Correcting it when the answer is wrong

You asked how to fill a form. The answer came back neat and confident. One line in it was plainly wrong, and you knew it was wrong.

So you typed three words. No, that is wrong. The next answer was different, and somehow worse than the first one.

This lesson gives you what to type instead.

Why “no, that is wrong” makes things worse

Those three words give the model almost nothing. It does not know which line you mean. It does not know what the correct fact is.

So it guesses again. It picks a line, changes it, and writes the answer once more in the same calm voice. Very often it changes the part that was already right.

Here is what that looks like in real life. You ask for five steps to fill a form. Step three is wrong, so you type no, that is wrong.

The next answer has a new step three. Step one is now shorter. Step four has quietly gone. You have lost ground, and you cannot see where you lost it.

There is real evidence for this. In 2025 a research team ran more than 200,000 test conversations across 15 different AI models. Answers were about 39 percent worse once the chat went past the first message. Their words were blunt. When the model takes a wrong turn, it gets lost and does not recover.

A vague complaint does not pull it back. It pushes it further away. You have to hand it something solid.

The correction that actually works

A good correction has three parts, always in this order. Name exactly what was wrong. Give the correct fact. Then ask for the whole answer again, from the start.

Here is the shape. Copy it and fill in your own words.

Stop. Two things in that answer were wrong.
1. You said the last date is 31 March. That is wrong.
2. You said I need a caste certificate. That is wrong.
The correct facts are these. The last date is on my college notice board.
The form needs only my photo and my class 10 marksheet.
Now write the whole answer again from the start, using these facts.
Do not just change one line.

Look at what that message is. It is a full request, not a complaint. It is the same skill from how to ask so you get a useful answer. A correction is just a fresh request with better facts inside it.

A good reply starts again at step one. It should read like this.

Here are the steps again, with your corrections.
Step 1. Keep your photo and your class 10 marksheet ready.
Step 2. Check the last date on your college notice board.
Step 3. Fill the form in capital letters.

Sometimes you get one short line saying sorry, and nothing else. That is not enough. Send it back and say: write the full answer again, all of it, from step one.

Two small habits make a correction land better. Point at the exact line, not at the whole answer. Repeat the words it used, so there is no doubt which part you mean.

The second habit is to correct facts only. If you also want the answer shorter or simpler, ask for that in a separate message. One message, one job.

Why a patched answer is dangerous

Say you only type this. The last date is wrong, please fix it.

The model fixes that one line. Everything below it stays exactly as it was. That is the trap.

The wrong date may have shaped the whole answer. The order of the steps. The plan it gave you. The papers on the list. All of that quietly keeps the old mistake inside it.

Now the page looks corrected, so you stop checking. It is like changing one entry at the top of a shop ledger (bahi khata) and never redoing the totals below. The page looks tidy and the sum is still wrong.

That is why you ask for the whole answer again. It costs you one message. It saves you from a tidy answer that is still broken.

When the chat is stuck, restart it

Some chats cannot be saved. You have corrected it twice and it is still going in circles. Do not correct it a third time.

Restart instead. Call it the restart move. It has two steps. It takes under a minute.

Step one. In the old chat, type this and send it.

Write everything I have told you in this chat as one clear request.
Do not answer it. Just write the request.

Step two. Copy what it writes back. Open a brand new chat. Paste it there, add anything it left out, and send.

Here is my full request.
[paste what the old chat wrote]
Extra facts it left out: I am in class 10 in a village in Uttar Pradesh.
Answer in five short steps.

The same research team that measured the 39 percent drop tells you to do exactly this. Starting a new chat with the same information can work much better than carrying on in the old one. Putting all your needs into one single request also helps.

Read the request before you send it. The model usually leaves something out. Add your class, your district, your subject, or whatever it dropped.

It feels like losing your work. It is the opposite. You keep the facts and you throw away the wrong turn. The new chat knows nothing about the argument you just had, and that is the whole point.

Three signs to close the chat and start over

You do not have to guess when a chat is finished. There are three clear signs.

It repeats itself. You get the same paragraph again with a few words swapped.

It has forgotten what you said. It asks again for your class, or your district, or the subject you gave it at the start.

The topic has changed. You began with a form and now you are asking about your exam. That is a different job. It needs a different chat.

Long chats also get weaker on their own, well before they are full. There is a whole lesson on that, one chat, one job.

If it changes its answer, you have proved nothing

Read this part twice. It is the piece most people get wrong.

You push back. The model agrees with you and gives a new answer. That feels like winning. It is not evidence of anything at all.

In a 2026 test, seven of the newest models were given a well-argued case for a wrong option after they had answered correctly. How often they switched sides ranged from about 18 out of 100 to about 97 out of 100. The argument did not have to be true. It only had to sound good.

Asking are you sure is just as weak. There is a study on self-checking. One leading model was tested on a set of school maths questions. Its score fell from about 95 correct out of 100 to about 89 after two rounds. When models were given feedback from outside instead, their scores went up.

So pressure is not the answer. Proof from outside the chat is the answer.

Outside proof means anything the model did not write. Your textbook. An official page. A notice on a board. A teacher. Type it into the chat and ask the model to compare.

Here is what you said: [paste its answer]
Here is what my book says: [type the exact lines from your book]
Where do these two disagree?
Which one is correct, and how can I check it myself?

This works because the model is now reading, not remembering. Giving it the page is far safer than asking it to recall the page. That idea has its own lesson, giving it your own book, notes or photo.

Where this goes next

You now hold four moves. Correct with facts and ask for the whole answer again. Never accept a patch. Restart a stuck chat instead of arguing with it. Treat a changed answer as a question, not as proof.

All four depend on one thing. You have to spot the mistake first.

That is a separate skill, and module 4 of this course is built around it. Start with why AI makes things up, so the mistakes stop surprising you. Then learn the two-chat test and three other checks, which take under two minutes. Along the way you will meet why it agrees with you far too easily. That habit of agreeing is the reason it flips when you push.

Before you move on, do the practice task below. Correct one real answer today, and watch what happens to the lines you did not mention.