From the Knowledge Garden — where we nurture ideas that bridge the gaps between policy dreams and classroom realities

Ishaan is fourteen, sitting cross-legged on his bed in Bengaluru, laptop balanced on a cushion, a history assignment open in one tab and a chatbot open in another. The question in front of him asks him to explain, in his own words, why the Non-Cooperation Movement collapsed in 1922. He reads it once. He opens the second tab. He types the question in, almost verbatim. Eleven seconds later, a clean, well-structured, entirely correct four-paragraph answer appears on his screen.

Ishaan does not copy it word for word, his school's plagiarism checker is good enough now that copying verbatim is a rookie mistake nobody makes twice. Instead, he reads the answer, absorbs its shape, and retypes it in his own words, restructuring two sentences and swapping out a few phrases so it reads as unmistakably his. The assignment is submitted eleven minutes after it was opened. It will receive a good grade. Ishaan could not, if you asked him tomorrow without the chatbot open, reconstruct more than the first sentence of his own answer.

Two floors down, his younger sister Diya, eleven, is doing something that looks almost identical from across the room, laptop open, chatbot open, homework in progress. But Diya is stuck on a math word problem, and instead of asking for the answer, she's asked the chatbot to give her a hint without solving it, because her school's AI literacy teacher spent a full period last month teaching exactly this distinction: the difference between using a tool to get unstuck and using a tool to get finished. Diya works through three wrong attempts, guided by increasingly specific hints, before she solves the problem herself. She could reconstruct her own reasoning tomorrow, because it actually happened.

Same tool. Same age of technology. Two entirely different outcomes, sitting in the same house.

This is, in miniature, the real question this piece sets out to answer, not "is AI good or bad for Indian students," a framing too blunt to be useful, but the sharper question buried inside it: what determines whether a fourteen-year-old with a chatbot ends up like Ishaan, or like Diya?

This piece continues the Knowledge Garden's exploration of Indian K-12 education, following The Tuition Trap, The Learning Poverty Number Nobody Talks About, The Substitute Economy, and Skill Versus Certificate. Those pieces traced a chain from teacher vacancies to foundational learning gaps to a graduate employability crisis rooted, in no small part, in a rote-memorisation culture that never quite built the reasoning skills employers now demand. AI arrives into exactly this system, at exactly this moment, and depending entirely on how it's used, it is either the single fastest way to finally close that reasoning gap, or the single fastest way to make it permanently, invisibly worse.

Unlike most of the structural issues this series has examined, decades-old teacher recruitment bottlenecks, a coaching economy built up over thirty years, this one is moving at a genuinely unusual pace, which is precisely why it deserves attention now rather than in a future retrospective piece.

In April 2026, the Central Board of Secondary Education, working with an expert committee chaired by an IIT Madras professor, formally mandated Artificial Intelligence and Computational Thinking (AI & CT) as a compulsory curriculum component starting from Class 3, beginning in the 2026–27 academic session, with Classes 9–10 following the year after. This is not a pilot programme or an optional elective anymore, CBSE had already offered AI as a skill subject for Classes 9–12 and run a 15-hour "SOAR" (Skilling for AI Readiness) module across more than 18,000 schools for Classes 6–8, but the 2026 mandate represents something structurally different: AI literacy embedded into daily instruction across subjects, for every child, starting at age eight, under the explicit framing of "AI for Public Good" set out by the Ministry of Education.

The scale of what this touches is worth sitting with. India runs one of the largest school systems on Earth, roughly 24.7 crore students across 14.7 lakh schools, taught by over one crore teachers, the same workforce whose recruitment and training gaps we documented in The Substitute Economy. A curriculum decision made in Delhi in April 2026 does not stay a policy document; within a single academic year, it becomes the daily classroom reality for tens of millions of children, most of them taught by teachers who have had, at best, months to prepare for a subject that didn't formally exist in their own training.

Meanwhile, adoption on the ground has already outpaced the policy. The India AI Impact Summit 2026 reported that roughly half of private-school students in Delhi already use generative AI tools multiple times a week, well ahead of any curriculum mandate, which is precisely the dynamic that makes this piece urgent rather than theoretical. Indian higher education institutions, per the FICCI-EY-Parthenon AI Adoption Survey 2025 covering 30 leading institutions, show 57% already operating with a formal AI policy in place and a further 40% actively developing one; on the ground, 53% already use generative AI to build learning materials, 40% have deployed AI tutoring systems, and 38% use AI for automated grading. Yet a separate 2025 study across 45 schools, 30 colleges and 15 universities found only 3% of institutions have deeply integrated AI into structured, intentional workflows, meaning the overwhelming majority of current use, at every level, is informal: a teacher opening a chatbot on their own initiative, a student like Ishaan quietly using one after homework hours, rather than a genuinely designed, school-wide approach to what AI literacy should actually look like.

That gap, widespread informal adoption racing ahead of structured, thoughtful implementation, is exactly the space in which the Ishaan outcome and the Diya outcome both currently live, side by side, often in the very same classroom.

It would be dishonest, and it would do a disservice to Indian parents and teachers genuinely weighing this question, to treat the concern about AI and cognitive decline as mere technophobia. The research base, while still young, is no longer merely anecdotal.

A 2025 study surveying 666 participants across diverse age groups and educational backgrounds, using both quantitative testing and in-depth interviews, found a significant negative correlation between frequent AI tool usage and critical thinking abilities, and, notably, that this relationship was mediated specifically by cognitive offloading: the well-documented tendency to delegate a mental task to an external tool and, in doing so, exercise the underlying mental faculty less over time. The same study found younger participants showed both higher dependence on AI tools and lower critical thinking scores than older participants, a finding with obvious, direct relevance to a country now introducing AI literacy to eight-year-olds.

This isn't an isolated finding. A broader 2025/2026 review of digital technology's cognitive effects concluded that while generative AI can genuinely support critical thinking under guided, reflective use, unstructured or excessive reliance shifts a learner's effort away from analysis and evaluation and toward simple acceptance of whatever the AI produces, with the reported cognitive gains clustering at lower-order tasks (recall, basic comprehension) while higher-order reasoning actually weakens under what researchers term "automation bias": the tendency to over-trust an authoritative-sounding output rather than evaluate it independently. A CHI meta-analysis of 17 separate studies, cited in this review, found that while AI use produces large overall learning gains on average, those gains are attenuated or even reverse into losses specifically for higher-order thinking skills, precisely because of this offloading effect.

Perhaps the single most vivid piece of evidence comes from outside the classroom entirely, but carries an obvious warning for it: a clinical study found that just three months after doctors began using AI assistance to help detect tumours in medical scans, their own unassisted ability to detect those same tumours had dropped by 6%, a striking, concrete illustration of a skill visibly atrophying once a tool reliably does the work instead. A separate, not-yet-peer-reviewed MIT study using EEG brain-activity monitoring during essay writing found that participants using a large language model showed measurably weaker neural connectivity during the writing task than those using a plain search engine or no external tool at all.

Translate this directly into Ishaan's bedroom. He has not learned to explain why the Non-Cooperation Movement collapsed. He has learned to retrieve, lightly rephrase, and submit an explanation generated by something else, a skill that resembles academic reasoning closely enough to earn a good grade, while being, in the ways that actually matter for his own long-term thinking ability, close to its opposite. Multiply Ishaan by a meaningful share of India's roughly 25 crore students, repeated across thousands of assignments over a school career, and the risk is not abstract: an entire generation could emerge more fluent than any before it at producing polished-looking work, and measurably less practiced at the underlying reasoning that work is supposed to represent.

This risk lands with particular force on a system this series has already shown is struggling with exactly this problem before AI entered the picture. Our piece on Learning Poverty documented that more than half of India's ten-year-olds cannot yet read a simple age-appropriate text with genuine comprehension. Our piece on graduate employability documented that the widest, most persistent gap employers report is not technical knowledge but reasoning, problem-solving and the ability to handle an unfamiliar situation, precisely the muscle that unguided AI use, per the research above, appears to weaken fastest. Introducing a powerful cognitive-offloading tool into a system that was already under-building reasoning skills, without simultaneously and deliberately teaching how to use that tool without offloading the thinking itself, is a genuinely serious risk, not a hypothetical one.

And yet, this is not a story with only one honest ending, and the research above, read completely rather than selectively, actually supports a second, more hopeful conclusion sitting right alongside the first.

The same body of research that documents AI's risk to critical thinking consistently, explicitly, distinguishes between two fundamentally different modes of use. One 2026 review put the distinction plainly: passive use leads to skill decay, but more structured, deliberate use may actually boost critical thinking and creativity. This is not a hedge or a footnote, it is close to the central finding across the literature. AI does not have a fixed cognitive effect. Its effect depends almost entirely on how it is used, and specifically on whether the human using it remains the one doing the reasoning, with AI supporting that reasoning, or whether the human has quietly handed the reasoning itself over to the tool.

This is exactly the distinction Diya's school built into a single lesson: the difference between a hint and an answer. It is a small pedagogical design choice, and it is also, per the research base above, close to the entire ballgame.

Used this way, AI's advantages for exactly the problems this series has already documented are substantial and specific, not vague or promotional:

It can be a tireless, infinitely patient Socratic partner for a child whose classroom has forty-plus students and one teacher. Our Substitute Economy piece documented, in granular detail, why individualised attention is structurally unavailable to most Indian schoolchildren, Rukmini simply cannot give a struggling nine-year-old and an advanced nine-year-old truly personalised instruction inside a single fifty-minute period, forty more times a week, for every subject. A well-designed AI tutor, used the way Diya's school taught her to use it, asking guiding questions rather than supplying answers, can approximate exactly the kind of individualised, level-appropriate scaffolding that our Learning Poverty piece identified as one of the most effective, and previously most resource-constrained, interventions available for closing foundational gaps.

It can meet a child in their own language, at a scale India's edtech sector has historically struggled to reach. Our earlier Knowledge Garden piece on the vernacular gap and the digital divide identified India's overwhelming English-first edtech landscape as a structural barrier for the roughly 70% of children who learn primarily in Indian languages. Large language models with genuine multilingual capability offer a plausible, previously unavailable path to delivering that same Socratic scaffolding in a child's mother tongue, rather than forcing translation through an English-first interface that was never designed with them in mind.

It can make the invisible visible for a teacher managing an unmanageable classroom. Used well, AI-assisted tools can flag, in real time, exactly which students are guessing rather than reasoning through a problem, exactly which concept a struggling reader is stumbling on, exactly where a class's collective understanding is thinnest, the same diagnostic information Rukmini, teaching three grades in one room, currently has no realistic way to gather at scale, but that the CBSE's own AI & CT framework, in its "AI for Public Good" framing, gestures toward as a genuine institutional goal rather than a private convenience.

Structured, deliberate use appears to genuinely build the reasoning skills this series has repeatedly identified as India's deepest gap. The same research base that documents cognitive offloading's risks also documents that instructional practices built around active, participatory engagement, group discussion, problem-based learning, guided AI use that requires justification and evaluation of the AI's own output rather than passive acceptance, are associated with measurable gains in exactly the critical thinking skills this series' employability piece found employers say Indian graduates lack most.

Read this way, AI is not inherently the villain of Ishaan's homework story or the hero of Diya's. It is closer to a mirror, or an amplifier: it will do, more efficiently than any tool before it, whatever the surrounding pedagogical culture asks of it. Pointed at a system that still rewards a polished final answer over a visible, evaluated reasoning process, precisely the rote-memorisation culture our earlier "Rote to Reasoning" piece examined, it will produce more Ishaans, faster and more convincingly than ever before. Pointed at a system deliberately redesigned around process, justification, and guided struggle, it has genuine, well-evidenced potential to produce more Diyas, at a scale and a cost no previous generation of edtech has managed.

Conversations with teachers navigating this transition in real time reveal a workforce discovering, largely without a rulebook, exactly which classroom practices tip a child toward Ishaan's pattern or Diya's.

A Class 8 CBSE teacher piloting elements of the new AI & CT framework ahead of full rollout described the single most effective change she's made this year: redesigning homework assignments to explicitly require students to show their reasoning process, including any AI interaction they used, rather than simply submitting a final answer. A student asked to submit their chat transcript alongside their answer, she found, behaves entirely differently than one submitting a polished paragraph with no visible trail, the visibility itself changes how the tool gets used, nudging students toward Diya's hint-seeking pattern rather than Ishaan's answer-retrieval pattern, without a single word of moralising about honesty.

A school leader involved in early AI & CT teacher training, meanwhile, flagged the scale of the readiness gap the policy timeline has created: with over one crore teachers nationally needing some form of orientation to a subject that did not exist in most of their own training, and rollout beginning within a single academic year, the risk of the mandate outpacing genuine teacher readiness is real and immediate, echoing, in a new and faster-moving form, precisely the implementation-lag pattern our Substitute Economy and Skill Versus Certificate pieces both documented in slower-moving policy areas like teacher recruitment and vocational education targets.

Given everything above, the practical question for every stakeholder in Indian education is not whether to allow AI into classrooms, that decision has effectively already been made, both by national policy and by tens of millions of students already using these tools informally, ahead of any policy at all. The real question is how to structure its use so that the balance tips toward genuine reasoning-building rather than genuine reasoning-erosion, at the scale of an entire national school system.

For teachers and schools: The single most evidence-backed intervention available, drawn directly from the research above, is redesigning assessment to make the reasoning process itself visible and gradeable, chat-transcript submission, "explain your AI interaction" reflection prompts, oral defence of written work, rather than continuing to grade only the polished final product a rote-oriented system has always rewarded. This is not a technology decision; it is the same pedagogical shift toward reasoning over recitation this series has argued for since its first piece, simply applied to a new tool.

For CBSE, NCERT and the Ministry of Education, as the AI & CT curriculum rolls out: The "AI for Public Good" framing already embedded in the national mandate is the right instinct; translating it into classroom practice will depend heavily on whether teacher training, promised alongside the curriculum, actually reaches the more than one crore teachers who need it before the mandate reaches their classrooms, precisely the recruitment-and-training-pipeline challenge this series' Substitute Economy piece already identified as India's persistent implementation bottleneck, now facing a compressed, single-academic-year timeline rather than a multi-year one.

For edtech companies building AI tools for the Indian K-12 market: The most responsible, and likely most differentiated, product design choice available is building tools that structurally default to Socratic scaffolding, hints, guiding questions, staged reveals, rather than direct answers, making the harder design choice the default one, rather than leaving that discipline entirely to a child's own self-restraint or a teacher's individual vigilance. This is also, not incidentally, a genuine market opportunity distinct from the answer-generating tools already crowding the space.

For parents: The distinction between Ishaan's use and Diya's use is rarely visible from outside a closed laptop, which makes this a harder parenting challenge than screen-time limits alone can solve. A more useful home practice than restricting AI access outright is periodically asking a child to explain, out loud and without the tool open, the reasoning behind a piece of AI-assisted homework, a low-cost, high-signal check that surfaces the Ishaan pattern quickly, in the same spirit as the direct reading-comprehension check our Learning Poverty piece recommended for catching literacy gaps early.

For students themselves, particularly older students capable of grasping the distinction directly: The research is unusually clear and worth stating to a student plainly rather than softening: asking an AI tool for a hint, a check on your own reasoning, or a challenge to your assumptions appears to build the exact skills employers and universities value most; asking it to simply produce the finished answer appears to erode them, measurably, over time. The tool is the same in both cases. The outcome is not.

For policymakers thinking beyond the current curriculum rollout: Given that roughly half of Delhi's private-school students were already using generative AI multiple times weekly before any curriculum mandate existed, the policy conversation cannot only be about what gets taught in a new mandatory subject, it must also grapple with the much larger, already-occurring, entirely informal use happening in every other subject, every evening, on personal devices the school has no visibility into at all.

Every major learning technology, from the printing press to the calculator to the internet search engine, has provoked some version of this exact debate, does this tool make us smarter, or does it make thinking optional? In most of those cases, history's answer turned out to depend less on the technology itself than on whether the surrounding educational culture adapted its methods to demand genuine understanding alongside the tool, or simply let the tool substitute for understanding altogether.

What makes this moment different, and genuinely higher-stakes, is the scale and speed at which it is arriving in India specifically, a national curriculum mandate reaching 24.7 crore students within a single academic year, layered on top of a system this entire Knowledge Garden series has already shown to be struggling, well before AI entered the picture, with foundational literacy, teacher shortages, and a rote-over-reasoning culture that graduate employers already report as India's costliest skills gap.

Ishaan and Diya are, right now, sitting in the same city, using the same technology, two floors apart in one house, and heading toward measurably different cognitive outcomes because of one small, deliberate pedagogical choice made by one teacher in one classroom. The genuinely encouraging finding buried in all the research above is that this choice is neither expensive nor mysterious. It does not require new infrastructure, new recruitment, or a new generation of teachers. It requires, as it has at every stage of this series, the harder and less glamorous work of designing how a tool gets used, asking for the reasoning, not just the answer, at the scale of an entire school system, before an entire generation's relationship with thinking itself is quietly decided by default rather than by design.

Further reading from the Knowledge Garden: