HSW-0262 · How Studying Works
Two learners see the same evidence.
Before the evidence appears, one learner makes a prediction.
The other watches a prediction that has been carefully matched to what they themselves were likely to predict.
The evidence that follows is identical.
Should learning be identical too?
A 2025 experiment with children suggests not necessarily.
Generating the prediction yourself can change the learning event even when you do not control the evidence that follows. The learner has committed to a model, experienced the task as their own action, and created a reference point against which the outcome can be interpreted.
This article owns the narrow job of self-directed prediction when the evidence itself is held constant. It does not claim that unguided discovery is superior to instruction, that hands-on activity always beats observation, or that agency alone guarantees learning.
Related owners remain separate: How Pretesting Works owns attempts before instruction; Prediction Error and One-Shot Memory owns the memory boundary around surprise; and Vicarious Learning and Memory owns learning by observing others more broadly.
Quick Answer
In a 2025 npj Science of Learning experiment, 95 children aged five to seven learned about water displacement. Children either generated predictions themselves or observed predictions from a fictitious peer. The observed predictions were designed to match each child’s prior beliefs closely, and both conditions saw the same outcomes. Children who made their own predictions performed better on a transfer test, with the authors interpreting the result as evidence that even a minimal form of agency can change conceptual learning. See Brod and colleagues, 2025.
The result is important because it controls a common confound. In many “active versus passive” comparisons, active learners get different information. Here the evidence was held much more tightly constant.
Sometimes the educational difference is not what information arrives, but whether the learner had to put a stake in the ground before it arrived.
1. Prediction Creates a Model Before the Answer Arrives
To make a prediction, a learner must expose at least part of a current model.
If asked which submerged object will displace more water, the learner has to use some belief:
- larger objects displace more;
- heavier objects displace more;
- material matters;
- both displace the same amount.
Before the result appears, the learner’s theory has become behaviorally visible.
That matters because feedback can now be interpreted relative to something specific.
2. Watching the Same Prediction Is Not Quite the Same Cognitive Job
If another learner makes the prediction, the observer can process it at several depths.
- “That seems plausible.”
- “That is probably what I would have said.”
- “I disagree.”
- “I have not committed either way.”
The observer may learn very well. Demonstrations and worked examples can be extremely powerful.
But observation does not guarantee that the learner has generated, committed to and evaluated a personal prediction.
3. What the 2025 Experiment Controlled
Brod and colleagues designed an unusually careful comparison.
The children in the active condition made their own predictions. Children in the yoked condition observed predictions from a fictitious peer. Those predictions were generated using a Bayesian model tuned to the individual child’s prior beliefs, so the observed predictions closely approximated what that child might have predicted.
Crucially:
- both groups saw the same underlying trial sequence;
- both groups saw the correct outcomes;
- the active learner did not choose which evidence would appear;
- the yoked learner still responded physically by confirming the displayed prediction;
- the main difference was whether the prediction was self-generated or observed.
This makes the study more informative than a simple comparison of “doing an experiment” with “watching a lesson”.
4. What the Study Found
Both groups learned substantially.
Posttest performance was high in both conditions and close to ceiling. The active group had a higher posttest mean, but the pre-to-post interaction was not significant, so that result should be interpreted cautiously.
On the transfer test, where ceiling effects were less severe, the active-prediction group performed better overall than the yoked group. The effect was modest, and not every transfer subtask showed a significant difference.
This is a useful scientific pattern: evidence for an advantage, accompanied by clear limits rather than a universal claim.
5. Agency Is One Candidate Mechanism
Children in the active condition reported a stronger sense of autonomy than those in the yoked condition.
The authors therefore discuss agency as one plausible mechanism. Making a prediction oneself can create the experience that “I am the one doing this cognitive action,” even though the learner does not control the experimental outcome.
But agency should not be treated as the only proven mechanism.
Self-generation may also increase:
- evaluation of competing hypotheses;
- attention to the relation between prediction and evidence;
- elaboration of the learner’s current causal model;
- ownership of the error when evidence contradicts the prediction.
The experiment does not cleanly isolate every one of these processes.
6. Prediction Is Not the Same as Choosing the Experiment
This distinction protects the result from overstatement.
The children did not choose which objects would be tested. They did not design the evidence sequence. They did not control the outcome.
They made a volitional prediction about predetermined evidence.
So the article is not evidence that fully open inquiry is always better than explicit instruction.
7. Prediction Is Not the Same as Hands-On Learning
The task was computerized.
Children were not physically conducting a water-displacement experiment themselves.
That means the result cannot be reduced to “touching objects improves learning”.
In fact, one strength of the study is that it removes much of the physical-activity difference and asks whether a smaller cognitive act—making your own prediction—matters.
8. Prediction Can Make Feedback Answer a Question the Learner Actually Asked
Feedback is easier to process when it resolves a live uncertainty.
If the learner has predicted:
“The heavier object will displace more water.”
then the observed result can be compared directly with that belief.
Without a prediction, the same result can remain merely another fact to receive.
This is one reason prediction can be useful without being magical: it changes the structure of the feedback problem.
9. Mathematics Example: Estimate Before Calculating
Before solving a percentage problem exactly, ask:
“Should the answer be closer to 20, 50 or 100?”
The prediction forces the learner to inspect magnitude and structure before running a procedure.
After calculating, the learner can compare the exact answer with the prediction.
If they disagree, the discrepancy becomes diagnostic: was the estimate wrong, the method wrong, or the arithmetic wrong?
10. English Example: Predict the Function Before Reading the Explanation
Before reading an analysis of a paragraph, ask the learner:
“Why do you think the writer repeats this image?”
The learner commits to an interpretive hypothesis.
The model analysis now becomes something to compare against rather than something to copy.
The learner can ask which textual evidence supports each interpretation and whether the initial prediction needs revision.
11. Science Example: Predict Direction Before Mechanism
Before showing a graph, demonstration or simulation, ask for a directional prediction:
“If temperature increases, what do you expect to happen to reaction rate?”
Then show the evidence and ask for the mechanism.
This sequence separates:
- prediction;
- observation;
- explanation.
That separation protects scientific reasoning from hindsight.
12. Why Hindsight Makes Observation Look Easier Than Prediction
Once the outcome is known, it often feels obvious.
This creates a problem for studying. A learner can watch a demonstration and think:
“Yes, of course that happens.”
But without a recorded prediction, there is no evidence that the learner could have anticipated the result from the correct model.
Prediction protects the before-state.
13. Prediction Quality Matters
A prediction can be too vague to teach much.
Weak:
“Something will change.”
Stronger:
“The larger fully submerged object will displace more water regardless of material.”
The stronger prediction identifies the variable and condition that the evidence can confirm or challenge.
14. The Prediction–Evidence–Revision Loop
- Predict. Commit to an outcome or explanation before feedback.
- Record confidence. Low, medium or high is enough.
- Observe. Describe what the evidence actually shows.
- Compare. Was the prediction correct, partly correct or wrong?
- Explain. Which model accounts for the evidence?
- Revise. Change the rule, not merely the answer.
- Transfer. Predict a new case where surface features differ.
The transfer step is essential. Otherwise prediction can become a one-question correction rather than conceptual learning.
15. When Watching Is Better
Observation can be superior when the learner does not yet have enough knowledge to generate a meaningful prediction or when random action would dominate the task.
Examples include:
- watching an expert model a complex procedure;
- studying a worked example before attempting a high-element-interactivity problem;
- observing a hazardous demonstration;
- learning what counts as relevant evidence before designing an inquiry.
Self-direction and explicit instruction are not enemies. Strong teaching decides when learner generation creates useful cognitive work and when demonstration carries necessary structure.
16. When Prediction Is Noise
Do not ask for a prediction merely because the activity needs engagement.
Prediction is weak when:
- the learner has no basis for choosing among outcomes;
- the answer is arbitrary rather than model-based;
- feedback arrives too late to compare meaningfully;
- the prediction is never revisited;
- the task rewards guessing without explanation.
A prediction should make a model inspectable, not merely generate another click.
17. Self-Directed Prediction vs Pretesting
Pretesting can ask for answers before instruction, often when learners do not yet know the target information.
Self-directed prediction is narrower. The learner generates an expected outcome or model before receiving evidence, and the comparison between prediction and evidence becomes part of conceptual revision.
18. Self-Directed Prediction vs Vicarious Learning
Vicarious Learning and Memory shows that observation can create substantial learning.
The 2025 prediction study does not overturn that. It identifies one condition under which self-generation produced an additional advantage even when the observed prediction was carefully matched.
19. Self-Directed Prediction vs Prediction Error
Prediction Error and One-Shot Memory owns whether surprise or expectation violation improves a particular memory event.
Self-directed prediction asks whether generating the expectation oneself changes conceptual learning compared with observing a matched expectation.
20. Parent and Tutor Guide
Before explaining a result, ask for a prediction when the learner has enough prior knowledge to make one.
- What do you think will happen?
- Why?
- How confident are you?
- What result would make you change your mind?
Then reveal or derive the evidence.
Do not punish the wrong prediction. Its educational value is that it makes the old model visible enough to revise.
21. Delayed Independent Check
After the learner has revised the concept, wait and present a fresh case.
- Change the surface features.
- Ask for a new prediction before any feedback.
- Ask for the mechanism.
- Record confidence.
- Only then show the outcome.
If the learner predicts correctly for the wrong reason, conceptual learning is not yet complete.
22. Evidence Boundary
The 2025 study involved 95 children aged five to seven in a computerized water-displacement task. The transfer advantage was modest. Some posttest analyses were complicated by ceiling performance, and the productive and explicit-concept transfer subtasks did not each show significant condition differences.
The authors explicitly note that generalisation to other domains and to hands-on experimentation requires future research.
Therefore, the study supports a careful educational proposition: when evidence is held constant, self-generating a meaningful prediction can sometimes deepen conceptual learning compared with observing a matched prediction. It does not prove that every self-directed activity is better than instruction.
23. Return: Make the Learner Put a Model on the Table
The power of prediction is not fortune-telling.
It is model exposure.
Before the world answers, the learner says what they think the world will do.
Ask for the prediction. Preserve it. Show the evidence. Compare honestly. Revise the rule. Then move to a new case and see whether the learner can predict again without the old example carrying them.
Continue through How Pretesting Works, Vicarious Learning and Memory, Prediction Error and One-Shot Memory, the How Studying Works Numbered Series Reading Index and the How X Works Hub.