Education and the art of adaptability

Most debates about assistive technology, AI in the classroom, or accommodations for disabled students share the same flaw: they ask the wrong question. We ask whether the tool works. We should ask what functional need it actually covers, and what it creates in return. Asking the wrong question gives wrong answers.

The tool does not restore what disability has taken. The wheelchair does not restore walking. AI does not restore fluency of thought. Accommodation does not restore equal conditions. Each covers a specific functional need, partially, and creates new problems in return. Identifying both is the condition for evaluating a tool honestly.

The functional need

When someone cannot walk, the problem is not, in general, the absence of walking as such. The problem is the impossibility of moving around: going from one place to another, meeting someone, crossing a space. Walking is the way most bodies solve this need. It is not the need itself.

The distinction between the functional need and the usual biological means of satisfying it changes everything. The tool does not have to reproduce biology. It has to cover the need. And in covering that need, it creates new ones.

The hammer example

Before getting to the more serious cases: the hammer. The functional need is "drive a nail". The hammer does it remarkably well. But it creates a new problem, known to everyone: hitting your fingers. The benefit is real, the trade-off too. The tool does not eliminate the risk; it shifts it.

This is true of all tools. With assistive technology, the trade-offs are simply more structural, and more consequential.

The wheelchair example

A wheelchair does not restore walking. It allows movement: on flat ground, in accessible buildings, outside of stairs without ramps. The need is covered, but under conditions.

First, a dependence on infrastructure: the wheelchair requires ramps, elevators, clear pavements, wide enough doors. A space not designed for it makes it useless. Accessibility is not a property of the user. It is a property of the building. Then the relationships change: one can no longer "walk alongside someone" in the literal sense. The height, the pace, the perceived obstacles are no longer the same. The tool changes the way one occupies shared space.

The keyboard example

The keyboard does not restore writing. It allows the production of signs corresponding to conventional symbols: letters, words, typography. That is not the same thing.

Handwriting allows tracing, hesitating, changing the shape of a letter, adjusting pressure. The keyboard produces standardized, identical, reproducible characters. The "a" I type is the same as yours. That is not a flaw. It is the nature of the tool. But what it covers is not what we had before: some things that writing permitted disappear.

In return, the keyboard requires infrastructure: a screen, a processing unit, an operating system. It does not function on its own. The autonomous tool has become dependent on an entire system.

The case of deep brain stimulation

I have been living since January 2025 with a deep brain stimulation implant to compensate for a motor disability (I write about it at length on this blog). The implant is not a cure, it is a tool. It partially covers the functional need and creates new constraints: a battery to be surgically replaced, regular adjustment sessions with a neurologist, medical follow-up that cannot be stopped. The infrastructure the tool requires is here an entire specialized medical system, which cannot be set aside or replaced easily.

And AI?

Some people already use generative AI as an assistive tool: for someone with motor or cognitive fatigue, it produces on their behalf; for someone with dyslexia, it rephrases; for someone with a hearing impairment, it transcribes.

The keyboard changed the form of writing, by standardizing characters. AI can change the content: if it formulates on my behalf, the result is my thought, but not necessarily my words. On top of this, the dependency is opaque: cloud infrastructure, proprietary model, transmitted data, possible unavailability. The tool is no longer in my hands.

It is in education that the question of the functional need is most urgent, and most poorly framed.

In education

For a standard assistive tool, the functional need is relatively clear: move around, write, communicate. In education, it is double.

There is the immediate need: produce a text, solve a problem, submit an assignment. And there is the developmental need: learn to produce a text, learn to solve problems. Most tools cover the first. Education is mostly about the second.

There is the tool that enables doing, and the tool that does in your place. The wheelchair brings the student to the examination room. It does not sit the exam in their place. The screen reader gives access to the text. It does not understand it in their place. Speech-to-text allows writing by speaking. The student is still the one formulating. AI can write the text in their place: the immediate need is covered, the developmental need bypassed.

This is not a matter of cheating. It is a matter of misidentified functional need. If the educational objective is "express a complex idea in writing" and the student's difficulty is motor, AI legitimately compensates the mechanics without touching what we want to develop. If the objective is "learn to build an argument", AI that builds the argument in their place compensates nothing. It replaces the target.

Disability, paradoxically, forces this clarity. When a dyslexic student asks for accommodation, the institution is forced to ask: is the objective "read this text" or "be able to read"? For some students at some stages, both. For others, only the first. AI asks the same question of everyone, all the time, without anyone having to declare anything. It is a massification of the problem, without the clarity that the accommodation request imposed.

The disabled student, and why I do not use "in a disability situation"

This is precisely what the term itself reveals. "In a disability situation" assumes that the functional need has been sufficiently studied to be entirely attributed to the situation, and that the right configuration of tools and accommodations can neutralize it. It is a promise that reality almost never keeps.

The term has a history: it was born of the social model of disability, which shifts responsibility from the body to the environment. There is something right in that: a deaf person in a world entirely in sign language is not disabled. But the institutional formulation goes further, and too far.

A profoundly deaf student does not recover spontaneous exchange in a lecture hall because they are given a note-taker. What they receive is a selective transcription, filtered by someone else, after the fact. That is not the same as hearing. A student with severe chronic fatigue does not recover energy because they are given extra time. Extended time stretches the duration of the exam, not their physical endurance. A student with severe anxiety disorder is not in the same conditions as others because they sit in a separate room.

In these cases, the gap narrows, it does not disappear. And in some configurations, even the narrowing is marginal.

The term also creates a practical problem in institutions: it authorizes considering that the obligation has been fulfilled once an accommodation has been granted. We provided the tool. We planned for the ramp. We printed in large font. The rest is the student's problem. This logic confuses accommodation with compensation, and compensation with equality.

I therefore use "disabled student" without quotation marks and without hedging. Not out of provocation, but because the term says what is: a person whose body or cognitive functioning creates real difficulties, which do not disappear because we changed the name.

This student carries, in an exam, a burden that others do not. There is the shared work: understand, mobilize, produce. And there is the invisible work: manage the accommodation tool, compensate for what it does not cover, monitor what the body is or is not doing, anticipate failure, manage the fatigue of the device on top of the fatigue of the exercise. This extra work is never counted. Sometimes, it is not even recognized as existing.

The tool that adapts

All the tools seen so far are stable. The hammer always does the same thing. The wheelchair does not decide where to go. The keyboard produces the same "a" regardless of the question. You can therefore clearly identify what they cover and what they do not do.

AI is different. It adapts to what you give it. It does not respond the same way twice for different requests, different users, different contexts. Can we then speak of an intelligent tool?

The word is a trap. AI does not understand. It produces responses that look like understanding. That is not the same thing, but in the relationship with the tool, the difference can disappear. A student who asks AI to explain a concept receives an explanation adapted to their formulation, their apparent level, their previous questions. The tool seems to understand where they are. That is not what is happening, but it is what you feel when using it.

This changes the relationship to the tool. Previous tools made their limits visible: the wheelchair stops at the bottom of the stairs, the keyboard cannot spell for you. With AI, the limit is shifting and opaque. It depends on how you solicit it, what you delegate to it, the habit you form with it. What it covers is not fixed: it expands as you rely on it more.

That is where the notion of adaptability becomes problematic in education. An adaptive tool in Vygotsky's sense (the scaffold) adapts to support, then withdraws to let competence take hold. AI adapts to respond, but it does not withdraw. The more you use it, the more fluid it becomes, the more you tend to use it again. That is not a design flaw: it is what a system designed to respond does. But a system designed to respond does not develop the capacity to formulate without it.

The challenge

Vygotsky described the zone of proximal development as the space between what one does alone and what one does with support. A good pedagogical tool, like a good teacher, works in that space, then progressively withdraws. The tools that compensate for disability do not withdraw, and that is legitimate: no one expects a wheelchair user to eventually manage without it. But a pedagogical tool that does not withdraw becomes a dependency, not learning.

AI does not only call into question the tools used in education. It calls into question the model.

Education is built on a hypothesis: we learn by doing. We learn to write by writing. We learn to reason by grappling with problems. Difficulty is not an obstacle to learning; it is its mechanism. What we overcome, we integrate. If AI takes charge of production, that effort is bypassed.

School survived the calculator: we simply stopped teaching square roots by hand. It survived the internet: we stopped requiring the memorization of encyclopedias. It will survive AI. But each time, the question is: what do we keep, what do we delegate, and why? With the calculator and the internet, the answer was relatively clear: we delegate mechanical tasks, we keep understanding. With AI, the boundary is less clear. AI does not replace a mechanical task. It replaces tasks we thought were cognitive.

What remains is not nothing: judgment, the ability to evaluate an answer, detect an error, decide. But these capacities develop by doing what AI does in our place. We cannot delegate production and keep the capacities that production develops, not without profoundly reorganizing what we teach, how, and in what order.

Covering the functional need, seeing what the tool does, seeing what it creates in return: that is necessary. It is not sufficient.

Disability, education, and tools raise, at bottom, the same question: what do we do with what cannot be compensated?

What the three domains share is the temptation to avoid this question. The term "in a disability situation" declares the problem solved by renaming it. The institution that provides an accommodation considers the obligation fulfilled. The school that adopts AI without reorganizing its objectives delegates production while believing it maintains learning. In all three cases, a tool was provided. Nobody checked whether it covered the need.

What gets measured, in all these cases, is what comes out: the student submits their work, the person moves around, the assignment is produced. What nobody looks at: the invisible extra work carried by the disabled student, the developmental need that was bypassed, the capacity to think without assistance that does not build up. Absences do not appear in assessments.

There is a question that tools cannot pose: what do we want the person to be able to do without the device? What are we willing to delegate, and what do we refuse? When AI formulates on our behalf, do we feel relieved or dispossessed? When the disabled student manages their accommodation on top of their work, do we see them?

These are not technical questions. They are questions about what we judge necessary to go through oneself.