The answer machine trap

Last night you finished twenty sums with an AI app open beside you. Every answer came out right, and the whole thing felt easy. Then the test came, your phone was in your bag, and your mind went blank.

What happened when a school actually tested this

Nearly one thousand students took part in a study at a school in Turkey. They were in classes 9 to 11. Researchers split them into three groups for maths practice.

The first group had no AI at all. They used their notes and their textbook, the way you always have. The second group had a plain AI chat app, the kind you already use. The third group had the same AI, set up with teaching rules.

During practice, both AI groups looked brilliant. The plain AI group scored about 48 percent higher than the group with no AI. The third group scored about 127 percent higher. Marks like that make a student feel unstoppable.

Then came the real exam. No phones, no app, only paper. The plain AI group scored about 17 percent below the students who never had AI at all. Not lower than they hoped. Lower than doing nothing. This work was published in June 2025.

Why the marks fell

Here is the whole reason in one line. The AI did the thinking, so those students got practice at reading answers, not at solving problems.

Reading a finished solution feels like understanding. Your eyes move down the steps and nothing feels hard. Nothing feels hard because nothing was hard for you. The machine did the work, and you watched.

The researchers also looked at how students typed. In the plain AI group, more than half of the first messages simply repeated the question or asked for the answer. That is a transaction, not a study session.

The students with teaching rules behaved differently. They sent more messages about each single problem. A hint leads to a question, and a question leads to another try. That back and forth is what practice actually is.

There is a name for what the plain AI group did. You save the mental effort today and pay it back later, with interest. Some researchers call it cognitive debt. In one small study, people wrote essays with an AI while their brain activity was recorded. By the fourth essay, that group could barely quote their own writing. It was a small study, and it was about writing, not maths. Take it as a warning, not as proof.

There is a second danger inside the same habit. The researchers typed “What is the answer?” into the plain AI on those same practice problems. It came back wrong about half the time. That test used an older model, and today’s models are better at school maths. But you cannot see the difference, because a wrong answer looks exactly like a right one.

The same app, used a different way

Now look at that third group again. On the real exam they were level with the students who had no AI. No gain, but no harm either. Same model, same school, same maths problems. Only the rules were different.

The teachers had given the AI two things. First, an instruction to give hints and never the final answer. Second, the correct solution and the mistakes students usually make on that problem. Rules like these are called guardrails.

So the tool is not the problem. The way it is used is the problem. That is good news, because the way you use it is the one part you control.

Two of the big apps now have a free study setting built in. ChatGPT added a study mode in July 2025. Gemini added Guided Learning in August 2025. Both ask you questions instead of handing over answers. That was still true when we checked in August 2026.

The contract you paste before you study

You can build the same rules yourself, in one message. Paste this at the top of every study chat, before your first question. Fill in the two boxes.

You are my teacher. I am a class 8 pass student in India.
I am studying [topic] for [class 10 board exam / SSC / railway exam].
Rules for this whole chat:
1. Never give me the final answer.
2. Ask me one question at a time. Then stop and wait for my reply.
3. Keep every reply to four short sentences or fewer.
4. If my step is wrong, tell me only which step is wrong.
5. Do not correct it for me. Ask me to try that step again.
6. Use rupees and Indian examples.
Start by asking me what I already know about [topic].

Then send your problem. A reply that follows the contract looks like this.

Good. You have written cost price = 2,450 rupees.
Now, what does 12 percent profit mean here?
Is the profit counted on the cost price or on the selling price?
Tell me your answer, and tell me why.

Notice what is missing. There is no answer and no full solution. In a classroom trial in Sierra Leone, researchers counted every message a well set up AI tutor sent. It asked a question in about three quarters of them. It gave a full solution in about two messages out of every hundred.

That gives you a simple test you can run in any chat. If the app is doing all the talking and none of the asking, your contract is not working. Send the rules again in the same chat, because clear instructions beat clever wording.

Two small habits make the contract easier to keep. Do one problem per chat, so the thread stays short and the app stays on the job. Ask for short replies, four sentences at most. Short replies are easier to read on a small screen, and they spend less of your data pack.

Use the answer at the back of your book

Here is the strongest trick in this lesson, and it costs nothing. Your NCERT exercises print the answers at the back. Give the AI that answer before you ask it anything.

The correct final answer is 2,744 rupees. It is from my textbook.
Do not tell me the answer or the method.
Give me one hint at a time until I reach 2,744 myself.
If a step of mine is wrong, say only that it is wrong.

This does two jobs in one line. The app cannot hand you the answer, because you already handed it over. And a wrong answer from the app cannot quietly become your wrong answer, because you know where you have to land.

NCERT textbooks can be downloaded free from the official NCERT website. Search for the book name, then open the government site rather than a copy of it. If your book has no answer key, use a solved example from inside the chapter instead. When there is no answer to anchor to, you have to test the answer yourself, and four quick checks show you how.

Three checks you can run on yourself

None of this works unless you can tell copying apart from learning. These three checks take about a minute each.

1. Cover the screen. Redo the same problem on paper, from the first line, with nothing to look at. If you cannot, you did not learn it. You read it.

2. Use the three strike rule. If you have asked for the same type of problem to be solved three times, stop asking. The fourth one you do alone, on paper, before the app is even open.

3. Count your messages. In the study, the students who learned sent many messages about one problem. If you sent one message and got one answer, that was not study.

The first check is the honest one, and it is the one people skip. It is also the only check that matches what the exam hall will ask of you.

What to do with this contract next

Save the contract in your notes app now, while you are thinking about it. You will use it for the rest of this module. The wording stays the same for maths, for physics and for chemistry. Only the topic and the exam change.

One honest point about time. In Ghana, about a thousand students used a maths tutor on WhatsApp for two 30-minute sessions a week. That ran for eight months, and researchers measured a real gain. There is no version of this that works in one night.

One more number is worth knowing. On a large free learning site, only about one student in eleven did even 30 minutes of practice a week. The site was free and open to every one of them. So the thing that decides your result is not which app you pick. It is whether you are still doing this in March.

You may be wondering where help stops and cheating starts. That is a fair question with a clear answer, and the lesson on cheating deals with it properly. For now, keep the contract, and keep the paper.