Start by conceding the point properly
സംശയം ഇപ്പോൾ ഉടനെ തീർക്കാം. അതൊരു നേട്ടമാണ്.
A student in 2019 with a doubt at eleven at night had three options: wait for the next class, ask a friend who might also be wrong, or give up on it. Today that doubt is answered in fifteen seconds, in Malayalam if you like, rephrased as many times as you need.
That is a real gain and I am not going to pretend otherwise. Anyone selling you a course by telling you AI is useless is hoping you have not tried it.
But look carefully at what got solved. Explanation got solved. And explanation was never the thing standing between a sincere student and a pass mark.
The four things it did not touch
Elsewhere on this page I argue that preparation is really four things: recall, spacing, mixing and feedback — wrapped in a system that survives your motivation. Run AI against that list honestly.
It will not decide what you study tomorrow
It answers what you ask. It has no idea what you have covered, what you are avoiding, or what the paper weights heavily. A tool that responds perfectly to your questions cannot save you from the questions you do not think to ask — and the syllabus you never open is exactly the one you would never think to ask about.
It will not notice that you have gone quiet
Nothing registers that you have not touched estimation for three weeks. There is no record, no grade slipping, no chapter sitting at zero. The tool is equally content whether you use it daily or twice a month, which makes it the opposite of the thing you actually needed.
It will not sit you down for fifty timed questions
You can ask it for practice questions, and it will produce them enthusiastically. What it will not do is impose the clock, the negative marking, or the discipline of finishing. And a question you answer at your own pace, with the option of asking a follow-up, is not testing the skill the exam tests.
And it will tell you a wrong code clause with total confidence
This is the one that costs marks, and it deserves its own section.
The confident-wrong-clause problem
Ask a general AI tool what the minimum bearing for a lintel is, and there is a fair chance it will tell you 100 mm as per IS 456:2000.
That citation does not exist. IS 456 has no lintel-bearing clause at all — clause 22.2 fixes effective span and nothing more. We checked the full text of the code specifically to be sure. The 100 mm figure is a real and widely used construction-practice convention; the code reference bolted onto it is invented.
And here is the mechanism that should worry you: the tool is not hallucinating from nothing. It is repeating a mis-citation that appears in guide books, in coaching notes and across the internet. It is confidently reproducing the average of what has been written, and on technical detail the average is often wrong.
A wrong answer that sounds uncertain costs you nothing. A wrong answer delivered with a clause number costs you the mark and the three years you go on believing it.Santhosh Sir
A textbook can be wrong too — and ours have been. The difference is that a textbook is one source you can check against another. A confident synthesis of every source at once removes the disagreement that would have warned you.
A second example, because one could be bad luck
Ask when vertical formwork to a column may be struck. The answer you will most often get is 24 to 48 hours.
That was correct — in the 1978 edition of the code. The current edition has said 16 to 24 hours since 2000. But an engine answering from what has been written will give you the figure that was written most often, and for a value superseded a quarter of a century ago, the old one still has a very large head start.
Frequency is not currency. That is a structural property of anything trained on accumulated text, and it is precisely the failure mode that matters in a field governed by revised standards.
But some AI tools do track you now
Fair objection, and it deserves a straight answer rather than a dodge. Tutoring tools that keep a record, notice what you are weak at and adapt do exist, and they will get better. That closes part of the gap I described above, and I expect it to close further.
Two things do not close by themselves, and neither is a limitation of the model:
- Alignment to this exam. Not civil engineering in general — the specific syllabus for the post you are sitting, weighted the way that paper weights it, phrased the way that examiner phrases things. That comes from working through the actual papers, which is labour, not capability.
- Verification. Somebody has to have opened the code and checked. That is not something a tool decides to do; it is something a publisher decides to pay for.
So the honest position is not "AI cannot do this". It is that the remaining work is unglamorous, specific to one exam, and does not get cheaper just because drafting did.
So how should you actually use it?
Not at all is the wrong answer. Use it for what it is genuinely excellent at, and do not use it for what it is structurally bad at.
| Use it for | Do not use it for |
|---|---|
| "Explain this again, more simply" | "What does IS 456 say about…" — verify the clause yourself |
| "Give me an analogy for this concept" | Any specific numeric value, limit or code table |
| "Why is my reasoning here wrong?" | Deciding what to study next |
| Rephrasing a definition you keep forgetting | Anything you will not later see tested under a clock |
The short rule: use it on the reasoning, not on the numbers. Reasoning you can check by following the argument. A number you cannot check without the code in front of you — which is exactly the situation where you will accept it.
What we do with it, since you may reasonably ask
We use AI heavily in building our material. It would be strange to write this page and pretend otherwise.
What we do not do is publish what it produces. Every number, limit and clause goes against the actual standard before it reaches a student, and that verification step is most of the work — it is also where we have caught errors that had been sitting in print for years. That process is its own article.
The tool made the drafting faster. It made the checking more necessary, not less. Those two facts arrived together, and only one of them gets advertised.