There is one test that tells you whether AI helped you with a math problem or quietly cost you something. Close the tab, take a clean sheet of paper, and do the same problem again from the first line. If you cannot, the homework got finished and the learning did not happen.
That is not a lecture about cheating. It is a measured result. Researchers ran a controlled experiment with close to a thousand high school students and found that the group given unrestricted AI help during practice later scored 17 percent worse on an exam than classmates who never had it at all.
What Happened When The AI Was Taken Away
The experiment ran in Turkish high schools in autumn 2023, across grades 9 through 11. Students got four 90-minute practice sessions covering four different math topics, and their classes were randomly sorted into three groups.
One group practiced with a standard ChatGPT-style interface. A second used a modified version built to behave like a tutor: it had the correct solutions loaded, and it was designed to hand out hints rather than the answer. The third group worked the way students always have, with a textbook and their notes.
During practice, the AI groups looked spectacular. The standard interface produced 48 percent more correct answers than the control group. The tutor version produced 127 percent more.
Then every session ended the same way: a 30-minute exam on that topic, with no help of any kind. On those exams the standard ChatGPT group scored 17 percent below the control group. The tutor group came out roughly level with it. The full paper, published in PNAS in June 2025, is worth reading if you want the statistics rather than the summary.
The short version: Same model. Same practice time. Same students, near enough. The only thing that changed was the shape of the help, and it decided whether the practice built anything.
Notice what makes this so easy to walk into. Nothing goes wrong at the time. The practice work comes back nearly perfect, which is exactly the feedback that tells you to keep doing what you are doing.
And the reckoning here was brutally fast. Not a final exam months out, but the same afternoon, half an hour after the laptops closed. Outside a controlled study that gap is usually weeks, which makes the cause far harder to connect to the effect.
Answer-Shaped Help is The Problem, Not AI
The tutor version in that study was not a different or better model. It was the same technology pointed at a different job: explain the route, withhold the destination.
So the useful question is not whether you are allowed to use AI for homework. Your school may have a policy, and you should follow it. The question that actually changes your grade is what you ask the thing to produce.
Ask for an answer and you get a transcription task. Your hand moves, your working memory stays idle, and nothing is stored. Ask for a method and you still have to run it yourself, which is the part that leaves a trace.
The Four Steps That Keep The Work On Your Side
This is the sequence to use on any problem set, with any tool. It costs about five extra minutes per problem you are stuck on.
1. Attempt it First, Even Badly
Write the first line before you open anything. A wrong first line is more useful than no line, because it tells you and the tool exactly where your understanding stops. Students who skip this step never find out what they actually did not know.
2. Ask for The Method, Not the Number
Type the instruction explicitly. Two that work on any assistant:
- “Do not give me the answer. Tell me which method this problem needs and why that one.”
- “Here is my working so far. Find the first line where I go wrong, and stop there.”
The second prompt is the strongest one in the list. It turns the tool into a marker instead of a ghostwriter, and it leaves the actual fixing to you.
3. Use a Tool That Shows Steps by Default
General chatbots will hand over a final answer unless you stop them, so the guardrail has to come from you every single time. Purpose-built solvers invert that. MathSolver is free, needs no account, and lays out each step with the reasoning attached, so the method is what you see first and the answer is just where the steps end. Photograph a handwritten problem or type it in, then use the follow-up questions to interrogate the step you did not follow.
Be honest with yourself about how you are using it, though. A step-by-step display that you scroll past is the same transcription task as before, in a nicer layout. The tool cannot make you do step 4.
4. Redo it Cold
Close everything and work the problem again on paper. This is the entire difference between the two groups in that study, compressed into one habit.
It also matches a result psychologists have replicated for decades. Retrieving something from your own memory is what fixes it there, more than rereading it does. Roediger and Karpicke demonstrated the effect on delayed tests in their 2006 study in Psychological Science, and every study-skills recommendation about self-testing traces back to that family of findings.
Why the Material Feels Easy Right Up Until the Test
Reading a clear solution produces a specific sensation. Everything follows, nothing is confusing, and it genuinely feels like understanding. That feeling is real, and it is not knowledge.
What you have experienced is fluency: the problem was easy to process because someone else did the hard part. Fluency and mastery feel almost identical from the inside, which is why students who have been reading AI solutions all term are so often shocked by an exam result.
The cold redo is the cheapest way to tell them apart. Two minutes of trying it yourself will tell you more about your readiness than an hour of nodding along to worked examples.
Where This Advice Runs Out
A few honest limits, because none of this is a guarantee.
The experiment above is one study, in one country, in one subject, with one generation of model. The authors themselves called for replication elsewhere before anyone treats it as a universal law. It is strong evidence about a mechanism, not a finished science.
AI tools also still make arithmetic mistakes while getting the method right, which is a strange failure mode to be on guard for. That is another argument for the cold redo: your recalculation is the check. No solver on the market independently verifies its own output, and any that claims perfect accuracy is overselling.
Free tools tend not to save your history either. MathSolver keeps chats in your browser rather than in an account, so clearing your cache clears your work. If a solution matters to you, copy it into your own notes.
And check your school’s rules before any of this. A policy that bans AI outright is not overridden by the fact that you used it thoughtfully.
The routine, in order
- Write the first line of the problem yourself, wrong or right
- Ask for the method or for your first mistake, never the final answer
- Read the steps to find the one you could not have produced alone
- Close the tab and redo the whole problem on paper
- If the redo fails, that topic goes on the revision list, not the finished pile
The One-Question Version
Every rule here collapses into a single check you can run in thirty seconds: after the help, can you still do it alone?
Students who can answer yes are using AI the way the tutor group in that study did, and their exam scores hold up. Students who cannot are the 17 percent group, and they will not find out until the paper is in front of them. The tool does not decide which one you are. The five minutes after you get the explanation do.
