By BarathVector Editorial — 2026-08-15

The first draft is often ugly. That is why it matters.

A child staring at a blank page must retrieve facts, arrange them, discover what is missing and decide what to say. The resulting paragraph may be clumsy. The struggle has nevertheless built something that a polished answer cannot deliver on the child's behalf: a path through the problem.

Generative AI can now erase that struggle in seconds. It can also explain a difficult idea, adjust a lesson to a learner's pace, translate instructions and give a student practice that a crowded classroom cannot always provide. Schools should not ban such a tool from the world children inhabit. They should decide when assistance begins.

The right default is simple: effort before assistance. For work meant to build core knowledge or judgement, the child should attempt the first draft, solution or explanation unaided. AI enters after there is thinking to examine.

Performance is not the same as learning

The most useful warning comes from an experiment, not a panic. In a study involving nearly 1,000 high-school mathematics students, researchers compared ordinary practice, access to a general-purpose GPT-4 assistant and access to a tutor designed with safeguards. The unrestricted assistant improved performance during practice, but students did worse when later tested without it. The guarded tutor largely avoided that loss. The results were published in the Proceedings of the National Academy of Sciences.

One study does not settle every subject, age or design. It does show why schools must measure independent performance, not merely the quality of AI-assisted homework. A student can submit a better answer while acquiring less ability to produce the next answer.

The design of the help matters. If a system provides the completed solution immediately, it rewards delegation. If it asks a question, reveals one hint and waits for an attempt, it can behave more like a tutor. A separate field experiment with 700 tutors and about 1,000 students found that an AI assistant for tutors improved student topic mastery by roughly four percentage points, with larger gains for lower-rated tutors. The working paper studied AI supporting a human instructor, not replacing the learner's work.

The lesson is neither “AI makes children stupid” nor “AI personalises education.” Both are slogans. Learning depends on task, timing, interface, teacher and what the assessment later asks the child to do alone.

India is right to teach the technology

CBSE has introduced an AI and computational-thinking curriculum for Classes III to VIII from the 2026–27 academic year. The official announcement stresses logical thinking, problem-solving and age-appropriate learning rather than treating AI as a specialist subject for older students.

That is the correct direction. Children need to understand that AI systems predict outputs from data, make confident mistakes, reflect biases and can be manipulated. They need practice checking sources, protecting personal information and declaring assistance. Ignorance is not safety.

But AI literacy is not constant AI use. A school can teach calculators while still requiring mental arithmetic. It can teach search while still asking pupils to remember. It can teach generative systems while preserving spaces in which a child reads, writes, draws and reasons without a machine completing the next step.

A first-effort rule

Schools should divide work into three visible modes.

In “human-only” mode, students build foundations: handwriting, vocabulary, mental calculation, recall, sustained reading and first attempts at analysis. Devices stay away. Assessment is designed to show what the student can do independently.

In “AI-assisted” mode, the learner submits an initial attempt before asking for help. The system may question an assumption, offer a hint, compare two structures or provide feedback. The student keeps a short record of the prompt, the response and what changed. The final submission includes a reflection on which advice was rejected and why.

In “AI-native” mode, students study the tool itself. They compare model answers, test bias, build simple applications or solve problems where using AI well is part of the objective. Evaluation then rewards verification, judgement and responsible use rather than pretending no assistance occurred.

The mode should appear at the top of every assignment. Ambiguity invites unequal enforcement and turns honesty into a disadvantage.

Age and privacy cannot be footnotes

UNESCO's AI competency framework for students calls for a human-centred progression covering ethics, techniques and system design. Its broader guidance on generative AI urges age-appropriate use, data protection and validation of educational tools.

UNICEF reported in June 2026 that children are adopting AI services more than three times faster than adults in its ten-country analysis, with at least 20 million children using them and about 13 million reporting homework use. UNICEF also says evidence about outcomes remains incomplete. That combination—rapid adoption and uncertain effects—argues for supervised experimentation, not complacency.

No child should have to disclose intimate family information to receive help with an essay. Schools should prohibit uploading identifiable student records, health details, photographs or unpublished work to consumer systems without an approved agreement. Younger pupils should use restricted, school-managed tools only when a teacher has defined the purpose. Vendors should not train on children's interactions by default or use them for advertising profiles.

The strongest countercase

The first-effort rule can become unfair. A child with dyslexia, a disability or limited proficiency in the classroom language may need assistance before producing conventional text. A teacher managing forty students may find AI feedback more available than individual help. And insisting on unaided drafts can preserve old disadvantages if affluent pupils receive private tutoring at home.

Exceptions are necessary. Accessibility support is not cheating, and the relevant “first effort” may be spoken, visual or made with assistive technology. Schools should define accommodations with the learner rather than enforce one format. Public systems should also prioritise AI tools for teachers and under-resourced classrooms so that restraint does not become another privilege.

The principle survives those exceptions: technology should expand the child's agency, not conceal whether learning occurred.

Assess the mind that remains

Every AI-assisted unit should end with some unaided retrieval, explanation or application. That need not mean more high-stakes examinations. It can be a short oral defence, an in-class paragraph, a worked problem or a teacher's question about the choices in the final submission.

Schools should track whether students become better without the tool over time. If assisted work improves while independent performance declines, the programme is producing attractive documents and weaker learners. If both improve, the tool is doing its job.

Children will live with artificial intelligence. The duty of education is not to recreate a world before it. It is to ensure that the machine arrives after curiosity, memory and judgement have been invited into the room.

Let the child make the first attempt. Then let AI help the child see it more clearly. The imperfect first draft is not wasted time. It is the visible beginning of a mind becoming its own.