How to Use AI as a Math Tutor—Find the First Wrong Step Without Revealing the Answer

The practical examples start with school algebra and cover mistakes with brackets, negative signs, fractions, and arithmetic, followed by other situations that need different checks. You can use the conversation workflow without installing anything. An optional section later explains a downloadable Python checker for typed one-variable linear equations. A chat prompt alone cannot guarantee a correct diagnosis or stop the tutor from revealing an answer.

Start with numbered work and a one-hint request

Include the original equation even if you have already simplified it. Without that starting point, a tutor can check the consistency of later lines while missing the mistake that created them. Give each line a stable label so the response can point to a precise transition.

Choose a text chat that your school or organization permits for this work, and paste the following request. Replace the example equation and step with your own attempt. If you are helping a student, let them write the revision before you submit it.

Help me check my algebra without solving the problem for me.

Compare the original equation with s1, then compare each pair
of consecutive steps in order. Identify the earliest transition
that does not preserve the same solutions.

Return the step label and one short question that helps me
recheck it. Do not write the corrected equation, calculate the
missing number, give my final answer, or discuss later mistakes.
Wait for my revised work.

If the work is valid so far, say that without inventing an error.
If notation or reasoning is unclear, ask me to clarify the
specific part before diagnosing it.

Original: 3(x + 2) = 15
s1: 3x + 2 = 15

All problems and conversations in this guide are constructed examples. The bracket example also appears in the downloadable attempt pack. The three pictured hints were produced by the offline script; other prompts and replies are suggested wording, not recorded AI responses or learner results.

A useful first response points to s1 and asks whether every term inside the parentheses received the outside factor. A response that supplies the new constant, writes the corrected line, or continues to isolate x has already exceeded this request. Try to identify the next change yourself before requesting a stronger hint.

Write your revision on paper first. Then resubmit the original equation and the updated numbered work with: “Check this revised attempt from the beginning. Give one hint only if a transition is still wrong.” If the first mistake has disappeared, the next mistake becomes the new target.

The prompt is a request for behavior. If an ordinary chat response reveals an answer, there is no separate display filter in this workflow to hide it. For a learner who must avoid seeing solutions, use a teacher-controlled system with reviewed hints or the offline template approach described later.

Decide what counts as the first wrong step

For an equation, a solution is a value that makes both sides equal. In this guide's supported algebra, two lines are equivalent when they have exactly the same solution set. The first wrong step is the first submitted line whose solutions differ from those of the preceding line.

This definition helps with several common disagreements. A line can look unfamiliar and still be equivalent. A correct final answer can follow invalid working. An unfinished solution can contain no error. The tutor needs to distinguish these situations before it chooses a hint.

Equation equivalence also has a limit: it checks the written mathematical result, not the explanation of how the learner reached it. For example, 2x = 4 and x + 2 = 4 happen to have the same solution. If the learner says the second line came from “dividing both sides by 2,” that explanation is wrong even though the solution sets agree. Ask the learner to write the operation; an equation-only check cannot establish that the stated operation was applied correctly.

For word problems, the first mistake may occur before any equation exists: choosing the wrong quantity for x, translating “less than” in the wrong order, or assigning incompatible units. For proofs, a claim may be true but unjustified. These require checks of meaning or reasoning that the included linear-equation parser does not perform.

Use progressively stronger hints on the same mistake

Start with a locating question. If it is insufficient, request a rule. If the learner still cannot proceed, offer an unsolved example with different numbers. Show one level at a time so the learner has an opportunity to respond.

For the bracket example, the offline script provides these hints:

H1: Recheck s1: did every term inside the parentheses receive the outside factor?

H2: At s1, apply the distributive property to every term before combining anything.

H3: Practice the same move on 2(y + 4) term by term, then return to s1.

View the three offline example hint screens

To request H2 in a chat, write: “I still cannot identify the change. Name one relevant rule for the same step without doing the calculation.” For H3, ask: “Give one unsolved practice example using a different variable and different numbers. Keep the original correction hidden.”

Read a response for correctness as well as restraint. “Check your work” is too vague to help; a complete solution provides too much. The useful middle is a clue tied to the exact operation at the identified step.

After the third hint in this workflow, show a smaller substep or ask a teacher for help. For distribution, the smaller substep might be drawing arrows from the outside factor to each term before calculating either product. If you need the rule explained from the beginning, ask for a lesson on a different equation, then return to your attempt. The three levels are a way to limit how much of the original work is disclosed, not a universal rule about how many hints every learner needs.

Match the hint to the kind of error

Choose a hint about the operation that failed. A reminder about balancing both sides will not fix a subtraction slip if the learner already made the same change on both sides.

A factor was applied to only part of a bracket

Original: 3(x + 2) = 15
s1: 3x + 2 = 15

The first break is at s1. The outside factor was applied to the variable term but not to the constant term. A locating question can ask which terms are inside the parentheses. H2 can name the distributive property. Neither needs to compute the missing product.

If the learner keeps making this mistake, ask them to mark the two multiplication relationships before writing another equation. That reveals whether they understand the scope of the bracket or are making an arithmetic slip after identifying both products.

A minus sign did not reach every term

Original: 8 - (x + 3) = 2
s1: 8 - x + 3 = 2

The first change mishandles the subtraction of the whole bracket. A useful question is: “At s1, did the minus sign affect each term inside the parentheses?” A rule reminder can ask the learner to think of subtracting the bracket as multiplying it by negative one.

This is different from changing -4x = 12 to x = 3, where the mistake concerns division by a negative number. Ask about the sign of the quotient. A sign hint should identify the relevant operation instead of treating every negative-number mistake as distribution.

An operation changed only one side

Original: 4x + 7 = 23
s1: 4x = 23

The attempt removes a term on the left without a matching change on the right. Ask: “At s1, what operation happened on each side?” This makes the learner inspect both sides without supplying the right-hand result.

A similar balance problem occurs when 5x + 2 = 3x + 10 becomes 2x + 2 = 3x + 10. Ask the learner to name the term they subtracted and locate its effect on each side. Avoid “move it across and change the sign” as the only explanation; it hides the operation that should preserve equality.

A division affected only part of the equation

Original: 2x + 4 = 12
s1: x + 2 = 12

The attempt divides the terms on the left by the same number while leaving the right unchanged. The clue should address the scope of division: “At s1, which parts of the equation received the divisor?”

This differs from a wrong numerical quotient. If both sides were divided and one result was miscalculated, ask the learner to recheck that one quotient. Giving a general balance hint in that situation can send them back to a rule they already applied correctly.

Like terms were combined incorrectly

Original: 2x + 3x = 20
s1: 6x = 20

The attempt preserves the variable but changes the sum of its coefficients incorrectly. Ask: “At s1, what operation joins the two coefficients?” If the learner used multiplication, a small analogy such as combining 2y + 4y can expose the difference between adding like terms and multiplying expressions.

If the learner instead has unlike terms, such as 2x + 3, ask what makes two terms alike before they combine anything. A term containing x and a constant cannot be combined into one coefficient of x.

The algebraic move is plausible but the arithmetic is wrong

Original: x / 3 + 1 = 5
s1: x / 3 = 3

The attempt has the shape of subtracting the same constant from both sides, but the right-hand arithmetic is inconsistent with that move. Ask: “At s1, can you recheck the subtraction on the right?” The learner should recompute it without being told the result.

Distinguish the expression x / 3 + 1 from x / (3 + 1). A photo, missing parentheses, or an unclear fraction bar can make these different expressions look similar. Confirm the transcription before labeling an arithmetic error.

A fraction was cleared without multiplying every term

Original: x / 2 + 3 = 5
s1: x + 3 = 10

The learner multiplied the fraction and the right side by the denominator, but left the separate constant unchanged. Ask: “At s1, which terms were multiplied when you cleared the denominator?” A rule reminder can ask them to apply the multiplier to the whole of each side.

Have them show the multiplication before simplifying. This lets you check which terms received it without doing the arithmetic for them. For several constant denominators, ask which common multiple they chose and whether it reached every term. If a denominator contains the variable, use the separate domain check described below.

Several later lines depend on an earlier mistake

If the original 2(x + 5) = 18 first becomes 2x + 5 = 18, later lines may follow consistently from that incorrect equation but cannot repair the lost equivalence. Focus the first hint on distribution at s1; reviewing the later subtraction and division immediately adds work that the learner will need to redo anyway.

The first wrong step may also come after a correct step:

Original: 3x + 6 = x + 14
s1: 2x + 6 = 14
s2: 2x = 6

Here s1 is equivalent to the original. The first break is s2, where the arithmetic after removing the constant is wrong. A response that marks s1 simply because it is the first line has failed to locate the mistake. Ask the learner to recheck the operation between s1 and s2, then resubmit the updated attempt. This is case M04 if you later want to reproduce it with the checker.

The method is different, unfinished, or already correct

An attempt may stop at 6x = 24 after starting from 6x - 4 = 20. The submitted transition is valid; the learner has not finished isolating the variable. Say that the work is correct so far. If they ask how to continue, ask which operation would undo multiplication by the coefficient without calculating it for them.

A learner may subtract a variable term before a constant, or simplify the opposite side first. Compare the mathematics before insisting on the order used in a model solution. If a teacher requires a particular method, review that requirement separately from equivalence.

When the final line already isolates x and the chain is valid, confirm completion without repeating the value. Ask the learner to check their own result in the original equation. Agreement with an answer key alone does not establish that every preceding step was valid.

The equation has no solution or holds for every value

As additional constructed examples, 2x + 4 = 2x + 5 can reduce to a false constant statement, while 2x + 4 = 2(x + 2) can reduce to an identity. A false constant statement has no solutions; an identity holds for every value in the supported domain. Neither situation requires inventing a numerical value for x.

If the learner has reached a constant-only line, ask: “Does this statement hold for no values, every value, or one particular value?” Have them explain their conclusion in words.

The downloadable checker can compare these solution sets, but may return no_error_yet even after a valid identity or contradiction. The optional software section explains why. The mathematical task ends when the learner correctly states which values, if any, satisfy the original equation.

Handle work the equation checker cannot judge

For a photograph, first ask the tutor to transcribe only the line it can read. Check whether a minus sign, exponent, or bracket was lost. If a symbol is unclear, rewrite that part yourself. Diagnosing an uncertain transcription can turn the tutor's reading error into a false accusation about the learner's work.

Check whether a decimal is exact or rounded

If an attempt changes 3x = 1 to x = 0.33, ask whether the learner intended an exact value or a rounded approximation. For an exact-answer task, that decimal does not preserve equality. If the task asks for two decimal places, it can be an appropriate rounded result. Check the rounding instruction before calling the work wrong.

Include the instruction in a chat request: “This problem asks for two decimal places. Check the exact algebra first, then review the final rounding separately.” Keep exact values through intermediate steps when the assignment permits it; early rounding can change later calculations.

The Python checker treats finite decimals as exact fractions and has no rounding tolerance. It cannot use the assignment instruction to decide that a rounded value is acceptable.

An inequality or cancellation needs an extra condition

For an inequality, multiplying or dividing both sides by a negative number reverses the inequality direction. In the attempted transition from -2x < 6 to x < -3, ask: “What happens to the direction when you divide by a negative number?” Keep the corrected inequality for the learner to write.

For rational expressions, canceling or multiplying by an expression involving the variable requires attention to where it is zero. A denominator cannot be zero. Dividing by a variable expression can also discard a possible solution: in (x - 1)(x + 2) = 0, dividing by x - 1 needs a separate check of the case where that factor is zero. Ask which condition the division assumes before accepting the shorter equation.

Both checks are outside the included parser. Ask a teacher or a suitable verifier to check the inequality or domain reasoning; changing the notation to make this parser accept it would remove part of the mathematical question.

The first mistake is in translating a word problem

Suppose the problem says “three less than twice a number is eleven,” and the learner starts with 2(x - 3) = 11. Before checking later algebra, ask: “Does the phrase subtract three before or after doubling the number?” Let the learner write the revised starting equation.

Have them define x and attach units to quantities where the problem needs them. If one length is in centimeters and another is in meters, ask how they put the measurements into the same unit before combining them. A correctly solved equation can still model the wrong relationship.

The included equation checker assumes that the original equation is already the right problem. A word-problem request should instead say: “First check how I translated the words into my equation. Give one hint about the first mismatch before reviewing the algebra.”

In a proof, ask which established fact justifies the first unsupported claim. A solver that returns the correct final result has not necessarily checked that justification.

If all you have is a final answer, there is no step sequence to diagnose. Ask the learner to reconstruct their working or attempt the problem again. Do not let a tutor invent intermediate lines and present them as the learner's reasoning.

Check the revision and whether the hint helped

Keep the previous attempt visible to the learner while checking the revised one from the original problem. This lets them compare their changes without assuming that the tutor's first diagnosis was correct.

Here is a constructed conversation using the earlier two-step example. It stops before the final answer:

You:
Original: 3x + 6 = x + 14
s1: 2x + 6 = 14
s2: 2x = 6
Find the first incorrect step and give one hint only.

Tutor:
s1 preserves equality. Recheck s2: what subtraction did you
intend to do on the right?

You:
I intended to subtract 6 from both sides. Here is my revision:
Original: 3x + 6 = x + 14
s1: 2x + 6 = 14
s2: 2x = 8
Check these steps again without writing the next line.

Tutor:
Both transitions now preserve equality. Your work is correct
so far. Continue with your own next step.

The learner supplies the revised equation and the intended operation. The tutor checks that change and waits. A response that continues the division would exceed the second request, even after correctly checking the revision.

If the same mistake remains, ask the learner what operation they intended. If they have repaired it and introduced a different mistake, restart with a locating question for that new transition. A longer hint aimed at the old step may now be irrelevant.

If the tutor calls valid work wrong, ask it to identify the operation it believes failed and check the two equations independently. For supported linear equations, the exact checker can help with equivalence. A calculator can check a numerical operation, but checking a few chosen values is not a general proof that two equations have identical solutions.

After the learner repairs the attempt, ask them to explain the rule in their own words and try another permitted practice problem with the same operation before requesting help. This gives you something more informative to inspect than “the original answer is now correct.” It is a suggested practice routine, not a measured learning result for this kit.

The tutor reveals the answer or keeps giving vague advice

If the answer has already appeared, a new instruction cannot make it unseen. Use a different practice problem involving the same operation, and attempt it before asking for help. You can still review the original mistake, but do not use the now-familiar answer as evidence that you can solve the next problem independently.

If the reply only says “check your work,” ask it to identify the pair of lines it compared and the operation to inspect, while keeping the correction hidden. If it keeps writing complete solutions, use a reviewed hint set or a human tutor for this task.

If an unfamiliar term makes the hint unusable, ask: “Explain that rule in simpler words using a different unsolved example.” For a beginner who cannot yet apply the rule, a short lesson on another problem may be more useful than another hint on the original.

The study on generative AI and high-school mathematics found that assisted practice performance and later unassisted performance could diverge, and that tutor safeguards mitigated the harm seen with unrestricted assistance. It supports checking what learners can do after help is removed. It does not establish the effectiveness of the prompts or code in this article.

Know when to ask a teacher and what to share

Check the rules for the actual assignment before sending it to an external service. Practice, graded homework, and restricted assessments may have different permissions. If help is prohibited, use the teacher's approved support route.

Remove names, grades, school or account details, and teacher comments that are unnecessary for checking the mathematics. If image use is authorized, crop identifying details from photographs. For minors, use the institution's approved account and data arrangements.

Ask a teacher when the input is unclear, you still cannot apply the rule after the practice example, or the suggested rule does not fit the step. Show the original problem, your latest numbered attempt, the hints already received, and the operation you intended. Keep the tutor's suggestion separate from your own explanation so the teacher can see what you understood.

Optional: run the downloadable exact checker

The sections above cover using the tutoring conversation. This section and the remaining implementation notes are for readers who want to inspect the code or build a controlled tutor. The downloadable tool is a command-line example, not a ready-made student website or chat application.

The evidence kit contains the tutor script, synthetic problems and attempts, hint policy, output schema, evidence gate, retained results, and README. Extract it into one folder. The offline checker and template hints use the Python standard library; they do not require an AI account or API call.

Inspect M04 with:

python3 ai-math-first-error-tutor.py diagnose \
  --problems ai-math-problem-pack.json \
  --attempts ai-math-student-attempts.json \
  --case M04

The returned object identifies first_error_found, step s2, and the two adjacent equations. It also contains a coarse error code and a signature, but no solved value.

The parser accepts integers, finite decimals, constant fractions, unary signs, parentheses, basic arithmetic, and one variable named x. Inputs such as 3x are supported. Use parentheses to make a fraction's scope explicit, such as (x + 2) / 3. It rejects nonlinear multiplication, powers, variable denominators, other variables, unknown symbols, and division by zero, with expression-length and exact-value limits.

One implementation detail affects diagnosis: the script parses the entire submitted attempt before comparing adjacent lines. If s1 already contains a supported mistake but a later line contains an unsupported power, the result is needs_review, not a diagnosis of s1. You may re-enter the clearly transcribed linear portion for a limited check, but the result then applies only to that portion.

The solution-set comparison is exact within this supported arithmetic. Error naming is heuristic. For example, removing parentheses makes the classifier consider a distribution mismatch; it does not independently prove that distribution caused the mistake. With 2(x + 1) = 6 changed to 2x + 2 = 8, the left-hand distribution is correct, but the right side changed. The current code still labels it distribution_mismatch. Inspect the adjacent lines before trusting the suggested rule.

Research on stepwise verification and remediation treats finding an error and generating useful feedback as distinct evaluation tasks. That distinction matters here: locating the right transition does not guarantee that the error label or hint explains it correctly.

Request an offline hint without skipping its review

Run one template hint for M04 with:

python3 ai-math-first-error-tutor.py hint \
  --problems ai-math-problem-pack.json \
  --attempts ai-math-student-attempts.json \
  --policy ai-math-hint-policy.json \
  --case M04 --level 1 --provider template

The hint policy defines the three levels and their limits. The command prints an output object and a leak_gate result. The object carries the step label, hint level, hint text, wait_for_revision: true, and reveals_answer: false. Those flags describe the requested behavior; they do not prove that the text is harmless or helpful.

The script includes next_hint_level, a helper that retains the current level for the same error, advances on a request for more help, and resets to level one when the error signature changes. At level three it stays at three. The command above accepts --level directly and does not persist a tutoring session or call that helper to enforce progression.

An application using this code must store the case, signature, and hint level itself, validate each request, and implement the handoff after an unsuccessful H3. Without that application logic, a caller can request level three immediately or repeatedly run the same hint command. The evidence gate tests the helper; it does not turn the command-line script into a complete interactive tutor.

Keep answer data separate from the model input

The problem pack includes answer data for the constructed problems: the original equation, expected solution, and corrected equations that a hint must not disclose. The attempt pack includes submitted steps and expected diagnostic labels. Both are publicly downloadable examples.

In an adapted system, keep actual answer keys and expected labels on the teacher or server side. Do not give learners or the hint model the full teacher-truth file. Author expected labels before evaluating a model, and review disagreements independently so the model is not grading against answers it invented itself.

The optional model receives only a small packet:

{
  "step_id": "s1",
  "hint_level": 1,
  "error_code": "distribution_mismatch",
  "previous_equation": "3(x + 2) = 15",
  "current_equation": "3x + 2 = 15"
}

This omits the stored solution, later steps, names, and grades. It reduces unnecessary exposure, but the model can still derive the answer from the equation. Separating files therefore reduces the material sent to the provider; it does not make answer leakage impossible.

Add AI phrasing only after the offline path works

The script's optional provider uses the Responses API and the snapshot gpt-5.4-mini-2026-03-17. The official model documentation lists that snapshot and support for Responses and Structured Outputs. The downloadable kit contains no retained live-model run, so the documentation does not establish that its particular API call succeeds in your environment.

To review that path, install the official OpenAI Python package in a separate environment, make your API key available as OPENAI_API_KEY outside the files, and change --provider template to --provider openai. Use synthetic cases first and retain the response before considering any learner-facing use. Live requests may incur API charges.

The implementation sends higher-priority tutoring instructions, the minimal packet, a strict five-field output schema, and an output-token limit. Structured Outputs constrains the response format; its content can still contain mistakes. A refusal, incomplete output, missing response text, or invalid parsed object needs an application-level stop path. The sample script does not provide a complete retry and recovery system for these outcomes.

It also sets store=False. Do not describe that setting as zero retention. OpenAI's data controls documentation distinguishes application state from abuse-monitoring logs and organization-level retention controls. Check the provider and your institution's agreement before sending learner work.

Understand what the leakage checks actually enforce

The local validator checks field names, step ID, an allowed level, word limits, stop flags, selected direct-answer phrases, equations in early hints, known corrected equations, and some solved-value patterns. It also rejects certain lists, newlines, and multiple question marks. These are useful checks of output shape and recognizable strings.

They do not provide a complete semantic review. A hint can satisfy the checks while using a wrong rule, describing the correction indirectly, or spelling the answer in words. During this editorial review, the validator accepted “Recheck s1: the value of x is three” for E01. It also accepted a Pythagorean-theorem hint for the same distribution error. These were hand-authored probes passed directly to the validator, not live-model outputs.

The first probe exposes a spelling-based leakage gap; the second exposes the lack of a relevance check. A PASS for either would not justify showing it to a learner. The Boolean reveals_answer: false is also a self-description, not an independent judgment.

The retained rejection screenshot demonstrates a known unsafe candidate being caught. It does not show that every possible paraphrase is blocked.

For strict withholding of answers, display only hints that a mathematics educator has reviewed for the particular problem and operation. If you permit model-generated phrasing, keep a separate review step before display and test semantic paraphrases, number words, alternate notation, irrelevant rules, and correct-but-overexplicit advice. Reject a failed candidate without displaying its text to the learner. A retry or fallback must pass the same review.

You must also decide who sees command-line output. The current hint command prints the candidate together with the gate result, including when validation fails. A learner interface should not show that raw object. It needs to display only approved text and keep rejected text in appropriately protected review logs.

Reproduce the evidence and read its limits

Run the evidence gate from the extracted kit folder:

python3 ai-math-evidence-gate.py

The downloadable gate passed again during this review. Its fixture set contains 24 synthetic attempts: seven complete correct attempts, one incomplete correct attempt, ten single-error attempts, four attempts with multiple errors, and two unsupported or malformed inputs. It produces 42 template hints for the 14 supported error cases and rejects eight built-in negative mutations.

The gate checks expected diagnoses, template output, hint-state behavior, selected parser boundaries, and deliberately damaging changes. The mutations include an incorrect expected step, direct-answer and corrected-equation leaks, level four, a stale signature, falsely accepted unsupported work, changed synthetic-data declarations, and an expanded hint policy.

Those checks establish agreement within the retained set. They do not establish resistance to every leak, accurate misconception labels outside the fixtures, persistent application behavior, or instructional quality. The additional probes in this review show why the built-in PASS needs that narrow interpretation.

The retained result file and README identify the fixtures as synthetic and report no live-model execution or qualified mathematics-educator review. The result's date is part of the original retained record; rerunning the gate does not update that historical date unless you explicitly create a new result.

For a new application, evaluate error localization, error explanation, and hint quality separately. Include valid alternative methods, correct incomplete work, multiple mistakes, identities, contradictions, misleading operation descriptions, and ambiguous input. Review whether each hint is correct, tied to the targeted transition, understandable, and sufficiently restrained. A leak count without a relevance check misses hints that conceal the answer but teach the wrong rule.

Before using this implementation with learners, assign an owner for answer data, approved hints, session progression, display filtering, and handoff. Use anonymous case IDs and keep unnecessary personal information out of model input and logs. Have a mathematics educator review the problems and hints, retain any dated model outputs, and investigate failures before approving instructional use in the intended setting.