How the Assistance Dilemma Works | Give Enough Help to Keep Learning Moving Without Taking Over the Thinking

eduKateSG Learning Node Series · 0052

Help can rescue a learner from a dead end. The same help, given too early or too completely, can remove the thinking the learner needed to practise.

The assistance dilemma is the problem of deciding when to help, how to help and how much of the problem to carry for the learner. It appears in tutoring, classrooms, worked examples, intelligent tutoring systems and now AI-assisted study.

The goal is not minimum help. It is not maximum help. It is the right assistance for the learner’s current state, followed by a return of responsibility.

Quick Read: Help Has Two Time Horizons

Assistance can improve current performance by making the present problem easier to complete.
Learning requires a second question: what will the learner be able to do later without the same support?
Too little help can leave a novice guessing, rehearsing errors or wasting effort on an inaccessible task.
Too much help can remove retrieval, method selection, planning, monitoring or execution that the learner needs to acquire.
The assistance dilemma therefore has two decisions: when is help needed? and what form of help preserves the most useful learner thinking?
Research in intelligent tutoring systems shows that adaptive policies can reduce unproductive help avoidance and improve later problem solving under some conditions.
AI makes the dilemma more important because high-quality answers can now arrive before the learner has attempted any of the work.

The best assistance does not merely make the problem go away. It changes what the learner can do when the assistance is gone.

The Homework Paradox

A student is stuck on Question 7.

A parent explains every step. The answer is completed in four minutes. Homework performance improves immediately.

Tomorrow, Question 8 uses the same idea with different numbers. The student is stuck again.

The help solved the task but may not have solved the learning problem.

Now reverse the situation. The parent refuses all help because “struggle builds character.” The student spends forty minutes trying random methods, becomes frustrated and finally copies the answer.

That did not build the learning either.

The assistance dilemma lives between these two failures.

The Original Research Question

Kenneth Koedinger and Vincent Aleven used the term assistance dilemma in research on cognitive tutors to describe a recurring instructional tension: more support can make learning activities easier and more efficient, yet learners also need opportunities to perform the cognitive work themselves.

Intelligent tutoring systems made the problem measurable because systems could vary hints, feedback, worked examples, problem solving and adaptation while tracking what learners did next.

The central question remains deeply human:

When should the teacher intervene, and what should the intervention reveal?

Performance and Learning Are Different Measurements

Assistance often improves the visible metric first: accuracy, completion, speed or reduced frustration.

But the deeper outcome may only appear later, when the learner faces a new problem without the same support.

This creates a measurement trap. A highly assisted learner can look excellent during practice because the system is carrying some of the task. A less assisted learner can look slower while building a capability that transfers.

Neither pattern should be assumed. Sometimes assistance truly accelerates learning by preventing wasted search. Sometimes reduced assistance produces productive effort. The design problem is finding which situation the learner is actually in.

The Two Decisions: When and How

The assistance dilemma can be separated into two control problems.

When: is the learner currently productive, recoverably stuck or unproductively stuck?
How: what is the smallest intervention likely to restore productive learning?

These decisions are related but not identical. A system may correctly detect that help is needed and still give the wrong kind of help.

Four Learner States

A useful practical model is to classify the moment before intervening.

State 1 — Productive: the learner is making legitimate progress. Do not interrupt merely because the work looks slow.
State 2 — Recoverably stuck: progress has paused, but the learner has a plausible next move or can benefit from a light prompt.
State 3 — Unproductively stuck: repeated attempts are not improving, the learner is cycling, guessing or reinforcing an error.
State 4 — Missing foundation: the task depends on knowledge the learner does not possess. A hint about the current step may be insufficient; instruction or repair is needed.

Good assistance begins by identifying the state, not by assuming every pause is the same kind of problem.

The Assistance Ladder

Help can be staged from low-information to high-information support.

Level 0 — Wait: allow productive thinking to continue.
Level 1 — Metacognitive prompt: “What have you tried?” or “What is the question asking for?”
Level 2 — Attention cue: point to the relevant information, representation or discrepancy.
Level 3 — Subgoal cue: name the intermediate objective without giving the operation.
Level 4 — Strategy hint: suggest a method family or comparison.
Level 5 — Partial step: demonstrate one critical move and return control.
Level 6 — Worked segment: show a larger portion of the route with explanation.
Level 7 — Full worked solution: demonstrate the complete procedure when the learner lacks the model needed to proceed.

The ladder is not a moral ranking. A full worked solution can be exactly right for a novice encountering a new procedure. The error is giving high-level assistance when lower support would have preserved useful thinking—or withholding high-level assistance when the learner has no viable model to work from.

Help Avoidance and Help Abuse

Learners can misuse assistance in opposite directions.

Help avoidance occurs when a learner needs support but refuses or delays asking, perhaps from pride, fear, poor metacognition or the belief that asking means failure.

Help abuse occurs when assistance is requested too early or repeatedly used as a shortcut before meaningful effort.

A well-designed learning system must support appropriate help seeking, not simply maximise or minimise requests.

What Adaptive Tutoring Research Adds

Research on data-driven tutoring systems has attempted to predict when learners are likely to make unproductive moves. Maniktala and colleagues developed a HelpNeed predictor for a logic tutor and tested an adaptive policy that proactively supplied hints when the model predicted help was needed.

Students in the adaptive condition showed lower help avoidance and more appropriate receipt of help during training. They also outperformed control students on the post-test, producing shorter, more efficient solutions in less time.

Later work by Alam and colleagues combined predictions of when help was needed with subgoal hints. The results suggested benefits particularly for lower-prior-proficiency learners and reduced help avoidance.

These studies do not produce one universal policy. They demonstrate something more important: the timing and form of assistance can be treated as an adaptive learning variable rather than a fixed teaching style.

Mathematics: Do Not Solve the Selection Problem for the Student

A learner faces a geometry problem and does not know whether to use similarity, Pythagoras’ theorem or trigonometry.

If the tutor says, “Use Pythagoras,” the immediate obstacle disappears—but method selection was part of the learning target.

A lower-level intervention might ask:

Which lengths are known?
Is there a right angle?
What relationship connects the unknown directly to the known quantities?

The assistance keeps the learner inside the decision rather than making the decision on their behalf.

Additional Mathematics: Separate Algebra Failure From Strategy Failure

A student may choose the correct differentiation strategy but make an algebra error. Another may manipulate algebra flawlessly but choose the wrong mathematical model.

These learners need different help.

The first needs local execution correction. The second needs method-selection support. Treating both with a full worked solution hides the diagnostic distinction.

Science: Hint Toward the Evidence, Not the Conclusion

In a science explanation question, a student writes a vague conclusion.

A teacher can give the correct explanation. Or the teacher can redirect attention: “Which observation in the diagram must your explanation account for?”

The second intervention preserves evidence selection and causal reasoning while narrowing the search space.

English: Assistance Can Accidentally Write the Sentence

When helping with composition, adults often improve a sentence so much that the sentence is no longer the student’s.

A more calibrated ladder can begin with questions:

What do you want the reader to understand here?
Which word is carrying the wrong meaning?
Can you split this sentence into two claims?
Which detail would make the scene concrete?

Only when the learner cannot reconstruct the sentence should the support move toward modeling.

Vocabulary: Do Not Give the Word Before Searching the Meaning

If a learner cannot retrieve a word, immediately supplying it solves recall but removes the retrieval attempt.

A cue can preserve more of the work: first sound, semantic category, contrast word, sentence context or partial spelling.

But if the word was never learned, escalating cues will not manufacture knowledge that does not exist. At that point, teach the word.

Assistance Dilemma Versus Hint Dependence

Hint Dependence owns the failure state in which learners become reliant on prompts and perform poorly when the prompts disappear.

The assistance dilemma is broader. It includes the decision before dependence exists: when should support arrive, what kind should it be and how should it adapt?

Assistance Dilemma Versus Worked Example Fading

Worked Example Fading owns a specific transition from highly completed examples toward independent problem solving by progressively removing worked steps.

The assistance dilemma includes fading but also asks whether a learner needs an example, a hint, a prompt, feedback, no intervention or direct instruction at a particular moment.

Assistance Dilemma Versus Expertise Reversal

Expertise Reversal owns the finding that instructional guidance useful to novices can become redundant or burdensome as expertise grows.

The assistance dilemma is the operational problem of using learner state to choose and adjust support. Expertise level is one important signal among several.

Assistance Dilemma Versus Fading

Fading owns the gradual removal of support as capability develops.

The assistance dilemma also includes escalation. Sometimes the right move is to add support because current difficulty is no longer productive.

The Diagnostic Pause Before Helping

Before intervening, ask four questions.

Does the learner understand the goal?
Does the learner possess the prerequisite knowledge?
Is the current strategy plausible?
Is another attempt likely to produce new information?

If the fourth answer is no, more unaided struggle may simply repeat the same failure.

Productive Difficulty Has an Information Test

Difficulty is productive when effort can plausibly update the learner’s model.

If each attempt reveals something—an incorrect assumption, a useful pattern, a narrowing of possibilities—the learner may be productively struggling.

If attempts are random, repeated without change or based on missing prerequisites, the difficulty is producing little educational information. Assistance should escalate.

Failure Mode 1: Rescue at the First Sign of Effort

A learner pauses for five seconds and an adult supplies the next step.

Repair: distinguish silence from stuckness. Give learners time to retrieve, plan and inspect before interpreting effort as failure.

Failure Mode 2: Withholding Help as a Principle

“They must discover it themselves” can become just as rigid as over-helping.

When prerequisites are absent or search is unproductive, clear explanation and worked examples can be more efficient and more humane.

Failure Mode 3: The Hint Solves the Wrong Bottleneck

A student is stuck because they do not understand a vocabulary term. The tutor gives a strategy hint. Nothing changes.

Repair: diagnose whether the bottleneck is language, knowledge, representation, method selection, execution or checking before choosing the intervention.

Failure Mode 4: Help Is Never Faded

The learner improves inside the supported environment, so the support is preserved indefinitely.

Repair: periodically test the capability under reduced assistance. Support should be justified by current need, not historical habit.

Failure Mode 5: Independence Is Tested Too Late

A student spends a month solving with hints and sees an unaided question only in the examination.

Repair: build unassisted checkpoints into practice. Independence is a skill state to measure, not a surprise condition saved for assessment day.

Cross-Domain Lens: Power Steering

Power steering helps a driver control a vehicle without deciding where the vehicle should go.

Good learning assistance should often work similarly: reduce unnecessary effort while preserving the decision that constitutes the skill.

If the system begins choosing the destination, route and steering inputs, the human is no longer practising driving.

Cross-Domain Lens: Medical Support

Support in medicine is matched to patient state. The aim is not to prove that less intervention is always better. It is to provide enough support to maintain function while addressing the underlying cause and, where possible, restore independent capacity.

Learning assistance follows a related control logic: stabilise when necessary, diagnose the cause, then return the function to the learner.

AI Makes the Assistance Dilemma Immediate

Generative AI can explain, outline, solve, rewrite, calculate and generate examples almost instantly. The cost of requesting high-level assistance has collapsed.

That makes an old tutoring question newly urgent: which parts of the cognitive work should the learner still perform?

A small 2025 preprint study of AI-assisted note-taking reported the best post-test performance under an intermediate-assistance condition and the lowest under a highly automated condition, even though participants preferred the easier high-assistance experience. The sample was small and the study is preliminary, so it should not be treated as a universal law. It does illustrate the measurement problem: convenience and cognitive benefit need not point in the same direction.

An AI Assistance Ladder for Students

First: ask AI to restate the task or clarify an unfamiliar term.
Next: ask for one question that directs your attention.
Then: ask for a subgoal or strategy family without the solution.
Then: show your attempt and ask what is wrong with the first failing step.
Only if needed: request a partial worked segment.
Finally: study a full solution, close it, and solve a parallel problem unaided.

This converts AI from an answer endpoint into an adjustable scaffold.

A 20-Minute Help Calibration Routine

Minutes 0–4: attempt the problem unaided and write the exact point of uncertainty.
Minutes 4–7: use one low-level prompt or cue.
Minutes 7–11: continue independently; escalate only if no new progress appears.
Minutes 11–14: if needed, inspect one worked step and explain why it is valid.
Minutes 14–17: finish the problem without further help.
Minutes 17–20: solve a near-transfer problem with assistance removed.

A Teacher Protocol

Define which cognitive decisions the lesson is supposed to teach.
Do not give away those decisions casually through hints.
Watch the learner’s attempts, not just the elapsed time.
Use the lowest assistance level likely to restore productive work.
Escalate quickly when prerequisites are missing.
Return control after each intervention.
Fade recurring assistance as performance stabilises.
Include unaided transfer checks.

A Student Protocol

Attempt before asking for the answer.
Name exactly where you are stuck.
Ask for the smallest useful hint.
After receiving help, close or hide it.
Restate the reason for the next step in your own words.
Complete the remaining work yourself.
Do a second problem without support.
If the second problem fails, you have evidence that the first assistance did not yet become learning.

A Parent Protocol

When a child asks, “How do I do this?”, avoid jumping immediately to a full explanation.

Ask what the question is asking.
Ask what they already know.
Ask where the first uncertainty appears.
Give one cue.
Wait for another attempt.
If the child lacks the prerequisite entirely, teach it clearly rather than stretching the struggle.
After the repair, ask for one independent repeat.

A Tutor Protocol for a Three-Student Class

Three learners make assistance calibration visible because the tutor can compare different help needs on the same task.

Student A is still productive: observe and do not interrupt.
Student B has selected the wrong representation: give an attention cue.
Student C lacks a prerequisite: pause and repair that prerequisite directly.
Then give all three a fresh problem with no prompts.

The class stops treating help as a reward or punishment. Help becomes a controlled response to diagnosed state.

What to Measure After Helping

Did the learner complete the current problem?
Could the learner explain why the assisted step was valid?
Could the learner continue after assistance stopped?
Could the learner solve a similar problem later?
Could the learner select the method without the original cue?
Did the amount of help required decrease?

The final question matters enormously. A learning system should expect assistance demand to change over time.

Canonical Owner Boundaries

This page owns the assistance dilemma: deciding when help should arrive, what form it should take and how much thinking it should leave with the learner.

Hint Dependence owns reliance on prompts that weakens independent performance.
Worked Example Fading owns progressive removal of worked solution steps.
Fading owns the general reduction of support as capability grows.
Expertise Reversal owns how guidance can become redundant as expertise increases.
Worked Examples owns learning from complete demonstrated routes.
This page owns the control problem that decides which of those forms of support, if any, should be used now.

Evidence and Limits

The assistance-dilemma literature is strongest in intelligent tutoring and structured problem-solving environments where help behaviour and later performance can be measured closely. It supports adaptive assistance as a serious design problem and provides evidence that predicting help need and choosing targeted hints can improve learning outcomes in some settings.

Human classrooms are messier. A learner’s emotional state, language, disability, fatigue, trust, prior experience and task stakes can all change what appropriate assistance looks like. Accessibility support should not be removed merely to manufacture difficulty. Safety-critical training can also require more direct intervention than an ordinary homework problem.

The principle should therefore remain functional: preserve the cognitive work that constitutes the learning target, remove unnecessary barriers, intervene when search becomes unproductive, and verify that responsibility can return to the learner.

The Return Path

Return to Question 7.

The student is stuck.

Do not ask only, “Should I help?”

Ask what thinking the learner still needs to practise, what obstacle currently blocks that thinking and what smallest intervention can reopen the route.

Then help.

And give the problem back.

The assistance dilemma is solved one moment at a time: enough support to restore productive learning, followed by enough independence to make the capability belong to the learner.

Use This Tomorrow

On the next difficult question, do not jump directly from “stuck” to “show me the answer.” Write the exact point where progress stopped. Ask for one cue that reveals less than the full method. Try again. Escalate assistance only when the new attempt produces no useful progress. After the problem is complete, solve one parallel question with all help removed. That second question is the test of whether assistance became learning.

Research and Further Reading

eduKateSG Learning Node Series · 0052 of the continuing series. Previous: 0051 — How Learner-Generated Examples Work. Continue through the Study & Learning Methods Hub and the wider eduKateSG Learning Hubs.