Personalization has become the central promise of educational technology: a system that adapts to you, that learns your pace, your style, your gaps, and adjusts accordingly. The promise is real and, in some forms, valuable. But there is a version of personalization that does something else entirely — that adapts not to what you need in order to learn, but to what you prefer in order to feel comfortable. These are not the same thing. In some cases, they are opposites.
A few years ago, I spent several months using an adaptive language-learning application that was, by most available measures, excellently designed. It tracked my performance across vocabulary, grammar, and listening comprehension. It adjusted the difficulty of exercises based on my error patterns. It identified which concepts I was struggling with and returned to them more frequently. The experience of using it was smooth, encouraging, and well-calibrated to my level in a way that self-directed study from a textbook was not.
At the end of those months, I could move through the application's exercises with considerable facility. My performance metrics were good. My streak was long. And when I tried to have a conversation with a native speaker of the language, I discovered that I was substantially less capable than I had felt during those months — that the fluency the application's exercises had produced was a fluency with the application's exercises, not with the language.
I have thought about this experience more carefully than the experience itself probably warrants, because I think it illustrates something that is both specific to adaptive learning systems and important for understanding what personalization can and cannot do.
The Promise and Its Two Versions
Personalization in education is not a new idea. Good tutors have always adapted to their students — adjusting pace, emphasis, and explanation based on what the student seems to understand. The research on one-on-one tutoring, most influentially summarised by Benjamin Bloom in his 1984 "2 Sigma Problem," shows that individual tutoring produces learning outcomes two standard deviations better than conventional classroom instruction. The tutored student performs better than ninety-eight percent of students taught conventionally. Personalization, in this sense, is the most effective educational intervention ever documented.
The question is what personalization is doing that produces this effect. Bloom's hypothesis was that the advantage came from the tutor's ability to maintain each student within their Zone of Proximal Development — the gap between what the student can do independently and what they can do with appropriate assistance — and to adjust the nature of the assistance in real time as the student's understanding developed. The adaptation was to the student's learning needs: to where the genuine gaps were, to which misconceptions needed correction, to how much challenge was appropriate for the current level of understanding.
This version of personalization is hard to replicate at scale, and educational technology has been trying to replicate it for decades. The question I want to examine is whether the personalization that is currently being replicated is this version — adaptation to learning needs — or a different version: adaptation to comfort preferences.
Comfort Personalization and Its Appeal
The distinction sounds clear in the abstract and is considerably murkier in practice, because comfort preferences and learning needs overlap significantly at the beginning of learning. A novice who finds material too difficult is not learning effectively; reducing difficulty to a level where the material is accessible is both meeting a comfort preference and meeting a learning need. The tutor who adjusts for the overwhelmed beginner is doing both things simultaneously.
The divergence occurs as expertise develops. What a learner needs, as understanding deepens, is increasing challenge: material that engages the developing competence rather than simply confirming it, problems that require the application of knowledge in unfamiliar contexts rather than the recognition of familiar patterns, questions that cannot be answered by retrieving practiced responses. What a learner often prefers, at this stage, is the opposite: the confirming exercise, the familiar pattern, the question they know they can answer correctly. The satisfaction of competent performance is real and motivating, and it feels like learning.
A personalization system that adapts to performance data faces a structural temptation: optimise for the signal that is easiest to measure and that produces the most positive user experience. That signal is successful task completion. A learner who completes tasks successfully, maintains a high accuracy rate, and feels good about the experience is a learner who will continue to use the system. A learner who encounters challenging material that they fail at, feel confused by, and need to struggle with may have a worse immediate experience even if they are learning more effectively.
The language-learning application I used was optimising for the first signal. It produced high task-completion rates because it had adapted the tasks to the level at which I could reliably complete them. The problem was that this adaptation had quietly slipped past what I needed to learn into what I was already comfortable doing. I was being challenged enough to maintain engagement but not enough to push the boundary of my actual competence. The personalization was working perfectly, and it was working in the wrong direction.
The Filter Bubble Problem, Applied to Learning
The concept of the filter bubble — developed by Eli Pariser to describe how algorithmic curation produces information environments in which people are primarily exposed to content that confirms their existing beliefs — has an analogue in personalized learning that has received considerably less attention than it deserves.
An adaptive learning system that adjusts content to the learner's demonstrated preferences is, in a specific sense, building an epistemic filter bubble around the learner's existing knowledge. It presents material the learner is likely to find accessible, relevant to their established interests, and consistent with their current understanding. This is pleasant and, up to a point, effective. It is also, systematically, the opposite of what genuine intellectual development requires.
Intellectual development — the kind that produces understanding rather than mere competence at familiar tasks — requires regular confrontation with material that does not fit existing frameworks. It requires the experience of existing knowledge being challenged, of categories that seemed adequate turning out to be inadequate, of connections that seemed obvious turning out to be wrong. This confrontation is uncomfortable, and it is productive precisely because it is uncomfortable: the discomfort signals that the existing framework is being genuinely tested rather than merely confirmed.
A personalization algorithm that filters for comfort will consistently undersupply this kind of confrontation. Not because it is badly designed, but because the signal it optimises for — user satisfaction, completion rates, positive engagement metrics — is negatively correlated with the experience it would need to provide. The learner who is most satisfied in the moment is often the learner who is being most productively confirmed in what they already know. The learner who is most unsatisfied — confused, challenged, aware of inadequacy — is often the learner who is most productively learning.
What Adaptive Learning Actually Requires
The research tradition that underpins genuine adaptive learning is considerably more demanding than the personalization currently deployed in most educational technology.
The most rigorous work on adaptive learning — associated with cognitive tutoring systems like those developed at Carnegie Mellon, and with the intelligent tutoring systems research of John Anderson, Kurt VanLehn, and their colleagues — aims to adapt not to demonstrated preference but to inferred knowledge state. The system attempts to model what the learner knows and does not know at the level of specific knowledge components, identifies which components are in a state of partial or incorrect understanding, and directs practice toward those components rather than toward the components the learner can already handle.
This is meaningfully different from most commercial personalization. Commercial systems typically adapt difficulty based on performance history — more errors produce easier material, fewer errors produce harder material — without attempting to model the underlying knowledge state that produces the performance. The performance-based adaptation is responsive to proxies rather than to the thing itself, and proxies that feel good — high accuracy, smooth progression — can be produced by routes that do not build the knowledge that would produce them independently.
The deeper problem is that even sophisticated knowledge-modelling systems face a challenge that personalization, in any form, cannot fully resolve: the challenge of knowing what a learner needs to encounter that is genuinely outside their current knowledge state. The system can model what the learner knows, but it cannot model what the learner does not know that they do not know — the conceptual gaps that are invisible precisely because the learner lacks the framework to recognise them as gaps.
This is the limitation that the best human tutors navigate not through data but through dialogue — through the kind of open-ended questioning that probes not just whether the learner can answer correctly but whether the learner understands what they are doing and why. The Socratic question "what do you mean by that?" is not a performance measure; it is an investigation of the thinking that produced the performance. Personalization systems measure performance. They do not, and currently cannot, investigate the thinking.
The Self-Assessment Problem
There is a further difficulty with personalization that is grounded in well-documented limits of human self-knowledge.
Many adaptive systems rely, in part, on learner self-assessment: reports of how confident the learner feels, which material seems familiar, which topics seem well understood. This reliance is understandable — self-assessment provides data that performance metrics alone cannot. It is also unreliable in a specific and well-documented way: the Dunning-Kruger research shows that the correlation between felt understanding and actual understanding is weakest precisely among learners with limited competence in a domain. Novices and early intermediates are systematically overconfident; they feel more certain of their understanding than their understanding warrants.
A personalization system that adapts to self-assessed confidence will therefore systematically underprovide challenge and remediation to the learners who need it most. The confident novice who reports high familiarity with a topic will receive less practice in that topic, less challenge, less confrontation with the gaps in their understanding — because the system has been told, by the learner, that the gaps do not exist. The learner who feels fluent with the application's exercises, who assesses themselves as having learned, who experiences high confidence and satisfaction, may be the learner who has been most systematically deprived of the confrontation that genuine learning requires.
This is what I think happened during those months with the language application. My self-assessed progress was genuine — I was improving at the exercises. My confidence was real — I felt capable. The gap between that confidence and my actual ability to communicate in the language was not visible from inside the experience of using the application, because the application had adapted so effectively to the level at which I could perform that it had stopped pushing against the boundary of what I actually needed to develop.
The Difficulty of Productive Discomfort
There is a specific kind of discomfort that is associated with genuine intellectual development, and it is worth naming it precisely because it is the discomfort that personalization systems most reliably eliminate.
When you encounter an idea that does not fit your current framework — that contradicts something you believed, that introduces a distinction you had not previously made, that requires you to revise an understanding you were comfortable with — the experience is unpleasant in a specific way. Not the unpleasantness of difficulty, exactly, but the unpleasantness of cognitive dissonance: the state of holding two things that do not fit together, without yet knowing how to resolve them.
This state is, according to the research on conceptual change and on productive failure (the work of Manu Kapur at ETH Zürich in particular), among the most educationally productive states a learner can be in. The dissonance that is not immediately resolved — that is held open rather than prematurely closed — drives the kind of deep processing that produces genuine conceptual revision rather than surface updating. The understanding that comes after genuine dissonance is qualitatively different from the understanding that comes from learning something entirely new: it reorganises existing knowledge rather than merely adding to it.
A personalization system that adapts to comfort will consistently avoid creating this dissonance. It will present material that fits the learner's existing framework, that confirms rather than challenges, that extends rather than revises. This is more pleasant than the alternative. It is also, specifically, the alternative that produces less conceptual depth.
Manu Kapur's research on what he calls "productive failure" demonstrates this with unusual clarity. In his studies, students who were asked to work on problems they could not yet solve — before receiving instruction on how to solve them — produced worse performance on the problems than students who received instruction first, but significantly better performance on subsequent transfer problems that required the application of the concepts in new contexts. The struggle to solve an unsolvable problem activated and organised relevant prior knowledge in a way that made the subsequent instruction more deeply understood. The prior failure was productive precisely because it had created the cognitive conditions that made the instruction maximally effective.
This is the educational dynamic that personalization systems cannot produce if they are optimised to avoid difficulty and failure. The productive failure requires the failure. The personalization that protects the learner from failure is protecting them from the mechanism.
Differentiation vs. Personalisation
The distinction between differentiation and personalisation is one that educational research makes carefully and that educational technology often collapses.
Differentiation is the adaptation of instruction to meet learners where they are — different entry points, different scaffolding, different pacing — in service of the same rigorous learning objectives. Every learner is working toward the same understanding, by paths appropriate to their current level. The objectives do not change; the route does.
Personalisation, in its commercial form, often goes further: it adapts not just the route but the destination, delivering content and challenges that match the learner's existing interests and preferences rather than a fixed curriculum. This sounds like an improvement on differentiation. It is, in many cases, a regression: it replaces a shared rigorous standard with a personalised comfortable one, and it optimises for engagement and satisfaction at the cost of the challenge that builds genuine competence.
The most important things to learn are often not the things that feel most immediately relevant. They are the things that expand the frame — the knowledge that makes new questions possible, the skills that open domains that were previously inaccessible, the conceptual tools that change what can be thought. A personalization system that curates toward existing interests and demonstrated preferences will consistently under-supply this kind of expanding challenge, because it is expanding precisely away from what the learner's history of preferences would predict.
A Closing Distinction
I do not want to leave the impression that personalization is without value, because this is not the argument. The argument is more specific: personalization as comfort-calibration — as the optimisation of the learning experience for immediate satisfaction, low friction, and high confidence — is not the same thing as personalization as learning-needs-calibration. The first is well-executed by most current adaptive systems. The second is rare, difficult, and requires a willingness to sacrifice exactly the metrics that make the first version commercially successful.
There is a version of adaptive learning that we think is worth building toward: a system that models not just what the learner can perform, but what the learner understands and does not understand, and that uses that model to identify the specific misconceptions and gaps that require the discomfort of revision rather than the comfort of extension. A system that deliberately introduces material the learner will find difficult and that tracks not just whether the difficulty resolves but how it resolves — whether the resolution represents genuine conceptual change or surface updating. A system that adapts, in other words, to what the learner needs rather than to what the learner prefers.
This system is harder to build, harder to evaluate, and less immediately satisfying to use. It is also, we believe, the only version of personalized learning that is honest about what learning actually requires. The language application I used was excellent at the first version. It knew what I could do and kept me doing it, slightly extended, in conditions that felt productive. What it could not do — and what no system that optimises primarily for satisfaction metrics can do — was tell me what I did not know I did not know, challenge me in the ways that would reveal the actual boundary of my competence, and provide the kind of uncomfortable confrontation with my own limitations that produces the deep encoding that makes learning durable.
Personalization, as a technical capability, is genuinely impressive. Personalization as an educational philosophy, without the constraints that genuine learning requires, is one of the most sophisticated ways yet devised to make people feel like they are learning without requiring them to do the work of it.
The measure of an adaptive system is not how well it learns what the user finds comfortable. It is how well it determines what the user needs to find uncomfortable — and then provides it in the form most likely to produce growth rather than avoidance. These are different optimisation problems. They require different design choices. And the difference between them is the difference between a system that serves the learning and a system that serves the feeling of learning.