Fundamentals
It is usually offered as a reason to let people pick how they like to be taught. That is not what the research says. What it says is more useful, and more demanding: people arrive at the same material from different starting points, with different amounts of room in their heads that day, and the teaching has to move.
Start here
In 1984 the educational psychologist Benjamin Bloom published a comparison between students taught in an ordinary classroom and students taught one-to-one by a tutor. The tutored students performed around two standard deviations better — roughly, an average student becoming a top one. It is one of the largest effects in education research, and it has an obvious problem attached: a private tutor for every learner is something almost no school and no company can pay for. Most of what has happened in education technology since is an attempt, with varying honesty, to get some part of that effect without that cost.
A fixed path is a guess about an average person who is not in the room.
What a tutor actually does
Watch a good tutor and very little of what they do is delivering material. They notice a pause. They ask a question to find out whether the problem is the step or the thing three steps back. They give a hint instead of the answer. They make it harder when someone is coasting, and they stop and rebuild when someone is lost. They judge, continuously, what this person needs next — and almost none of that is visible in the content itself. Adaptive learning is the attempt to do some of that judging at a scale one human cannot reach.
What the research supports
These are not our results. They are among the better-established findings in learning science, they are decades old, and they are the standard any adaptive system — ours included — should be judged against.
Work that is too easy teaches nothing and work that is too hard teaches nothing, and the band in between is narrower than most courses assume. The psychologist Lev Vygotsky called it the zone of proximal development: the range of things you cannot yet do alone but can do with help. Its position is different for every learner, and it moves as they learn — which is precisely why a fixed course cannot stay inside it.
You can hold only a few new things in mind at once. John Sweller’s cognitive load theory draws the practical conclusion: when an explanation demands more than that, understanding does not degrade gracefully, it collapses. Much of what feels like "not being good at this" is an explanation that asked too much at once — and the fix is to change the explanation, not the learner.
The most uncomfortable finding in the field. Robert Bjork’s work on desirable difficulties shows that conditions which make learning feel harder and slower — spacing practice out, being tested rather than re-reading, mixing topics instead of drilling one — reliably produce better recall later. Re-reading feels productive because the material becomes familiar. Familiarity is not knowledge, and confusing the two is the most common way people waste study time.
Not a nice-to-have on top of the teaching. Deci and Ryan’s self-determination theory identifies three needs that sustain effort: some genuine say in what you are doing, a growing sense of being capable at it, and not feeling alone while you do it. Reinhard Pekrun’s work adds the other half: how much control a learner feels, and how much they value the work, shapes the emotions they bring to it — and those emotions measurably change how well they learn.
What the research does not support
The most popular belief about individual differences in learning is also the one with the weakest evidence behind it. It is worth being direct about, because it is still widely sold.
The idea is familiar: each person is a visual, auditory or kinaesthetic learner, and teaching matched to their type will work better. Surveys consistently find most educators believe some version of it.
The trouble is that testing it requires a specific result — learners taught in their preferred style should outperform learners taught in another, and the advantage should reverse when you swap the styles. In 2008 a group of cognitive psychologists led by Harold Pashler reviewed the literature for exactly that pattern and found almost no properly designed studies showing it, alongside several that looked for it and found nothing. People certainly have preferences. The evidence that teaching to those preferences improves learning is not there.
We mention it for two reasons. It is a genuinely useful thing for a reader to know. And it sets a boundary on what "adaptive" ought to mean: not asking someone to pick a style from a dropdown, but responding to what they are actually doing.
So what should change
Most systems described as adaptive reorder content — the same lessons, arranged differently. Reordering is the least of it. What a tutor varies is the teaching.
Making the work harder for someone cruising and easier for someone drowning — the sweet spot, tracked as it moves rather than guessed once at enrolment.
How much is worked through with the learner and how much is left to them, and — the part that gets forgotten — taking that support away again as they become able to do without it.
Some ideas can be approached from an example, a rule, a picture, or a problem you try first and fail at. These are genuinely different routes to the same understanding, and they are not equally good for every learner on every day.
What comes next, how much of it, and when something is revisited rather than assumed to have stuck. Spacing and review are among the best-supported findings in the field and among the least often designed for.
Adaptive means the teaching changes, not the playlist.
How we approach it
Our approach follows from the findings above, and from one view about what a system is allowed to do when it is making judgements about a person.
The teaching decisions should be explainable. When something chooses to give you an easier problem or take your hints away, you should be able to find out why, in words. We think that matters for a plain reason: a decision nobody can explain is a decision nobody can challenge — not the learner, not their teacher. It also keeps us honest, because a system whose reasoning has to be shown cannot lean on reasoning that would not survive being read.
What the system believes about a learner belongs to the learner. Anything that adapts to a person is forming a picture of them, and a picture that includes how they respond to difficulty is more personal than a test score. So it should be theirs to look at, correct and delete — not something quietly held about them, and not something their manager can browse.
We have worked with educational psychologists at PLUS Salzburg and JKU Linz since before we started building, which is the main reason this page can tell you what the evidence does not support as readily as what it does.
Being straight about it
Adaptive learning has a long history of promising more than it delivered, so it is worth saying where the line falls. Approximating a tutor is not the same as having one. A system can notice that you are struggling and change what it does about it; it cannot look at your face, and it does not know that your week has been terrible. It is a useful thing to put alongside teaching, and a poor thing to put in place of it. Anyone telling you their software replaces a teacher is selling you something the research does not support either.
Where this leads
Deciding that someone needs more support, or a different way in, still leaves the question of what that lesson should actually look like — which is an older craft with a literature of its own.
Read next: didactics, on how a lesson is built and why its order carries so much of its meaning. The structure all of this reads from is the knowledge graph, and the technology underneath is Atlas.
Judge it for yourself
It is the fastest honest test, and it works on ours too. If the answer is a shrug or a score with no reasoning attached, you have learned something useful.