AI for studying

Math solvers and homework help: how to get unstuck without losing the learning

Stuck on homework? Five steps to help that teaches: attempt, name the stuck point, get a hint not an answer, find the wrong line, redo it cold.

To get unstuck on a math problem without losing the learning, attempt it first and write down exactly where you stopped. Then get a hint for that step, not the whole answer. Compare your working to a verified solution and find the one line that is wrong. Redo the problem from a blank page, and log the error. A step-by-step answer helps after your attempt, as a worked example. Before it, it is copying with extra steps. This page is the method, not a solver: nothing here solves a pasted problem, and the example is our own.

Two kinds of stuck

Homework help starts with naming which kind of stuck you are. There are two, and they need different help.

Stuck on a step. You know the method. You started the problem, you got three lines in, and one line will not come. You know what the answer should look like; you cannot get there. This is the common kind, and it is the cheap kind to fix: one hint for one step.

Stuck on the method. You read the problem and nothing starts. You do not know which rule applies or what the first line is. This needs a worked example of the same kind of problem, read slowly, then an attempt of your own.

Both kinds need your attempt first, even the second. An attempt that goes nowhere still tells you, and whoever helps you, where the gap is. "I don't know where to start" is a feeling; "I don't know what to do with the −3 in front of the log" is a place. The rest of this page turns the feeling into the place, and then fixes the place.

Why the one-message answer costs you later

A solver app or a chatbot will give you the finished solution in one message. It feels like help. The research says it is help for tonight's homework and a cost on the test.

Bastani and colleagues (2025) ran a field experiment with nearly a thousand high-school students doing maths practice. Students who could ask an unrestricted AI assistant anything solved more practice problems than students with no AI: their practice scores were 48 per cent higher. Then everyone took the exam without the tool. The unrestricted group scored 17 per cent lower than the students who had never had AI at all. A second group had an assistant built to coach, giving hints and not final answers; their practice scores were 127 per cent higher and their exam scores were no worse than the no-AI group's. The students who had been handed answers had practised asking, not solving, and the exam asked them to solve. The four-part test for any study tool is built on this study.

The same point was known before AI. Sweller and Cooper (1985) found that students learning algebra from worked examples, studied carefully, learned the procedure faster than students who only solved problems. A worked example teaches when you study it as an example: you follow each step, you ask why it is there, and then you do one yourself. The same text handed over as "the answer" and copied into the homework is read, not learned. Which of the two it is depends on what you do before and after you see it, and that is what the worked-example method for maths is about.

So the question is never whether to see a step-by-step solution. It is when.

The five steps to get unstuck

This is the method for help with math problems that leaves you able to do the next one. It works for help from a textbook, a classmate, a teacher or a tutor, human or AI.

  1. Attempt it, and write the exact point you stopped. Work the problem as far as you can on paper. When you stall, write one line under your working that names the stall: "I don't know what to do with the −3 in front", "I combined the logs and got an answer that does not check", "I don't know which rule starts this". If you cannot write a single line of working, write what kind of problem you think it is.
  2. Ask for a hint for that point only. Not "help me solve this math problem", which invites the whole answer. Ask the specific question: "Which log law turns a sum of two logs into one log?" Ask the textbook's worked example, a classmate, the teacher at lunch, or a tutor that hints. If you are asking an AI tutor, say so in the first line: "Do not give me the answer. Which rule applies to line 2?"
  3. Compare your working to a verified solution and find the first line that differs. Once you have finished, or once a hint has got you to the end, put your working beside a solution you trust: the textbook's, the teacher's, or one that was checked against a verified answer key. Read down both. Stop at the first line that is different. That line, or the one before it, is the error. Say the rule it broke in your own words.
  4. Redo the problem from a blank page. Close the solution. Close your old working. Do the whole problem again from the question alone. If you cannot, you did not learn it yet; you recognised it. The redo is where the learning happens, for the same reason recall beats rereading (Roediger and Karpicke 2006): producing the answer is the learning, and reading an answer is not producing it.
  5. Log the line and the rule, and redo it in two days. One row in an error log: the problem's type, the wrong line, the rule in your words. Then do a similar problem cold two days later. Metcalfe (2017), reviewing the research on learning from errors, found that an error followed by corrective feedback on the specific mistake, and then a chance to try again, is learned from; an error that is simply replaced by the right answer is not. The two-day return is the spacing that makes the fix hold.

The first two steps are what stop homework help from becoming copying. The last three are what make it learning.

"Check my work, find the line": the alternative to a solver

A solver does one thing: it produces a solution to the problem you typed. It does not read your working. It cannot tell you that your line 2 is wrong, because it never saw your line 2.

Finding the line is the thing you actually need when you are stuck on a step, and it is the thing you can do yourself with a textbook. Put your page beside the book's solution. Read both, line by line, top down. The first place they differ is where to look. Then ask: what rule did my line use, and what rule did theirs use? Write the difference in one sentence. That sentence is the whole lesson of the problem, and it is what goes in the error log. Chapter 4 of the free playbook, "Find the line", is the error log as a page to print.

A tool can do the same, and the thing to ask of it is what it compares against. A general chatbot asked "check my math" will read your working and generate a judgement, and that judgement is a prediction. It will sometimes agree with a wrong line, and sometimes invent a correction to a right one. A checker that compares your answer to a verified answer key, computed and checked by code, is a different kind of tool: it can tell you that your answer is wrong because it is wrong, not because it looks wrong. Ask any tool which of the two it is before you trust its mark.

When a step-by-step answer helps, and when it replaces the thinking

A math solver that shows steps is not good or bad. The timing is.

It helps It replaces the thinking
After your attempt, as the check in step 3 Before any attempt, as the first thing you look at
For a brand-new procedure, as a first worked example you then redo Copied into the homework as your work
When you redo the problem cold afterwards When you cannot tell whether it is right
When you can check it against something verified When it is the fifth one tonight and you have redone none

The right-hand column has a second problem beyond learning: the answer is often wrong. A chatbot produces a likely-looking solution, not a computed one, and a photo solver misreads a handwritten 7 as a 1 or an x as a times sign. In our reading of 320 one- and two-star public reviews of seven answer and flashcard apps (2026-10-18), the most frequent complaint, in 105 of them, was a wrong answer the student had trusted. A student who cannot yet do the problem is exactly the student who cannot tell the wrong solution from the right one, which is why the check in step 3 has to be against something verified, not against another generated answer.

Worked example: a log equation, the wrong line, the redo

A Grade 12 Advanced Functions (MHF4U) student is on the idea called solving logarithmic equations, and the homework problem is ours: solve log₂ x + log₂(x − 2) = 3.

The attempt. The student writes:

  • Line 1: log₂ x + log₂(x − 2) = log₂(2x − 2)
  • Line 2: log₂(2x − 2) = 3, so 2x − 2 = 2³ = 8
  • Line 3: 2x = 10, so x = 5

Then the check: log₂ 5 + log₂ 3 = log₂ 15, and log₂ 15 is about 3.9, not 3. It does not check.

Step 1, the stuck point written down: "Combined the logs, got x = 5, it doesn't check. The combining line must be wrong."

Step 2, the hint, for that point only. The student asks a classmate, or a tutor told not to give the answer: "Which log law turns a sum of two logs into one log, and what does it do to the arguments?" The hint back: "The sum of two logs with the same base is the log of the product. You add the logs; you multiply what is inside."

Step 3, compare to the verified solution. The textbook's working:

  • log₂[x(x − 2)] = 3
  • x(x − 2) = 2³ = 8
  • x² − 2x − 8 = 0
  • (x − 4)(x + 2) = 0, so x = 4 or x = −2
  • x = −2 is rejected, because log₂(−2) is undefined: the argument of a log must be positive, and with x = −2 both logs in the original equation have negative arguments.
  • x = 4.

Check: log₂ 4 + log₂ 2 = 2 + 1 = 3. It checks.

The first line that differs is line 1. The student's line added the arguments, x + (x − 2) = 2x − 2; the rule multiplies them, x(x − 2). Everything after line 1 was done correctly on the wrong equation, which is why the arithmetic looked fine and the check failed.

Step 4, the redo, cold. Solution closed, old page turned over. The student writes log₂[x(x − 2)] = 3, gets x² − 2x − 8 = 0, factors, rejects −2, and checks x = 4. About four minutes.

Step 5, the log. One row: "Log equations. Sum of logs: multiply the arguments, don't add them. Also: reject any root that makes an argument negative." Redo a similar problem on Thursday. The same idea sits on the MHF4U course map beside the exponential equations it leads back to, which is where Thursday's similar problem comes from.

The AI's part, if there was one, was a hint and a check. The working on the page is the student's.

Homework help for a parent: what to do when your child is stuck

If you searched "homework help" for your child, the best thing you can do is not solve it. Most parents reach for the answer because the child is upset and the answer ends the upset. It also ends the learning.

Instead, ask one question: "Where exactly did you stop?" Make them point at a line. If they cannot point, the stuck is the method kind, and the fix is to find the worked example in the textbook for that type of problem and read it together, slowly, asking "why did they do that?" at each step. Then close the book and have them try the homework problem again.

If they can point at a line, the stuck is the step kind. You do not need to know the maths. Ask them to say what rule the line used; ask them to find the same rule in the textbook; let them compare. Then have them redo the whole problem on a clean page, and write one row in an error log. You have given them the five steps without doing any of them.

Where to get hints that teach

In rough order of how available they are at 9 pm:

  • The textbook's worked examples. The example before the exercise set is almost always the same kind of problem. Cover its solution and do it; then compare.
  • A classmate who explains without writing. Ask them to talk you to the next line, not to hand you their page.
  • The teacher's office hour or lunch help. Bring the attempt and the stuck line, not the blank page. A teacher who can see where you stopped can help in two minutes.
  • School tutoring. Most schools run a free help period or peer tutoring. Same rule: bring the attempt.
  • An AI tutor that hints. One that asks what you tried, hints before it answers, and checks against something verified. The five questions that tell an AI tutor from an answer bot take ten minutes to run and are worth running before you rely on one.

Whatever the source, the first words out of your mouth are the stuck line, and the question is for a hint, not the answer.

The method on one page

  1. Attempt, and write the exact point you stopped.
  2. Hint for that point only, from a textbook example, a classmate, a teacher or a tutor that hints.
  3. Compare your working to a verified solution; find the first line that differs; say the rule.
  4. Redo the whole problem from a blank page.
  5. Log the line and the rule; redo a similar problem in two days.

A step-by-step solution helps after your attempt, as a worked example you then redo, and when you can check it. It replaces the thinking before any attempt, copied into the homework, when you cannot tell if it is right, or when it is the fifth tonight.

Questions students ask

How do I get help with math homework without cheating?

Attempt the problem and write down exactly where you stopped, then ask for a hint for that step only, from the textbook's worked example, a classmate, a teacher or a tutor that hints. Compare your working to a verified solution, find the wrong line, redo the problem cold, and log the error. What you hand in is your redo.

Is it bad to use a math solver for homework?

Before you try the problem, yes, for your learning: in a controlled study, students who practised with unrestricted AI answers scored lower on the exam without them. After an honest attempt, a step-by-step solution is a worked example: read it, close it, redo the problem. And check it, because solver and assistant answers are often wrong.

How do I find where I went wrong in a math problem?

Put your working beside a solution you trust and read down both until the first line that differs. That line, or the one before it, is the error; say the rule it broke in your own words. Then redo the whole problem from a blank page without looking at either.

Why do step-by-step math solvers give wrong answers?

General AI assistants generate a likely-looking solution rather than computing a verified one, and photo solvers misread handwriting and symbols. Students then trust the output because it looks complete. In our reading of public reviews of answer apps, the most common complaint was a wrong answer the student relied on.

What is the best homework help for math?

A hint for the exact step you are stuck on, followed by your own redo. The textbook's worked examples, a teacher's help period and a tutor that asks what you tried before it answers all do this. Anything that gives you the finished answer first is the weakest help, however fast it is.

Put tonight's studying on rails

YSR Studies turns your notes, lectures and course into cards that come back before you forget, a tutor that makes you try first, and a map of what you know. Free, five hours a month, no card asked.

Open YSR Studies

Read next

Sources

  1. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
  2. Sweller, J., & Cooper, G. A. (1985). The use of worked examples as a substitute for problem solving in learning algebra. Cognition and Instruction, 2(1), 59–89. https://doi.org/10.1207/s1532690xci0201_3
  3. Metcalfe, J. (2017). Learning from errors. Annual Review of Psychology, 68, 465–489. https://doi.org/10.1146/annurev-psych-010416-044022
  4. Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
  5. Our reading of 320 one- and two-star public app-store reviews of seven answer and flashcard apps, 2026-10-18. The most frequent complaint theme, in 105 of the 320, was a wrong answer the student had trusted. No app is named here.