Introduction

Artificial intelligence is altering the economics of education in India because a growing share of explanation, summarisation, translation, question generation and first-stage feedback can now be produced at very low marginal cost. The significance is not simply that AI systems have become more capable. Access to premium or near-premium AI services has also been distributed through consumer channels that are already used by millions of Indian households. Reliance Jio currently promotes an 18-month Google AI Pro benefit for eligible subscribers, while Airtel’s earlier Perplexity Pro promotion offered a 12-month subscription to eligible customers during its redemption period. OpenAI separately offered a 12-month ChatGPT Go promotion in India before ending that promotion on 21 January 2026.^[1][2][3]

These commercial arrangements should not be misunderstood as permanent public entitlements or proof that premium artificial intelligence has become universally free. Eligibility conditions, promotional periods, telecom plans and product terms can change. They nevertheless reveal an important economic direction. General-purpose AI capability is increasingly being bundled with products that households already purchase. When the cost of obtaining explanations and generated study material falls, the scarce element in education shifts. The expensive part becomes less about producing another answer and more about ensuring that the learner understands it, practises independently, identifies errors, follows a disciplined timetable, receives appropriate intervention and can demonstrate genuine mastery.

This shift has major implications for schools, universities, coaching institutions, teachers, families and education technology companies. A learner may have access to an AI system that can generate a constitutional law explanation, solve an algebraic problem, translate a science concept into Bengali or Hindi, create a mock examination and provide instant feedback. Yet the availability of such functions does not automatically establish that the learner has acquired durable knowledge. The difference between performance with assistance and learning that survives after assistance is removed is becoming one of the central questions in education policy.

What AI changes in education economics

Traditional educational products frequently bundle several distinct services into one price. A coaching fee may pay simultaneously for lectures, notes, test series, teacher access, administrative supervision, peer competition, doubt clearing and the reputation of the institution. Generative AI separates some of these functions. Explanations, examples, model answers, translations, revision plans, flashcards and large numbers of practice questions can be generated almost instantly. As the marginal cost of producing this content falls, institutions that continue to charge primarily for the scarcity of notes or recorded lectures face increasing competitive pressure.

The effect resembles other information markets in which digital technology reduced the price of reproduction but did not eliminate demand for trusted professional services. Education has an additional complication because the consumer often cannot immediately determine whether the product has worked. A student may enjoy a clear explanation and still fail to recall the concept one week later. A parent may observe frequent app use without knowing whether independent problem-solving has improved. A university may receive polished assignments without knowing whether the student formed the argument. Educational value is therefore not identical to the quantity or apparent quality of information supplied.

This creates an important distinction between answer production and learning production. An answer-production system is judged by whether it returns useful material quickly. A learning-production system must be judged by whether the learner develops knowledge or skill that persists, transfers to unfamiliar problems and can be demonstrated without inappropriate assistance. The latter requires instructional design, measurement and often some form of accountability. Artificial intelligence can participate in that process, but the existence of a capable model does not establish that the surrounding educational system has been designed well.

Access is expanding faster than effective use

ASER 2024 provides an important picture of digital access among rural adolescents. Among surveyed children aged 14 to 16, 89.1 per cent reported a smartphone available at home, 82.2 per cent reported that they could use a smartphone, but only 31.4 per cent of smartphone users reported owning one personally. The same data showed gender differences in individual ownership, with boys more likely than girls to report having their own device.^[4] These distinctions matter. A device present in the household is not the same thing as continuous personal access, privacy or freedom to use it for study.

ASER is especially valuable because its digital component included task-based assessment rather than relying exclusively on confidence or self-report. Older children were asked to perform practical digital tasks. This helps distinguish basic operational capability from mere exposure to technology. Yet even task-based digital literacy is not equivalent to the capacity required for responsible use of generative AI. A learner may know how to search, upload an image or use a messaging application and still lack the ability to determine whether a generated legal citation exists, whether a scientific explanation is internally consistent or whether a fluent answer has silently changed the meaning of a source.

The emerging divide is therefore layered. The first layer concerns connectivity, devices and affordability. The second concerns foundational literacy and numeracy. The third concerns information judgement: whether a learner can inspect evidence, distinguish a source from a generated claim, recognise uncertainty and ask a question precise enough to reveal rather than conceal misunderstanding. Artificial intelligence can reduce barriers for strong learners while increasing dependency among learners who lack these foundations. The same tool can therefore narrow one inequality and widen another.

Language creates both an opportunity and a risk. AI systems can offer explanations in Indian languages and can translate difficult material into simpler forms. For students who have historically faced a shortage of high-quality instructional content in their preferred language, this may be significant. At the same time, quality is unlikely to be uniform across languages, domains and dialects. Indian institutions should not assume that performance demonstrated in English automatically transfers to Bengali, Hindi, Marathi, Tamil, Nepali or other linguistic settings. Accuracy needs to be measured independently.

What experimental evidence actually shows

Research published during 2025 provides a useful warning against simplistic claims that artificial intelligence either inevitably improves learning or inevitably weakens it. A randomised controlled trial involving 194 students in an undergraduate physics course at Harvard compared a purpose-built generative AI tutor with an active-learning classroom lesson covering the same material. Students using the AI tutor learned more during the intervention, completed the material in less time and reported stronger engagement and motivation.^[13] The result demonstrates that a carefully designed AI tutor can produce substantial benefits in a particular educational setting.

The details of that study are as important as the headline. The intervention was not merely unrestricted access to a general chatbot. The tutor was deliberately designed around pedagogical principles, relevant content and structured prompts. It therefore supports a narrower proposition: a well-designed AI instructional system can be effective. It does not establish that any chatbot, used in any subject, by any age group, will outperform a teacher. External validity remains important because university physics students in one institution differ from primary-school pupils, competitive-examination candidates and students studying law, medicine or literature.

A different field experiment by Hamsa Bastani, Osbert Bastani and colleagues examined nearly one thousand high-school mathematics students. Students with access to GPT-4 performed better while they could use the system, but those using a relatively unrestricted interface subsequently performed worse when AI access was removed than students who had never received that assistance. A more carefully designed tutoring version, which provided teacher-designed hints and avoided simply revealing solutions, largely mitigated the negative learning effect.^[14] The study provides unusually clear evidence that immediate assisted performance and underlying skill acquisition can move in different directions.

Taken together, these findings shift the important policy question from whether AI should be used to how AI should be used. A system that asks guiding questions, withholds final answers until appropriate, diagnoses common errors and requires the student to generate intermediate reasoning may support learning. A system that instantly completes every task may improve short-term output while reducing the cognitive work required for skill acquisition. The educational unit of evaluation should consequently be the entire interaction design, not simply the underlying model name.

Where human teaching retains value

The strongest case for human teaching is not that humans will always possess more information than machines. That proposition is already difficult to defend for many factual tasks. Human value lies increasingly in observation, judgement, responsibility and social context. A teacher may notice that Arjun repeatedly avoids a particular type of mathematics problem, that Riya can reproduce a legal rule but cannot apply it to facts, or that a student who usually participates has abruptly stopped submitting work. Such observations derive from continuity of relationship rather than from the generation of another explanation.

Human teaching also carries an accountability function. When a coaching institution promises weekly testing, the value is partly that someone has decided what should be tested, established conditions, reviewed performance and intervened after poor results. AI can automate portions of each stage, but an institution that accepts responsibility for the final system creates a different product from a general chatbot that simply responds to whatever the learner asks. Parents and students may therefore continue paying for education even when AI is inexpensive because they are purchasing a reliable process rather than information alone.

Social and motivational factors are equally significant. Education involves delayed rewards. Students must often practise material that is difficult, repetitive or temporarily discouraging. A generated study plan cannot by itself ensure that the plan is followed. Technology may improve reminders, personalisation and responsiveness, but sustained effort frequently depends upon expectations created by teachers, peers, families and institutions. This is especially important in examination systems where months of structured preparation precede a single high-stakes test.

Schools are moving from software use towards computational thought

CBSE Circular Acad-15/2026 introduced a Computational Thinking and Artificial Intelligence curriculum framework for Classes III to VIII from the 2026 to 2027 academic session. The framework is aligned with the National Education Policy 2020 and places emphasis on computational thinking, problem-solving, logical reasoning, digital literacy and responsible technology use.^[5][6] This is a significant curricular change because it moves AI education below the secondary stage and treats computational thinking as a general capability rather than a specialised subject reserved for older students.

The distinction between teaching artificial intelligence and teaching with artificial intelligence is essential. Young learners do not need to memorise the names of current commercial models. They need to understand sequencing, classification, patterns, decomposition, evidence, error and the difference between a confident answer and a justified answer. Many of these skills can be developed without continuous screen use. An unplugged classroom exercise in which children design instructions, test them and identify where the instructions fail may teach computational thinking more effectively than passive exposure to an AI application.

Curriculum publication is an institutional commitment, not evidence of implementation quality. Teacher preparation, timetable allocation, language support, assessment design and access to appropriate materials will determine whether computational thinking becomes an intellectual habit or another chapter to memorise. Evaluation should therefore distinguish implementation inputs from learning outcomes. Purchasing devices or establishing an AI laboratory may be necessary in some settings, but these activities should not be used as substitutes for evidence that children have acquired the intended capabilities.

Higher education must examine learning processes, not only final submissions

India’s higher education system is operating at very large scale. The Ministry of Education released AISHE reports for 2022 to 2023 and 2023 to 2024 in July 2026. The 2023 to 2024 report recorded approximately 4.50 crore students in higher education, compared with 4.46 crore in 2022 to 2023. It also reported a national Gross Enrolment Ratio of 30 for the 18 to 23 age group and 17.32 lakh faculty members.^[7] At this scale, generative AI creates an assessment problem that cannot be solved by individual vigilance alone.

A conventional take-home assignment once carried an implicit assumption that producing several pages of coherent analysis required significant student effort. That assumption is no longer reliable. A student can now generate a plausible essay, citation list, problem solution or case summary within minutes. The educational response should not be to assume that all polished work is dishonest. It should be to redesign assessment so that the evidence supporting a grade is stronger. Institutions increasingly need to assess both the final product and the student’s demonstrated command of the process.

A stronger assessment system can combine controlled and open-tool components. Students may be permitted to use AI for brainstorming or language support but required to maintain a source record, identify which suggestions they rejected, explain their reasoning and demonstrate the same principle in a new problem. A short viva can establish whether a student understands a submitted dissertation chapter. A law student who submits a written answer may be asked to apply the same doctrine to a modified factual situation without AI assistance. A programming student may be asked to explain why the code works and repair an unseen error.

Academic integrity cannot be reduced to an AI detector score

The University Grants Commission’s 2018 regulations on academic integrity remain important because they establish institutional mechanisms for plagiarism and define levels of plagiarism in academic work.^[8] Generative AI, however, creates a different evidentiary problem. Text similarity tools ask whether wording overlaps with existing material. AI detectors attempt to infer how text was produced from statistical features of the writing. These are not equivalent questions, and a similarity percentage should not be interpreted as an AI probability.

Research evaluating AI-generated text detectors has found substantial reliability problems. A study in the International Journal for Educational Integrity tested multiple detection systems and concluded that the tools produced both false positives and false negatives and were not sufficiently reliable to serve as evidence of academic misconduct on their own.^[15] Detector performance also changes as models, writing styles and editing methods change. A score generated today may reflect a proprietary method that an institution cannot independently inspect.

Fair procedure therefore requires contextual evidence. When authorship is questioned, an institution should examine the assignment instructions, student’s drafts, sources, version history where available, oral explanation and any disclosed use of technology. The student should have an opportunity to respond. An AI-detector output may be one investigative signal, but converting it directly into punishment creates both evidentiary and fairness problems.

Student data turns the educational tutor into a legal and governance question

Generative AI systems used in education can process considerably more than a student’s name and email address. A conversation may reveal academic weakness, disability, family circumstances, health concerns, religion, location, examination anxiety or other personal information. The data may also include uploaded assignments, photographs, voice recordings or information about minors. The value of an AI tutor therefore cannot be assessed only by the quality of its answers. Institutions need to understand how the system handles personal data.

India’s Digital Personal Data Protection Act, 2023 defines a child as an individual below eighteen years of age. Section 9 provides for verifiable parental consent before processing children’s personal data and restricts processing likely to cause detrimental effects, tracking or behavioural monitoring of children, and targeted advertising directed at children, subject to statutory exceptions and exemptions.^[9] The legal timing is particularly important. A commencement notification issued on 13 November 2025 provides that sections 3 to 17 of the Act, which include section 9, commence eighteen months from the publication of that notification.^[10]

The Digital Personal Data Protection Rules, 2025 were also notified in November 2025 with phased commencement. Rules dealing with several substantive operational requirements are scheduled to commence eighteen months after publication.^[11] As of 19 September 2026, institutions therefore need to distinguish provisions already in force from duties scheduled to commence later. The phased legal timetable should not, however, be mistaken for a reason to postpone privacy engineering. Education systems that begin collecting large volumes of children’s data now may create future compliance and trust problems if retention, consent and deletion processes are not designed from the beginning.

UNESCO’s guidance on generative AI in education similarly emphasises privacy protection, age-appropriate use, institutional validation and a human-centred approach.^[12] UNESCO guidance is not Indian statutory law, but it provides a useful governance framework, especially where technology is developing faster than institutional rules. Educational institutions should consequently evaluate safety and pedagogical fitness before large-scale deployment rather than assuming that popularity in the consumer market establishes suitability for children.

The education market is moving from access to accountable outcomes

Telecom bundling provides a visible example of how quickly the price of AI access can change. Jio’s current promotion distributes Google AI Pro through eligible mobile arrangements.^[1] Airtel’s 2025 Perplexity promotion made a 12-month Pro subscription available to eligible customers during a redemption period that ended in January 2026, although customers who redeemed it could continue for the applicable benefit period.^[2] OpenAI’s 12-month ChatGPT Go promotion in India also ended in January 2026.^[3] These examples demonstrate why an education business built primarily around reselling access to a model faces strategic risk.

The more durable business model is likely to combine interchangeable AI infrastructure with educational systems that are difficult to commoditise. These include locally aligned curriculum design, trusted teachers, supervised examinations, parent communication, structured revision schedules, validated question banks, individual diagnosis, disability support and credible outcome measurement. Model providers may change, but the educational job remains.

This also changes how commercial claims should be read. A company may advertise millions of users, large numbers of questions answered or many hours of engagement. These are product metrics, not automatically learning metrics. A subscription’s stated retail price is also a commercial reference point rather than evidence that every learner receives equivalent educational value. Education purchasers should ask whether the product improves performance on independently designed assessments, whether benefits persist when AI assistance is removed and whether gains occur across different student groups.

Opportunities for India

India’s scale creates unusual opportunities for responsible educational AI. The combination of large student populations, widespread smartphone access, substantial linguistic diversity and uneven teacher availability makes personalised digital support potentially valuable. A high-quality system could provide additional practice to rural learners, help teachers generate differentiated exercises, translate difficult concepts, support students who require repeated explanations and extend learning beyond classroom hours. The opportunity is especially significant where the alternative is not an expert private tutor but no individual support at all.

India’s multilingual environment also creates a strong case for domestic evaluation capacity. AI products should be tested for factual accuracy, cultural context, readability and harmful error in Indian languages rather than merely translated from English benchmarks. Legal, medical and civic information requires particularly careful evaluation because plausible errors can create consequences outside the classroom. Public institutions and universities could develop open benchmark sets covering school subjects, professional education, competitive examinations and regional languages.

Teacher augmentation is another opportunity. Artificial intelligence can assist with lesson planning, first-stage feedback, question generation, translation and administrative drafting. Time saved on repetitive tasks may be redirected towards individual diagnosis and discussion. The benefit is not guaranteed because poorly integrated technology can also create additional checking work. Institutions should therefore measure actual teacher time before and after deployment rather than assuming automation has reduced workload.

Assessment reform may become one of the most valuable national applications. India conducts examinations and university assessments at enormous scale. AI can help generate question variants, classify common errors and provide formative feedback, while human-controlled systems preserve examination validity. The objective should not be to automate judgment indiscriminately. It should be to use technology where it increases feedback frequency while retaining human responsibility where a decision affects progression, certification or disciplinary consequences.

Research opportunities for Indian universities

Indian universities should treat educational AI as an empirical research field rather than merely a technology-adoption programme. The strongest research questions concern learning outcomes, equity, language, assessment validity, teacher workload, privacy and institutional governance. India offers conditions that are difficult to reproduce elsewhere because the same intervention can be tested across urban and rural settings, multiple languages, different school systems and wide variations in educational resources.

Study design should anticipate rapid model change. Researchers should record the exact model, access date, system configuration and relevant prompts because a commercial service may change during a longitudinal study. Primary outcomes should be specified in advance where feasible. Attrition should be reported. Subgroup analysis should examine whether the intervention benefits strong and weak students similarly. Comparison groups should receive credible alternatives such as teacher feedback, additional practice or peer tutoring rather than no support at all.

A structured evaluation framework for educational AI

Schools, universities and coaching institutions need a repeatable method for deciding whether an AI product should be adopted. The following framework is analytical rather than a statutory standard. It combines learning science, assessment validity, privacy governance and operational questions. Its purpose is to prevent institutions from choosing products solely on the basis of impressive demonstrations or commercial reputation.

Limitations, competing viewpoints and methodological cautions

The argument that human accountability will remain valuable should not be converted into a claim that present institutional arrangements are inherently superior. Human teaching can be inconsistent, expensive, inaccessible or poor in quality. Some students may receive more patient and individually tailored explanation from an AI tutor than from an overcrowded classroom. Artificial intelligence may therefore replace some services that are currently delivered by humans, especially where those services consist largely of standard explanation or repetitive practice.

A second viewpoint is that future AI agents may become capable of monitoring progress, adapting instruction, contacting parents, generating assessments and maintaining long-term student profiles. If these capabilities become reliable and socially accepted, some functions presently described as human accountability may also be automated. That possibility should be treated as a forward-looking hypothesis rather than an established outcome. Technical capability does not itself resolve questions of legal responsibility, consent, child safety, institutional trust or certification.

Third, experimental evidence remains limited relative to the diversity of education. A positive result in university physics does not establish the same effect in primary literacy. A finding from mathematics does not automatically apply to legal reasoning, history or creative writing. Models also change rapidly. A study that tests one model version may be difficult to reproduce after the provider updates the system. Educational AI therefore requires continuous rather than one-time evaluation.

Fourth, measuring learning is itself difficult. Examination scores can improve because students become familiar with a question format rather than because conceptual understanding has deepened. Self-reported engagement can rise even when retention does not. Platform usage can reflect novelty. High completion rates may simply indicate easier tasks. Strong evaluation should use multiple measures, including delayed testing, transfer, explanation, error diagnosis and unaided performance.

Finally, there is an equity tension. Restricting AI in order to preserve traditional assessment may disadvantage students whose peers use it privately. Permitting unrestricted use may advantage students with better devices, stronger prompts, paid plans or greater digital literacy. Institutions therefore need explicit rules and accessible provision rather than leaving AI use to informal household advantage.

Future directions

The next phase of educational AI is likely to be less about whether students can obtain a chatbot and more about how AI becomes embedded into learning systems. General-purpose models will continue to compete on capability and price, while educational institutions will compete on the quality of the surrounding structure. This may include validated curricula, supervised assessments, local-language support, teacher intervention and evidence that students can perform independently.

Schools should develop age-appropriate AI literacy rather than a collection of temporary product tutorials. Universities should redesign assessment around process evidence and transfer. Coaching institutions should use AI to increase practice and feedback while protecting the value of teacher diagnosis. Government policy should support independent evaluation, multilingual benchmarks, teacher development and responsible procurement rather than measuring progress merely by licences purchased or devices distributed.

Privacy and governance will also become more central as AI systems retain longer educational histories. A tutor that remembers several years of a student’s performance could become extremely useful, but the same capability creates a detailed behavioural profile. Questions of retention, access, correction, consent and portability will therefore move from technical details to central educational governance issues.

The strongest future systems may combine artificial and human intelligence deliberately. AI can handle high-frequency explanation, adaptive practice and administrative work. Teachers can focus on judgement, misconceptions, motivation, difficult cases and the social dimensions of learning. Assessment can use technology for frequent low-stakes feedback while reserving controlled conditions and human review for high-stakes certification. Such a division of labour is more plausible than either complete automation or an attempt to preserve education exactly as it existed before generative AI.

Conclusion

Artificial intelligence is changing education first by changing the price of educational language. Explanations, summaries, translations, practice questions and draft feedback can now be generated at a scale that would previously have required substantial human time. Telecom promotions in India illustrate how rapidly access can be bundled into ordinary consumer services. Yet cheap access to intelligence-like output does not make education free because education is not simply the delivery of answers.

Experimental research already shows both sides of the change. Carefully designed AI tutoring can improve learning, but unrestricted assistance can raise performance during practice while weakening later independent performance. The lesson is therefore not that AI is inherently beneficial or harmful. The design of the educational interaction matters. Systems should be judged by durable learning, transfer, independent reasoning and safety rather than fluency or engagement alone.

For India, the opportunity is substantial. Widespread smartphone access, large student populations, multilingual demand and expanding AI curricula create conditions in which technology can extend support far beyond the traditional classroom. The corresponding responsibility is equally substantial. Schools and universities need better assessment, stronger privacy governance, transparent rules and independent evidence. Commercial providers need to demonstrate outcomes rather than merely access.