Artificial Intelligence in the Classroom: A Teacher’s Honest Guide to What AI Can and Cannot Do

Artificial Intelligence in the Classroom: What It’s Genuinely Good For, and What It Cannot Replace

A contributing voice from educational research | Specialist in curriculum, pedagogy, and school leadership

Artificial intelligence tools are arriving in Indian classrooms faster than guidance about using them wisely, and the research that does exist points to a specific, useful distinction rather than a blanket verdict either way. Decades of research on structured, narrow AI tutoring tools, predating today’s generative AI, has consistently found real learning benefits in well-defined domains like mathematics. Newer research on open-ended generative AI use, particularly for tasks like writing, raises a genuinely different concern: tools that make work easier can also remove the effortful struggle that learning science identifies as necessary for durable learning. Those are two different claims, and treating AI in the classroom as one undifferentiated question misses the distinction that actually matters.

Key facts

• A substantial, decades-old body of research on intelligent tutoring systems, structured software that gives step-by-step feedback in well-defined domains such as mathematics, has found these tools can approach the effectiveness of human one-on-one tutoring in some studies, a genuinely well-established finding that predates today’s generative AI tools by many years.
• Learning science has long identified “desirable difficulties”, a concept associated with researcher Robert Bjork, effortful retrieval and genuine struggle during learning generally produce more durable, longer-lasting learning than passively receiving a correct answer. This is directly relevant to evaluating generative AI tools: a tool that supplies a complete, correct answer removes exactly the kind of effort this research identifies as necessary for real learning to stick.
• This creates a genuine, evidence-based distinction worth building classroom policy around: AI used as a structured practice partner, giving feedback, asking guiding questions, checking work incrementally, has a stronger evidence base than AI used as an answer-generator for open-ended tasks like essay writing, where the tool can substitute for the cognitive effort the task was designed to require in the first place.
• India’s CBSE has introduced AI-related curriculum content and skill modules in recent years, reflecting a genuine national policy direction toward AI literacy, though this piece cannot verify the current, specific scope or grade-level details of that curriculum without a working search tool this session.

The useful framework is not “AI good” or “AI bad.” It’s what the tool is actually being asked to do

A maths tutoring tool that checks a student’s work step by step and asks them to try again is asking for effort and giving feedback on it. A chatbot that writes a finished essay on request is removing the effort entirely. Both are “AI in the classroom,” and treating them as the same question is exactly the mistake this research helps a teacher avoid.

This distinction matters because it gives a teacher an actual decision rule, not just a general caution. The question to ask about any specific AI use is not “is this AI” but “does this use require the student to do the cognitively effortful part of the task, or does it do that part for them.” Structured practice and feedback tools tend to fall on the first side; open-ended answer generation for graded, skill-building work tends to fall on the second.

Five principles for using AI tools well in a secondary classroom

1. Diagnose which category a specific AI tool or use case actually falls into before deciding whether to allow or encourage it. A tool used for structured practice and feedback is a different proposition than the same underlying technology used to generate a finished piece of work a student was meant to produce themselves.
2. Protect the effortful, “desirable difficulty” part of learning specifically. If an AI tool removes the struggle a task was designed to require, whether that’s working through a maths problem or drafting an argument, it may be reducing learning even while producing a technically correct or well-written result.
3. Use AI tools to extend teacher feedback capacity, not replace the relationship. Automated feedback on structured practice can free a teacher’s time for the relational, judgement-based parts of teaching that a tool cannot replicate, rather than substituting for the teacher-student relationship itself.
4. Be explicit and consistent with students about which tasks are meant to involve effortful struggle and which are appropriate for AI assistance. Ambiguity about this distinction is likely to produce inconsistent student behaviour more than a clear, stated policy would.
5. Start with one specific, well-defined use case, structured practice in one subject, rather than a blanket classroom AI policy. Given how different narrow tutoring-style uses and open-ended generation uses are in terms of evidence, piloting one clearly-defined use is more informative than adopting or banning AI wholesale.

Frequently asked questions

Is there actually good research showing AI tutoring tools help students learn? Yes, for a specific category: decades of research on structured, narrow intelligent tutoring systems in well-defined domains like mathematics has found real learning benefits, in some studies approaching the effectiveness of human one-on-one tutoring. This is a different claim from the newer, less settled question of open-ended generative AI use for tasks like essay writing.

Does letting students use AI to help with homework actually hurt their learning? It depends specifically on what the AI is doing. If it’s replacing the effortful part of the task, generating a finished answer or essay the student was meant to produce themselves, learning science’s “desirable difficulties” research suggests this can reduce the durable learning that effortful struggle would have produced. If it’s giving structured feedback on the student’s own attempt, the evidence is more favourable.

How can a teacher tell the difference in practice? Ask whether the specific use requires the student to do the cognitively effortful part of the task themselves, working through a problem, drafting their own argument, or whether the AI is doing that part for them. The first category has stronger research support; the second is where the genuine concern about undermining learning applies.

What’s the most useful first step for a school without a clear AI policy yet? Piloting one specific, well-defined use case, ideally a structured practice-and-feedback tool in a subject with clear right and wrong answers, rather than either banning AI entirely or adopting it without any distinction between different types of use.

What to do:

1. Name one student in your class who is most affected by how AI tools are currently being used, well or poorly, in your subject, and write down one specific thing you will try differently for them this week, not a general policy, but a named action for a named child.
2. Search DIKSHA or ask your Block Resource Coordinator for any resource specifically on AI use in secondary classrooms in your state context. If nothing exists, note that gap; the absence of a resource is itself useful information.
3. Spend five minutes this week observing whether a specific student’s AI use, where it’s happening, looks like structured practice or like outsourcing the effortful part of a task, not teaching, just watching and noting.
4. Tell one colleague about the distinction between AI as a structured practice tool and AI as an answer-generator, and discuss where your school’s current, informal policy actually falls.
5. At the end of the week, ask yourself what you would try differently next week, and what you would look for to know whether it is working.

Note on sourcing: web search was unavailable during this session, so current, specific studies, statistics, and details of CBSE’s AI curriculum initiatives could not be verified live. The core research described here, intelligent tutoring system effectiveness and the “desirable difficulties” concept associated with Robert Bjork, reflects well-established, broadly known findings in learning science, but a reader preparing this for publication should verify specific, current citations and Indian policy details before treating them as fully sourced.

Support resources: NIPUN Bharat FLN assessment tools (via your state SCERT or DIET); NCERT teacher guides; Azim Premji Foundation open-access research; DIKSHA platform content in 36 languages; Tele-MANAS mental health helpline: 14416 or 1800-891-4416; CHILDLINE: 1098.