How it learns
Deciding what to teach next,
and proving it worked.
Aptiz AI does not try to be a model that knows more. It is a teaching layer that sits on top of a general model and answers a different question: given this student, in this state, facing this error — what is the single best thing to do next, and did it actually work? This page is the method in full.
01 · The teaching loop
Every submitted task runs the same six steps. The loop is the product — the model underneath is a component of it.
Understand the task
Parse the problem, the course it belongs to, and the goal it is meant to serve.
Form a diagnosis
Analyse the answer and the working to produce a diagnosis that carries an explicit confidence. When confidence is too low to act on, the system asks one clarifying question rather than guessing.
Generate candidates
Produce several genuinely different teaching moves for that one diagnosis — not several rewordings of one explanation. This is the step most systems skip.
Rank and deliver
Score each candidate for expected learning gain given this student’s profile and current state, then deliver the highest-scoring one.
Verify the learning
Check mastery with a new task or a restatement. Following an explanation is not evidence of having learned it. This step is what stops the loop from optimising for the feeling of understanding.
Update the model
Write the result back into the student’s knowledge state and teaching preferences, then decide whether to continue, review, step back, or move on.
The first pass is pull: a student arrives with a problem they got wrong. After that the system goes push — it holds a knowledge state, so it can decide on its own what needs reviewing and when, without waiting to be asked. The demo on the homepage runs one full turn of this loop on a real error.
02 · What the system knows about a student
Two models, updated on different clocks. Collapsing them into one is how personalization degrades into a tone setting.
The stable profile
How this student tends to understand things. Preferred explanation length, hint strength, response to examples versus derivations, language level.
Updated slowly, across many sessions. Held as a live estimate that behaviour can overturn — never as a fixed label a student gets stuck with.
The dynamic learning state
What this student knows right now. Mastery per topic, error type, confidence, hint reliance, fatigue and frustration signals.
Updated after every single step. This is the input that decides what happens next.
Both are correctable. A student state that cannot be argued with is a log, not a model — and the difference shows up the first time the system is wrong about someone.
03 · What a teaching move actually is
A teaching move is one concrete teaching decision. It is the unit the system generates, ranks and chooses between — and the reason its output differs from a rewritten paragraph. Six moves cover most of what a good teacher does at the moment a student is stuck:
| Move | Reaches for it when |
|---|---|
| Follow-up question | The diagnosis is uncertain and one answer would settle it. |
| Counter-example | The rule is nearly right; showing where it breaks corrects it faster than restating it. |
| Analogy | The formal statement is not landing and a concrete image can carry the structure. |
| Worked example | There is a real concept gap and the student needs the whole shape before the parts. |
| Step hint | The student can get there and only needs the next rung — with the hint faded next time. |
| Prerequisite check | The error is downstream of something older that was never solid. |
A teaching strategy is a set of these moves and the order they run in. Strategies that hold up under evidence get kept, described, and reused on students in comparable states; the ones that do not get cut. The Founding Pilot starts every learner on three of them.
04 · The three things this gets confused with
"AI tutor" now covers three products that work nothing alike. Here is where each one stops.
vs. a general AI chatbot
A general chatbot waits for the student to ask a good question, then produces one answer and stops. That puts the hardest judgement — knowing what you don’t know, and what to do about it — back on the person least equipped to make it. Aptiz AI infers the next step from the material, the errors and the behaviour, generates several candidate moves, picks one, verifies it, and updates. The chatbot is a powerful engine with a manual gearbox; a student who already knows how to drive gets a lot out of it, and one who doesn’t gets a fluent explanation of something they weren’t stuck on.
vs. an adaptive question bank
Adaptive practice platforms built on a question bank and a knowledge graph choose the next question from a fixed library. That is a real capability, and it is a different one: the teaching method is whatever the item author wrote, and it never changes. Aptiz AI chooses the next teaching method from generated candidates — follow-up, counter-example, analogy, worked example, hint strength, ordering — which means the method itself is something the system can get better at, not a constant it inherits.
vs. a human tutor
A good tutor has method instincts that beat any current system. The problem was never the teaching — it is that those instincts cannot be recorded, cannot be compared across students, and cannot be afforded by most families. No one can hold dozens of students × hundreds of topics × months of responses in their head and reliably answer "which explanation actually worked, for whom, and did it hold a week later?" Aptiz AI turns that judgement into something written down and comparable. People stay in the loop for quality calibration, which is the part they are genuinely better at.
05 · What counts as evidence that learning happened
One standard: unprompted transfer, still there after a delay.
Learning evidence requires all three, in order: there was clear evidence of an initial error; a teaching intervention happened; and the student then demonstrated mastery on a new task they were not walked through. Miss any one and nothing has been shown.
The reason to be strict is that every softer metric can be moved without teaching anyone anything. Watch time, completion rates, streaks and satisfaction all improve when a product becomes more pleasant, which is a different achievement.
Not accepted as evidence
Accepted
A new task, solved without prompting, in a form the student has not been walked through — and still solvable when it comes back days later.
Why the bar is set there
Benjamin Bloom’s 1984 "2 sigma problem" reported that students taught one-to-one with mastery methods performed around two standard deviations above conventional classroom instruction — and framed the open question as whether any scalable method could approach that. It is the clearest historical statement of what personalized, mastery-based teaching is worth, and it is why the target here is mastery rather than engagement. It is not a prediction about this product: that figure came from human tutors under study conditions, and nothing about it transfers to a piece of software by default. Any claim we make about learning gains has to come from our own data, held to the standard above.
06 · Why recording is the foundation, not the product
A pile of chat logs is not an advantage. It is text. What compounds is a structured causal record, written the same way every time:
what kind of student,
in what knowledge state,
facing what goal and material,
got which teaching move,
produced what behaviour and errors,
and did the result improve.
Kept in that shape, the record does two things no person can: it forgets nothing, and it compares across hundreds of students at once. That is what makes "which method works for which state" an answerable question instead of an opinion.
Precisely because recording is something AI already does better than any person, none of our effort goes there. All of it goes into the one layer that is still unbuilt: choosing correctly.
07 · Where the system stops
Knowledge is strong, not infallible
General models cleared the usability bar for explanation and content. They still get things wrong, so reliable sources and a verification step stay in the loop rather than being assumed away.
Profiles are estimates, not labels
How a student learns is held as a running estimate that their behaviour can overturn at any point. Nobody gets sorted into a learning type and taught that way forever.
People stay in the loop
The AI makes the day-to-day teaching calls. People set quality bounds, handle low-confidence cases and exceptions, and review method effects — the judgement work, not the daily tutoring.
And one boundary about this page: it describes the method, not a results claim. Every figure shown anywhere on this site is illustrative until a learner’s own data replaces it.
Next
See a single turn of the loop run on a real error.