The case for AI personalisation in education is a genuinely strong one: right content, right level, right pace, calibrated to the individual student rather than the median of a class. Anyone who has spent time in a classroom watching students mark time or misbehave while they wait for peers to catch up, or fall through gaps in provision designed for someone else, needs no convincing of the difference this could make. An education sector that has taken an increasingly evidence-based approach in recent decades cannot in good conscience ignore tools that have a demonstrable impact on learning.
The enthusiasm is understandable, but it is obscuring a weakness in the approach that I fear will erode intellectual resilience for a generation of students. AI personalisation is exceptionally good at calibrating content and practice to the individual. It is less well suited to the parts of learning that require other people.
Content acquisition is not enough
Content acquisition is the intake and processing of material: facts, concepts, frameworks, arguments, and methods. Intellectual formation is what the person becomes in the process: how they learn to reason, hold a position, engage with disagreement, and recover from being visibly wrong. In short, acquisition allows you to pass exams, whereas formation allows you to hold your own in the real world.
Content acquisition is optimised by removing friction: adaptive pacing, individualised feedback, material calibrated precisely to where a student is. AI does this exceptionally well, because that is what the tools are designed to do. Intellectual formation, on the other hand, requires friction. Not that of content pitched at the wrong level, but that which comes from peers: positions that push back, conversations the student does not control, and the experience of being wrong in front of people. Remove that and you remove the very thing that is doing the formative work.
The obvious counter-argument is that AI can simulate disagreement, and increasingly well. Anyone who has used Claude, ChatGPT, Gemini or any other mainstream chatbot as a thinking partner knows they can do this well with the right instructions and prompts. It would be dishonest to suggest otherwise, particularly given that I have written at length about the value of AI as a professional sparring partner.
The question is not whether AI can challenge your thinking. It can, and will do so more fluently and patiently than most people you will argue with. But that fluency and patience are precisely the problem. The friction that forms young people is neither fluent nor patient. It is uncomfortable, resistant, and can’t simply be shut down or redirected with a couple of words.
The stakes are the point
Even if an AI conversation effectively dismantles your argument and leads to a genuine change in understanding, there is no social record of having been persuaded. Nobody saw you change your mind, and you can tell the world (and in many cases yourself) that the outcome was what you thought all along. In most cases you can even delete the chatbot’s memory of the exchange to ensure it stays that way. Advocates point out that this makes AI a very low-risk environment for students to explore and test ideas. That is certainly true, but a social learning opportunity is left on the table.
When a peer disagrees with you in a classroom, the consequences are real and visible. You are being called out in front of people who will see you again tomorrow, whose opinion of you matters, who have their own investment in the question. Being wrong in that situation leaves you with two options, neither of them comfortable: concede and change your mind in public, or continue to defend a weak position and live with the consequences. Learning to navigate these moments is part of growing up; intellectual and social resilience are not built in comfortable conversations.
The problem runs deeper than the absence of social stakes. Even in the moment of the exchange itself, students know that an AI counter-position can be disagreed with, ignored, or abandoned by closing the tab, all without consequence. Every mainstream chatbot reinforces this: each one carries a reminder that it can make mistakes and should not be relied upon uncritically. It is hard to feel the full force of a challenge from something that opens with a disclaimer about its own reliability.
A peer who pushes back is genuinely invested in the debate, whether because they think they are right, think you are wrong (the two do not always go hand in hand) or relish playing devil’s advocate. That investment is what gives real disagreement its weight. Being challenged by someone who genuinely means it is how young people discover what they actually think.
Students do not just feel more pressure when they argue in front of peers. They think differently. The presence of an audience that matters, people whose opinions they care about and whose judgement they will carry with them, changes not just how they present an argument but how they construct one. AI can simulate the exchange. It cannot simulate the room.
Complement, not substitute
Deployed deliberately, AI personalisation can achieve a level of differentiated provision that a busy teacher with 24 or more students in front of them could only dream of otherwise. Every student working at the right level and pace, without the adult in the room having to manage dozens of individual tracks simultaneously. That frees classroom time for the collaborative thinking that produces formation, but only if schools are explicit about what these tools are for and what they are not.
What needs protecting, actively and by design, are the experiences in which formation happens. Shared texts that give a class common intellectual ground (you cannot have a genuine argument about a book nobody else has read because the system decided something slightly different was a more precise fit for their profile). Structured debate that puts positions under real peer pressure, not the patient and ultimately deniable pushback of a chatbot. Socratic seminar in which students learn to hold a line, adapt under questioning, and change their minds in public. These are far more than nice-to-haves that are secondary to getting through the curriculum content. They are what produces the person who will go out into the world, not the anonymous candidate number who will sit the exam.
I don’t for a second believe that school or group leaders would deliberately deprioritise these things. But it has long been a truism in education that the measurable quietly displaces the important all too easily if the two are not perfectly aligned. Content acquisition is still the thing most commonly and concretely measured by most of the world’s education systems. For decades, in schools with limited resources and pressure to achieve top exam results above all else, it has often crowded out intellectual formation. If we aren’t careful, AI personalised learning will push the focus even further in that direction.
Treating faster individual progress as a substitute for shared intellectual experience, rather than a complement to it, risks leaving students behind as people in the name of taking them forward as productive units. Nobody is setting out to produce this outcome, but the commoditisation of education is nothing new and good intentions that don’t fit with commercial ambition rarely survive contact with the market.
The measure of whether we get this right will not be how much content students can acquire, or how efficiently their individual progress can be maximised. It will be whether they can hold a position when a room full of peers disagrees with them, change their minds in public without losing themselves, and recover from being visibly wrong in front of people who matter. AI can give them the ammunition they need to engage with the arguments worth having. But a sandbox in which they know they can avoid losing altogether should not be the space where they learn to do so.