A student opens a mathematics platform.
Question 1 is easy.
Correct.
Question 2 is harder.
Correct.
Question 3 is harder again.
Wrong.
The platform lowers the difficulty.
Is that intelligent tutoring?
Not necessarily.
A system can vary difficulty and still know remarkably little about what the learner is thinking. It may know only that the last answer was wrong. It may not know whether the learner misunderstood place value, misread the question, selected the wrong operation, made a transcription error, guessed, or knew the method but ran out of working memory.
An intelligent tutoring system aims at a more ambitious job. It tries to maintain a model of the learner, select or generate a learning task, interpret the student’s response, provide feedback or help, and update what it believes the learner needs next.
In other words, good adaptive learning is not simply:
right → harder; wrong → easier.
It is a loop:
estimate learner state → choose task → observe response → diagnose → intervene → update estimate → choose next task.
That loop has been studied for decades in educational technology, cognitive tutors and adaptive-learning research. In current school practice, the language now overlaps with AI, personalised learning and digital platforms. The hype can make it difficult to see the underlying mechanism.
Current Education Endowment Foundation evidence is useful here because it separates product enthusiasm from evaluated implementation. The 2025 EEF efficacy trial of Maths-Whizz, an online tutoring programme for Year 5 pupils, found an average positive effect equivalent to about one additional month of mathematics progress compared with business-as-usual provision. The programme included adaptive software, but also diagnostic assessment, teacher dashboards, implementation support and school routines. EEF’s current EdTech research agenda likewise identifies adaptive learning, AI and the protection of valuable cognitive processes as active areas for research rather than solved problems.
The central question is therefore not:
“Can software personalise learning?”
Of course software can personalise something.
The real question is:
“What is being personalised, on the basis of what evidence, toward what learning model, and with what role left for the student and teacher?”
The 50-second answer
An intelligent tutoring system works by trying to approximate some of the decisions a skilled tutor makes.
It needs at least four interacting parts:
- Domain model: a representation of the knowledge or skills being taught.
- Learner model: an estimate of what this learner knows, misunderstands, forgets or can currently do.
- Tutoring policy: rules or models for deciding what task, hint, explanation or feedback should come next.
- Interface and evidence stream: the environment through which the learner acts and the system observes performance.
When the learner responds, the system updates its estimate and chooses the next instructional move.
The quality of the system therefore depends on more than whether the algorithm is “AI”. It depends on:
- whether the domain model represents meaningful learning;
- whether learner data are diagnostic rather than merely plentiful;
- whether feedback supports thinking rather than bypassing it;
- whether task selection balances practice, challenge and review;
- whether the teacher can interpret and use the system’s information;
- whether students actually use the system enough and as intended;
- whether the learning transfers beyond the platform.
A tutoring system is intelligent educationally only to the extent that the loop improves learning decisions.
A flashy interface with weak diagnosis is still weak tutoring.
1. The tutoring problem: instruction is a sequence of decisions
A human tutor rarely teaches by delivering one fixed explanation from beginning to end.
The tutor asks a question.
The student answers.
The tutor notices hesitation.
They ask for a smaller step.
The learner makes a sign error.
The tutor decides whether it is a careless slip or a misconception.
They give a prompt.
The learner corrects it.
The tutor gives another problem with a slight variation.
The learner succeeds.
The tutor increases independence.
Teaching is therefore partly a control problem. The instructor repeatedly observes the learner and selects the next move.
Traditional classroom teaching solves this at scale through lesson design, questioning, feedback, grouping and teacher judgement.
An intelligent tutoring system attempts to automate part of this decision loop for an individual learner.
That is ambitious because the system never observes “understanding” directly.
It observes traces.
Answers.
Response times.
Hint requests.
Sequences of actions.
Error patterns.
Revisions.
Sometimes written or spoken explanations.
From those traces, it must infer something about an invisible state.
That inference is the centre of intelligent tutoring.
2. The domain model: what exactly is being learned?
Before a system can personalise learning, it needs a model of the subject.
Consider fractions.
A weak domain model might simply contain a chapter called “fractions”.
A stronger model decomposes the territory:
- fraction as part–whole;
- fraction as number on a number line;
- equivalence;
- comparison;
- common denominators;
- addition and subtraction;
- multiplication;
- division;
- ratio relationships;
- common misconceptions.
Now task performance can be linked to more specific capabilities.
The system may know that a learner can add fractions with the same denominator but repeatedly treats the denominator as additive when denominators differ. That pattern suggests a different intervention from a student who knows the procedure but makes multiplication errors.
The granularity matters.
Too coarse, and the system only knows “weak at fractions.”
Too fine, and the model becomes administratively complex or statistically unstable.
Good domain models therefore represent skills, concepts, dependencies and likely error patterns at a resolution useful for teaching.
This is not purely a technical problem. It is curricular.
What counts as a meaningful unit of mathematical knowledge?
Should explanation and procedure be separate?
Should the model represent representations such as number lines?
How should misconceptions be encoded?
An adaptive platform quietly contains answers to these education questions whether or not its marketing page mentions them.
3. The learner model: an estimate, not a mind reader
The learner model is the system’s current belief about the student.
This can include estimates of:
- mastery of specific skills;
- probability of knowing a concept;
- common misconceptions;
- fluency;
- forgetting risk;
- response confidence;
- persistence;
- help-seeking behaviour;
- pace.
Different systems use different techniques: rule-based models, Bayesian knowledge tracing, performance-factor analysis, item-response methods, deep learning, embeddings, large language models or combinations.
The technical method matters, but one principle matters even more:
the model is an estimate under uncertainty.
A correct answer does not prove mastery.
The student may have guessed.
A wrong answer does not prove ignorance.
The student may have misclicked.
A fast response may indicate fluency or impulsivity.
A slow response may indicate careful reasoning or confusion.
Therefore learner modelling works through repeated evidence, not single-event certainty.
This is why well-designed adaptive systems use patterns across tasks and often mix item types rather than declaring mastery after one success.
The teacher should interpret dashboard estimates the same way: as evidence worth considering, not as an X-ray of the child.
4. Knowledge tracing: estimating what persists beneath performance
Knowledge tracing asks a subtle question:
Given the learner’s sequence of responses, what is the probability that the underlying knowledge is present?
Suppose a student answers:
correct, correct, wrong, correct, correct.
A raw accuracy measure says 80%.
A knowledge-tracing model may ask more:
- How difficult were the items?
- Did the wrong answer match a misconception?
- Was a hint used?
- How much time passed between attempts?
- Was the latest evidence more informative than the earliest?
The system then updates an internal mastery estimate.
This matters because learning is latent. Performance is noisy.
Human teachers solve the same problem informally when they say:
“She got that right, but I’m not convinced she understands it yet.”
or:
“He made one arithmetic slip, but the method is secure.”
The machine version attempts to formalise that judgement.
For a dedicated eduKateSG owner on the broader mechanism of inferred mastery, see the surrounding How X Works learning estate rather than treating a dashboard score as equivalent to knowledge itself.
5. Task selection: adaptive does not mean permanently harder
The most visible feature of adaptive software is task selection.
The learner succeeds, and the next task changes.
But a sophisticated tutoring policy may choose among several goals:
- introduce a new skill;
- revisit a prerequisite;
- increase challenge;
- provide additional practice;
- space previously learned material;
- test transfer;
- diagnose a suspected misconception;
- build fluency;
- reduce frustration after repeated failure.
These goals can conflict.
If the system always maximises immediate success, the learner may receive easy work.
If it always maximises difficulty, motivation and accuracy can collapse.
If it always pursues the estimated weakest skill, the learner may get trapped in endless remediation and lose contact with current curriculum.
If it never revisits mastered content, forgetting can erase the gain.
Adaptive sequencing is therefore a policy problem.
What learning objective should be optimised over what time horizon?
That question is educational before it is algorithmic.
6. Feedback: tell, hint or ask?
Suppose a learner gets an equation wrong.
The system could say:
Incorrect. The answer is 12.
Fast.
Or:
Check what happens to the sign when you subtract 5 from both sides.
More diagnostic.
Or:
Which operation would undo +5 while preserving equality?
More generative.
Or it could show a worked example, ask the learner to compare solutions, highlight the first incorrect step, or delay the hint until another attempt.
The feedback decision matters because help can either support cognition or replace it.
A tutoring system that reveals every answer immediately may produce smooth progress through content while reducing the need to retrieve, diagnose or persist.
A system that withholds help too long may generate unproductive failure.
The familiar assistance dilemma appears in digital form: how much help is enough?
eduKateSG has a separate owner for How the Assistance Dilemma Works | Give Enough Help to Keep Learning Moving Without Taking Over the Thinking. Intelligent tutoring systems operationalise that dilemma repeatedly at scale.
7. Step-based tutoring: why the final answer can be too late
Some intelligent tutors analyse intermediate steps rather than only final answers.
This can be powerful.
Imagine solving:
\(3x + 5 = 20\)
A learner writes:
\(3x = 25\)
A final-answer-only system waits until the end.
A step-based tutor can intervene immediately:
“If 5 is added on the left, what inverse operation preserves equality?”
The learner corrects:
\(3x = 15\)
The system has caught the misconception near its source.
This resembles a human tutor looking over the learner’s working.
It can also produce richer diagnostic data. If a student repeatedly makes the same invalid transformation, the system can distinguish that from random arithmetic mistakes.
The challenge is technical and curricular: the system must recognise multiple valid solution paths without falsely flagging unconventional but correct reasoning.
That is harder than checking one number.
8. Worked examples and faded assistance
Novices often benefit from worked examples because they reduce the search burden while making a correct method visible.
An intelligent tutor can personalise when to move from:
worked example → partially completed example → prompted problem → independent problem.
This is sometimes called faded guidance.
A learner who has already mastered the method should not be forced through unnecessary scaffolds.
A learner who is still guessing should not have the scaffold removed merely because three screens have passed.
Adaptive systems can, in principle, use learner evidence to time the fade.
In practice, the quality depends on the learner model and task design.
This illustrates a recurring theme:
adaptivity is only as good as the evidence used to adapt.
9. The engagement problem: a system can be accurate and still go unused
A technically excellent tutor produces no benefit if students do not use it.
This is one reason real-world implementation matters.
Usage depends on:
- timetable allocation;
- device access;
- login reliability;
- teacher expectations;
- student motivation;
- internet stability;
- interface quality;
- perceived relevance;
- school follow-through.
EEF’s Maths-Whizz trial is useful precisely because it evaluated a programme in real schools rather than a short laboratory demonstration. The intervention included recommended weekly usage, diagnostic placement, teacher monitoring and implementation support.
Even then, actual usage varied.
This is not a side issue.
In educational technology, implementation is part of the treatment.
A system may be highly effective per hour used and still have weak overall impact if learners rarely log in. Another may have modest per-task benefits but strong curriculum integration and therefore accumulate more learning.
Schools should ask both:
Does the tool work when used?
and:
Can our students and teachers actually use it at the intended dose and quality?
10. The teacher dashboard: evidence or decoration?
Adaptive platforms often provide dashboards with mastery percentages, traffic-light colours, recommended groups and skill gaps.
These can be useful.
They can also become data wallpaper.
A dashboard helps only if the teacher can answer:
- What does this metric mean?
- How is it calculated?
- How uncertain is it?
- What teaching decision should it change?
- Can I inspect the underlying student work?
Suppose a dashboard says:
Fractions: 62% mastery.
What should the teacher do?
Without decomposition, perhaps nothing.
If the system shows:
equivalence secure; comparison unstable; repeated denominator-addition misconception, the information becomes actionable.
The dashboard should compress evidence without hiding its educational meaning.
Teacher professional judgement remains essential because software sees only the interactions it records. It may not know that the student was exhausted, received help from a sibling, misused a calculator, could explain orally, or had never been taught a representation used by the platform.
11. Adaptivity can personalise the wrong thing
Imagine a platform that perfectly adapts arithmetic difficulty.
The student’s real problem is interpreting word problems.
The system may become increasingly precise about the wrong construct.
This is why domain and assessment validity matter.
Personalising task difficulty is not the same as personalising learning need.
A learner may need:
- different representation;
- vocabulary support;
- more retrieval;
- less scaffolding;
- conceptual explanation;
- motivation support;
- teacher conversation;
- a richer problem.
A system whose only adaptive dimension is easier/harder cannot solve every version of “personalised learning”.
That is not necessarily a criticism. Narrow tools can be excellent. The problem appears when narrow adaptivity is marketed as comprehensive individualisation.
12. Personalisation can reduce shared curriculum if poorly bounded
Suppose two students use an adaptive platform for months.
The system keeps Student A on prerequisite content because mastery estimates remain low.
Student B advances far ahead.
Adaptivity has increased individual fit.
It may also have created diverging curriculum access.
This can be appropriate temporarily. A learner missing foundations needs repair.
But schools need a return path.
Otherwise adaptive remediation can become a digital form of permanent low expectation.
Ask:
How does the system reconnect a learner to age-appropriate or course-appropriate curriculum?
Who decides when the learner is ready?
Can the teacher override the sequence?
Does the student still receive rich whole-class instruction?
Adaptive tutoring should supplement a coherent curriculum, not quietly replace it with an endless individual queue.
13. The data problem: more clicks do not equal more understanding
Educational software can collect extraordinary quantities of data.
Clicks.
Hover time.
Keystrokes.
Response latency.
Hint use.
Error transitions.
Item sequences.
That creates a temptation:
if we can measure it, it must tell us something important.
Not necessarily.
High-dimensional behavioural traces need valid interpretation.
A long response time can mean:
- deep thinking;
- distraction;
- confusion;
- device trouble;
- leaving the room.
Educational data become useful when tied to a model of cognition and validated against meaningful outcomes.
The strongest tutoring systems are not simply data-hungry.
They are theory-constrained.
14. AI changes the interface before it necessarily changes the pedagogy
Large language models make tutoring systems feel more human.
Students can type natural questions.
The system can generate explanations.
It can paraphrase.
It can create examples.
It can ask Socratic-style questions.
It can discuss a mistake in conversational language.
This is valuable.
It also creates a critical distinction:
conversational fluency is not the same as pedagogical intelligence.
A chatbot may sound supportive while giving too much help.
It may generate a plausible but incorrect explanation.
It may fail to maintain a reliable learner model across sessions.
It may answer the student’s immediate question without considering curriculum sequence.
It may solve the task the student should solve.
Therefore modern AI tutoring needs the same old educational architecture:
What is the learner supposed to know?
What evidence do we have?
What is the smallest useful intervention?
What should happen next?
The interface is new.
The tutoring problem remains.
15. Cognitive offloading: when helpful AI removes the learning
Consider a student writing an essay.
They ask an AI tutor:
“Give me three arguments, evidence and a conclusion.”
The system provides excellent material.
The student assembles it.
The product improves.
Did the student learn to generate arguments?
Maybe not.
This is cognitive offloading: a tool performs part of the mental work.
Offloading is not inherently bad.
Calculators offload arithmetic.
Spellcheck offloads detection.
Maps offload route memory.
Good tools free cognition for higher-value work.
The educational question is which cognition should be offloaded now?
EEF’s current EdTech research agenda explicitly identifies the need to understand which cognitive processes technology should protect and which can be delegated.
An intelligent tutor should therefore sometimes refuse to be too helpful.
It may ask the learner to attempt before receiving a hint.
It may provide a cue rather than a solution.
It may require retrieval.
It may hide worked steps until needed.
Pedagogical intelligence includes strategic restraint.
16. A Primary mathematics example
A Year 5 learner struggles with fraction comparison.
The system presents:
Which is larger: 3/5 or 4/7?
The learner chooses 4/7 because 4 > 3 and 7 > 5.
A shallow system says:
“Incorrect. Try again.”
A stronger tutor detects a likely whole-number reasoning error.
It might present a number-line representation and ask:
“If both fractions are between 0 and 1, does having the larger denominator automatically mean the fraction is larger?”
Then it gives a diagnostic comparison:
1/2 or 1/8?
If the learner still chooses 1/8, the misconception is clearer.
The system can then provide a short explanation or representation and retest with a fresh pair.
Notice what made the system intelligent educationally:
- error classification;
- diagnostic item selection;
- representation change;
- retest.
Not merely “adaptive difficulty”.
17. A Secondary algebra example
A learner solves:
\(2(x+3)=14\)
They write:
\(2x+3=14\)
The system detects incomplete distribution.
Instead of showing the correction, it presents:
\(3(a+4)\)
and asks the learner to expand it.
If the learner writes \(3a+4\), the system has confirmed the pattern.
It then uses an area representation or repeated addition:
\(3(a+4) = (a+4)+(a+4)+(a+4)\)
The learner reconstructs:
\(3a+12\)
Then returns to the original problem.
The tutor has moved sideways to diagnose a component skill before returning.
A human tutor might do exactly the same.
18. A language-learning example
Intelligent tutoring is not limited to mathematics.
A language system can model:
- vocabulary knowledge;
- spelling;
- grammar;
- pronunciation;
- reading fluency;
- error patterns.
Suppose a learner repeatedly writes:
He go to school every day.
A strong tutor distinguishes this from random error and targets third-person singular agreement.
It may present contrastive sentences:
I go.
They go.
He ___.
The learner supplies goes.
Then the system varies the verb and context, later spacing the item.
Again, the value lies in structured diagnosis and adaptation.
19. A writing tutor faces a harder problem
Writing is less constrained than algebra.
There may be many valid answers.
Quality depends on argument, evidence, organisation, style, audience and purpose.
An AI writing tutor can generate useful feedback, but validity becomes harder to guarantee.
Does the system know the intended curriculum?
Can it distinguish a deliberate stylistic choice from an error?
Does it over-correct toward generic prose?
Does it provide so much revision that the text becomes the system’s rather than the learner’s?
For open-ended domains, teacher oversight becomes even more important.
Automation can increase feedback volume.
It does not automatically increase feedback quality.
20. Intelligent tutoring is not the same as personalised learning
Personalised learning is a broad educational ambition.
It can include:
- adapted pace;
- different resources;
- student choice;
- targeted teacher support;
- different pathways;
- individual goals.
An intelligent tutoring system is one technological mechanism that can support some forms of personalisation.
It is not synonymous with the whole idea.
A human teacher can personalise without software.
Software can adapt without delivering meaningful personalisation.
21. Intelligent tutoring is not the same as a question bank
A question bank provides items.
An adaptive question bank changes which item comes next.
An intelligent tutor ideally adds diagnosis and instructional response.
Ask:
Does the system know why the learner got the answer wrong?
Can it choose a hint based on that diagnosis?
Can it update a skill model?
Can it decide when to revisit?
If not, “intelligent tutor” may be mostly branding.
22. Intelligent tutoring is not a replacement for teachers
A platform sees a narrow slice of the learner.
The teacher sees:
- class discussion;
- motivation;
- peer interaction;
- oral explanation;
- misconceptions across tasks;
- fatigue;
- curriculum context;
- student history.
A tutor can provide individual practice at scale.
A teacher can coordinate the education.
The best implementation question is therefore often not:
“Will this replace teaching?”
but:
“Which tutoring decisions can software handle reliably so teacher attention can move to the decisions software cannot?”
23. The implementation question schools should ask first
Before asking about AI architecture, ask:
Where in our learning system will this tool live?
During class?
Homework?
Intervention?
Revision?
Independent practice?
Who monitors usage?
What happens when a learner gets stuck?
How does platform data change classroom teaching?
What curriculum does the platform assume?
What happens if the platform and classroom sequence diverge?
What is the expected weekly dose?
Implementation architecture often matters more than feature count.
24. Evidence: what does the Maths-Whizz trial actually tell us?
The Education Endowment Foundation’s 2025 efficacy trial of Maths-Whizz evaluated an online tutoring programme in Year 5 mathematics across 63 primary schools in England. The programme used an initial assessment to place pupils, adaptive online tutoring, teacher dashboards and implementation support. EEF reports a positive average impact equivalent to approximately one additional month of progress in mathematics.
That is encouraging.
It should not be inflated.
One evaluated programme does not prove that all intelligent tutoring systems work.
The effect belongs to the programme as implemented, not simply to “AI” or adaptivity in isolation.
Usage varied.
Schools received training and support.
The age group and subject are specific.
Results from one system may not transfer to another system with a different domain model, interface, tutoring policy or implementation model.
The responsible interpretation is:
well-implemented adaptive tutoring can improve mathematics outcomes, and programme-level evidence matters more than generic claims about adaptive technology.
25. What EEF’s current EdTech agenda adds
EEF’s current research agenda for EdTech identifies several live questions:
- how adaptive learning should be used;
- how AI tools affect learning;
- which cognitive processes should be protected from technological offloading;
- how teachers can use technology effectively;
- how implementation conditions shape impact.
This is a healthy sign.
The field is not treating “AI in education” as one intervention.
That phrase hides many different mechanisms.
A tutoring bot.
An automated essay scorer.
A teacher planning assistant.
An adaptive question sequencer.
A worked-example generator.
A retrieval scheduler.
These tools can affect different parts of the learning process.
Evaluation should therefore be mechanism-specific.
26. Common failure mode: mastery percentage theatre
The platform shows:
Mastery: 93%
Everyone feels informed.
But what does 93% mean?
Probability of answering similar items correctly?
Coverage of skills?
Model confidence?
Recent accuracy?
A proprietary score?
Teachers and students need interpretability.
A precise-looking number can create false confidence if its construct is vague.
Ask vendors to explain metrics in educational terms.
27. Common failure mode: the platform optimises engagement rather than learning
Digital systems are excellent at making experiences sticky.
Points.
Animations.
Streaks.
Levels.
Avatars.
These can support motivation.
They can also optimise screen time.
Educational success is not minutes spent in the platform.
It is change in capability.
Schools should distinguish:
engagement metric from learning metric.
A learner can spend forty delightful minutes and learn little.
Another can complete fifteen difficult minutes and make meaningful progress.
28. Common failure mode: adaptive remediation becomes a tunnel
The system detects weakness.
It sends the learner backward.
Another weakness appears.
Backward again.
Months later, the student is still working below current curriculum.
This may be appropriate for severe foundational gaps.
It may also be a system with no re-entry logic.
Every adaptive intervention should ask:
How does the learner get back to the main road?
29. Common failure mode: help becomes answer supply
A student can learn how to farm hints.
Click hint.
Click next.
Reveal step.
Reveal answer.
The system records progress.
The learner records little memory.
Hint systems need productive friction.
Possible designs include:
- require an attempt before hint;
- give conceptual rather than procedural hint first;
- fade hint detail;
- follow a hint-assisted item with an unassisted transfer item;
- model hint dependence explicitly.
The target is not “minimum struggle”.
It is maximum useful cognition per unit of help.
30. Common failure mode: teachers trust the system more than the student
Dashboard says weak.
Student says, “I know this.”
The teacher says, “The computer says you don’t.”
This is poor use of data.
Test the discrepancy.
Give the learner a fresh problem.
Ask for an explanation.
Inspect the platform items.
Maybe the student guessed earlier.
Maybe the model is stale.
Maybe the learner understood after the last session.
Maybe the platform measures a slightly different skill.
Data should trigger inquiry, not end it.
31. If you are a learner
Use intelligent tutoring software as a practice partner, not an oracle.
When you get something wrong, do not click immediately past the feedback.
Ask:
What kind of error was this?
Did I not know the concept?
Did I choose the wrong method?
Did I make a careless execution error?
Did I misunderstand the language?
When the system gives a hint, try to finish the rest without further help.
When a mastery score changes, test yourself away from the platform.
Can you solve a new problem on paper?
Can you explain the idea?
Can you use it in a different form?
The platform is strongest when its evidence survives outside the platform.
32. If you are a parent
Do not judge an adaptive programme by how futuristic it looks.
Ask:
- What subject and age group?
- What evidence supports this programme?
- What does it adapt?
- What happens after an error?
- Can a child game the hints?
- How much weekly usage is intended?
- What does the teacher see?
- How does this connect to school curriculum?
- What does the child do away from the screen?
Watch the learner occasionally.
Not to police.
To understand the interaction.
Is the child thinking before clicking?
Reading feedback?
Using hints immediately?
Guessing?
Explaining?
The same software can produce different learning depending on use.
33. If you are a teacher
Decide what job the system does.
Fluency?
Diagnostic practice?
Intervention?
Homework?
Retrieval?
Then evaluate it against that job.
Use dashboard information to ask questions, not merely form groups.
Periodically inspect raw student work.
Check whether platform progress transfers to classroom tasks.
Teach students how to use hints productively.
Protect curriculum breadth.
And monitor who gets stuck in remediation pathways.
If the software saves you time, spend some of that time on the things software cannot see well: oral explanation, motivation, rich discussion, misconceptions in context, relationships and transfer.
34. A procurement test for schools
Before buying an intelligent tutoring system, ask the vendor:
- What is the domain model?
- How are skills defined?
- How is learner mastery estimated?
- What evidence changes the next task?
- How are misconceptions represented?
- What happens after repeated failure?
- How are hints designed?
- How is forgetting handled?
- Can teachers inspect and override recommendations?
- What independent evaluation exists?
- What implementation conditions were present in successful trials?
- How is student data protected?
If the answers focus entirely on “AI-powered personalisation” without explaining the instructional model, keep asking.
Educational intelligence should be describable in educational language.
35. The ordinary-weekday test
It is Tuesday at 10:40.
A student has completed twenty adaptive questions.
The dashboard is green.
The teacher asks one problem on paper.
The student solves it.
Then the teacher asks:
“Why does that method work?”
The student explains.
Then:
“Here is the same idea in a different context.”
The learner adapts.
Now the green dashboard has educational meaning.
The intelligent tutoring system did not replace the teacher’s judgement.
It created evidence worth checking.
Frequently asked questions
What is an intelligent tutoring system?
It is educational software designed to provide adaptive, learner-specific instruction or practice by modelling the subject, estimating learner state and selecting feedback or tasks in response to performance.
Is adaptive learning the same as AI?
No. Adaptive systems can use simple rules, statistical models, machine learning or AI. A system can be adaptive without using modern generative AI.
Does an intelligent tutor replace a teacher?
Current evidence supports using technology as part of a wider teaching system, not as a universal replacement for teachers. Human teachers provide curriculum coordination, rich explanation, social learning, motivation, judgement and context that software may not capture.
Do intelligent tutoring systems work?
Some programmes have positive evaluation evidence. Effects vary by programme, subject, population, usage and implementation. The EEF Maths-Whizz trial found an average positive effect equivalent to about one additional month of mathematics progress in Year 5 pupils. That finding should not be generalised to all systems.
What makes feedback intelligent?
Not fluency of wording alone. Good feedback responds to the learner’s actual error or state, gives the right amount of support and improves the next attempt.
What is a learner model?
An internal estimate of what a student knows, misunderstands or is ready to learn, based on observed performance and other data.
Can students game adaptive systems?
Yes. They may guess, overuse hints, seek external answers or exploit task patterns. Good systems and good implementation include checks for independent performance.
Is more personalisation always better?
No. Excessive individualisation can fragment shared curriculum, trap students in remediation or reduce collaborative learning. Personalisation needs curricular boundaries.
What should parents look for?
Independent evidence, transparent instructional logic, connection to curriculum, productive feedback and signs that learning transfers away from the screen.
What should schools protect from AI offloading?
Any cognitive process that is itself a learning goal: retrieval, explanation, method selection, argument generation, calculation where fluency matters, error diagnosis and sustained reading. Tools can offload lower-value work, but that boundary should be deliberate.
Evidence and caveats
The Education Endowment Foundation’s Maths-Whizz Efficacy Trial, published in 2025, is one current high-authority programme evaluation relevant to adaptive tutoring. The randomised controlled trial involved 63 primary schools and Year 5 pupils. EEF reports a positive effect on mathematics attainment equivalent to approximately one additional month of progress.
The intervention was a specific programme with defined implementation support. It should not be treated as proof that “AI tutoring” in general produces the same effect.
EEF’s current EdTech research agenda continues to identify adaptive learning, AI, teacher practice and cognitive offloading as areas requiring further research. This is especially important as generative AI changes interfaces faster than causal evidence accumulates.
Several general evidence caveats matter:
- programme effects depend on usage and implementation;
- adaptive systems vary greatly in instructional sophistication;
- short-term platform performance may not equal durable learning;
- open-ended subjects are harder to model reliably than constrained problem spaces;
- learner models are probabilistic and can be wrong;
- teacher integration influences whether digital evidence changes instruction;
- data privacy and algorithmic fairness require separate scrutiny.
The responsible conclusion is not that intelligent tutoring systems are a solved technology.
It is that adaptive tutoring can be educationally useful when the software embodies a credible learning model, produces diagnostic evidence, gives appropriately calibrated help and sits inside a well-implemented human teaching system.
The deeper idea: intelligence is in the loop
The phrase intelligent tutoring system encourages us to look for intelligence inside the software.
Look instead at the loop.
What does the system think the learner knows?
What evidence changed that belief?
Why did it choose this problem?
Why this hint?
What happens after the hint?
Does the learner become more independent?
Does the teacher learn anything useful?
Does knowledge survive when the screen disappears?
If those questions have strong answers, the technology may genuinely strengthen tutoring.
If not, difficulty adaptation and conversational fluency are not enough.
The smartest educational software is not the system that does the most thinking.
It is the system that knows which thinking still needs to belong to the learner.
Sources and further reading
- Education Endowment Foundation. Maths-Whizz Efficacy Trial, published 2025. https://educationendowmentfoundation.org.uk/projects-and-evaluation/projects/maths-whizz-efficacy-trial
- Education Endowment Foundation. Maths-Whizz programme findings. https://educationendowmentfoundation.org.uk/projects-and-evaluation/projects/maths-whizz-2024
- Education Endowment Foundation. Research Agenda theme: EdTech. Current priorities include adaptive learning, AI and cognitive offloading. https://educationendowmentfoundation.org.uk/projects-and-evaluation/research-agenda-themes-priority-areas/research-agenda-theme-edtech
- Education Endowment Foundation. Using Digital Technology to Improve Learning. https://educationendowmentfoundation.org.uk/education-evidence/guidance-reports/digital
- eduKateSG. How the Assistance Dilemma Works | Give Enough Help to Keep Learning Moving Without Taking Over the Thinking