The label problem
Personalisation systems work by building a model of the learner. The trouble starts when that model hardens: a slow start in algebra becomes a classification, the classification changes what the student is shown, and the narrowed material makes the classification look accurate. The loop closes quietly and nobody in it is doing anything wrong.
A student preparing for a competitive examination is not a fixed profile. They are a person having a difficult fortnight, or a good one, with a syllabus that does not care either way.
Adapt to what the student did this week. Do not adapt to what the system decided about them last term.
Rules worth holding
- 01Let the model forget. Recent evidence should outweigh old evidence by design, not by accident.
- 02Keep every door open. A student can always reach harder material, whatever the recommendation says.
- 03Explain the change. If the path shifts, the student and the teacher should be able to see why in one sentence.
- 04Measure recovery, not just progress. The useful signal is how quickly a student comes back after a bad topic.
The teacher is not a fallback
Adaptive systems are often designed as though the teacher is what happens when the software gives up. Inverted, the design gets better: the system does the tracking and the sequencing, and hands a teacher the two or three students whose pattern is worth a conversation this week.
Accessible education is not the same as automated education. The goal is to make a good teacher's attention go further, in places where there are not enough of them.
