea: learners might move through material at different speeds while receiving responses tailored to their performance.
B. F. Skinner later developed teaching machines grounded in behavioural psychology and programmed instruction.⁴ Students progressed incrementally through carefully structured material and received reinforcement after correct responses. These devices were not intelligent in the computational sense, yet they established several principles that later educational technologies would inherit, including self-paced learning, immediate feedback and structured progression.
The emergence of Artificial intelligence as an academic field is usually associated with the Dartmouth Summer Research Project of 1956. John McCarthy, Marvin Minsky, Claude Shannon and other researchers proposed that aspects of intelligence might be described precisely enough for machines to simulate them.⁵ Education soon became an attractive domain because teaching requires diagnosis, reasoning, adaptation and decisions about what a learner should encounter next.
One of the earliest influential computer-based educational environments was PLATO, developed at the University of Illinois.⁶ PLATO supported interactive lessons, testing and communication and anticipated functions that later became standard in digital learning environments. It was not an AI system in the contemporary sense, but it demonstrated that computers could become interactive educational spaces rather than merely calculating devices.
Intelligent Tutoring Systems and Adaptive Learning
Intelligent tutoring systems represented a more direct application of artificial intelligence to education. Instead of simply presenting predetermined material, such systems attempted to represent both subject knowledge and the learner’s current understanding and then choose instructional responses according to performance. A typical intelligent tutoring system included a model of the subject, a model of the learner’s knowledge, a pedagogical strategy and an interface for interaction.⁷ The conceptual advance was important: the system attempted not only to determine whether an answer was correct, but to infer why a learner was making a particular error and what intervention might help.
An influential early example was SCHOLAR, developed by Jaime Carbonell to teach South American geography.⁸ The system attempted to engage students in dialogue and adapt its responses according to what they appeared to know. Later intelligent tutoring systems became particularly effective in structured domains such as mathematics, science and computer programming, where learner states and misconceptions could be modelled with greater precision.
Research later showed that well-designed tutoring systems could produce meaningful learning gains, although effectiveness varied by subject, design and implementation.⁹ This body of work established a durable proposition in educational AI: some benefits associated with one-to-one tutoring might be delivered computationally at scale, even if a machine tutor did not reproduce the full pedagogical and relational role of a human teacher.
Adaptive learning extended the same principle. Rather than requiring every learner to progress through identical material at the same rate, adaptive systems alter content, pacing or difficulty according to performance. Students who struggle with a concept can receive additional explanation or practice, while those who demonstrate mastery can progress more rapidly.
Machine Learning and Digital Education
During the 1990s and 2000s, Artificial intelligence increasingly shifted from rule-based expert systems toward machine-learning methods capable of identifying patterns in data. At the same time, schools and universities were becoming more digital. Learning management systems, online assessments, student information systems and educational platforms began generating large quantities of information about learner behaviour.
This contributed to the growth of educational data mining and learning analytics. These fields use digital educational data to identify patterns associated with performance, misconceptions, engagement and potential dropout.¹⁰ The development of the internet further changed the educational environment. Search engines and digital libraries weakened dependence on physical textbooks and local libraries, while online learning platforms allowed courses to reach international audiences. Massive Open Online Courses (MOOC) during the early 2010s demonstrated that instruction could operate at unprecedented scale. MOOCs also exposed a crucial distinction that access to information is not equivalent to learning. Completion rates were often low and learners still needed motivation and a sense of progression.¹¹
AI appeared to offer one response to this limitation. If digital education could reach very large populations, intelligent systems might provide forms of personalised support that human instructors could not provide individually. The long-standing educational challenge therefore shifted from distributing content to sustaining learning around that content.
The Generative AI Turning Point
The release of ChatGPT marked a turning point because educational AI moved out of specialised systems and into ordinary natural-language interaction. Earlier tools were often embedded within particular platforms, subjects or institutional processes. Generative AI placed a general-purpose conversational system in front of the learner. This sharply reduced the barrier to use. In a single conversation, a student could request an explanation of a difficult concept, ask for a simpler version, demand an analogy, generate a practice test, receive feedback and then move into a related topic. The same interface could summarise readings, translate material, generate code, propose research questions or produce a complete essay. The educational significance lies not only in capability but in breadth: many academic functions that previously required separate tools or human assistance became accessible through one interface.
The OECD has described generative AI as potentially transformative because it can support more autonomous learning while simultaneously challenging conventional homework, assignments and assessment.¹² UNESCO has likewise noted that generative AI has developed faster than the regulatory and institutional responses of many educational systems.¹³ This mismatch matters because students do not wait for policy cycles. Students can adopt a new tool within days, while institutions may take months or years to change assessment rules, curricula, teacher training and governance.
Generative AI brings this tension directly into everyday learning. The same system that can explain a difficult idea to a learner at midnight can also produce the learner’s assignment. The educational issue is therefore not access to the tool itself, but what cognitive work the learner continues to perform when the tool is used.
The Current Educational Landscape
Generative AI is now used by students for information retrieval, writing support, revision, explanation, brainstorming and feedback. Its adoption is especially significant because much of it has occurred independently of institutional planning. Students often encountered these tools first in their everyday digital lives and only later within formal educational policy. Initial institutional responses were dominated by fears of plagiarism and academic misconduct. Some schools and universities restricted access to generative systems. Such approaches became difficult to sustain as generative capabilities were incorporated into search engines, productivity suites, browsers and learning platforms. A prohibition aimed at one application does little when similar capabilities are available across a student’s digital environment.
The debate has consequently moved from a binary question – whether students should use AI – toward questions of purpose, disclosure, assessment design, data governance and AI literacy. This shift is more demanding than prohibition because it requires institutions to distinguish productive assistance from inappropriate delegation. It also requires educators to decide which cognitive processes a task is intended to develop before deciding where AI should or should not be used.
AI as a Personal Tutor
Personalised tutoring remains one of the strongest educational cases for artificial intelligence. Benjamin Bloom’s influential work on the “2 Sigma Problem” showed that individual tutoring could produce substantially stronger educational outcomes than conventional classroom instruction.¹⁴ The obstacle has always been scalability: one-to-one attention is difficult to provide across large classes and even harder to distribute equitably across different income groups and regions.
AI can partially relax that constraint. An AI tutor can vary explanations, generate practice questions, adjust difficulty, translate material and provide immediate feedback. It can remain available outside school hours and, once deployed, operate at low marginal cost. For learners who cannot afford private tuition or who live where specialist teachers are scarce, that availability could be educationally significant.
The phrase “personal tutor” can overstate what current systems understand. A generative model may adapt its language to a learner without possessing a reliable model of that learner’s conceptual state. It can provide a fluent explanation while missing the misconception underneath the question. It may also produce incorrect information with confidence. Personalisation of tone is not the same as diagnosis of understanding. A more realistic role for AI is to provide additional instructional capacity within a wider learning design, rather than to act as an autonomous replacement tutor. It can widen access to explanation and practice, but important misconceptions still need routes to human diagnosis and correction. Fluency should not be confused with pedagogical understanding. No doubt, AI is likely to change teaching more substantially than many earlier technologies, but the more plausible transformation is a redistribution of work rather than the disappearance of the teacher. The OECD has identified the potential for AI to reduce repetitive workload and allow educators to devote more time to pedagogy and interaction.¹⁵ Used well, this could shift teacher time away from administrative production and toward diagnosis, discussion, feedback and student support.
Teachers notice hesitation, interpret behaviour, motivate reluctant learners, resolve conflict, exercise judgement and provide continuity across time. They create expectations and classroom cultures. They also carry responsibilities that a model cannot meaningfully assume, particularly when educational decisions affect a child’s development, opportunity or welfare.
AI may increase the importance of the teacher by changing what expertise looks like in practice. Teachers will need to decide when AI adds value, design tasks that preserve necessary intellectual effort, interpret machine output and identify errors or omissions. Strong subject knowledge becomes more, not less, important when educators must evaluate material that they did not produce themselves.
Assessment and Academic Integrity
Assessment is perhaps the area most directly disrupted by generative AI because conventional education has often inferred knowledge from produced work. A student submits an essay, report or piece of code and the institution assumes that the artefact provides evidence of the student’s understanding. Generative AI weakens that inference.
A competent piece of academic prose can now be generated rapidly. AI can propose arguments, restructure language and imitate formal academic style. The resulting work may not fit conventional definitions of plagiarism because it need not reproduce an existing source. The issue goes beyond copying. Institutions must decide whether submitted work still provides credible evidence of an individual student’s capability.
This changes the central assessment question. Rather than asking only whether AI was used, educators increasingly need to ask what the learner understands and which parts of the demonstrated capability genuinely belong to the learner. That may increase the value of oral examinations, presentations, project defence, supervised writing, annotated drafts and iterative assessment in which students explain how their reasoning developed. AI can also be incorporated into assessment rather than treated only as a threat to it. Students might be asked to generate an AI response, identify factual and conceptual weaknesses, verify evidence, expose hidden assumptions and construct a better answer. Such tasks recognise that future competence may involve evaluating machine-generated knowledge as much as producing information independently.
Assessment should measure what it says it measures. If an assignment is designed to test argument construction, outsourcing the argument defeats the purpose. If the goal is to test critical evaluation of external material, AI may legitimately become part of the task. Assessment design should therefore begin with a clear statement of what the task is intended to test.
AI Literacy and Critical Judgement
The widespread presence of AI makes AI literacy an educational objective in its own right. That literacy involves much more than knowing how to formulate prompts or obtain polished output. Students need to understand what AI systems can do, how they can fail and when their outputs require verification.
Generative systems can fabricate information and students must therefore learn to compare sources, inspect evidence, distinguish fact from interpretation and recognise when a plausible answer is not a reliable answer. UNESCO’s AI Competency Framework for Students identifies a human-centred mindset, ethics of AI, AI applications, and AI system design as its four competency dimensions.16
Generative AI lowers the cost of producing an answer. The educational challenge is to ensure that students can still determine whether that answer is accurate, relevant and well supported.
Personalisation and the Future of Learning
AI could weaken one of the deepest structural assumptions of mass schooling: that large groups of learners must move through broadly similar material at broadly similar rates. Future educational systems could maintain longitudinal models of a learner’s strengths, misconceptions, prior knowledge and progress, then vary explanation, sequence and difficulty accordingly.
A system might recognise that a student struggles with symbolic algebra but understands visual geometry particularly well. It could use that strength to reshape explanations, alter practice and adjust pacing. Personalisation of this kind could reduce boredom for advanced learners and frustration for those who need more time, keeping students closer to an appropriate level of intellectual difficulty.
The educational promise is inseparable from a governance problem. A system capable of modelling a child over many years could also accumulate an unusually intimate educational profile: areas of weakness, behavioural patterns, interests, language ability, attention patterns and perhaps inferred emotional states. The boundary between useful personalisation and persistent surveillance could become narrow.
Personalisation should therefore be treated as a design choice with limits. Educational institutions need to determine what information is genuinely necessary for learning, how long it should be retained, who can access it and which inferences should remain off limits. More personalised systems are not automatically better educational systems.
Accessibility and Inclusion
AI can make education more accessible through speech recognition, text-to-speech conversion, translation, captioning and simplified explanations. Students with visual, auditory or learning disabilities may benefit from material being presented in alternative formats. Multimodal AI could extend this further by allowing learners to move between text, speech, image and interactive representations according to need.
Accessibility, however, is not guaranteed by technical capability. Systems often perform best in languages, accents and contexts that are strongly represented in training data. Learners working in less-resourced languages may receive weaker explanations, poorer speech recognition or less culturally appropriate examples. A system that works impressively in a dominant language may be much less capable elsewhere.
Meaningful inclusion requires systems to work across linguistic, cultural and accessibility differences in practice. The relevant question is not whether AI can support inclusion in principle, but whether the systems actually deployed work reliably for the learners who most need that support.
AI and the Global South
The potential significance of educational AI may be greatest in parts of the Global South where schools operate under persistent constraints: shortages of specialist teachers, crowded classrooms, uneven access to learning materials and limited availability of private tutoring. In such settings, even imperfect digital support can have a different marginal value than it does in already well-resourced systems.
AI could support teachers with lesson planning, content adaptation and explanation. Learners without access to private tuition might obtain supplementary academic assistance through relatively inexpensive devices. Multilingual systems could also widen access to material for students whose strongest language differs from the language of instruction. These possibilities make educational AI potentially important not only as an efficiency technology but as an access technology.
The same systems, however, can create new forms of dependency. Leading foundation models are concentrated within a small number of companies and countries, and their training data disproportionately reflects globally dominant languages, institutions and cultural contexts. A learner may gain unprecedented access to explanation while receiving that explanation through systems whose assumptions, classifications and knowledge priorities were largely determined elsewhere. This is more than a question of access. Students may gain powerful educational tools while becoming dependent on models developed elsewhere. If educational systems become dependent on a small number of externally developed models, greater access to knowledge may coexist with reduced control over how that knowledge is represented, ranked and interpreted.
For the Global South, access alone is not enough. Countries also need local capacity in languages, datasets, educational content, evaluation benchmarks, teacher capability and technological expertise. Countries do not need to build their own frontier models. But they do need the ability to question, adapt and govern the systems through which students increasingly encounter knowledge.
Hallucination and Bias
One of the most important educational limitations of generative AI is its capacity to produce inaccurate information fluently. Students have traditionally associated linguistic coherence, speed and confidence with expertise. Generative AI weakens that association because an incorrect answer can appear polished, structured and authoritative. The problem is not merely that AI sometimes makes mistakes. Human sources also contain errors. The distinctive educational risk is that the interface can collapse the visible difference between knowledge, inference and fabrication. A student may receive each in the same confident tone unless the system explicitly signals uncertainty.
This makes the ability to judge evidence increasingly important. Learners need habits of asking where a claim comes from, what evidence supports it, whether independent sources agree and what degree of uncertainty remains. Paradoxically, unreliable AI may teach an important lesson: verification cannot be skipped.
Artificial intelligence systems learn from human-generated data, and those data reflect historical inequalities, institutional practices and cultural assumptions. Bias becomes especially consequential when AI is used not simply to support learning but to influence decisions about learners.
Predictive systems could affect admissions, assessment, student placement, or decisions about which learners are considered at risk. The OECD has emphasised that algorithmic systems require governance capable of identifying discriminatory outcomes and protecting educational fairness.¹⁷
Efficiency is not a sufficient justification for automated decision-making. A statistically useful prediction can still be inappropriate when applied to an individual, and opaque correlations may reproduce disadvantage without making the mechanism visible. Educational institutions remain responsible for the consequences of systems they choose to deploy. Human oversight must involve real authority, not ceremonial approval. If an educator is expected merely to approve an algorithmic recommendation, responsibility has been preserved in name but not in practice. Fairness requires the capacity to challenge, override and explain consequential decisions.
Privacy and Student Data
Educational AI systems may collect unusually detailed information about learners. A persistent tutor could know a student’s academic weaknesses, patterns of error, interests, pace of progression and potentially behavioural signals gathered across years of interaction. Such information can improve personalisation, but it can also create highly intrusive profiles. Children warrant especially strong protection because they cannot reasonably be expected to understand the long-term implications of extensive digital data
collection or negotiate meaningfully with powerful technology providers.UNESCO has therefore emphasised privacy, age appropriateness and human-centred governance in its recommendations concerning generative AI in education.¹⁸ Institutions need clear rules concerning ownership, access, retention, secondary use and commercial exploitation of student data. Data should not be collected simply because a system can collect it. A technology that improves learning by a small margin while creating a permanent and opaque behavioural record may impose a cost that education is not justified in accepting.
Cognitive Offloading and Intellectual Dependency
The deepest educational challenge raised by Artificial intelligence may not be cheating, privacy or even factual error. It may concern the development of the human mind itself. Technology has always allowed people to externalise cognitive tasks: writing externalised memory, calculators reduced the burden of arithmetic, and search engines reduced the need to retain large quantities of retrievable information.
Generative AI extends cognitive offloading into activities much closer to the centre of intellectual work. A learner can ask a system to formulate an argument, choose examples, create a structure, rewrite a weak paragraph and produce a conclusion. Each individual use may appear harmless. The developmental question arises when assistance becomes habitual and the learner repeatedly delegates the very process through which capability is supposed to form.
Learning often requires productive difficulty. Constructing an argument, retrieving information from memory, testing an explanation or remaining with a problem after an initial failure can be uncomfortable because these activities demand cognitive effort. That effort is not incidental to learning; in many cases it is part of the mechanism by which understanding and skill are built.
AI creates a subtle risk because it can improve the immediate quality of output while reducing the underlying development of the person producing it. A student may submit better prose while becoming less capable of structuring an argument independently. A programmer may produce working code faster while understanding less of the logic. Efficiency at the level of the task can therefore conceal weakness at the level of the learner.
This does not imply that cognitive offloading is inherently harmful. Experts routinely use tools to remove low-value work and focus attention on higher-order problems. The key distinction is developmental stage and educational purpose. Offloading a skill that has already been acquired is different from offloading the process through which that skill is being acquired.
The Changing Value of Knowledge
The widespread availability of AI raises an increasingly common question: why should students learn information that machines can retrieve or generate instantly? A superficial response is that factual knowledge matters less when external systems can provide answers on demand. That conclusion confuses access to information with the possession of understanding.
Knowledge may become more important in an AI-rich environment because it provides the foundation for judgement. A person with little historical knowledge cannot easily recognise fabricated history. A student without mathematical understanding cannot reliably evaluate an AI-generated calculation. Someone unfamiliar with a scientific field may be unable to distinguish a plausible explanation from one that violates basic principles.
Critical thinking is sometimes discussed as though it were a general skill independent of content. In practice, judgement depends heavily on what a person already knows. Verification requires reference points. Questioning an answer requires some understanding of what might be missing, inconsistent or improbable. AI therefore does not eliminate the need for knowledge; it changes one of the reasons knowledge matters.
Curriculum design must reflect this shift. Education should not respond to generative AI by abandoning foundational knowledge in favour of generic “AI skills” or prompt techniques. Learners still need disciplinary concepts, vocabulary, methods and historical context. Students still need knowledge of their own if they are to recognise when an external system is incomplete, misleading or wrong.
Education may place greater emphasis on interpreting, connecting, contextualising and challenging information. These higher-order activities still depend on disciplinary knowledge. In an age of abundant answers, judgement becomes more important precisely because it requires a sufficiently rich internal model of the world.
From Answering Questions to Asking Better Questions
Traditional educational assessment has often rewarded students for producing correct answers to questions created by teachers. Generative AI changes the economics of that arrangement because acceptable answers can increasingly be produced immediately. The value of formulating meaningful questions may therefore rise.
Students will need to determine which problem matters, what assumptions are embedded within an answer, what evidence is absent and what alternative explanations remain possible. The OECD has suggested that generative AI may encourage education to place greater emphasis on inquiry, ambiguity and distinguishing fact from opinion.¹⁹
Questioning should not be reduced to “prompt engineering”. A technically effective prompt can still pursue a trivial question. What matters more is problem formulation: deciding what deserves investigation, where the problem begins and ends, and what would constitute a satisfactory answer.
The Future Teacher
The future teacher is unlikely to compete with AI on speed of information retrieval or first-draft content production. The comparative advantage of the teacher will increasingly lie in interpretation, mentorship, motivation, ethical reasoning and the design of meaningful learning experiences.
AI may handle repetitive elements of lesson preparation and administration while educators devote more attention to discussion, diagnosis and feedback. But this shift will not happen automatically. Institutions could just as easily use AI to increase administrative throughput, standardise instruction or expand class sizes without improving the quality of teacher-student interaction.
Teacher preparation must keep pace with these changes. Educators need enough understanding of AI to know where it can strengthen learning, where it creates dependency and how to verify its outputs. AI competence should become part of teacher education rather than remain an optional technology skill available only to enthusiasts.
Teachers will also have to defend intellectual standards. As polished output becomes easier to produce, assessment will need to place greater weight on evidence, reasoning, explanation and ownership of thought. This gives good teaching a different kind of importance: it must make learning visible even when performance can be assisted by machines.
Universities and Higher Education
Artificial intelligence also challenges the traditional value proposition of universities. Historically, universities concentrated knowledge, expert teachers, libraries, laboratories and professional networks in particular places. The internet weakened the scarcity of information. Generative AI may now weaken the scarcity of explanation. Students can increasingly obtain sophisticated explanations outside formal classrooms and receive immediate help without waiting for office hours. This does not make universities obsolete, but it changes what is scarce within higher education. Expert mentorship, intellectual community, laboratory access, collaborative research, professional networks and trusted certification may become relatively more important than routine content delivery.
Assessment credibility may become especially valuable. As high-quality written output becomes easier to generate, institutions that can credibly demonstrate what an individual actually knows and can do may acquire greater value. Universities may matter less as exclusive providers of explanation and more as trusted places for development, verification and intellectual community.
This also creates pressure on pedagogy. If a lecture merely repeats information that students can obtain more flexibly from an AI tutor, its value will be questioned. Higher education will need to offer what a general-purpose model cannot easily reproduce: sustained intellectual relationships, authentic projects, access to expert communities, contested discussion and credible opportunities to demonstrate independent capability.
Lifelong Learning
Artificial intelligence may also accelerate a transition from episodic education toward lifelong learning. The traditional sequence of education, qualification and employment assumed that a substantial body of knowledge could be acquired early in life and then applied across a relatively stable career. Rapid technological change makes that model less reliable.
The future pattern is more likely to involve repeated cycles of learning, work, adaptation and reskilling. AI tutors could identify gaps, recommend resources, create practice environments and provide personalised support throughout an individual’s career. Education could consequently become less associated with a particular age or institution and more closely integrated with professional activity. This development would also blur the boundary between formal and informal learning. A worker may learn through an AI assistant embedded in daily tasks rather than through a discrete course. The educational challenge will be to ensure that convenience does not replace depth and that new capability can still be assessed and recognised credibly.
A Neohumanist approach to education takes this discussion to a deeper level. It does not only focus on knowledge or skills but also on what kind of human being should education help to shape. In its formulation, humanism is extended towards universalism, embracing concern not only for other people but for animals, plants and the wider living world. Education therefore combines intellectual development with empathy, ethical discernment, ecological consciousness and service. It seeks to free the mind from narrow social, cultural and geographic sentiments through rational inquiry while expanding the learner’s
sense of responsibility beyond the self.This also means resisting the reduction of learners to future workers, examination scores or increasingly detailed data profiles. A Neohumanist education sees the learner as a developing moral and intellectual being whose identity should not be confined by nationality, religion or other inherited boundaries. Education, in this sense, is partly an emancipatory process: it enables people to examine the assumptions shaping their thinking while expanding the circle of beings towards whom they feel concern. In an AI-rich environment, educational success therefore cannot be measured only by speed, personalisation or output quality. AI can support learning, but it cannot determine its ultimate purpose. Technology should enlarge the learner’s capacity for independent judgement, compassion and responsible action rather than substitute for the development of conscience, relationship and meaning.
Human-Centred Governance
The future of AI in education should not be determined entirely by technical capability or commercial incentives. Education serves purposes beyond efficiency. It contributes to citizenship, culture, critical reasoning, ethical development, social participation and the formation of independent adults.
AI governance in education should begin with educational objectives, not with a list of tasks that technology can automate. Human agency, transparency, privacy, fairness, accountability and inclusion should be treated as design requirements rather than afterthoughts. Humans must remain responsible for consequential educational decisions, and algorithmic recommendations should support professional judgement rather than silently replace it.
Possible Futures
The future of AI in education is unlikely to follow a single path. Three institutional patterns are already plausible, and each places a different value on human capability.
In the first, AI becomes a substitution technology. Students delegate increasing amounts of difficult intellectual work to machines, while institutions respond by expanding detection and restriction. Output quality may remain high even as confidence in what students can independently do declines. The defining problem in this pathway is a widening gap between performance and capability.
In the second, AI becomes primarily an efficiency technology. Lesson preparation, marking and administration become faster, but curriculum, assessment and classroom structures change little. Institutions gain productivity without reconsidering which forms of learning remain valuable when explanation and content production are inexpensive.
In the third, AI becomes a capability technology. Students use it to test ideas, encounter alternative explanations and receive targeted support, while still demonstrating reasoning that does not depend on the system. Teachers use AI to extend pedagogical capacity, but retain authority over learning goals, assessment and consequential decisions. These pathways are not determined by model capability alone. They depend on institutional choices about assessment, teacher preparation, data governance and the kinds of intellectual effort that education deliberately preserves. The central design question is therefore not how much AI can be added to education, but what kind of learner an AI-rich education is intended to produce.
Human in the Lead
Much AI governance literature uses the phrase “human in the loop” to describe systems in which a person remains involved in an automated process. In education, that formulation is too weak. A human can remain in a loop while the machine effectively determines the direction, frames the choices and narrows the available alternatives.
Education requires a stronger principle: the human should remain in the lead. AI may recommend a learning pathway, but educators should determine the purpose of learning. AI may produce an answer, but students should decide whether the answer deserves acceptance. AI may identify patterns in educational data, but people must determine what actions are appropriate and what values those actions serve.
Education cannot be judged only by speed, cost or output quality. Its purpose includes developing people who can act, judge and take responsibility. A student who produces excellent work only with machine support has demonstrated a different capability from one who can understand, defend and evaluate that work independently.
“Human in the lead” connects governance with pedagogy. It requires real human authority over consequential decisions and meaningful intellectual agency for the learner. The objective is not to keep people ceremonially present around AI, but to ensure that AI remains subordinate to educational purposes set by people.
Conclusion
Artificial intelligence in education is the product of a long historical development rather than a sudden technological rupture. Teaching machines introduced self-paced instruction and immediate feedback. Intelligent tutoring systems introduced learner modelling. Machine learning and learning analytics enabled educational systems to identify patterns in student data. Online education expanded reach, while generative AI has now made conversational and content-generating intelligence widely available.
The post-2022 transformation is nevertheless significant because AI has moved directly into the intellectual activity of students and teachers. The potential benefits are substantial. AI can improve personalisation, accessibility, tutoring, teacher productivity and lifelong learning. It may extend educational support to learners who currently lack specialist instruction and help teachers redirect time from repetitive tasks toward higher-value interaction. For parts of the Global South, these capabilities could be especially important where educational demand exceeds available specialist capacity.
The risks are equally consequential. AI can weaken the credibility of conventional assessment, encourage cognitive dependency, reproduce bias and create intrusive forms of educational surveillance. It may also deepen global inequalities if language coverage, model ownership and technological capability remain concentrated. These risks cannot be solved by technical accuracy alone because many arise from the way
AI changes human behaviour and institutional incentives.
The central educational task is to decide which human capabilities must still be cultivated as machines take on a growing share of intellectual work. That shift increases the value of judgement. An education system fit for the AI age must teach students to think both with AI and without it. Artificial intelligence should extend human intellectual possibility without displacing the development through which people become capable of judgement. That is the practical meaning of keeping the human in the lead, and it should define the future relationship between AI and education.
Notes
¹ Digital Education Council, Global AI Student Survey 2024 (2024).
² Organisation for Economic Co-operation and Development, OECD Digital Education Outlook 2023: Towards an Effective Digital Education Ecosystem (Paris: OECD Publishing, 2023).
³ Sidney L. Pressey, “A Simple Apparatus Which Gives Tests and Scores—and Teaches,” School and Society 23 (1926): 373–376.
⁴ B. F. Skinner, “The Science of Learning and the Art of Teaching,” Harvard Educational Review 24, no. 2 (1954): 86–97.
⁵ John McCarthy, Marvin L. Minsky, Nathaniel Rochester and Claude E. Shannon, “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence” (1955).
⁶ Donald L. Bitzer, Peter G. Braunfeld and W. K. Lichtenberger, “PLATO IV: An Economical Computer Based System for Education,” Proceedings of the National Computer Conference 42 (1973).
⁷ Hugh Burns and Charles G. Capps, “Foundations of Intelligent Tutoring Systems: An Introduction,” in Foundations of Intelligent Tutoring Systems (Hillsdale, NJ: Lawrence Erlbaum Associates, 1988).
⁸ Jaime R. Carbonell, “AI in CAI: An Artificial-Intelligence Approach to Computer-Assisted Instruction,” IEEE Transactions on Man-Machine Systems 11, no. 4 (1970): 190–202.
⁹ Kurt VanLehn, “The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems,” Educational Psychologist 46, no. 4 (2011): 197–221.
¹⁰ Ryan S. J. d. Baker and Kalina Yacef, “The State of Educational Data Mining in 2009: A Review and Future Visions,” Journal of Educational Data Mining 1, no. 1 (2009): 3–17.
¹¹ Katy Jordan, “Initial Trends in Enrolment and Completion of Massive Open Online Courses,” International Review of Research in Open and Distributed Learning 15, no. 1 (2014).
¹² OECD, OECD Digital Education Outlook 2023.
¹³ Fengchun Miao and Wayne Holmes, Guidance for Generative AI in Education and Research (Paris: UNESCO, 2023).
¹⁴ Benjamin S. Bloom, “The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring,” Educational Researcher 13, no. 6 (1984): 4–16.
¹⁵ OECD, OECD Digital Education Outlook 2023.
¹⁶ UNESCO, AI Competency Framework for Students (Paris: UNESCO, 2024).
¹⁷ OECD, OECD Digital Education Outlook 2023.
¹⁸ Miao and Holmes, Guidance for Generative AI in Education and Research.
¹⁹ OECD, OECD Digital Education Outlook 2023.