Many AI learning systems treat support as a simple ladder:
Hint 1 → Hint 2 → Hint 3 → Answer
But learner support is not really a ladder.
Support is a decision.
The right intervention depends on why the learner is stuck.
A learner may need:
- a reminder of a criterion,
- clarification of the task,
- a worked example,
- retrieval support for something previously learned,
- reduction of irrelevant complexity,
- feedback on a misconception,
- or simply more time to think.
Giving “more help” is not always the right response. Sometimes it removes exactly the decision the learner needs to practise.
What I would love to see
An AI learning layer that works more like:
OBSERVE → INTERPRET → INTERVENE → RE-ELICIT → FADE → TRANSFER → STATE UPDATE
Instead of assigning a fixed hint_level, the system would use evidence from learner performance to choose the smallest useful intervention.
The goal would be:
Support access to the task without taking over the learner’s decision.
Then the system asks again.
If the learner succeeds, support is gradually faded.
If the learner can perform only with support, that is stored differently from independent performance.
If the learner succeeds later in a new context, that becomes evidence of transfer.
Why persistent learner state matters
A useful learner model could distinguish between:
- independent performance,
- supported performance,
- recurring barriers,
- support dependency,
- successful fading,
- transfer to a new task or context.
This would let AI support change over time instead of treating every interaction as a fresh attempt.
Example
Suppose a learner repeatedly chooses between two grammatical forms incorrectly.
The system should not automatically reveal the rule.
First it might determine whether the learner:
- cannot recall the form,
- cannot identify the relevant contextual cue,
- knows the rule but applies the wrong criterion,
- understands the distinction in isolation but not in live context,
- or can recognise the answer but cannot produce it independently.
Those are different learning problems and should lead to different interventions.
The broader product idea
Current models are already capable of interpreting rich learner evidence.
What seems to be missing is a product layer that turns that capability into a controlled adaptive learning workflow.
Not:
wrong answer → bigger hint
but:
evidence → diagnosis → minimum sufficient support → new attempt → fading → transfer
This could be useful not only for language learning, but for tutoring, professional training, mathematics, coding, writing, and other domains where the goal is eventually independent performance.
A principle I would love to see built into AI learning systems:
Give the smallest amount of help that can restart productive thinking.