HSW-0155 · How Studying Works
First attempt: you recall 4 out of 10 and predict that next time you will remember about 5.
Second attempt: you recall 7 out of 10 and predict that next time you will remember about 6.
Third attempt: you recall 9 out of 10 and still feel as though the material is not quite secure.
Your performance is improving faster than your confidence.
Underconfidence with practice is the metacognitive pattern in which judgments of learning underestimate actual memory performance after repeated study–test cycles, even as objective recall improves.
It is counterintuitive because students are usually warned about overconfidence. Yet repeated practice can produce the opposite calibration problem: learners may continue acting as though material is less secure than it has become.
The 50-Second Read
- Practice can outrun confidence. Recall may improve more quickly than judgments of learning.
- The effect often emerges after multiple cycles. Early predictions can be roughly calibrated or even overconfident, followed by later underconfidence.
- Retrieval practice helps twice. It improves memory and provides evidence that can improve metacognitive monitoring.
- But confidence may update too slowly. Learners can remain anchored to earlier, weaker performance.
- Underconfidence wastes time. Already-secure material can receive unnecessary restudy while weaker material waits.
- Do not solve this with positive thinking. Solve it with performance evidence.
- Use delayed, independent checks. Confidence should rise only when capability survives without support.
1. The Strange Shift From Overconfidence to Underconfidence
The original underconfidence-with-practice literature found a robust pattern across repeated study–test cycles.
On the first cycle, judgments of learning were often reasonably calibrated or slightly overconfident.
On later cycles, actual recall improved more than predicted. The learner became underconfident.
The classic 2002 Journal of Experimental Psychology: General paper documented this pattern across several conditions and materials. See Comparing objective and subjective learning curves: Judgments of learning exhibit increased underconfidence with practice.
2. Objective and Subjective Learning Curves Can Separate
Imagine two curves across three cycles.
| Cycle | Actual recall | Predicted recall |
|---|---|---|
| 1 | 45% | 50% |
| 2 | 70% | 60% |
| 3 | 88% | 70% |
The learner is clearly improving.
The internal forecast is improving too, but not enough.
The calibration error comes from the gap between those curves.
3. Why This Matters for Study Control
Metacognition is useful only if monitoring changes control.
If a learner underestimates what is already secure, they may:
- restudy mastered items;
- delay moving to new material;
- avoid harder transfer tasks;
- repeat comfortable retrieval unnecessarily;
- misallocate scarce revision time.
Underconfidence is therefore not merely an emotional issue. It can become an allocation error.
4. Retrieval Practice Provides Metacognitive Evidence
Repeated testing does more than strengthen memory.
It gives the learner concrete information:
- which items were retrieved;
- which were missed;
- how quickly retrieval occurred;
- which errors repeated;
- which items became easier across cycles.
A 2022 Metacognition and Learning study extended underconfidence-with-practice research to action memory and found underconfidence in later cycles after both verbal and enacted encoding. Retrieval practice increased judgments more quickly than repeated restudy, reducing underconfidence relative to study–study conditions. See The underconfidence-with-practice effect in action memory.
5. Why Confidence Updates Too Slowly
Several explanations have been proposed.
- Anchoring: the learner begins with a confidence estimate and adjusts upward too little as learning improves.
- Past-test dependence: earlier weak performance continues to influence later judgments.
- Foresight bias changes: repeated testing reveals what future retrieval actually feels like, changing calibration.
- Scale behaviour: percentage judgments may reflect relative confidence rather than literal subjective probability in a simple way.
Research comparing explanations has supported a substantial role for anchoring and insufficient adjustment. See The contributions of anchoring and past-test performance to the underconfidence-with-practice effect.
6. Underconfidence With Practice vs the Stability Bias
The Stability Bias is broader: learners often predict too little future change in memory.
Underconfidence with practice is one specific pattern that can emerge across repeated learning cycles when actual performance rises faster than the learner’s judgments.
The stability bias helps explain the general forecasting error. UWP describes the observed calibration pattern across cycles.
7. Underconfidence With Practice vs Delayed Judgments of Learning
Delayed Judgments of Learning asks when predictions should be made so that the cue resembles future retrieval more closely.
UWP asks how calibration changes across repeated practice cycles.
A delayed judgment can improve prediction quality and still exist inside a multi-cycle learning process where confidence updates imperfectly.
8. Underconfidence vs Low Self-Efficacy
These should not be collapsed.
Underconfidence with practice is a specific calibration phenomenon in memory prediction.
Low self-efficacy is a broader belief about capability to perform actions or reach outcomes in a domain.
A learner can show UWP on a memory task without having generally low academic self-belief.
9. Mathematics: Improvement Can Arrive Before Identity Updates
A student who has struggled with algebra for months may begin to improve after a targeted repair.
The evidence changes first:
- fewer sign errors;
- faster manipulation;
- more correct starts;
- better mixed-question performance.
The self-model may lag behind:
“I’m still bad at algebra.”
The educational job is not to replace evidence with reassurance. It is to show the learner the repeated independent receipts that justify updating the forecast.
10. English: Writing Growth Can Be Underestimated Because Quality Is Multi-Dimensional
Writing rarely improves as one obvious percentage.
A learner may still feel uncertain while:
- task fulfilment improves;
- paragraph structure stabilises;
- sentence errors decline;
- evidence becomes more relevant;
- revision becomes faster.
Because the whole performance still feels effortful, confidence may not update enough.
Use component evidence and full-task evidence together.
11. Science: A Better Model Can Make Many Questions Easier at Once
Science understanding can improve nonlinearly.
Once a causal model becomes coherent, several question types become easier simultaneously.
The learner may still be anchored to the earlier phase when every question felt separate.
A mixed test can provide the evidence needed to update confidence.
12. The Mastery-Creep Problem
Underconfidence can create a strange study failure:
the definition of “mastered” keeps moving upward.
- First, 80% recall is not enough.
- Then 90% is not enough.
- Then one hesitation is treated as proof of weakness.
- Then the learner keeps reviewing material that already meets the real performance standard.
This is not high standards. It is a control criterion detached from actual task demands.
13. Use a Declared Exit Criterion
Before study begins, define what qualifies as secure enough for the current phase.
Example:
Move this topic to maintenance when I can retrieve 8/10 after a delay, solve two mixed problems without help, and explain the main method choice.
Once the criterion is met, confidence is not allowed to keep moving the goalpost without new evidence.
14. Retrieval Practice Reduces Both Ignorance and Uncertainty
Repeated retrieval provides two kinds of information.
- Learning evidence: what can actually be recalled.
- Monitoring evidence: how accurately the learner can predict that recall.
This is one reason practice tests are more informative than rereading. They do not merely strengthen memory. They expose the learner to the conditions of memory use.
15. But One Good Test Is Not Enough
Underconfidence should not be “corrected” by declaring mastery after one strong performance.
Use repeated receipts:
- immediate retrieval;
- delayed retrieval;
- changed cues;
- mixed application;
- timed or independent performance where relevant.
Confidence should rise with robust evidence, not with encouragement alone.
16. The School Route: Students Need Feedback on Calibration, Not Only Answers
A teacher can ask students to predict performance before a quiz and compare prediction with outcome afterward.
Over several cycles, the learner can see whether confidence tends to lag behind improvement.
The conversation changes from:
“Believe in yourself more.”
to:
“Your last three delayed tests were 78%, 84% and 86%. Your predictions were 60%, 65% and 68%. Your forecasting model needs to update.”
17. The Systems Route: Monitoring Has Latency
A sensor can lag behind a changing system.
Metacognitive confidence can behave similarly.
The learner state improves, but the internal estimate updates slowly.
If control decisions are based on the lagging estimate, resources keep flowing to a problem that is already largely repaired.
18. The Financial Route: Underconfidence Overfunds Low-Risk Assets
Imagine a portfolio manager who keeps treating a stable asset as high risk despite repeated evidence to the contrary.
Capital remains trapped there while genuine risk elsewhere goes underfunded.
In study terms, underconfidence can overfund already-secure topics.
The fix is not reckless divestment. It is evidence-based reallocation.
19. The Learning Route: Move Secure Material to Maintenance
Acquisition and maintenance need different budgets.
Once a topic has robust evidence of mastery:
- reduce frequency;
- increase spacing;
- keep occasional retrieval checks;
- redirect active study time to weaker material.
This lets confidence catch up through later successful receipts without paying the full acquisition cost repeatedly.
20. The Education Route: Teach Students That Confidence Is an Estimate
Confidence should neither rule nor be ignored.
It is an estimate that needs calibration.
Students should learn to say:
I feel uncertain, but the last three independent tests show that this topic is currently secure. I will move it to maintenance and check again later.
That is mature self-regulation.
21. The Training Route: Confidence–Performance Tracking
For one topic across three cycles, record:
| Cycle | Predicted recall | Actual recall | Gap |
|---|---|---|---|
| 1 | 55% | 50% | +5 over |
| 2 | 62% | 75% | −13 under |
| 3 | 70% | 90% | −20 under |
If the learner repeatedly underpredicts later-cycle performance, study allocation should stop obeying confidence blindly.
22. The Improvement Route: Train the Update, Not the Feeling
You do not need to force yourself to feel confident.
You need to improve the update rule.
- Prediction before test.
- Actual performance after test.
- Identify direction of calibration error.
- Adjust the next prediction.
- Adjust study allocation only after checking robustness.
Over time, the internal model should learn from its own forecast errors.
23. The World Route: Expertise Includes Knowing When a Problem Is No Longer the Main Problem
Professionals must reallocate attention as conditions change.
A repaired engineering fault should not keep receiving emergency resources. A stabilised patient should not be treated as though the original crisis remains unchanged. A resolved software incident should move into monitoring.
Learning systems need the same ability to downgrade repaired risk.
24. The Scale Question Matters
Research has also questioned whether percentage judgments of learning always behave like literal subjective probabilities.
A 2015 Memory & Cognition paper found that underconfidence with practice appeared with scale judgments but not in the same way with binary betting decisions, suggesting that the format of the judgment matters. See Judgments of learning index relative confidence, not subjective probability.
This is a useful reminder: metacognitive measurements are themselves tasks with assumptions.
25. What Not to Do
- Do not equate uncertainty with weakness when repeated performance says otherwise.
- Do not keep restudying secure material solely because it does not feel perfect.
- Do not raise confidence from encouragement without evidence.
- Do not declare mastery from one successful immediate test.
- Do not assume underconfidence with practice appears identically across every judgment format or learning task.
- Do not let earlier poor performance anchor the learner model forever.
26. Evidence Boundary
Underconfidence with practice is a robust finding in repeated learning paradigms, but its magnitude and explanation vary across materials, judgment formats and learning procedures. Retrieval practice appears to provide useful metacognitive evidence, and anchoring accounts explain part of the pattern, but no single mechanism should be treated as universally complete.
The safe educational lesson is that learners should compare confidence with repeated independent performance and allow study allocation to update when evidence shows that capability has improved.
27. Return: Confidence Is Allowed to Lag. The Plan Is Not.
You do not need to feel transformed the moment learning improves.
Confidence can catch up later.
But the study plan should respond to evidence now.
When repeated independent performance rises faster than your confidence, trust the receipts enough to move secure material to maintenance and spend scarce time where the evidence still shows a real weakness.
Continue through The Stability Bias, Delayed Judgments of Learning, Labor-in-Vain and the How Studying Works Numbered Series Reading Index.