I used an AI chatbot for my studies for a month. It never once said "I'm not sure."
Where it was brilliant, where it quietly lied to me, and the rules I follow now.
Kabir Joshi11th Sample
9 Oct 2026 · 5 min read

Photo: Mister rf · CC BY-SA 4.0 · via Wikimedia Commons
Let me say this first so nobody calls me a boomer in the comments: I like these tools. I use one almost every week when I code, and I'm not going back. But last month I ran an experiment on myself. For about four weeks I tried to use an AI chatbot for everything study-related, and every time it got something wrong or fishy I wrote it in a note on my phone. I wanted to find the edges.
The note got long. Here's what I learned.
Where it was honestly brilliant
- Explaining the same thing five ways. I asked it why a capacitor stores energy using a water-tank picture, then a crowd-in-a-corridor picture, then "like I'm twelve". The third one clicked. A textbook only gets one shot at explaining something; this thing gets as many as I have patience for.
- Making me practise. "Ask me ten questions on Newton's laws, one at a time, don't tell me the answer until I try." This is the best use I found. It's like a friend who never gets bored of you.
- Cleaning up my own notes. I paste my notes and say "reorganise this into an outline, don't add anything." It doesn't need to know anything for that, so it can't lie much.
- Debugging. Python error messages are a lot. Pasting the error and asking "what does this mean in plain words" saves me twenty minutes.
Where it lied (with receipts)
1. The arithmetic that looked perfect. I asked for the range of a projectile launched at 30° with . It wrote the correct formula,
and then plugged in instead of . So I got about 20.4 m instead of about 35.3 m. Everything around the mistake looked so tidy, units, rounding, a neat boxed answer, that I almost copied it. I only caught it because the number felt small for a 20 m/s throw.
2. The book that doesn't exist. I asked for a source for a claim about enzymes. It gave me an author, a title and a chapter number. I searched the library website, then the open web. Nothing. The book is invented. It sounded like a book that should exist, and apparently that was enough.
3. "According to NCERT..." I asked what NCERT says about a topic and it described a "Figure" and a "Box" with total confidence. My actual book has no such box. It doesn't have my textbook open in front of it. It writes something that sounds like NCERT, which is not the same thing at all.
4. Code that passes my tests and still breaks.
def avg_last(marks, n):
return sum(marks[-n:]) / n
I'd asked for "the average of the last n marks". It worked on my three test cases. Then avg_last([70, 80], 5) returns 30.0, because it divides by 5 even though there are only two marks. The right answer is 75. When I asked "what edge cases did you miss?" it spotted the problem instantly. So it can fix it, but only if you poke it.
The scary part isn't the mistakes
Mistakes are fine. My own notes have mistakes. The scary part is that it sounds exactly the same when it's right and when it's wrong. A teacher says "hmm, not sure, let me check." A book has an index and a printed date. The chatbot has one tone, calm and sure, for everything.
And most of the errors in my note, I caught only because I had already studied that topic. Which means you can catch it only on things you already know. For a student that's a bit of a trap, no? The topics where you most want help are exactly the ones where you can't tell if the help is wrong.
Prompts that actually worked for me
| Instead of... | I now ask... | Why it's better |
|---|---|---|
| "Solve this problem" | "Give me only the first step as a hint" | I still do the thinking |
| "Is my answer right?" | "What is wrong with my answer?" | It tends to agree with you if you ask the first way |
| "Explain photosynthesis" | "Explain it, then ask me 3 questions to check I got it" | Turns reading into practice |
| "Give me a source" | "Don't give sources. Tell me the keywords to look up in my book." | No invented books |
My four rules now
- Try first, ask second. If I haven't struggled with a problem for at least fifteen minutes, I've robbed myself of the part where learning actually happens.
- Method, not answer. Hints beat solutions.
- Every number gets redone by hand. Takes two minutes. I also make it write units at each step, because a wrong unit is the easiest lie to spot.
- Facts get checked in NCERT or with a teacher. Names, dates, "quotes", anything it cites is a rumour until I find it myself.
One more thing
No chatbot sits next to you in the exam hall. Your own brain does. If I lean on the tool for a month and my own legs forget how to walk, I'd rather find out annoyed at myself now than at 9am on exam day.
If you've caught one lying, put it in the comments. I'm collecting them. My most embarrassing one is still the invented book.
12 responses
Sign in to join the conversation.
[+][–]Meera Patel11thSample· 10d ago
This is very well put. The line about only being able to catch errors in topics you already know is the real issue, and I haven't seen many people say it so clearly. I would add one more habit: when it gives a numerical answer, estimate the magnitude yourself first. Even a rough "this should be somewhere around 30 to 40 metres" is enough to make a wrong 20.4 feel suspicious before you trust it.
[+][–]Rohan Sharma11thSample· 10d ago
the invented book is so funny 💀 mine once cited a "journal" that sounded SO real. i went looking for it for 10 mins like a detective. never again
[+][–]Arjun Nair12thSample· 10d ago
Nice trick for the projectile one that needs no calculator: for the same speed, launch angles and give the same range, because . So 30° and 60° must give the same answer. If a chatbot gives you different ranges for the two, you've caught it. Free sanity check.
[+][–]Yash Tiwari11thSample· 10d ago
wait bhai maximum range 30° pe hota hai na? that's what I remembered
[+][–]Arjun Nair12thSample· 10d ago
45°. At 45°, , which is the biggest value sine can take, so that's the maximum range. 30° and 60° are equally far from 45° on either side, that's why they tie.
[+][–]Yash Tiwari11thSample· 10d ago
ohhh the complementary thing makes sense now. thanks bhai, noted
[+][–]Harsh Gupta12thSample· 10d ago
"It never once said I'm not sure." Bro that's also my entire personality in every class discussion. At least I'm honest about being wrong afterwards
[+][–]Siddharth RaoDropperSample· 10d ago
Drop-year perspective: I used one of these last year to generate extra practice questions, and it was useful. But it once told me a wrong value for something in physiology and I memorised it for almost a week before my own notes contradicted it. The lesson I took: never learn a fact from it that you haven't seen in your book. Use it for questions and explanations, not for the facts you're going to be tested on.
[+][–]Kabir Joshi11thSample· 10d agoAuthor
that's exactly the thing, the fact you memorise from it gets stuck harder because it sounded so sure. thanks for sharing, "questions yes, facts no" is a good short version
[+][–]Tushar PawarDropperSample· 10d ago
Edge cases! Joke I tell myself: write the tests before you ask it for the function. If it fails your tests, you caught it in 10 seconds instead of 10 days. Also try asking it for the failing input, "give me an input that would break this" works weirdly well
[+][–]Pooja Iyer12thSample· 9d ago
omg I asked it for a mnemonic for the order of something in the cell cycle and it gave me a lovely rhyme in the WRONG order 😭 I was so proud of it too. I now make my own mnemonics, it's half the fun anyway
[+][–]Ishaan Bose11thSample· 9d ago
lol same energy, mine confidently told me a reaction product in chemistry and I was like "this is why I hate this subject" and then it turned out the AI was the one who was wrong. small win for me





