AI in Education
Intelligent Tutoring Systems: How AI Actually Teaches One Student at a Time
An intelligent tutoring system is software that teaches a single student directly, adapting each next question or hint to what that student just got wrong. It works through four linked parts: a domain model, a student model, a tutoring model and an interface. This article explains how those parts function, reviews what the research actually shows about learning gains, covers Indian deployments including Mindspark and the NCERT-linked adaptive tools, and sets out what schools and teachers should check before buying one.
Admin ·
An intelligent tutoring system is software that teaches a single student directly, adapting each next question or hint to what that student just got wrong. It works through four linked parts: a domain model, a student model, a tutoring model and an interface. This article explains how those parts function, reviews what the research actually shows about learning gains, covers Indian deployments including Mindspark and the NCERT-linked adaptive tools, and sets out what schools and teachers should check before buying one.
What an intelligent tutoring system actually is
An intelligent tutoring system (ITS) is software that teaches one student at a time, deciding what to ask next based on what that student just answered. It does not play a video and move on. It tracks which specific skill you are weak in, feeds you problems at the edge of your ability, and gives a hint the moment you stall instead of a red cross at the end.
That is the whole idea. Everything else is engineering detail.
The distinction from a normal e-learning course matters. A course is a fixed path for everyone. An ITS builds a different path for every learner, which is why the same class of 40 children can be working on 40 different problems on the same afternoon. Researchers have studied this since the 1970s, starting with systems like SCHOLAR and later the Cognitive Tutor for algebra, which by the mid-2000s was running in thousands of American schools. So this is not a 2023 invention that arrived with chatbots. AI language models made ITS cheaper and more conversational, not new.
The four parts under the hood
Almost every ITS, from a 1990s research prototype to a 2024 LLM-based tutor, is built from the same four components. Knowing them helps you evaluate any product a vendor puts in front of you, because a weak system is usually weak in one specific part.
The student model is the part that decides whether the system is genuinely intelligent or just shuffling questions. Old systems used Bayesian Knowledge Tracing, which estimates the probability that you have mastered a skill after each attempt. Newer ones use deep learning variants. If a vendor cannot tell you what their student model tracks, they are selling a question bank.
| Component | What it holds | What breaks if it is weak |
|---|---|---|
| Domain model | The subject broken into small skills, plus how to solve each problem | Feedback is generic, cannot show the correct next step |
| Student model | A running estimate of what this learner knows and misunderstands | Questions repeat or jump in difficulty for no reason |
| Tutoring model | Rules for when to hint, when to reteach, when to move on | Hints arrive too early, so students game the system |
| Interface | How the student works and how the teacher sees the data | Children need instruction time just to use the tool |
Does it work? What the evidence says
Yes, with a size that is worth taking seriously but not worth exaggerating. A large 2014 review by Kurt VanLehn compared human tutors, step-based ITS and standard classroom teaching, and found step-based tutoring systems performed roughly on par with human one-to-one tutors, both clearly ahead of no tutoring. Effect sizes in the ITS literature usually land somewhere between 0.3 and 0.8 standard deviations depending on subject and design. Maths and programming do best, because answers are checkable and skills decompose cleanly.
India has one of the better field studies anywhere. The J-PAL evaluation of Educational Initiatives' Mindspark centres in Delhi, published in 2019 by Muralidharan, Singh and Ganimian, gave middle-school students 90 minutes a day of after-school time with an adaptive learning system plus instruction. After four and a half months, students gained about 0.37 standard deviations in maths and 0.23 in Hindi over the control group. What made it interesting was that gains were largest for the weakest students, the ones sitting several grade levels behind the syllabus. That is exactly the group a single teacher with 50 children cannot reach.
The honest caveat: that study had a real programme, trained staff, attendance tracking and paid facilitators. A tablet handed to a child with no structure does not reproduce it.
Where this is showing up in Indian education
Adaptive practice has quietly become the most common form of AI in Indian schooling, even where nobody calls it AI. Diksha, the national platform, carries adaptive question sets linked to NCERT chapters. Private platforms run mastery-based maths and science practice for CBSE and state boards. Several state governments have run remedial programmes on adaptive tools after learning loss during the 2020 and 2021 school closures, because the core problem, students in Class 7 who read at Class 3 level, is precisely what an ITS handles better than whole-class teaching.
Spoken English and pronunciation tutors are the second big category, and they work because speech recognition gives instant, objective feedback that a large class cannot. Coding is the third. Nearly every Indian coding-for-kids platform now auto-grades submissions and nudges the learner towards the specific line that failed.
What is still rare in India is deep integration with what happens in the classroom the next morning. A teacher in a Pune municipal school told an evaluation team that she liked the dashboard but had no free period to read it. That is the real bottleneck, not model quality. If you are also comparing this with the platform side of school technology, the difference between an ITS and course delivery software is covered in learning management systems ai.
How to evaluate or pilot one without wasting a year
Most ITS pilots in Indian schools fail for boring reasons: no device time, no electricity backup, no one owning the data, no link to the term test. Fix those before worrying about the algorithm.
Run a small pilot with one grade and one subject, and insist on a baseline test. Without a baseline you will have no idea whether the tool helped or whether those children were simply good at maths.
One last thing worth saying plainly. An intelligent tutoring system replaces drill and diagnosis, not teaching. It cannot tell that a child stopped trying because her father lost work, and it cannot make a bored fourteen-year-old care about quadratic equations. The teacher still does the hard part. The software just makes sure she knows exactly who needs her on Monday.
- Ask what the student model tracks, skill by skill, and ask to see a real dashboard with real data, not a demo account
- Check syllabus mapping against your board and your language of instruction, not just CBSE English
- Confirm the minimum viable dose: most evidence-backed programmes use 30 to 90 minutes a week per subject, consistently
- Budget for device time and one person who owns the timetable slot, otherwise usage collapses by week six
- Test the offline mode on a low-end Android device with 3G, because that is the reality in most schools
- Set one measurable outcome before you start, such as movement in a diagnostic test after 12 weeks
FAQs
1. What is the difference between an intelligent tutoring system and an LMS?
A learning management system stores and delivers content, tracks enrolment and marks completion, the same way for every learner. An intelligent tutoring system decides what each individual student should attempt next based on a running model of what they know. Many schools use both, with the ITS handling practice and the LMS handling delivery and records.
2. Is ChatGPT an intelligent tutoring system?
Not by itself. A general chatbot can explain and answer questions well, but it does not maintain a persistent model of your mastery across weeks or control the sequence of what you practise. Several vendors now build proper tutoring systems on top of large language models, adding the student model and curriculum structure that the raw chatbot lacks.
3. How much do intelligent tutoring systems cost for an Indian school?
Pricing is usually per student per year and varies widely, from a few hundred rupees for basic adaptive practice to several thousand for full programmes with facilitator training and dashboards. The bigger hidden cost is device time and a staff member responsible for running the sessions, which is where most pilots quietly fail.
4. Which subjects work best with AI tutoring?
Mathematics, programming, physics problem-solving and language mechanics such as grammar and pronunciation, because answers can be checked step by step and skills break into small trackable pieces. Essay writing, history interpretation and anything needing judgement or discussion still work better with a human, though AI can give useful first-pass feedback.
5. Can intelligent tutoring systems work without internet?
Yes, several Indian platforms run offline on Android tablets or in computer labs and sync student data when a connection is available. If you are piloting in a rural or low-bandwidth school, test the offline build on the actual devices you own before signing anything.
6. What jobs exist in building intelligent tutoring systems?
Instructional designers who break syllabus into skill maps, learning engineers and data scientists who build knowledge-tracing models, content authors who write items with graded hints, and implementation specialists who train teachers. Indian edtech firms and NGOs working on foundational learning hire across all four, and the skill-mapping roles often go to former teachers.