The Curriculum Arrived Before the Computers: India's AI Education Gamble

Outlook Business gave its issue of 3 August 2026 to a single question: is India preparing its schoolchildren for the AI age? The package runs to ten pieces, and it is unusually valuable, because the editors did something most magazines avoid. They assembled people who disagree with each other.

The credits, since they matter. The main feature and the interviews with Victor R. Lee (Stanford), Ashish Sood (Delhi's Education Minister) are by Zenaira Bakhsh. Rukmini Banerji (Pratham, ASER) is interviewed by Fozia Yasin; Arvind Virmani (former NITI Aayog member and former Chief Economic Adviser) by Parth Singh; and Alakh Pandey (Physics Wallah) by Shashank Bhatt. There are essays by Karthik Raman of IIT Madras, who chaired the panel that designed the new curriculum, Dr Jitendra Nagpal on the psychological dimension, career counsellor Jitin Chawla on jobs, and school principal Rajiv Kumar Sharma from the implementation end. Read them at outlookbusiness.com.

What follows is not a summary. It is that package taken apart, every figure checked against the government's own data, and the disagreements pushed until they become usable in an answer. Two numbers in the feature turned out not to match UDISE+, and one state appears in the wrong list. More importantly, the package contains a real policy fight over whether money is the constraint at all, and that fight is worth more to you than any single fact in it.

Cross-paper relevance

• GS2 — Education policy and the NEP, the 6%-of-GDP debate, federalism and the Concurrent List, Centre-State capacity gaps, equity in public service delivery, public-private partnership in education.
• GS3 — AI governance and ethics, the digital divide, sovereign AI capability, skilling, and technology's effect on employment.
• GS1 — Indian society: inequality of opportunity, rural-urban divides, language as a barrier to mobility.
• GS4 — Ethics: the limits of simulated empathy, the teacher-student relationship as moral rather than transactional, academic integrity, and technology grounded in human values.
• Essay — Demographic dividend versus demographic liability; technology and inequality; "education is the most powerful weapon."

Part I — What India actually built

On 1 April 2026, the Ministry of Education and CBSE launched a Computational Thinking and Artificial Intelligence curriculum for Classes 3 to 8, applicable from the 2026-27 academic year, releasing a curriculum document, a teachers' handbook and learning material. It followed the India AI Impact Summit at Bharat Mandapam, 16 to 20 February 2026, the first global AI summit hosted in the Global South, drawing over 100 countries and more than 20 heads of state.

Sharma, writing as a serving CBSE principal, adds the context that makes this less abrupt than it appears: CBSE has offered AI as an elective skill subject from Class 6 for the past three years, and reports rising demand from students and parents choosing it over other options. The Class 3 extension formalises and deepens something already under way. His historical comparison is apt, that this resembles the introduction of computers to Indian schools in the late 1980s and early 1990s, with one difference that he puts well: computing arrived as one more subject, whereas AI acts on every other subject.

The obvious objection is that this is too much, too early, for classrooms as varied as India's. Raman, who chaired the design panel, answers it directly, and his defence is the most intellectually substantial thing in the package.

The design. Classes 3 to 5 receive embedded computational thinking, delivered through puzzles, worksheets and existing subjects rather than a separate computing period, aligned chapter by chapter with the maths textbook. Classes 6 to 8 layer AI concepts and project-based learning on top. Computational thinking here means four specific habits: decomposition, pattern recognition, abstraction and algorithms. Sharma notes the fuller formula CBSE has adopted, computational thinking plus AI plus ethics plus "unplugged" learning, the last being the part that matters most for schools without devices, with the stated goal of AI-literate learners by 2030.

The pedagogy. Can a Class 6 child meet supervised, unsupervised and reinforcement learning? Raman's answer is to map the three onto experiences a child already has. A teacher collecting named notebooks is working with labelled data. Sorting unnamed notebooks by handwriting is finding structure without labels. A chess program that improves after each game is learning from rewards. No equations, no code, three names attached to three familiar things.

The principle is Jerome Bruner's, from The Process of Education (1960): the hypothesis that "any subject can be taught in some intellectually honest form to any child at any stage of development." Bruner also supplied the mechanism, the spiral curriculum, in which an idea is revisited at rising complexity rather than taught once. India's National Curriculum Framework prescribes exactly this. So Class 6 meets these ideas as awareness, Class 7 returns through AI's domains, and by Class 8 a student trains a small model. Three passes, each more honest than the last.

The ladder. Raman's most portable idea is a four-rung hierarchy worth memorising, because it disciplines a whole category of loose commentary:

1. AI literacy — knowing what these systems are and are not
2. AI skills — using and interrogating them well
3. AI knowledge — the mathematics and engineering underneath
4. AI development — building such systems

Schools must own rungs one and two and open a window onto three; rung four looks after itself. His statement of the goal is the line to remember: the ceiling has never been the problem, and the job of a national curriculum is to raise the floor. The constraint in Indian classrooms is not ability or aptitude but exposure, and exposure is precisely what a curriculum can manufacture.

Two further points travel well. First, AI cannot live inside the computer-science period. Machine translation belongs in the language class, classification is already taught in biology, geography runs on data and maps, and no civics discussion of social media is complete without asking how recommendation systems decide what teenagers see next. Second, there is a bridge problem: a child entering Class 7 in 2026 never rode the lower turns of the spiral, so every school adopting this during the decade has a cohort to ramp up.

Sharma adds an inversion worth borrowing for an essay. For decades the project was teaching machines to think like humans. A curriculum that trains children to decompose, abstract and algorithmise is, in part, teaching humans to think like machines, and it is worth pausing on what that implies rather than treating it as self-evidently good. He also flags the practitioner's immediate headache, that students already use chatbots aggressively for assignments and teachers increasingly cannot distinguish original work from AI-assisted work. That is an academic-integrity problem arriving faster than the assessment system can adapt.

So the policy is real, early by world standards, and thoughtfully designed. The question is what it lands on.

Part II — What it lands on

Here I part company with the feature slightly, and the correction runs in the same direction as its argument.

The feature states that 63% of schools have functional computers and 67% have internet. UDISE+ 2024-25, the Ministry of Education's own release published September 2025, reports otherwise:

• 64.7% of schools have computers, up from 57.2% in 2023-24, but only about 58% have functional computers. "Has a computer" and "has a working computer" are different questions, and roughly seven percentage points of India's schools sit in that gap.
• 63.5% have internet, up from 53.9%. Not 67%.
• By management, government schools stand at 58.6% internet against private schools at 77.1%.
• About 94,000 schools still have no electricity, against 93.6% that do.

The dataset covers 14.71 lakh schools and 24.69 crore students, down from 24.80 crore the previous year, and records more than one crore teachers for the first time since UDISE+ began in 2018-19, precisely 1,01,22,420.

The corrected picture is therefore worse than the reported one. Roughly two in five schools cannot reliably run a computer-based lesson. That is the floor beneath a Classes 3 to 8 AI curriculum.

In fairness to the design, the sequencing is unusually robust to this. Embedded computational thinking through puzzles and existing subjects needs no device, and the explicit inclusion of unplugged learning suggests the designers anticipated exactly this constraint. The infrastructure gap bites hardest at Classes 6 to 8, where students are meant to train a model and handle real datasets.

The divide is a map, not an average

All figures UDISE+ 2024-25.

A child in a Delhi government school and a child in a West Bengal government school receive the same CBSE circular and do not live in the same country. On internet access the ratio is better than five to one.

A correction the other way. The feature groups Chhattisgarh with Bihar and West Bengal as states where under a third of schools have functional computers. UDISE+ 2024-25 places Chhattisgarh at roughly 89% with computer facilities and about 93% functional, near the top of the table rather than the bottom. Bihar and West Bengal do belong there. If you cite state examples, use the two that hold.

The GS2 proposition this licenses: a uniform national curriculum delivered onto radically non-uniform state capacity does not produce uniform outcomes. It widens gaps, because the schools best placed to exploit a new curriculum are those that already had the most. Policy equal in form can be regressive in effect.

Sharma makes the same point from inside the system, noting that smart panels, tablets and self-paced platforms require investment and teacher training that strain what he calls a fragile school ecosystem, so not every school can afford to comply.

Part III — Delhi, and the limits of the best case

Sood's interview is useful precisely because Delhi is the easy case: a city-state, the national capital, top of every infrastructure table. What a well-resourced government names as its priorities tells you what binds even when money does not.

His stated dual focus is foundational literacy and numeracy alongside digital penetration, aiming to get schools NIPUN Bharat certified on FLN outcomes while building smart classrooms, language labs, ICT labs and digital libraries. Note that even in Delhi, FLN comes first.

On specifics he claims ICT labs in 175 schools where computers had been non-functional for a decade, built to CBSE standards with up to 40 machines each; smart digital libraries with Chromebooks; and an "AI Grind" programme in which five lakh students worked on ten city-centric problem statements across environment, healthcare and traffic, using A3 sheets rather than devices. These are a minister's claims about his own administration, which I have not independently verified, so attribute them as such.

Two things in his account are analytically interesting regardless.

His justification for starting early is defensive, not economic. He frames AI literacy around children encountering fakes and deepfakes and needing to judge whether a face they see online is real. That is an information-integrity argument, and it sits closer to Lee's position than to any skilling rationale.

He names a training pathology worth remembering. In his account the same handful of teachers are sent for training repeatedly, so capacity never broadens. His fix is to design modules and publish schedules six months ahead using the summer vacation, and to strengthen Delhi's Vidya Samiksha Kendra. "The same few teachers, again and again" is an excellent concrete example of implementation failure for a governance answer.

On the three-language question he takes the straightforward line that NEP 2020 settles it by providing that primary education be in the mother tongue or local language.

Part IV — The teacher question

India crossed one crore school teachers in 2024-25. Everything above depends on them, and the evidence here is the most uncomfortable in the package.

The Centre for Teacher Accreditation (CENTA) ran a nationwide survey in 2025, timed to five years of NEP 2020:

• Over 70% of teachers already use AI, rising to about 75% among those with more than three years' experience. Adoption is not the problem.
• The commonest use is lesson planning (about 60%); roughly 26% generate classroom activities.
• 67% rated their own AI expertise at 6 or above out of 10, averaging 7.
• But only 57% correctly answered a basic question testing a common misconception about AI.
• 84% reported apprehensions, led by job displacement in education (34%) and accuracy (23%).

Read the third and fourth bullets together. Confidence is running ahead of competence. That is harder to fix than ignorance. A teacher who knows she does not understand a tool is cautious with it; a teacher who wrongly believes she does will pass its errors on with authority, and in a classroom authority is what makes an error stick.

Four contributors converge here from different directions, which is as strong as qualitative evidence gets:

• Raman, from curriculum design, reports appetite rather than resistance, sharp questions and real adaptability, but insists a webinar is not training. Development must be sustained and embedded in the subjects teachers already teach, not delivered as a one-time inoculation. CBSE has made "Computational Thinking and Understanding AI" the training theme for the current year.
• Banerji, from four decades of measurement, argues teacher training is periodic and one-size-fits-all, and so does not equip a teacher for the specific composition of her own classroom.
• Sood, from administration, describes the same few teachers being trained repeatedly.
• Pandey, from the private sector, argues the bigger bottleneck is not student access but teacher capacity, and that the useful application of AI is reducing the repetitive load of lesson preparation, evaluation, gap identification and answering the same doubt many times over. His concrete image is a teacher walking into class already knowing which concepts confused most students yesterday and who needs attention today.

Raman's standing advice deserves quoting because it inverts the usual assumption: design every activity for the student in the room with the least prior exposure, not the one the syllabus imagines. That student is not the limit on the class, she is the starting line, and the design's whole point is that the line keeps rising.

Part V — The floor beneath the floor

No answer on this topic is complete without the prior question, and here Banerji's interview is indispensable.

The numbers. ASER 2024 found the share of Class 3 children in government schools able to read a Class 2 text rose from 16.3% in 2022 to 23.4% in 2024, a real post-pandemic recovery credited partly to NIPUN Bharat. It also means roughly three in four Class 3 children in government schools still cannot read a Class 2 text.

The government's own PARAKH Rashtriya Sarvekshan 2024, conducted by NCERT in December 2024 and released July 2025, tested 21.15 lakh students across 74,229 schools in 781 districts, replacing the old National Achievement Survey. National language average was 64%, up from 62% in 2021 but below 2017's 66.7%, with a steep decline by grade in mathematics: Grade 3 at 60%, Grade 9 at about 37%, the lowest for any subject at any grade.

Why has this stayed broken? Banerji's diagnosis is the most structurally useful material in the package, with four moving parts:

1. Children arrive unequal. Some have had preschooling, some none, and all meet the same Class 1 curriculum. NEP 2020 accordingly argues for universal, quality early childhood education.
2. Teaching aims at the top. Instruction is delivered to the whole class as though it were homogeneous, which Abhijit Banerjee in Poor Economics calls teaching to the top of the class.
3. The curriculum is overambitious and linear. A fixed syllabus must be finished each year and expectations harden each grade, so a child who starts behind falls further behind permanently. Lant Pritchett and Amanda Beatty call this the negative consequences of an overambitious curriculum.
4. Nobody notices in time. Families value schooling but often do not realise their child has fallen behind.

Her conclusion: unless catch-up and remediation are built into the primary grades, NEP's foundational goal cannot be met.

The remedy is known. Pratham's Teaching at the Right Level (TaRL) groups children by current learning level rather than age or grade. It is among the most rigorously evaluated education interventions anywhere: six randomised evaluations across seven Indian states, run with J-PAL over roughly fifteen years, producing some of the largest effect sizes measured in the education literature, with reach in the tens of millions across India and Africa.

That yields a clean line: the remedy for India's learning crisis is not unknown, it is known, tested and under-adopted. Banerji's live experiment is whether AI can help teachers do the grouping and matching TaRL requires, which would be differentiated instruction at scale.

Her scepticism about the fashionable framing matters equally. She resists treating AI as a personalised tutor, noting evidence on the effectiveness and durability of learning gained that way is still accumulating, and calls that a narrow view of what technology can do. Her preferred framing is AI as a friendly guide and companion, whose real prize is learning to learn through curiosity and choice rather than faster movement along the same linear track.

Part VI — What AI literacy should actually mean

Three contributors, working in completely different systems, arrive at the same conclusion. Their agreement is the strongest signal in the package.

Lee rejects the idea that AI literacy is prompt engineering, because interfaces change and a curriculum built on operating today's tools has a short shelf life. What endures is understanding how systems are trained on data to detect and predict patterns, judging whether AI suits a task or is a poor substitute for it, and grasping risks beyond the immediate task.

Raman makes the same move from the other side: AI education is not tool training. Students already experiment with chatbots and image generators, useful exposure, but the tools stay sealed black boxes, admired rather than understood, and prompting skill will never open the box.

Chawla, advising students on careers, independently reaches the same place: AI literacy is becoming as basic as computer literacy once was, students need not become AI engineers, but they must know what AI can do, where it fails, and how to use it responsibly. His formulation of the risk is the sharpest in the set: without subject knowledge, AI can create the illusion of competence.

A Stanford education professor, an IIT Madras curriculum designer and a working career counsellor independently rejecting the most popular definition of AI literacy is a convergence worth citing.

Lee's sharpest contribution reframes the divide itself. Asked whether a new digital divide is opening between those who can use AI and those who cannot, he argues the operating overhead is low and falling, since vendors are commercially motivated to make these tools easy. Anyone can type a request in plain language. The divide he fears is between users who understand that AI can miss things, edge cases, user errors, hallucinations, and users who accept whatever the system returns as correct and superior.

If that is right, the policy implication for India is severe. Hardware alone will not close the gap. Sending computers to Bihar is necessary and nowhere near sufficient. A well-equipped school that teaches children to trust AI output produces worse citizens than a poor school that teaches them to doubt it. The scarce input is not compute but a teacher who can model scepticism, which returns us to that 57% figure.

Lee's remaining points fold in neatly. On critical thinking, capacities we stop practising will weaken, and the danger is that the tasks we historically used to build judgement, writing an analytical essay, working to an original solution, are precisely the tasks AI does passably. His prescription is to shift assessment from valuing the product to valuing the process, and to treat AI as a choice rather than an imperative, with deliberate limits at younger ages. On good use, he points to project work, real datasets, prototypes and activities where students actively test the tool's limits, plus a genuine upside for children in rural areas and small schools who can reach subjects their school does not offer.

Raman reaches the same conclusion about assessment by another route, and states it more bluntly. If computational thinking and AI are examined through three-mark "write a short note" questions, the system will manufacture rote learners and reward the coached child over the curious one. A child who recites a definition of machine learning has memorised a string. A child who trained a model to sort photographs of her school's waste, documented what confused it and had a classmate review the work has learned something no examination can extract. Projects, portfolios and peer review are not soft alternatives to rigour; here they are the rigour.

Sharma's academic-integrity problem is the same coin's other face. If assessment continues to reward a finished product that AI can generate, teachers will keep struggling to tell original work from assisted work. Changing what is assessed is the only durable answer to cheating, and it is more effective than any detection tool.

Lee's closing inversion is the line to build an essay around: rather than AI transforming education, education should transform how we think about AI, stripping away the apparent magic to see it as impressive computing, useful for some tasks and limited for others.

Part VII — The human question, and why it is a GS4 answer

Nagpal supplies the dimension the technology debate usually omits, and it maps directly onto Ethics.

He opens by conceding AI's genuine strengths, which is what makes the argument credible. AI is never irritated by a question asked a fifth time. It adapts to pace, gives immediate feedback, works across many languages, is always available, and reduces the embarrassment that stops a child from asking. For students who learn differently, lack confidence, or live far from good schools, these are real democratising gains.

Then the limit. Education was never only the transfer of information. Schools are social environments where children acquire empathy, resilience, persistence, ethical reasoning and self-regulation. Before a child solves a maths problem she learns to trust, cooperate and feel secure. Teachers notice small shifts in behaviour, steady a child after failure, mediate disputes and make a classroom emotionally safe, and those interactions shape mental-health trajectories, not merely test scores.

His central distinction is the one to carry into GS4. Conversational AI can mimic empathy convincingly, detecting distress in language and returning comforting words. But it cannot feel compassion, cannot care, and cannot bear responsibility. What resembles emotion is prediction over words. An adult holds that distinction easily. A young child, facing an increasingly human-seeming system, may not, and the line between a real relationship and a simulated one can blur.

From this he raises the questions worth quoting in an ethics answer: what happens to a child's interpersonal development if emotional affirmation is sought mainly from a machine rather than from adults, and what is lost in collaboration and negotiation if classroom discussion gives way to one-to-one exchanges with software?

His resolution is not rejection but proportion: context, not the technology, decides the outcome; the notion of AI replacing teachers is the central misconception; and the object of education is not only informed individuals but kind and psychologically resilient citizens.

Pandey, from a very different vantage, lands in the same place: AI can only partly substitute for the human relationships that shape a child's education, particularly the motivation and mentorship teachers provide and the bonds students form with each other. His constructive version is that if AI absorbs routine personalisation, classrooms become more interactive, with teachers spending less time delivering content and more discussing ideas, mentoring and nurturing curiosity.

Part VIII — Whose context is the machine trained on?

The feature's strongest original argument concerns cultural context, and it survives scrutiny.

The reasoning: learning works better when new ideas are anchored in familiar contexts, and models trained overwhelmingly on Western data carry Western defaults in examples, assumptions, emphases and humour. The historical rhyme it reaches for is Macaulay's Minute on Indian Education, 2 February 1835, which proposed forming a class of persons Indian in blood and colour but English in taste, opinion, morals and intellect.

Keep the analogy, but state the mechanism precisely, because this is where a good answer separates from a rhetorical one. Macaulay's was a deliberate assimilationist design, announced as such and therefore politically contestable. Model bias is an emergent property of training data: unintended, diffuse, and correspondingly harder to identify or resist. "AI is the new Macaulay" is a slogan. Explaining why an unintended default is harder to contest than a declared policy is analysis.

Banerji sharpens the language dimension into something practical. Most tools are built in English or Hindi. In many Indian states English is introduced early in primary grades, yet teachers are often uncomfortable teaching it and children have little exposure outside school. Her question is whether AI can strengthen the language a child already uses and then employ first-language oral mastery as a bridge to a second. Her honest assessment is that AI's potential for learning across languages is largely undiscovered for young children, and what research exists is not yet available at scale or at a price that reaches primary classrooms.

Pandey turns this into a design brief, and it is the most concrete formulation in the package. Democratisation, he argues, does not mean giving everyone the same AI tool; it means every child having an equal opportunity to learn well. If products assume a premium subscription and uninterrupted high-speed internet, they will serve only a thin slice of students. What India needs is AI that is affordable, fluent in Indian languages, functional on inexpensive devices, frugal with data, and usable where connectivity is intermittent. He also states the foreign-model risk plainly: education is bound to language, culture, curriculum and local context, so depending entirely on models built for other education systems risks tools that do not understand Indian students or reflect Indian classrooms. His assessment that the internet's democratisation promise is roughly half fulfilled, access solved while learning stayed one-size-fits-all, is a good line for an introduction.

India's policy response is real and dateable. Under the IndiaAI Mission (approved 2024, outlay ₹10,371.92 crore), the government has backed 20 indigenous foundation-model proposals, 12 large and 8 small language models. In February 2026, Sarvam open-sourced two models trained on IndiaAI compute, of roughly 30 billion and 105 billion parameters, optimised across 22 Indian languages, alongside BharatGen's Param2. Compute has scaled to roughly 38,000 GPUs with a further 20,000 announced, and about 93.18 lakh GPU hours sanctioned across 237 projects.

The framing to use: sovereign AI is normally justified on national security and economic competitiveness. The education case is the underrated third pillar, because the cognitive environment children grow up in is shaped by whose examples the machine reaches for.

Part IX — The fight about money

Now the disagreement, the most valuable thing in the package and the part most aspirants will miss.

The consensus position. NEP 2020 reaffirmed the target of 6% of GDP on education, a benchmark traceable to the Kothari Commission (1964-66). In June 2026 the Parliamentary Standing Committee on Education, Women, Children, Youth and Sports, chaired by Digvijaya Singh, again flagged the shortfall, noting combined Centre and State spending of roughly 4.1% of GDP and urging effort toward 6%. (A precision note: both 4.06% and 4.12% circulate, for different reference years and bases. Cite it as "roughly 4.1% of GDP against NEP's 6% target, per the Parliamentary Standing Committee, June 2026," and you are safe on the number and right about the gap, which has persisted for six decades.)

Virmani's dissent. He argues the 6% target is close to meaningless as stated:

• These benchmarks derive from cross-country averages in World Bank indicators, without reference to the specific problem being solved or the benefit-cost ratio of the proposed solution.
• They ignore the constitutional division of responsibility. Health is a State subject; education was a State subject until moved to the Concurrent List; and every Finance Commission already adjusts for disparities in state income and tax capacity.
• Most damaging, his own work finds minimum reading ability is uncorrelated with Net State Domestic Product. If learning outcomes do not track state income, money cannot be the sole binding constraint.
• His proof by example: Uttar Pradesh, Jharkhand and Odisha significantly improved minimum learning outcomes over the past decade with little or no extra expenditure.
• His general complaint is that India generates a new scheme for every political problem and very little research on whether any of it raised income or welfare. His illustration: evidence points to sanitation and sewerage as a dominant driver of child stunting, yet state bureaucracies keep funding nutrition programmes instead of fixing sewage.

Verify the constitutional point, because it is examinable. Education moved from the State List to the Concurrent List by the 42nd Amendment, 1976, on the recommendation of the Swaran Singh Committee, and now sits at Entry 25 of List III. Virmani's "until the mid-1970s" is correct, and naming the amendment makes it land harder.

How to use this disagreement. Do not pick a side and preach. The examiner-pleasing move is to hold both: the 6% target is a legitimate statement of national priority and a real fiscal gap, and spending alone is demonstrably insufficient, on the evidence that outcomes are uncorrelated with state income and that three low-income states improved without extra money. The synthesis is that allocative efficiency and outcome measurement deserve at least as much attention as the aggregate, and that a target without a delivery mechanism is an aspiration, not a policy.

Virmani's structural metaphor is also worth borrowing. He describes the first thirty post-Independence years as building a "double cylinder": a large higher-education cylinder atop a schooling cylinder, which he calls unstable and in need of correction. Since schooling is primarily a State responsibility while the Union has more direct authority over higher education and R&D, his prescription is that States shift from credentials to learning while the Centre continues to push AI as a job skill. The two are not mutually exclusive.

Pandey adds the delivery-model question, which is a standard GS2 sub-theme. Asked whether AI should reach government schools through public infrastructure or private competition, he argues public systems supply reach and continuity while private players supply agility and iteration, and that partnership with the student at the centre beats either alone. His caution from the last edtech cycle is worth quoting in any answer on regulation: education runs on trust, earned slowly and lost quickly, and growth should be an enabler rather than the purpose. His proposed test for any product is disarmingly simple, whether students are actually learning better because of it.

Part X — Jobs, skilling and the dividend

The package's final thread is the one students actually ask about, and Chawla, Virmani and Pandey between them give it a spine.

Chawla's core distinction is that AI is not replacing professions in a single stroke, it is replacing tasks first, especially those that are repetitive, predictable, data-heavy or pattern-based. He is careful not to stop at the comfortable version: AI will also take on analysis, comparison, recommendation and strategy, so the reassurance that "only routine work is at risk" is too weak to rely on.

The roles he expects to shrink in their current form are specific and worth citing: data entry, basic transcription, routine customer support, manual report creation, simple content rewriting, first-level screening and scheduling support. His point is not that people in them have no future, but that the role must be upgraded. Someone who only enters data is exposed; someone who cleans it, interprets it, uses AI on it and turns it into decision support is not. Someone who only rewrites content is exposed; someone who understands communication strategy, audience behaviour and narrative is not.

He also names the roles being created, and the observation that matters is that they are not only for coders: AI product managers, workflow specialists, AI trainers, data curators, AI ethics and policy professionals, human-AI interaction designers and sector-specific consultants, drawing on business, psychology, law, healthcare, finance, education, design and ethics. For a GS3 answer on employment this is the useful corrective to the assumption that AI-era jobs are a programming monoculture.

His five student skills compress well: understand AI rather than fear it; learn to ask better questions; build genuine domain knowledge; strengthen critical thinking, because AI sounds confident when it is incomplete, biased, outdated or wrong; and develop the human capacities machines do not have, trust, empathy, communication, creativity, leadership and ethical judgement. His illustration is the one to remember, that a patient needs reassurance and not only a diagnosis.

For schools he proposes three changes that align neatly with Raman on assessment: a real-world project every term, formal teaching of communication and professional readiness rather than leaving it to college placement season, and a portfolio of evidence so that a student leaves school with projects, writing, research and demonstrated skills rather than marksheets alone.

His closing formulation is the sharpest sentence in the whole package, and it belongs in an essay: the real danger is not that AI replaces every job, it is that people who use AI well may replace those who do not.

Virmani supplies the macro case, and he is the most optimistic voice here. On skilling, he assesses India's vocational education and training system as roughly a tenth of the size its per-capita income warrants. Comparative data supports the direction: India's formally skill-trained share of the workforce is in the low single digits on most estimates, against figures in the seventies and eighties for Germany, Japan and South Korea, though definitions of "formally skilled" differ enough between sources that the exact numbers should be cited as approximate. His proposals are concrete: make internships and apprenticeships central, follow Germany in making on-the-job training commercially attractive to firms, permit and encourage CSR funds for training, phase out degree-based recruitment in favour of defined job-specific skills, and fix the shortage of trainers through collaboration with Germany, Australia and Canada while raising the social status of trainers and skilled workers.

On AI and jobs he pushes against the prevailing pessimism, reading recent research as showing employment rising at firms that use AI intensively while remaining broadly unchanged at modest users, with specific tasks eliminated but workers who learn the tools reallocated to different ones.

On the dividend, his argument is that India's demographic position is structurally unlike China's, the EU's, Japan's or the Anglosphere's: India's share of the global labour force will rise while theirs falls, positioning India as a supplier of human capital across skill levels. From which follows the sentence worth carrying: countries with shrinking populations will use AI to replace labour, while India should use it to raise the quality of its workers, developing the human-AI interface as a competitive advantage, including in health, education and government.

Set that against Mehrotra's warning, quoted in the feature, that failing to provide access to computing and AI will turn the demographic dividend into a demographic liability, and you have the two poles of the debate, both defensible, inside one magazine issue.

For students working through these questions practically rather than for the exam, the SAARTHI Career Portal is a school-run career-guidance resource of the kind Chawla is arguing every school should offer. (Noted for transparency: the portal is a JavaScript application that did not render for automated inspection, so this description rests on its stated identity rather than a full review of its contents.)

What to take into the exam hall

1. The policy is ahead of the plumbing. Curriculum launched 1 April 2026 for Classes 3 to 8. About 58% of schools have a functional computer, 63.5% have internet, per UDISE+ 2024-25. Both facts are true at once, and the tension is the answer.
2. Use the state spread, never the national average. Delhi at 99.9% against Bihar 25.2% and West Bengal 25.1% on computers is the sharpest illustration of unequal state capacity under a uniform national scheme.
3. The teacher finding is the non-obvious one. Over 70% already use AI, 67% self-rate 6+ out of 10, only 57% cleared a basic misconception check. Confidence exceeding competence is harder to fix than a training gap.
4. Foundational learning is the binding constraint. ASER 2024: 23.4% of Class 3 government-school children read a Class 2 text, up from 16.3% in 2022. PARAKH 2024: Grade 9 mathematics about 37%. And the remedy is known: TaRL, six randomised evaluations across seven states with J-PAL.
5. Redefine the divide. Following Lee, Raman and Chawla, AI literacy is not prompt engineering, and the gap that will matter is between those who can evaluate AI output and those who accept it. That makes teacher capability and critical thinking the scarce inputs, not hardware.
6. Assessment decides everything. Raman's point that three-mark definitional questions manufacture rote learners is the most practical criticism in the package, it answers the cheating problem better than any detection tool, and it generalises well beyond AI.
7. Hold the money debate on both sides. Roughly 4.1% against NEP's 6% is a real gap; and outcomes are uncorrelated with state income, with UP, Jharkhand and Odisha improving without extra spending. Education sits on the Concurrent List by the 42nd Amendment, 1976, Entry 25.
8. Design for the constraint, not the ideal. Pandey's brief, affordable, multilingual, low-data, works on cheap devices and intermittent connectivity, is the most concrete equity test available, and it applies to any technology-led welfare intervention.
9. Do not omit the human dimension. Nagpal's distinction, that AI can simulate empathy but cannot feel compassion or bear responsibility, is a GS4 answer in a single sentence.

The demographic-dividend argument is usually made as arithmetic: a young population is an asset if it is skilled. The AI turn sharpens it into something more precise. The dividend now depends less on how many children pass through school and more on whether they leave able to judge what a machine tells them. On current data, we are teaching roughly a quarter of Class 3 to read while teaching all of Class 3 to think computationally. Both sentences have to become true together, and only one of them is a technology problem.

For rolling coverage as the 2026-27 rollout proceeds and the next UDISE+ and ASER rounds land, see our sister site Ujiyari. For the exam foundations underneath this piece, work through our notes on education and the NEP, AI governance, ethics and policy, the digital divide and federalism.

Bharat

Sources and credit. Prompted by the Outlook Business package of 3 August 2026 on AI and school education: the main feature and the interviews with Victor R. Lee and Ashish Sood by Zenaira Bakhsh; the interview with Rukmini Banerji by Fozia Yasin; with Arvind Virmani by Parth Singh; with Alakh Pandey by Shashank Bhatt; and essays by Karthik Raman (IIT Madras), Dr Jitendra Nagpal, Jitin Chawla and Rajiv Kumar Sharma. Their text is theirs and is not reproduced here. Positions attributed to them in this piece are summarised from those interviews and essays and credited by name; Sood's claims about Delhi's ICT labs, Chromebooks and the "AI Grind" programme are a serving minister's account of his own administration and are flagged in the text as unverified. All supporting data was verified independently against the sources below. UDISE+ 2024-25, Ministry of Education, released September 2025, via PIB and the published report: 14.71 lakh schools, 24.69 crore students (from 24.80 crore in 2023-24), 1,01,22,420 teachers, 64.7% of schools with computers against about 58% functional, 63.5% with internet (58.6% government against 77.1% private), about 94,000 schools without electricity, and the state-wise spread for Delhi, Kerala, Andhra Pradesh, Uttar Pradesh, Bihar, West Bengal and Chhattisgarh. Ministry of Education and CBSE: Computational Thinking and AI curriculum for Classes 3 to 8, launched 1 April 2026 for academic year 2026-27, panel chaired by Karthik Raman, aligned to NEP 2020 and NCFSE 2023. India AI Impact Summit, Bharat Mandapam, 16 to 20 February 2026. Jerome Bruner, The Process of Education (1960), for the quoted hypothesis and the spiral curriculum. Centre for Teacher Accreditation (CENTA) nationwide survey, 2025: adoption above 70%, lesson planning about 60%, classroom activities about 26%, 67% self-rating 6 or above averaging 7, 57% correct on the misconception item, 84% reporting apprehensions. ASER 2024 (Pratham): Class 3 government-school reading of a Class 2 text at 23.4%, up from 16.3% in 2022. PARAKH Rashtriya Sarvekshan 2024, NCERT, conducted December 2024 and released July 2025: 21.15 lakh students, 74,229 schools, 781 districts, national language average 64%, Grade 3 mathematics 60% against Grade 9 at about 37%. Teaching at the Right Level: Pratham and J-PAL, six randomised evaluations across seven Indian states. Abhijit Banerjee and Esther Duflo, Poor Economics, for "teaching to the top of the class"; Lant Pritchett and Amanda Beatty on overambitious curricula. Parliamentary Standing Committee on Education, Women, Children, Youth and Sports, chaired by Digvijaya Singh, action-taken report, June 2026, on education spending of roughly 4.1% of GDP against NEP 2020's 6% target; Kothari Commission (1964-66) for the origin of that benchmark. 42nd Constitutional Amendment, 1976, on the recommendation of the Swaran Singh Committee, moving education to the Concurrent List, Entry 25. IndiaAI Mission, MeitY, approved 2024 with an outlay of ₹10,371.92 crore: 20 foundation-model proposals, Sarvam's two open-sourced models of February 2026 across 22 Indian languages, BharatGen's Param2, roughly 38,000 GPUs with 20,000 more announced, and about 93.18 lakh GPU hours across 237 projects. Macaulay's Minute on Indian Education, 2 February 1835. Comparative skilling figures are cited as approximate because definitions of "formally skilled" differ by source. Two corrections to the feature, made on the government's own data and flagged in the text: its figures of 63% functional computers and 67% internet do not match UDISE+ 2024-25, which gives about 58% functional and 63.5% internet; and Chhattisgarh, listed there among states below one-third on functional computers, stands at roughly 89% with computer facilities and about 93% functional, while Bihar and West Bengal do fit that description.