India's AI curriculum starts at Class 3, with puzzles not chatbots

The obvious way to teach a child about AI is to give the child an AI. India has just published a school curriculum that does close to the opposite, and the reasoning behind it deserves a look from parents here, not least because so many Singapore families have a cousin, a niece or a nephew sitting in one of those classrooms.

Read past the headline and the chatbots vanish

On 1 April 2026, the Education Ministry launched CBSE’s Computational Thinking and Artificial Intelligence curriculum for Classes 3 to 8, effective from the 2026-27 session, according to The Tribune’s report of the launch. Minister of State Jayant Chaudhary framed the point as children becoming “thoughtful creators rather than passive technology users”.

Then open the document itself. For Classes 3 to 5, roughly ages eight to eleven, it allocates 50 hours a year, and those hours are not a new subject with a new textbook. They are folded into Mathematics and The World Around Us, delivered through worksheets and what the curriculum calls “fun math games, puzzles, and exercises”. Assessment is written tests with puzzles, group activities such as treasure hunts, and a Teacher Observation Journal, with schools told to move away from traditional summative testing towards continuous, competency-based work.

AI as most people mean the word turns up later. Classes 6 to 8 get 100 hours a year, split between advanced computational thinking, an introduction to AI literacy, and interdisciplinary projects. The document is unusually blunt about the sequence: “AI is introduced later, once the pre-requisite knowledge of CT is built for understanding AI.”

So a Class 3 child in India will spend her AI hours finding patterns, breaking problems into steps and arguing about which method to use. Screens are not banned at this stage, and the competencies do include getting familiar with the parts of a computer, internet safety and block-based coding such as Scratch. But she will not spend those hours prompting anything.

The constraint that shaped the design

Some of that is conviction. Some of it is arithmetic.

Government figures for 2023-24 put functional computers in 57.2% of Indian schools and internet access in 53.9%, as the Education Ministry reported. The averages hide worse. A 360info analysis by Namesh Killemsetty, a public policy academic at O.P. Jindal Global University, notes that Bihar and West Bengal sit below 25% for computer access while Punjab has equipped 95.6% of its 19,243 government schools, across a system of roughly 250 million students in 1.5 million institutions.

A curriculum that assumed a device per child would have been a curriculum for a minority. CBSE wrote one that does not assume it: schools pick their own platform, and the framework leans on “free and open-source tools to ensure accessibility and adaptability for all institutions”. Fifty hours a year of pattern puzzles will run on paper in a room with no power socket.

Here is the part we find genuinely interesting. The version built for scarcity may be the better version. Decomposition, pattern spotting, algorithmic thinking: these are the capacities that outlast whichever model is fashionable in three years. A rollout with money for a laptop per child would have been sorely tempted to buy the laptops first and think about the pedagogy afterwards.

Singapore reaches the same answer from the opposite direction

Our own ministry got here through a completely different set of pressures, and landed in nearly the same place. In a parliamentary reply on 6 May 2026, MOE said that for Primary 1 to 3, “schools will not assign any work that requires them to use AI directly”. From Primary 4 pupils may use tools designed for education, under teacher supervision, because by then they “would have developed some executive functioning skills like planning, task initiation, and ability to evaluate their own thinking”. MOE added that “schools will also teach students when they should not use AI, so that they do not take short cuts to get answers without actual learning”.

India is working around patchy infrastructure at enormous scale. Singapore has near-universal devices and is worried about over-reliance. Two very different problems, one shared conclusion: build the thinking first, hand over the tool second. When two systems with almost nothing in common arrive at the same sequencing, that is worth more than either one saying it alone.

What the plan still has to survive

Teachers, mainly. A survey by the Centre for Teacher Accreditation of more than 5,000 educators, published in August 2025 to mark five years of the National Education Policy, found around 70% already using AI in their work, most often for lesson planning. It also found 67% rating their own AI skills as above average while only 57% could correctly identify a basic AI misconception. Confidence is running well ahead of knowledge, which is exactly the wrong order for the adult in the room.

The curriculum is also softer than the coverage suggests. Its own wording calls the time allocation “suggested”, adding that it is “flexible, and schools can go beyond the allocated time depending on the needs of the school”. CBSE is one board among many, so what reaches a Class 3 classroom in a district with no computer lab depends entirely on whether the worksheets and the teacher training actually turn up. That is an implementation question, and implementation is where curricula usually go to die.

To its credit, the document names the risk out loud. It asks for ethics teaching that helps students “verify information, detect bias, and recognise when AI reliance crosses into cognitive delegation or cognitive over-reliance”. We have written before about cognitive debt and about teaching children to fact-check an AI; it is rare to see a national curriculum put the phrase in writing.

The bit you can borrow tonight

You do not need CBSE, or MOE, or a subscription to anything. The design principle is free: the thinking comes before the tool.

Give a child a problem she cannot one-shot, and let the struggle sit there for a while. Ask her to say the steps out loud before she touches a screen. When she does use an AI, ask her afterwards how she would check it. That is most of what those 50 hours are, minus the worksheet.

We should say plainly that we have a stake in this view. Mentus AI builds mentors designed to make a child work rather than hand over answers, so a curriculum arguing that thinking precedes the tool is a curriculum arguing our case. Read us accordingly. The evidence still points where it points, and so, it turns out, do two education ministries with nothing else in common.