On Thursday, March 12, I will participate as an invited expert in the International Seminar “Educational Innovation with Generative Artificial Intelligence in Virtual Learning Ecosystems,” organized by Universidad de La Salle (Mexico and Colombia) and Universidad de Guadalajara. The session I will contribute to carries a title that sounds both promising and somewhat dangerous: Designing Personalized Learning Experiences Mediated by Generative AI.
It is promising because personalization is perhaps the most seductive promise that artificial intelligence currently offers to education. It is dangerous because, when misunderstood, that promise can easily produce the opposite of what it intends.
I am writing this entry before the seminar rather than after it because I want to think in public. The questions posed by the organizers deserve answers that cannot be improvised in a roundtable discussion. They require time, reflection, and writing. As I have argued before in this space, writing is itself an act of thinking that machines may simulate but cannot truly perform.
The question no one wants to ask
What do we mean when we talk about personalized learning in contexts mediated by generative AI?
This was the first question proposed for the seminar discussion. The easy answer is that personalization means adapting content, strategies, and learning pathways to each student’s pace, level, and learning style. Educational platforms promise it. Academic papers describe it. Conferences celebrate it.
Yet something in this narrative feels strangely familiar. It resembles the same conceptual trap I discussed previously when reflecting on teaching thinking with artificial intelligence: the confusion between personalization and optimization.
Personalizing learning is not the same as optimizing the distribution of content. An algorithm that detects that a student struggles with fractions and then presents more exercises on fractions is not personalizing anything. It is recalibrating a vending machine. It simply provides more of the same, adjusted to a detected deficiency.
Real personalization would ask why the student struggles with fractions. Is the difficulty related to numerical representation? To language? To motivation? To previous experiences with measurement in everyday life? Those questions are not generated by an algorithm. They are asked by a teacher who knows the student, who has spoken with them, and who has noticed that brief moment of disconnection in their expression that no analytics dashboard can capture.
There is another dimension that is often ignored when discussing personalization in higher education: the structural conditions of teaching. Courses with forty or fifty students are expected to achieve the same learning outcomes in the same time frame as courses with ten or fifteen. There is no algorithm capable of resolving that pedagogical tension.
I often use a metaphor in class that seems increasingly accurate. In a previous post about the Socratic question, I described generative AI as a shop where ready-made answers are sold, paraphrasing Saint-Exupéry. Algorithmic personalization is something similar: a shop offering preassembled learning routes.
These routes may be carefully packaged, labeled, and adapted to the learner’s “size.” Yet a route is not a path. A path is made by walking. It includes detours, unexpected encounters, and even dead ends that later turn out to be intellectually productive. Algorithmic routes tend to eliminate these elements in the name of efficiency. In doing so, they risk eliminating learning itself.
If personalization removes all friction, what remains to be learned?
Flexible trajectories without losing direction
The second question posed in the seminar is more practical, and therefore more difficult: how can generative AI support the design of flexible and customizable learning pathways without compromising curricular coherence?
I have spent several weeks designing e-learning courses for a graduate program in which I teach, and this tension is far from theoretical. It is the everyday dilemma of anyone designing an online course: how far can flexibility go before the structure collapses?
One useful approach I have developed with my students involves what I call a “verb map.” Instead of starting from the formal language of a competency statement in the syllabus, we analyze what the student actually does minute by minute within the learning design. They read. They click. They select. They drag. They compare. They write.
If more than sixty percent of those verbs are passive, the learning environment becomes little more than a digital magazine with buttons, no matter how much adaptive AI is added to the interface.
At this point it is helpful to recall what Thomas Fawns describes as an “entangled pedagogy.” From this perspective, pedagogy, technology, and institutional conditions cannot be separated into independent components; they operate as an interdependent system of practices and constraints (Fawns, 2022). When personalization is treated as a purely technical problem to be solved by better algorithms, we overlook this entanglement. What we end up optimizing is not learning but platform behavior.
Generative AI can support flexible learning trajectories in several concrete ways without replacing pedagogical judgment. It can generate alternative versions of an activity for students with different entry levels. It can provide preliminary feedback on written work when no tutor is immediately available. It can help build question banks with more meaningful distractors. It can even assist instructors in designing prompts that cannot easily be solved by another prompt.
In all these cases, however, the core design decisions remain human. Generative AI is an instrument, not the conductor of the orchestra. As in jazz, what determines quality is not the technical sophistication of the instrument but the listening capacity of the musician.
Curricular coherence is not lost because multiple pathways exist. It is lost when the destination is unclear. If the competency is well defined—and that remains a human task—students may arrive through different routes. What cannot change is the learning destination itself.
The data we need and the data we do not
Another question raised by the seminar concerns the data required to personalize learning ethically.
Here we should be honest. Most of the data collected by educational platforms is not useful for personalizing learning. It is useful for personalizing user experience, which is something entirely different.
Knowing how many minutes a student spent on a page does not tell me whether they understood the material. Knowing how many times they clicked on a resource does not indicate whether they processed it critically. Knowing that a quiz was completed in three minutes tells me that the student was fast, not that they learned something.
The data that genuinely supports meaningful personalization is data that captures student thinking rather than platform behavior. An argumentative essay. A question formulated in a discussion forum. A decision made in a branching scenario. A persistent misconception that reveals an alternative conceptual model.
These forms of evidence are qualitative, messy, and difficult to process algorithmically. That is precisely why most platforms prefer clicks.
Before collecting any data, the essential pedagogical decision is to determine what is actually worth knowing about a student in order to support their learning. Anything beyond that risks becoming surveillance disguised as analytics.
In an institutional policy document on generative AI that I developed recently, I proposed a compass for responsible use based on four principles: equity, accountability, transparency, and security. The same compass applies here. If a data point does not contribute to equity in learning or cannot be explained transparently to the student, it probably should not be collected.
The risk no one puts on the slide
Discussions of algorithmic personalization often highlight benefits while ignoring risks. At least three deserve explicit attention.
The first is the risk of cognitive bubbles. If an algorithm consistently presents students with material aligned only with their current level and interests, they may never encounter ideas that challenge their assumptions. Yet deep learning frequently emerges precisely from those moments of intellectual discomfort. Bjork and Bjork describe such situations as desirable difficulties (Bjork & Bjork, 2011; 2019).
The second risk is metacognitive dependency. When a platform determines what the student should study, when, and for how long, students may never develop the capacity to regulate their own learning. The ability to ask oneself “What do I need to learn next, and how should I approach it?” may be one of the most important competencies in contemporary education. A system that makes those decisions automatically may save effort while simultaneously undermining autonomy.
The third risk is political and economic. Adaptive learning platforms are not neutral tools. They belong to companies with business models. The data they collect has market value. Algorithmic decisions inevitably reflect the biases embedded in training data and design assumptions.
Personalizing learning through such systems without asking who benefits economically from that personalization is comparable to celebrating that water is free without asking who owns the pipes.
Measuring what matters
How should the impact of AI-mediated personalization be evaluated?
The honest answer is that we still do not know how to do this well.
Student satisfaction, short-term performance on quizzes, or time spent in a platform are convenient indicators, but they are weak proxies for learning. What matters is whether students can transfer knowledge to new contexts, develop independent judgment, and formulate questions they were previously unable to articulate.
Any evaluation of AI-mediated personalization should therefore compare what students can do with AI assistance and what they can do without it. The goal is not to prohibit AI but to make visible the difference between assisted thinking and autonomous thinking. If that distance becomes too large, personalization may be replacing rather than supporting learning.
In the verb-mapping exercise I use with my e-learning students, one question functions as a simple diagnostic: would this activity work equally well if an AI completed it instead of the student?
If the answer is yes, the activity is probably not evaluating learning. It is evaluating compliance. And generative AI is the most compliant student imaginable.
What I will bring to the seminar
The seminar will not be an optimistic presentation about the future of AI-driven personalization. Instead, I plan to bring uncomfortable questions, concrete examples from instructional design where generative AI helped and where it interfered, and a number of mistakes I have made along the way.
The conviction that has gradually taken shape over these months is simple: personalizing learning is not a technical problem that can be solved by better algorithms. It is a pedagogical problem that begins with better questions.
Generative AI can be an extraordinary tool for supporting personalization, but only when the person using it has a clear understanding of what learning actually means. That definition will never come from a platform. It emerges from the interaction between the teacher who designs, the student who asks questions, and the conversation between them that no algorithm can anticipate or replace.
References
Bjork, R. A., & Bjork, E. L. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher et al. (Eds.), Psychology and the real world (pp. 56–64). Worth Publishers.
Bjork, R. A., & Bjork, E. L. (2019). Forgetting as the friend of learning: Implications for teaching and self-regulated learning. Advances in Physiology Education, 43(2), 164–167. https://doi.org/10.1152/advan.00001.2019
Fawns, T. (2022). An entangled pedagogy: Looking beyond the pedagogy–technology dichotomy. Postdigital Science and Education, 4, 711–728. https://doi.org/10.1007/s42438-022-00302-7
Perkins, M., Furze, L., Roe, J., & MacVaugh, J. (2024). The Artificial Intelligence Assessment Scale (AIAS): A framework for ethical integration of generative AI in educational assessment. Journal of University Teaching and Learning Practice, 21(6). https://doi.org/10.53761/q3azde36
Pratschke, J. (2024). Generative AI and education. Springer.
