HSW-0287 · How Studying Works
A student writes, “Current is used up by the bulb.”
The teacher crosses it out and writes the correction. The student can repeat the corrected statement five minutes later.
Three days later, a new circuit question appears. The student reasons as though current is still being consumed.
The correction was remembered.
The old model still influenced the inference.
That distinction matters. Learning is not fully repaired when a learner can recognise that an earlier statement was wrong. Repair is stronger when the corrected model, rather than the old model, governs later explanation, prediction and decision.
This article owns that narrow job: how an obsolete or false explanation can continue to influence reasoning after explicit correction, and how study should test whether the replacement actually controls later thinking. It does not replace the broader How Error Correction Works, Source Monitoring, Knowledge Decommissioning, or Multiple-Choice Lure Memory. Here the question is what happens after a correction has been given and apparently understood.
The direct answer
Research on the continued-influence effect shows that people can continue to rely on retracted information in later reasoning even when they remember the retraction. A 2026 Cognitive Research study with 74 participants tracked eye movements while people read causal misinformation and retractions. Retractions increased processing effort, and correction format influenced how information was integrated; higher reading proficiency and verbal working-memory capacity were associated with less misinformation susceptibility. See Wong and colleagues, 2026.
A 2025 Quarterly Journal of Experimental Psychology paper reported two experiments with 441 participants. Alternative explanations reduced continued reliance, and combinations with forewarning or inoculation strengthened the immediate effect. Yet over seven days, reliance on misinformation increased again despite memory for the correction. See Buczel and colleagues, 2025.
Another 2025 eye-tracking study found that correction formats which reminded participants of the misinformation before presenting corrective details produced better belief updating and memory for corrections than correction-only formats in that experiment. The reminder appeared to direct attention toward the contrast between false and true information. See Wellons and Wahlheim, 2025.
A correction is not finished when the learner can repeat the new sentence. It is finished more convincingly when the new model wins the next inference.
1. Correction memory and corrected reasoning are different outcomes
Ask a learner, “Was the original claim wrong?” They may answer yes.
Then ask a new question whose solution depends on rejecting the old claim. The learner may still reason through the obsolete explanation.
The first test checks memory for the correction. The second checks whether the correction controls inference.
2. Why can the old explanation remain useful to the mind?
An explanation does more than store a sentence. It connects causes to consequences.
If the old explanation filled a causal gap, deleting it creates a hole. A bare “that was wrong” may remove confidence in the statement without providing a replacement route for reasoning.
When a later question demands a causal bridge, the old route can remain accessible because it still supplies structure.
3. Retraction is subtraction; repair usually needs replacement
Compare two corrections:
- Subtraction: “No, friction is not what makes the object keep moving.”
- Replacement: “No. In the idealised case, continued motion does not require a continuing net force; a net force changes velocity. Friction instead changes the motion by opposing it.”
The second correction gives the learner something usable when a new problem asks what causes acceleration.
4. A replacement explanation must fit the same causal slot
It is not enough to add a true fact nearby.
If the wrong idea answered “why did this happen?”, the correction should answer that same question more accurately. Otherwise the old model retains functional advantage.
5. The contrast itself can become a learning object
Students are often told never to repeat a misconception because repetition may increase familiarity.
That rule is too simple. The 2025 reminder study found benefits when false information was clearly labelled and directly contrasted with true corrective details. The educational implication is not “repeat myths often.” It is that a carefully marked contrast can help encode what changed.
6. Label the old model as old
When revising a wrong note, do not leave the obsolete statement sitting beside the corrected one with equal visual status.
Use an explicit state transition:
OLD MODEL — wrong because ______ → REPLACEMENT MODEL — supported because ______
The learner should be able to explain both the error and the repair.
7. Do not ask only “What is correct?”
Ask four questions:
- What did I believe?
- Why was it wrong or incomplete?
- What replaces it?
- What new prediction changes because of the replacement?
The fourth question tests whether correction has become reasoning.
8. Delayed regression is a serious study problem
The 2025 two-experiment study matters because immediate improvement was not the whole story. Over seven days, participants increasingly relied on the misinformation again even while remembering the correction.
For studying, that means same-session correction is weak evidence. The repaired model must be tested later.
9. A correction can decay differently from the old explanation
The learner may retain “teacher said that was wrong” while losing the detail that made the replacement useful.
When later reasoning begins, the old causal association may be easier to retrieve than the newer, less-practised alternative.
This is one reason corrected knowledge deserves retrieval practice of its own.
10. Correct the relation, not only the answer
If a student writes the wrong numerical answer because they used the wrong formula, changing the final number does not repair the relation between problem structure and method choice.
Trace the first incorrect relation:
feature noticed → rule selected → operation performed → conclusion
Repair where the chain first diverged.
11. Mathematics case: the square-root trap
A learner internalises √(a+b) = √a + √b because several early examples seem to behave nicely.
The teacher says the identity is false and gives a counterexample. That is useful.
A deeper repair explains the structural reason: square root is not a linear operation over addition. Then the learner must discriminate valid transformations from invalid ones in fresh expressions after a delay.
12. Science case: replace the causal mechanism
Suppose a learner explains seasons by saying Earth is much closer to the Sun in summer.
A correction should not stop at “distance is not the reason.” It should rebuild the explanatory chain around axial tilt, angle of incoming sunlight and day length, then ask the learner to predict opposite seasons across hemispheres.
13. English case: wrong interpretation can survive a corrected quote
A student may misread a character’s motive, then be shown textual evidence against the interpretation.
If the learner merely memorises the “correct” interpretation, the old narrative may still guide later paragraphs. Repair requires rebuilding the evidence-to-inference chain and testing it against a new passage or moment in the text.
14. Vocabulary case: false friends need contrast
When two similar words are repeatedly confused, replacing one definition may not be enough.
Build a contrast set: meaning, collocation, grammatical behaviour, one near-example and one nonexample. Then choose between the words in changed sentences.
15. Why copying the corrected answer is weak evidence
Copying ensures the page contains the right information.
It does not establish that the learner can inhibit the old route and generate the corrected route independently.
Correction work should end with source-closed reconstruction or application.
16. The wrong-answer notebook can accidentally rehearse errors
Error logs are useful when they encode repair. They are less useful when they become museums of wrong answers.
For each error, record:
- the trigger that produced the old response;
- the first wrong relation;
- the corrected relation;
- one discriminating cue;
- one fresh retest date.
17. Correction format should expose the difference
The 2025 eye-tracking study suggests that reminder-based corrections can direct attention toward the true corrective details when the old false claim is clearly marked.
For learning, use contrast deliberately:
Wrong relation: X causes Y. Corrected relation: Z causes Y; X was correlated/irrelevant/a consequence instead.
The purpose is not drama. It is discriminability.
18. Do not overlearn the error while teaching the correction
Contrast is useful only if the corrective information receives enough attention and retrieval.
Repeatedly displaying a catchy false claim with a tiny correction can increase familiarity with the wrong material. Keep the correction prominent, explicit and meaningfully connected to the old claim.
19. The “why wrong?” step is not optional for conceptual errors
Procedural slips sometimes need only a small check.
Conceptual errors are different. If the learner cannot explain why the old model fails, they may not know when to reject it in a changed problem.
20. Replace global confidence with local confidence
“I understand circuits now” is too broad.
Ask which repaired relation the learner can now defend. Confidence should attach to tested claims, not to the feeling that a correction session went smoothly.
21. Use a near-miss problem
After correction, present a problem where the old rule almost seems applicable.
If the learner can identify why the old route is tempting and still select the corrected one, the repair is more discriminating than simple repetition of the original example.
22. Use a changed-cue problem
Do not retest only with the exact words used in the correction.
Change the surface: diagram instead of prose, new numbers, different context, reversed question, or a request for prediction rather than definition.
The corrected model should survive cue change.
23. Use a delayed inference test
Several days later, do not ask, “What did the teacher correct?”
Ask a new question that requires the corrected model. This is closer to the 2025 finding that remembered correction and misinformation reliance can diverge over time.
24. The correction should enter retrieval practice
Students often retrieve the original content many times but retrieve the correction only once.
Give the replacement model repeated opportunities to win:
- retrieve the corrected principle;
- explain why the old rule fails;
- apply the corrected principle;
- discriminate a near-confusable case;
- return after delay.
25. AI can correct an answer without repairing the learner
If an AI system replaces a wrong paragraph with a polished correct paragraph, the document improves immediately.
The learner may not.
Ask the learner to identify the first wrong assumption, reconstruct the replacement without the tool and solve a new case. Better output is not evidence that the obsolete mental model has been retired.
26. Search corrections can fail when the old source remains more memorable
A vivid video, memorable teacher phrase or catchy misinformation claim can remain more retrievable than a dry correction.
Do not compete only at the level of truth labels. Build a replacement explanation that is coherent, retrievable and connected to the same cues that used to activate the wrong model.
27. Parent guide: ask for the changed prediction
After a child says, “I know I was wrong,” ask:
- What do you believe now?
- Why is the old idea wrong?
- What would you predict differently now?
- Can you solve a fresh example without looking?
This keeps the conversation on repaired reasoning rather than shame about the mistake.
28. Tutor guide: correction should end with learner-controlled explanation
A tutor can make a correction feel clear by doing all the causal work.
Before moving on, reduce answer-giving support. Let the learner generate the corrected chain, answer a discriminator question and explain why a tempting alternative fails.
29. Accessibility supports do not invalidate correction evidence
A learner may use text-to-speech, a formula sheet, a visual organiser, simplified language or other legitimate support.
The independence question is whether the learner owns the target distinction under appropriate accommodation—not whether every support has been removed.
30. When a simple correction is enough
- The error was a clear transcription slip.
- The learner already has the correct model and notices the inconsistency immediately.
- The wrong answer does not reveal a recurring causal or procedural rule.
- A fresh independent item confirms the repair.
Do not turn every typo into a conceptual intervention.
31. When correction persistence deserves deliberate repair
- The same misconception reappears under new wording.
- The learner remembers the correction but predicts from the old model.
- The error has a coherent causal story behind it.
- The wrong route is highly familiar or repeatedly reinforced.
- Feedback improves immediate work but not delayed performance.
32. Evidence boundaries
The continued-influence literature often uses controlled misinformation narratives or headline tasks. A school misconception is not identical to misinformation encountered in news. The educational applications here are mechanism-informed proposals that should be tested with actual student performance.
The 2025 reminder study measured immediate belief updating and correction memory; its authors explicitly note that it did not test long-term downstream durability. The 2025 seven-day work, by contrast, directly highlights delayed regression but uses different paradigms and techniques.
Reminder-based correction should therefore not be turned into a universal rule that every false statement must always be repeated. The balance between contrast, familiarity and corrective attention depends on format and task.
33. Evidence ledger
- Wong et al., 2026, Cognitive Research: Principles and Implications: original eye-tracking study, 74 participants; examined real-time processing of causal misinformation and retractions and linked correction processing to later inferential judgments.
- Buczel et al., 2025, Quarterly Journal of Experimental Psychology: two experiments, total N = 441; alternative explanations and combined techniques reduced continued influence initially, but reliance increased over a seven-day interval despite memory for correction.
- Wellons & Wahlheim, 2025, Cognitive Research: Principles and Implications: original eye-tracking belief-updating experiment; reminder-based corrections improved false-belief reduction and memory for corrections relative to correction-only formats under the tested conditions.
34. The repair protocol
- Name the obsolete model. Do not leave it implicit.
- Locate the first wrong relation. Find where reasoning diverged.
- Explain why it fails. Use evidence, a counterexample or a violated condition.
- Install a replacement. Fill the same explanatory slot.
- Contrast old and new. Make the discriminating feature explicit.
- Close the source. Reconstruct the corrected model.
- Use a changed-cue problem. Test generalisation.
- Delay the next test. Check whether the replacement still controls inference.
35. Return: make the correction win the next inference
Students are not defective because an old idea returns after correction.
A wrong model can be familiar, coherent and well connected. A new correction may initially be little more than a sentence attached to “teacher says”.
Retire the old relation explicitly. Explain why it fails. Build a replacement that answers the same question. Contrast the two. Retrieve the replacement. Apply it under changed cues. Return after delay. The correction becomes learning when the new model—not the remembered fact that a correction occurred—guides what the learner does next.
Continue through How Error Correction Works, Source Monitoring, Knowledge Decommissioning, Multiple-Choice Lure Memory, the How Studying Works Numbered Series Reading Index and the How X Works Hub.