You Got It Wrong. Now What? Using an AI Tutor on Your Own Mistakes

TL;DR. Most students review a practice quiz by scrolling the results, nodding at the right answers, and closing the tab. That converts a diagnostic into a participation trophy. A wrong answer is data about a specific broken link in your understanding, and the repair depends entirely on which link broke. This piece walks through the four reasons people miss questions, the prompts that get an AI tutor to diagnose rather than re-explain, and a concrete review loop you can run in fifteen minutes. SimpleQuizMaker has an AI Tutor built into the results screen so you can ask about a specific question without retyping it.

The review problem nobody names

Here is the sequence almost everyone follows after a practice test.

Take the quiz. See the score. Feel something about the score. Scroll through the questions you missed. Read the correct answer. Think "ah, right, of course." Close the tab.

That whole sequence takes about ninety seconds and produces almost nothing. The feeling of "ah, right" is the problem — it is recognition, and recognition is not retrieval. Seeing the correct answer next to the question makes the connection feel obvious in a way it absolutely will not be three days from now when the answer is not sitting there in green text.

Cognitive psychologists call this the *fluency illusion*: the ease of processing something is mistaken for the strength of knowing it. Rereading produces enormous fluency and very little durable memory. It is the single most common study mistake, and reviewing a quiz by reading the answer key is just rereading wearing a costume.

The fix is not to review harder. It is to review differently — to treat each wrong answer as a small investigation with a specific question: *what, precisely, went wrong here?*

Four ways to miss a question

Not all wrong answers are the same failure. Sorting them is the highest-leverage thing you can do, because the four types need four completely different responses, and using the wrong repair is why review often does not stick.

1. The knowledge gap

You did not know the fact. There was nothing to retrieve. You had never encountered the term, or you encountered it once in a lecture and it never made it into memory.

This is the easiest type to fix and the least interesting. You need encoding, not analysis: read the material, make a card, schedule it for review. Do not spend twenty minutes philosophizing about a gap that just needs a flashcard.

2. The retrieval failure

You knew it. You could not get to it. This is the tip-of-the-tongue miss, and it feels different from a gap — there is a frustrating sense of *almost*, and when you see the answer you get a jolt of recognition rather than a shrug.

Retrieval failures do not mean you need to reread. They mean the memory exists but the path to it is weak. The repair is more retrieval practice, spaced out — and specifically retrieval with different cues than the ones you originally learned with, so the fact is reachable from more than one direction.

3. The misconception

You knew something confidently, and the thing you knew was wrong. This is the most dangerous type because it does not feel like a gap at all. You were not guessing. You were sure.

Misconceptions are structural. They usually come from an over-generalized rule ("i before e"), a half-remembered heuristic applied outside its range, or an analogy that worked in chapter two and quietly stopped working in chapter six. Rereading the correct fact rarely dislodges one, because the misconception has its own internal logic that the correct fact does not address.

This is the type where an AI tutor earns its keep, and we will come back to it.

4. The process error

You knew the material, you retrieved it correctly, and you still got it wrong. You misread "not," you inverted the fraction on the last step, you picked the answer to the question you expected instead of the one that was asked, you ran out of time and guessed.

Process errors are not knowledge problems and treating them as knowledge problems wastes hours. The repair is procedural: slow down on negations, write out the step you keep skipping, do a timed set to build pacing. If you restudy the content, you will fix nothing, because the content was never broken.

Why the four-way split matters

Consider a student who misses six questions on a biology practice test and spends the evening rereading the chapter.

If two misses were process errors, that reread does nothing for them. If three were retrieval failures, the reread actively hurts — it builds fluency without building retrieval strength, so the student walks away *more* confident and no more capable. If one was a misconception, the reread probably glides right over it, because the sentence that would correct it reads as agreeing with what they already believe.

That is a full evening of studying with close to zero return, and the student has no way to know, because the subjective experience of rereading is pleasant and productive-feeling.

Sorting first turns that evening into: two procedural notes, three cards scheduled for spaced retrieval, and one genuinely hard conversation about a belief that is wrong. Maybe forty minutes, and every minute of it aimed at something real.

Getting an AI tutor to diagnose instead of lecture

The default behaviour of any language model asked about a missed question is to explain the correct answer. That is the least useful response available, because you already have the correct answer — it is printed on the results screen.

What you actually need is a diagnosis of your answer. Why was the thing you picked attractive? What would have to be true for it to be right? What does choosing it suggest about the model in your head?

The difference is entirely in how you ask. Compare:

> "Explain why the answer to question 4 is B."

against

> "I answered C on question 4. I picked it because I thought the enzyme was consumed in the reaction. Where exactly does that belief break down, and what would I have to believe instead for B to follow?"

The first gets you a textbook paragraph. The second gets you a targeted correction of a specific wrong model, which is the only thing that dislodges a misconception.

Prompts that work, by failure type

For a suspected misconception:

> "Here is the question, here is what I picked, and here is my reasoning. Do not tell me the right answer yet. Tell me what rule I seem to be applying, and where that rule stops being true."

Asking it to withhold the answer is the important part. It forces the response toward the structure of your thinking rather than the content of the answer key.

For a retrieval failure:

> "I knew this but could not produce it under time pressure. Give me three different ways to cue this fact — a definition cue, an example cue, and a contrast with something it is easily confused with."

Multiple cues is the goal. A fact you can only reach from one direction is a fact you will lose.

For a knowledge gap:

> "I have never seen this concept. Give me the one-sentence version, then the version with the exception that makes it non-trivial, then one question that would tell me whether I actually understood the second version."

The escalation matters. The one-sentence version is what you can hold; the exception is where exams live.

For a process error:

> "I understood this and still got it wrong. Here is the question and my work. Identify the step where the error entered, and tell me what a checkable habit would look like for that specific step."

Do not ask for a content explanation here. Ask for a procedure.

The follow-up that does the real work

Whatever the tutor says, the next message should be some version of:

> "Now ask me a question that would distinguish whether I have actually fixed this, and do not tell me if I am right until I answer."

This is the whole game. An explanation you nodded along to is worth very little; a question you answered correctly from memory is worth a great deal. Ending every tutor exchange with a retrieval attempt converts a passive explanation into an active one.

A fifteen-minute review loop

Here is the loop, concretely, for a quiz where you missed six questions.

Minutes 0-3: sort without analyzing. Go through the six misses and tag each one: gap, retrieval, misconception, process. Do not fix anything yet. Do not look up anything yet. Just sort. The tag is your best guess at what happened in your head at the moment you answered, and your memory of that moment decays fast, which is why this step comes first.

Minutes 3-5: dispatch the gaps. For every gap, make a card and move on. No discussion. A gap is not interesting; it is inventory.

Minutes 5-8: dispatch the process errors. Write one line per error describing the habit that would have caught it. "Circle the word NOT before reading options." "Write the units on every intermediate line." These lines are worth more than any amount of content review, and they are worth rereading before your next practice set.

Minutes 8-15: spend everything left on the misconceptions. This is where the tutor conversation goes. One misconception properly dismantled is worth more than five facts memorized, because a misconception generates an unbounded number of future wrong answers — it is not one broken fact, it is a broken rule that will keep producing them.

Retrieval failures get handled by your spaced-repetition schedule, not by this session. Add them to the deck and trust the schedule.

What this looks like in practice

A student preparing for a chemistry exam takes a twenty-question practice quiz and misses five.

Two are gaps — she has genuinely never learned the solubility rules for sulfates. Cards made, ninety seconds, done.

One is a process error: she inverted a ratio in the final step of a stoichiometry problem, which she has now done three times. Her habit line: "before the final multiply, write what unit the answer should be in." That line goes at the top of her next practice set.

One is a retrieval failure on a definition she has seen a dozen times but could not produce under time pressure. Into the deck, scheduled, forgotten about for now.

And one is a misconception, which is where she spends the remaining ten minutes. She believed that a catalyst shifts equilibrium position. It does not — it changes the rate at which equilibrium is reached, not where the equilibrium sits. That is a genuinely reasonable thing to believe wrongly, it is internally consistent, and it will produce wrong answers on any question touching equilibrium or kinetics until it is dismantled.

She asks the tutor what rule she appears to be applying. It names the confusion between rate and position. She asks for the distinguishing question. She answers it. She gets it right. That is the evening's real work, and it took ten minutes.

Where this sits in the product

SimpleQuizMaker's AI Tutor exists because of exactly this workflow. After a quiz, each question you missed has an explanation panel, so you can ask about that specific question without copying it into a separate chat window and reconstructing the context by hand. The tutor already has the question, your answer, and the correct answer.

There is also a standalone chat if you want to work through something that is not tied to a specific quiz — a concept you are stuck on before you even test yourself on it.

Both draw on the same monthly message allowance: **20 messages a month on the free plan, 150 on the Student plan**. That is deliberate. Twenty messages is enough to run the loop above on several quizzes, which is the point — the loop is short. If you are burning through a hundred messages a week, you are probably asking for explanations rather than diagnoses, which is the failure mode this whole article is about.

Everything here works with any AI tutor, including a general-purpose chatbot with the question pasted in. The prompts are the part that matters, not the product.

When the tutor is wrong, and how to notice

An AI tutor is a fluent explainer, and fluency is not accuracy. It will occasionally state something false with exactly the same confident cadence it uses for things that are true, which makes the error hard to catch precisely when you are least equipped to catch it — you are asking because you do not know.

Three habits keep this from costing you.

Ask for the mechanism, not just the verdict. A wrong answer dressed as a right one usually collapses when you ask *why* one step follows from another. "Because that is the rule" is a warning sign. A real explanation names the thing that makes the rule work, and if the model cannot supply that, the claim is worth checking.

Watch for confident specificity on details. Models are most reliable on the shape of a concept and least reliable on exact figures, dates, named studies, and edge cases. If the tutor tells you the mechanism of enzyme inhibition, that is likely sound. If it tells you a specific enzyme's optimal pH to one decimal place, verify it against your course material before you memorize it.

Trust your textbook over the tutor when they disagree. Not because textbooks are infallible, but because your exam is written against your course's material. Even in the rare case where the model is right and the course is wrong, the course is what you are being graded on — and a disagreement is itself worth surfacing to your instructor, which is a better use of it than silently picking a side.

The deeper point: an AI tutor is most valuable in the place where hallucination is least likely, which is diagnosing the structure of your own reasoning. Asking "what rule am I applying, and where does it stop being true?" is a question about the argument you just supplied. There is far less room to invent than in "list the five causes of X."

Using it with other people

Two adaptations are worth mentioning because they change what the loop is good for.

In a study group, the sorting step works better out loud and with someone else doing the sorting. Describe what you were thinking when you answered; let a partner tag it as gap, retrieval, misconception, or process. People are consistently better at spotting someone else's misconception than their own, because from the inside a misconception does not feel like a belief — it feels like the way things are.

As a teacher, the four-way split is the most useful thing you can teach about test review, and it takes about ten minutes of class time to introduce. It also changes what you do with your own item data: if most of a class picked the same wrong option, that is almost never thirty independent knowledge gaps. It is one shared misconception, or a question that is broken. Both are worth knowing, and telling them apart is what item analysis is for.

The one habit worth stealing

If you take a single thing from this: end every review of a wrong answer by answering a question, not by reading an explanation.

Reading is recognition. Answering is retrieval. Recognition feels like learning and is not. Retrieval feels like work and is.

The students who improve fastest between practice tests are not the ones who review the most questions. They are the ones who, for each question they review, end up producing an answer from memory rather than nodding at one on a screen. That is the entire difference, and it costs nothing except the discomfort of being asked something you might get wrong twice.

Related reading