Terminology
Because “AI” is the mainstream umbrella term used in media and everyday speech, I use it for readability, without implying that AI and LLMs are synonymous. My critique does not refer to “all AI”, nor narrow tools like calculators or spellcheckers, but LLM-based systems that generate fluent, school-valid artifacts at scale.
I recently attended the Mathematics Biennial in Gothenburg (as audience). The program was noticeably AI-heavy: AI for teaching, AI for assessment, AI for “personalized learning”, AI as the obvious direction forward. What struck me wasn’t that people were curious — curiosity is healthy. What struck me was how little critical discussion there was about what teaching is, what learning is, and what happens when schooling is reorganized around systems that can produce the appearance of competence at industrial scale.
I teach math, science, programming, and technology in high school. I’m also an abolitionist when it comes to LLMs in learning. Not because “technology is bad”, and not because teachers should deny reality. Because education is not (or shouldn’t be) a production line. And LLMs are a production-line technology.
This essay is meant as a tool: a language teachers can use in staff rooms, planning meetings, union discussions, and policy debates. I’ll ground it both in critical pedagogy and in research and policy that already raises serious alarms. I’ll end where these discussions often stop too early: energy, water, and militarism/surveillance.
Re-Grounding: What Teaching Is (Freire, hooks) and What “Market School” Does to Learning
Freire: The “Banking Model” and the Politics of Treating Students as Containers
Paulo Freire’s critique of the banking model of education is not a historical artifact. It is a live diagnostic tool for contemporary education.
In the banking model:
- knowledge is treated as a thing that can be deposited,
- students are treated as passive recipients,
- education becomes a mechanism for adaptation and control rather than liberation and agency (Freire, 1970).
Even if we believe we’ve moved beyond this (we haven’t), much of schooling still behaves like banking: content coverage, standardized outputs, measurable objectives, time pressure, and compliance. When education is already organized like banking, LLMs look like a miracle — because LLMs are exceptionally good at moving symbols (words, explanations, solutions) around.
But Freire’s point isn’t simply that banking is “boring” (although it is). It is that banking is oppressive, because it trains people to accept knowledge as something delivered by authority, rather than something made in dialogue with the world and each other.
An LLM is, in a strange way, the perfect banking teacher: endlessly patient, always available, and always ready to “deposit” fluent text. Because it speaks in smooth, confident language, it can easily be mistaken for understanding. That is not a side effect — it is the mechanism (how it is trained).
bell hooks: Engaged Pedagogy, Education as the Practice of Freedom
bell hooks offers a complementary re-grounding. Engaged pedagogy insists that education is relational, embodied, and ethically charged. It is not merely the transfer of information but a practice of freedom — where students and teachers are both implicated in the work of becoming (hooks, 1994).
This matters because the selling point of LLMs is often framed as “support”, “help”, “access”, or “equity”. hooks pushes us to ask: support for what kind of classroom? Help toward what vision of the learner?
A classroom built around LLMs easily slides toward a classroom built around:
- individualized consumption,
- private interaction with a machine,
- reduced peer dialogue,
- reduced dependence on human relationships,
- and a subtle redefinition of learning as “getting an acceptable output”.
If education is a practice of freedom, then “freedom” is not the freedom to outsource thinking. It is the freedom to develop the capacity to think, speak, and act with others in the world.
Illich: Institutions That Confuse Process With Product
Ivan Illich’s Deschooling Society helps name something teachers often feel but struggle to articulate: institutions can teach people to confuse the process of learning with institutionalized outputs (Illich, 1971).
LLMs intensify exactly that confusion. They make it easier than ever to hand in something that looks like learning — without the learner having done the work that produces understanding.
That creates a crisis not only of assessment, but of meaning. If students learn that “success” is submitting fluent output, then schooling becomes training in compliance with output standards. That is market logic, not education.
The Structural Claim: Why AI (in My Opinion) Is Being Pushed Even When It Undermines Learning
We need to say the quiet part out loud:
AI is being pushed in education not because it improves learning, but because it improves institutional throughput, measurement, and control — goals that often conflict with the slow, demanding cognitive work that learning requires.
This is the mismatch:
- Schools say they value thinking, curiosity, understanding, independence.
- Systems reward efficiency, legibility, standardized outputs, and risk management.
LLMs promise to make students “produce more”, teachers “grade faster”, and administrators “see more data”. Even if those promises were true, they clearly point in the wrong direction. Education is not fundamentally a throughput problem.
When research shows reduced engagement or “cognitive debt”, see (Kosmyna et al., 2025), those findings are not anomalies. They reveal the structural conflict between learning and productivity.
A Practical Teacher Tool: the Learning Integrity Test (or LIT)
To keep this discussion concrete, here is a quick tool you can use for any assignment in math, science, or programming. I’ve somewhat loosely based it on the work of (Grinschgl et al., 2021) and (Risko & Gilbert, 2016).
The Learning Integrity Test
For any task (and this could be applied to any subject, not just math, science and programming), ask:
-
What is the learning goal: capacity or product?
- Capacity: reasoning, explaining, proving, modeling, debugging, interpreting, arguing, revising.
- Product: a formatted report, a translation, a cleaned-up paragraph, a cover letter.
-
What is the thinking step that is the point of the task? Name the step precisely. Examples:
- Math: choosing a strategy; representing a situation (diagram, equation, function, model); noticing invariants/constraints; proving/justifying (why this method works, not just that it works); checking reasonableness and units; debugging errors; generalizing and connecting ideas across contexts.
- Science: critically reviewing information and argumentation; distinguishing scientific from non-scientific claims; making evidence-based judgements in issues around environment, climate, sustainable development, and health; planning and carrying out systematic investigations (research question, risk assessment, execution, evaluation, communication).
- Programming: decomposing problems; designing data representations; writing tests (and using failure as information); debugging systematically (reproducing bugs, isolating causes, inspecting state); reading and using documentation; reasoning about correctness, complexity, security, and edge cases; explaining tradeoffs and design choices.
-
Does the LLM do that step? If yes, then the tool is not supporting learning. It is substituting for it.
A simple slogan that helps in discussion:
If the tool does the thinking step, the student doesn’t get the learning.
”But Calculators Exist”
Yes. The point is not “no tools”. The point is: different tools belong to different pedagogical moments.
A calculator can offload arithmetic once students have foundational numeracy, and we can still assess conceptual understanding. But an LLM’s core capability is to generate the very artifacts that stand in for understanding across subjects: explanations, solutions, essays, lab reports, code. That means it can offload the very steps students still need to build.
What the Actual Research Suggests (and Why We Should Adopt a Precautionary Stance)
Cognitive Offloading, False Mastery, Cognitive Debt
LLMs are uniquely effective at producing fluency: coherent explanations, confident reasoning, and polished structure. In schooling, fluency is dangerous because we often mistake it for understanding. This is not a minor issue — it creates false mastery: students feel competent because they can submit (or read) a coherent response, but they have not built the internal knowledge structures required to reconstruct, transfer, or defend that knowledge under even mild pressure (an oral follow-up, a new context, a delayed test, or a slightly changed problem).
In other words, LLMs don’t just “help students write”. They can replace the cognitive work that produces learning, leaving behind a product that looks like knowledge while the underlying capacity is missing. The result is a predictable pattern teachers already recognize: students become better at delivering outputs and worse at demonstrating ownership of ideas, methods, and reasoning.
A central empirical anchor here is Kosmyna et al.’s study on LLM-assisted essay writing (Kosmyna et al., 2025). Across repeated sessions, they report differences consistent with lower engagement during LLM-assisted writing (including differences in measured brain connectivity), more homogenous writing within the LLM condition, lower self-reported ownership, and difficulty accurately quoting one’s own text. Yes: it’s a preprint, and yes: it focuses on essays. But that’s exactly the point — essay writing is one of the clearest cases where “the output” is supposed to be the trace of thinking. If the tool produces the trace without the thinking, we shouldn’t be surprised when recall, ownership, and independent reconstruction weaken.
So the issue isn’t (just) that “students might cheat”. The issue is: LLMs industrialize cognitive offloading in the very domains (writing, explaining, reasoning) that school relies on to cultivate durable understanding.
The “Shadows” Beyond Hype: a Structured Map of Harms
Al‑Zahrani’s paper matters here for a different reason. It offers a shared vocabulary for harms that teachers often experience as “a messy pile of problems”. It doesn’t claim to settle the causal question “AI harms learning” with one decisive experiment. Instead, it synthesizes concerns across the literature into a structured set of risk areas — loss of human connection, privacy and security concerns, algorithmic bias, lack of transparency, reduced critical thinking/creativity, unequal access, reliability, and the consequences of AI-generated content (Al-Zahrani, 2024).
That matters in practice because schools are currently being sold a narrow story: “AI = efficiency + personalization”, while the main classroom-level worry gets reduced to “cheating”. This framework helps us widen the lens to what educators actually end up managing: relationship erosion, surveillance incentives, inequity, epistemic unreliability, and deskilling pressures. In other words, it helps move the conversation from “How do we catch students?” to “What kind of educational system are we building — and who does it serve?”
UNESCO: GenAI (LMMs) Is Not the Solution, and Human Agency Must Be Protected
UNESCO’s guidance is unusually direct. GenAI should not be framed as the solution to education’s fundamental challenges; the determining factor is human capacity and collective action, not technology (UNESCO, 2023). UNESCO explicitly warns about risks to human agency, equity, privacy, and cognitive development, recommends validation of pedagogical appropriateness, and highlights environmental costs.
This matters because it undercuts the “inevitability” narrative. If we don’t have robust evidence of long-term educational benefit — and we do have meaningful evidence of risk — then pushing LLMs into learning is not innovation. It is uncontrolled experimentation on children.
Why “Personalization” Is a Trap
“Personalized learning” is one of the most effective rhetorical engines of AI adoption because it sounds like care. Who could be against meeting students where they are?
But the word personalization hides a fork in the road. One path is pedagogical and relational. The other is managerial and extractive. They are not the same thing, and treating them as equivalent is how the trap works.
Human Personalization (Ethical, Contextual, Accountable)
A teacher personalizes learning through relationship and responsibility:
- knowing the student (including their history, strengths, constraints, and needs),
- adjusting demands in dialogue (negotiated expectations, not imposed pathways),
- using professional judgment that can be explained and contested,
- being accountable to the student, the class, and shared norms of fairness.
This kind of personalization is slow, situated, and human. It does not require turning the student into a data profile.
Algorithmic Personalization (Profiling, Optimization, Sorting)
Platform “personalization” typically means something else:
- profiling students through continuous data collection,
- opaque and unaccountable decision-making (“the system recommends…”),
- optimization for what is measurable (clicks, time-on-task, completion, compliance),
- pressure toward sorting/tracking—because once you quantify learners, you inevitably rank them, route them, and manage them.
Even when marketed as “support”, algorithmic personalization often functions as behavior management at scale. It translates education into inputs and outputs that administrators can monitor, vendors can sell, and systems can standardize.
So this isn’t merely a privacy issue. It is a political shift in the meaning of education:
- from an interpersonal, relational practice of teaching and learning,
- to an individualized system of prediction, nudging, and control.
A staff-room line that is calm but refuses the euphemism:
Personalization that requires surveillance isn’t personalization — it’s profiling.
And a second line, if you want it sharper:
When someone sells “personalized learning”, ask: personalized for the student’s growth — or personalized for the institution’s management?
Subject-Specific: What This Looks Like in Math, Science, and Programming
Math
If a student types “Solve this” and receives a step-by-step solution, the student has received a performance artifact, not the capacity to:
- choose methods,
- check assumptions,
- detect errors,
- generalize,
- or explain why a method works.
Even worse: the student may learn that math is primarily about formatting steps into a correct-looking sequence. That is the opposite of mathematical literacy.
Abolition stance here is not “ban all tech”. It is: don’t embed a tool whose main strength is producing solution-shaped text into the learning process.
Science
In science, the core “thinking work” is not just producing correct explanations. It is being able to critically review information, distinguish scientific from non-scientific claims, and use a scientific perspective to reason and take a stance on societal issues — especially environment, climate, sustainable development, and health.
And when we do practical work, the report is meant to be a trace of a systematic investigation: the question you formulated, how you planned and risk-assessed, what you did, how you evaluated limitations/uncertainty, and how you communicated conclusions responsibly.
If an LLM writes the report (or the argumentation), the student doesn’t learn scientific literacy or scientific working methods; they learn to generate science-shaped text — exactly the kind of surface plausibility the subject is supposed to train them to critique.
Programming
In programming, LLMs can produce “working” code quickly. That is exactly why they can undermine learning:
- decomposition,
- debugging,
- testing,
- reading documentation,
- building mental models of execution.
If students learn programming with constant code generation, they may become dependent on a tool for the very skill they are supposed to develop: making programs from problems.
You can still teach LLMs as an object of study: the political economy of software automation, bias, verification, and security. But putting it inside the learning loop for novices is like replacing practice with performance.
Abolition
Here is my proposed (non-negotiable) boundary:
LLMs should not be integrated into student learning activities (writing-to-learn, problem solving, explanation, inquiry, and practice).
Students will still encounter them, of course. But we do not have to institutionalize them, normalize them, or build pedagogy around them.
We can teach:
- about LLMs,
- about verification,
- about propaganda and deepfakes,
- about automation and labor,
- about surveillance and data extraction,
- about energy and water costs, without treating the LLM as a legitimate substitute for thinking in school tasks.
The Wider Ethical Horizon: Energy, Water, and Militarism/Surveillance
Even if someone believes the pedagogical risks are manageable, the ethical concerns remain — and teachers should not be pressured into ignoring them.
Energy and Climate
It’s tempting to treat AI’s footprint as a side issue, something separate from pedagogy. But the climate crisis did not pause when commercial LLMs became mainstream. If anything, LLM adoption is being built into a future of higher baseline electricity demand and expanded data-center infrastructure.
LLMs are not “just software”. They are industrial infrastructure:
- large-scale data centers,
- high-power chips,
- continuous cooling,
- and (crucially) a business model that aims for billions of daily interactions.
The International Energy Agency has documented the rapid rise in attention to AI-related electricity demand and the growing significance of data centers in energy systems planning (International Energy Agency, 2024). What matters for educators is not only the number for a single query, but the trajectory: schools are being asked to normalize a tool whose business success depends on massive scale.
MIT Technology Review’s reporting makes this concrete in ways teachers can actually use. The per-query impact may look small, but the industry’s direction is toward ubiquitous integration, reasoning models, agents, and always-on inference — driving larger and more opaque energy demand. They also stress a key political point: the public often cannot see or verify the energy accounting because major model providers operate as “black boxes” and disclose too little (O’Donnell & Crownhart, 2025).
So when schools adopt LLMs at scale, we’re not only choosing a classroom workflow. We are participating in the normalization of an energy-hungry infrastructure in a time when education already has a civic responsibility to address climate and sustainability as real, urgent conditions of our students’ lives — not as an optional theme.
Water Scarcity
Data centers do not just consume electricity. They consume water — often fresh, potable water — to keep hardware from overheating. Water is not an abstract resource; it is a local and regional constraint. And water scarcity is already a political reality in many places, including in Europe’s drought-affected regions.
Research has proposed methods to estimate the water footprint of AI models and argues this is a systematically overlooked cost (Li et al., 2023). Reporting has also emphasized that AI infrastructure can consume enormous volumes of water for cooling operations (O’Donnell & Crownhart, 2025). UNESCO explicitly calls for attention to environmental costs and sustainable targets for AI providers (UNESCO, 2023).
For schools, the ethical issue is not whether an individual teacher used a chatbot once. It is whether we accept — and help normalize — institutional dependence on infrastructures that externalize environmental costs and demand continuous expansion.
A sharper way to put it:
When we institutionalize LLMs as everyday schooling infrastructure, we are turning “learning” into a driver of data-center growth — during a climate and water crisis that schools are supposed to help students understand and respond to.
That is not neutral. It is curriculum in the deepest sense. It teaches what society is willing to spend energy and water on, and what we treat as non-negotiable.
Militarism and Surveillance
AI is not only an “education technology”. It is a security technology. It travels across domains: from ad targeting, to policing, to border enforcement, to military decision support.
Human Rights Watch has documented the Israel Defense Forces (IDF) use of surveillance technologies, AI, and other digital tools in Gaza hostilities, warning these tools increase civilian harm, rely on faulty data and inexact approximations, and raise grave legal and ethical concerns (Human Rights Watch, 2024). Whatever one’s broader political framing, the structural point is unavoidable: AI systems are already part of life-and-death decision infrastructures. This is not an abstract “sci-fi” scenario; it is contemporary reality.
So we should ask, as educators, as unions and communities:
- Who owns the systems we normalize in school?
- What other markets do they serve?
- What does it mean to train students into dependency on commercial AI ecosystems entangled with surveillance and militarism?
Local Models vs Commercial Models
There is a real ethical difference between:
- using a local model on hardware you control (clear boundaries, no data leakage, less lock-in), and
- integrating commercial AI products tied to data extraction and broader security markets.
But the pedagogical problem remains. If the tool substitutes student thinking, it still undermines learning. Local control reduces some harms; it does not transform an LLM into a learning-friendly tool.
What All This Means (and Why This Is Urgent)
Put together, the pattern is hard to ignore. LLMs are being pushed into education not because they are good for learning, but because they are good for institutional throughput, measurement, and control — and because they fit a market-school logic where education is judged by outputs, efficiency, and compliance. Meanwhile, the costs fall where they always fall: on students’ cognitive development (through substitution and false mastery), on teachers’ professional autonomy (through deskilling and platform dependence), and on the public (through energy, water, and political-ethical externalities).
If we accept LLMs as normal learning tools, we quietly change what school is. We turn writing, reasoning, investigating, and explaining into tasks students can outsource — then we pretend the resulting output represents learning. That is not modernization. It is the abolition of education from the inside: keeping the building and the grades while hollowing out the human work of becoming capable, critical, and free.
This is why I don’t think abolition is extreme. It is proportional. It is a defense of the core purpose of school: that students actually do the thinking, together, in the world, with teachers — not that they submit the correct-looking result.
Freire warned us about education that trains adaptation; hooks insisted on education as a practice of freedom; Illich warned about institutions that mistake products for learning.
LLMs intensify the worst tendencies of institutional schooling — because they are perfectly aligned with what systems reward: output, efficiency, compliance, and legibility. That is why they are being pushed. Not because they make students wiser, freer, or more capable.
If we want students who can think, we cannot build schools that teach them not to.