ChatGPT Doesn't Know What CBSE Wants
Quick take
My answer was correct. I got two out of five. AI knows your subject — it has never seen your marking scheme, and that gap is where marks quietly disappear.
Explore this topic
Article body
ChatGPT Doesn't Know What CBSE Wants
My answer was correct. I got two marks out of five.
I'd asked an AI to explain a Physics numerical, understood it completely, and written what it gave me — a clean, confident solution that arrived at the right number in about four lines. My teacher circled the whole thing and wrote "steps?" in the margin.
That's when it clicked. The AI knew Physics. It did not know the exam. Those are two entirely different things, and nobody had told me they could come apart.
The subject and the exam are not the same subject
An AI model has read an enormous amount of physics, chemistry, history and maths. On the actual content, it's usually solid.
But your board exam isn't testing whether you understand the content. It's testing whether you can produce a specific answer, in a specific format, that a specific marking scheme rewards. That scheme is a document. It lists what earns marks. And the model has never seen it.
Correct and rewarded are different categories. Only one of them is graded.
This isn't a CBSE-only problem — ICSE and every state board have their own conventions. But CBSE is where I've watched the most marks quietly disappear.
Where the gap actually shows up
| Step marking | In Maths and Science you're marked per step, not per answer. AI compresses working because it's explaining, not scoring. Right answer, missing steps, lost marks |
| Value points | Marking schemes list the specific points an examiner ticks. A beautifully written answer that misses three listed points scores below a plain one that hits all five |
| Length by mark | A 2-marker and a 5-marker want different lengths. AI writes until the idea is complete. Examiners are reading hundreds of scripts against a scheme |
| NCERT phrasing | Board questions are set from NCERT. AI draws from the whole internet, so it uses the textbook's idea with different vocabulary — and the scheme was written around the textbook's |
| Deleted chapters | Syllabus gets rationalised. Models trained on older material happily teach you topics that are no longer in your syllabus at all |
| Diagrams | Labelled diagram, pencil, ruler — sometimes carrying a full mark on its own. AI cannot draw your diagram and won't remind you it's worth marks |
*None of these are the AI being wrong. Every one of them is the AI answering a different question from the one your paper is asking.
The deleted chapter problem is the sneaky one
The others cost you marks. This one wastes your time, which is worse when you're three weeks out.
Syllabuses get trimmed. Chapters get removed or reduced. A model has no idea which version of the syllabus you're sitting — it just knows the subject as it appeared across everything it was trained on. Ask it to help you revise and it may cheerfully walk you through a topic that isn't on your paper.
Check the current syllabus, not the internet's memory of it.
Ground it every time: paste the actual chapter or your syllabus into the chat and say "only use this."
Never ask "what's in the syllabus?" — that's the one question it has no reliable way to answer for your board and your year.
Your source of truth is the current syllabus from your board's site and your NCERT textbook. Not a chatbot, and not last year's senior.
Use the marking scheme as the answer key
Here's the thing almost nobody exploits, and it's free.
CBSE publishes sample question papers and their marking schemes on the academic site every year. That document is the examiner's instructions. It shows you exactly how marks are split, what the expected value points are, and how much writing each mark is worth.
Read three of those and you'll understand your exam better than a year of AI explanations will teach you. Previous year papers do the same job.
The examiner's instructions are published. Almost nobody reads them.
And once you have a real marking scheme, AI becomes genuinely useful again — because now you can give it the standard instead of asking it to guess one.
What it's still good at
I'm not telling you to stop using it. I use it most days. Just point it at the right jobs:
- Explaining what you didn't follow in class — patient, unlimited, never annoyed at the fourth "wait, why?"
- Rewording a definition until it makes sense in your head — then memorise the textbook's version, not the AI's
- Quizzing you — paste your chapter, ask it to test you until you fail. Being questioned beats being explained to
- Checking your working — write the answer yourself, then ask where it's weak
- Building a revision schedule from your own dated syllabus
Notice the pattern. Every one of those uses it as a tutor — something that helps you produce the answer. None of them use it as an answer key. That distinction is basically the whole post.
Quick Tips
- Paste the chapter in — grounded answers beat remembered ones, every time.
- Write full steps anyway — even when you can see the answer. Steps are marks.
- Read three marking schemes — free on your board's site, and worth more than most guidebooks.
- Match length to marks — a 5-marker isn't an essay and a 2-marker isn't a sentence.
- Memorise the textbook's wording for definitions, not the AI's nicer version.
- Never trust it on syllabus scope — that's the one thing it genuinely cannot know.
Go download one marking scheme tonight
Pick your worst subject, find last year's sample paper and its marking scheme, and read how the marks are actually split. It takes twenty minutes and it will change how you write every answer for the rest of the year.
The AI knows your subject. Only the scheme knows your exam.Comments 0
Keep reading
Similar blogs by topic
ChatGPT Made Up a Book for My Project. I Almost Submitted It.
It invented the author, the publisher, even the page numbers. Here's what actually goes on inside these things — LLMs, hallucination, RAG, MCP, agents — explained the way someone should've explained it to me before I nearly failed history.
It Took Me Two Minutes to Prove My Family Group Chat Wrong
Deepfakes got good enough that squinting at videos stopped working. Here's the two-minute check that still does — reverse image search, lateral reading, and the one scam that will actually reach your family.
I Scored 94%. Then Someone Asked What I'd Actually Made.
Fifteen years of school and I couldn't point at one thing I'd built. Here's why marks stop being information once everyone has them — and how to have real proof of work by next month.