Gains:
- Ability to use artificial intelligence in problem solving in three different roles as strategy consultant, solution producer and solution auditor, appropriate to the risk level of each.
- Ability to systematically check each step of a solution by numbering, rule querying, arithmetic checking, checksums and listing all cases.
- Being able to verify both the result and the solution with independent methods, understanding that the correct result does not mean the correct path.
The solution to a mathematical problem is a chain of interconnected steps; If any link in the chain is broken, the result is invalid. AI is very good at building this chain quickly — but it's up to humans to monitor the chain's strength. In this unit, you will learn how to use AI as a problem-solving partner and, more importantly, how to systematically control each step produced. The goal is to take advantage of the speed of AI while not missing any mistakes.
First, a definition: Step-by-step verification is checking whether each intermediate step of a solution is correct one by one. In mathematics, "Is the result correct?" The question is not enough; "Is each intermediate step correct?" The question is essential. Because sometimes wrong steps lead by chance to the right result (an unstable solution), sometimes a single arithmetic error in a correct strategy spoils the result.
Three different roles of AI in problem solving
You can use AI in problem solving in three different ways, and each requires a different validation:
- Strategy consultant: "How do I solve this problem?" Here, AI suggests a method. The risk is low because you do the calculations yourself.
- Solution generator: "Solve this problem." The risk is high because both the strategy and the calculation come from AI; Every step must be supervised.
- Solution checker: "Check my solution." It is medium risk; AI can find bugs, but it can also make up bugs that don't exist or miss the real bug.
The safest approach is usually to combine 1 and 3: get the strategy from the AI, do the calculation yourself or with SymPy, then have the AI check and cross-check.
Step by step: how to check a solution
1. Number each step. Asking the AI for the solution in numbered steps allows you to then say, "There's an error in step 3."
2. At each step "which rule?" sir. Have the AI write down which algebraic rule, theorem, or identity is used in each pass. If the name of the rule is wrong, the step is also questionable.
3. Check the arithmetic separately. Even if the algebraic logic is correct, number errors are common. Review each addition, multiplication, and sign change.
4. Try bounds and exceptions. Test the result at simple values like x=0, x=1. If the formula is nonsensical at these values, there is an error.
5. Go back in reverse. Check the result of the derivative by integrating, substituting the root of the equation, multiplying the factorization back.
6. Look for size/unit consistency. If the units do not add up in applied problems (e.g. adding meters and seconds), the formula is incorrect.
Tip: "Verification by reverse operation" is the most powerful verification tool in mathematics and is almost free. If you found the root of an equation, substitute it; If you took an integral, take its derivative; If you did a factorization, multiply back. This 30-second check prevents hours of wrong direction.
The power of provisioning comes from its independence from AI. When you substitute a root, you make a deterministic check that is completely independent of what the AI does in the solution process: either the equation is satisfied or it is not, there is no in between. That's why verification is much more valuable than the AI saying "sure, right"; because the AI may repeat the same mistake when validating its own solution, but independent verification is outside this loop. When you can deliver a critical result in two different independent ways (e.g. both substitution and SymPy), your confidence increases exponentially.
A classification of errors
The table below summarizes the types of errors you most commonly see in AI solutions and the way to catch each one.
Error type
example
capture path
sign error
Writing −(a−b) = −a−b
Opening and checking my name manually
arithmetic error
7×8 = 54
Confirmation with calculator/SymPy
wrong rule
forget the chain rule
"Which rule?" to ask
Field/definition error
get log(negative)
Checking the domain
Skipped status
Ignoring discriminant < 0
List all statuses
Made-up theorem
non-existent theorem name
Confirmation from source
three mini cases
Case 1 — Right result, wrong path. A student asked the AI a limit question. The AI found the result to be 3 — which was correct — but in the intermediate step it used L'Hôpital's rule where it could not be applied (without the 0/0 uncertainty). The student asks "which rule applies and why?" When asked, the AI accepted the mistake and reached 3 again in the correct way. The result was the same, but the path was not solid; The journey was also scored in the exam.
Case 2 — Omitted case. An engineer, |2x − 4| = 6 absolute value equation was solved by YZ. YZ only solved 2x − 4 = 6 and gave x = 5; It skipped the 2x − 4 = −6 state (x = −1). The engineer asks "how many states are there in absolute value?" When I asked, the second root appeared. The correctness of both roots was confirmed by substitution.
Case 3 — Cross-check found 3 errors. A teacher had the AI do a 10-step trigonometric identity proof. Then he compared the same proof step by step with SymPy's simplify function. He found that the AI applied identity incorrectly in 3 out of 10 steps. Total check time 12 minutes; The trust gained is priceless.
Four copyable templates
1) Numbered, rule-based solution:
Solve this problem: [problem]. Write the solution in NUMBERED steps. In each step, state the name of the rule/theorem you used in parentheses. Do not skip any step by saying "obviously"; show each transition.
2) Don't force all states:
Solve this problem: [problem]. List ALL POSSIBLE cases when solving (e.g. absolute value, quadratic, domain constraints). Do not skip any cases; If you skip it, justify why you skipped it.
3) Requesting provision:
You found: [result]. Now VERIFY this result by the inverse operation:[differentiate / substitute / multiply back]. Show the result of the checksum and tell whether it is consistent with the original problem.
4) Error-oriented control:
JUST look for errors in the solution below. Check each step in order; when you find the first error, write down which step, what type of error (sign/arithmetic/rule) and what the correct one is. Creating a new solution.Solution: [here]
Weak prompt / Strong prompt
Weak: "Solve x² − 5x + 6 = 0."
Result: "x = 2 and x = 3" — may be true, but you won't notice if the odd case is missed or the factorization is wrong.
Strong: "Solve the equation x² − 5x + 6 = 0 by both factoring and the discriminant formula; show that the two methods give the same roots. Then provide each root by substituting it into the equation."
Result: Two independent methods verify each other, ensuring the result is accurate. If there is an error, the two methods will conflict and you will see it immediately.
Common mistakes
- Just looking at the result. The right result does not mean a solid path. Also check the road.
- Skipping provisioning. A 30-second substitution/back derivative check catches most errors; Don't neglect it.
- Not listing situations. Omissions in absolute value, quadratic, root and domain constraints are the most common errors.
- Trusting the AI to say "no bugs". AI in the role of auditor may miss the real error; The human eye is essential in critical work.
- Being satisfied with one method. If possible, solve and compare using two independent methods.
Caution: When having the AI check your solution, it may sometimes make up a non-existent "bug" and dissuade you from your correct solution. So when the AI shows an error, you also check whether that error is really an error. The auditor can also be wrong.
In summary
In problem solving, AI plays three roles: strategy advisor, solution generator, solution auditor. The safest approach is to take the strategy from AI, do the calculations with independent tools, and systematically audit each step. Numbered steps, "which rule?" question, checksum (inverse operation), limit cases, and listing of all cases are the basic tools for catching errors. The right result does not mean a solid path; Check both the result and the path.
Application task
Choose a quadratic equation or an absolute value equation. Solve YZ with templates 1 and 2. Then: (a) check that all cases have been covered, (b) ensure each root by substituting, (c) solve and compare again by a second method, if possible. If you find an error, note what type it is (sign/arithmetic/rule/skipped condition).
checklist
- [ ] I took the solution in numbered steps.
- [ ] I confirmed the rule used in each step.
- [ ] I have verified that all contingencies have been addressed.
- [ ] I checked the result by inverse operation (checking).
- [ ] If possible, I compared it with a second method.
- [ ] I also independently verified the "errors" the AI showed.