S

Introduction

My first serious attempt to think clearly about the relationship between education and digital technology came in my doctoral work. Looking back now, I can see that thesis as an early effort to understand something that has only since become more urgent. At the time I was trying to make sense of the rise of information and communication technology in English secondary schools. What I found, though I did not yet have the language for all of it, was not simply a story about computers arriving in classrooms. It was a story about what schools are for, who gets to decide that, and what happens when a whole educational system begins to speak in the language of technology, efficiency, investment and economic competition. This new book begins from that earlier attempt. It returns to the same questions in a new moment, one shaped not simply by digital technology in general but by artificial intelligence in particular.

My doctoral thesis focused on England, especially the period from the mid 1990s to 2010, when the New Labour governments made digital technology a central feature of school reform. I was interested in a puzzle that many people in education will recognise. Huge sums of money were spent on technology for schools. There was enthusiasm from ministers, advisers, businesses and many educational organisations. Computers, networks, software and platforms became increasingly visible in classrooms and corridors. Yet the promised transformation of teaching and learning never arrived in the straightforward way the policy language seemed to predict. Schools changed, certainly. Teachers worked differently, leaders managed differently, pupils encountered a more digital environment. But the deep educational gains that had been promised remained elusive, patchy or difficult to identify. I wanted to know why.

To answer that question, I had to step back from the technology itself and look at the larger setting in which it had been introduced. One of the most important things I came to see was that the arrival of digital technology in schools was part of a much broader reform movement. It was tied to the remaking of public services in England. Schools were being asked to think of themselves in new ways. They were being drawn into a world of targets, audit, measurable outcomes, performance indicators, competition, market language and managerial routines. In that wider setting, technology was never just a tool. It was also a symbol. It came to stand for modernity, for progress, for the future, for adaptation, for survival in a changing world.

This mattered because the debate about schools in England had never simply been about efficiency. One of the foundations of my thesis was an older question about educational purpose. I became interested in the long argument within English education between what might be called vocational and non-vocational visions of schooling. On one side sits the view that schools exist mainly to prepare young people for work and economic life. On the other sits the view that schools exist to initiate young people into forms of thought, culture, judgment, self-understanding and citizenship that cannot be reduced to economic usefulness. Michael Oakeshott’s contrast between the school as market and the school as monastery gave me a helpful shorthand for this division. I did not use it because it settled the argument, but because it exposed it. It helped me see that digital policy was not entering neutral ground. It was entering a field already shaped by deep disagreements about what education is for.

This was the first move of the thesis. It was to show that any discussion of technology in schools has to begin with a discussion of educational purpose. Once I saw that, I could also see something else. Much of the literature on information and communication technology in education, whether optimistic or critical, was already accepting a certain background picture of the world. It assumed that we live in an “information society” or a “knowledge economy”, that global competition is the defining fact of our age, and that schools must therefore remake themselves accordingly. The terms varied, but the assumption underneath them was remarkably stable. Technology was taken to be central to this new reality, and schools were expected to respond.

My thesis did not simply reject those terms. Instead, I treated them as part of the discourse I wanted to understand. I read policy documents, reports and commentary from the period and noticed how often certain words appeared. “Globalisation.” “Knowledge economy.” “Information society.” “Transformation.” “Innovation.” “Skills.” “Modernisation.” Even when these concepts were vague or sociologically suspect, they still had power. They organised the language of reform. They told schools what sort of world they were supposedly living in and what sort of institutions they were now expected to become.

To understand how that language worked, I drew on Stephen Ball’s work on policy sociology. Ball helped me see that policy is not just a set of official statements handed down from above. Policy is a process. It is made through documents, speeches, reports, guidance, agencies, advisers, school leaders, inspectors and practitioners. It moves through institutions. It is interpreted, enacted, adapted and sometimes resisted. This was enormously helpful because it meant I did not have to confine my attention to formal government policy papers alone. I could look instead at the wider family of texts that practitioners actually used and that, in practice, shaped what schools did.

I also drew on critical discourse analysis, especially the work of Norman Fairclough, Ruth Wodak, Teun van Dijk and others. That approach gave me a way of reading policy language as something more than information. It allowed me to ask how texts build authority, how they define problems, how they allocate responsibility, whose voices they amplify, whose voices they marginalise, and how they make a particular picture of reality seem natural and unquestionable. This was important to me because the documents I was reading often sounded factual and neutral, while in fact they were carrying a great deal of ideology.

The most original part of the thesis, I think, came from bringing Daniel Bell into the discussion. Bell’s name is usually associated with post-industrial society, but what mattered to me was not only his account of economic change. It was his claim that modern society cannot be understood through the economic realm alone. He argued that social life is organised through different realms, which I translated for my purposes into the techno-economic, the political and the cultural. Each has its own principles. The techno-economic is driven by efficiency, productivity and instrumental reason. The political concerns legitimacy, participation, public consent, authority and justice. The cultural concerns meaning, identity, self-realisation, imagination and the life of the spirit. Bell’s point was that these realms do not collapse neatly into one another. Change in one does not automatically dictate change in the others.

This turned out to be a powerful lens for reading educational policy. It allowed me to ask whether the discourse introducing digital technology into schools was really educational at all, or whether it was in fact an economic discourse wearing educational clothes. It also let me see what was absent. Critical discourse analysis is very good at showing what is present in a text and how it operates. Bell helped me ask what had gone missing. Where were the political values of democratic participation, shared voice, public argument and legitimacy? Where were the cultural values of meaning, personal growth, care, imagination and the cultivation of the whole person? If they appeared at all, were they being treated as ends in themselves, or were they being folded back into the service of economic goals?

To move from theory to evidence, I carried out two case studies. I asked a group of practitioners and local actors involved in the implementation of digital policy in North London secondary schools to nominate the documents they found most influential. This mattered because I wanted to understand policy not just as written in Whitehall, but as encountered on the ground by people trying to make it work. In the first case study, which covered the period from the mid 1990s to 2005, the documents reflected the early and expansionary phase of digital reform. They were full of confidence. Technology was described as transformative. Schools were told that they needed to adapt to a new age. The vocabulary of the knowledge economy and the information society was everywhere. Teachers were often described in deficit terms, as people who needed to be trained up so that technology’s promise could finally be realised.

Pupils were frequently portrayed as disengaged by traditional schooling but potentially reactivated through technology. Business appeared as a natural partner. Government appeared as the visionary sponsor. The school was increasingly imagined as an institution that should align itself with the demands of a global economy.

What struck me in this first case study was how completely the discourse was dominated by the techno-economic realm. Even when documents appeared to speak about learning, inclusion, participation or even enjoyment, those values were typically drawn back into the language of efficiency, skills and economic competitiveness. Bell’s framework allowed me to see that very clearly. Teaching was valued because it could support technological adaptation. Inclusion was valued because it could widen the skilled labour pool. Participation was valued because it could feed the modern economy. The cultural and political goods of education were present mainly in translated form. They had been annexed.

The second case study covered the period from 2006 to 2010. Here the mood shifted. Technology was no longer arriving. It had arrived. The issue was now how to manage it. The documents selected in this period were narrower, more technical, more managerial and more explicit about audit, procurement, self-review, infrastructure, safeguarding and value for money. This in itself was revealing. Once technology was established in schools, the discourse moved from promise to maintenance, from evangelism to administration. The same techno-economic logic remained dominant, but it now focused more on financial accountability and risk.

This second phase taught me something important. The logic that had originally justified the spread of technology did not disappear once technology was in place. It intensified. Schools were now expected not only to use digital systems, but to justify their spending on them, manage their risks, document their impact and align them with a language of measurable return. Safeguarding emerged as a major theme. So did procurement. In one sense this was understandable. Once schools are full of networks, devices and platforms, questions of safety and cost become unavoidable. Yet the way these issues were framed reinforced my broader conclusion. The discourse still worked overwhelmingly in the register of the techno-economic realm. Even risk was managed through the language of efficiency and systems control.

By the end of the thesis I felt I had reached a more settled answer to the original question. The problem had never just been that schools did not use the technology well enough, or that teachers needed more training, or that hardware had once been weak and then improved. The deeper issue was that the entire policy discourse had been shaped by one governing logic. It assumed that educational value could be understood mainly in terms of economic need, system efficiency and technological modernisation. It used digital technology as both symbol and instrument of that logic. It turned schools toward vocational ends and treated those ends as if they exhausted the meaning of education.

I also concluded that this mattered a great deal. It mattered because policy language does not just describe schools. It helps produce them. It changes the kinds of practices that are valued, the kinds of people who are listened to, the kinds of evidence that count, the kinds of futures that seem possible. When digital technology entered schools through this discourse, it helped strengthen a particular answer to the question of what schools are for. That answer placed work, competition, skills, performance and economic adaptation at the centre. It left less room for slower educational goods, for non-vocational study, for personal formation, for democratic voice and for the cultivation of judgment, imagination and belonging.

That doctoral work was my first attempt to think seriously about all this. It was not my last word. If anything, I now see it as an opening move. It gave me a method, a language and a set of questions that I have carried with me ever since. It taught me to look beyond the surface claims made for technology and ask what picture of education is being smuggled in with them. It taught me to read policy as culture and as power. It taught me that schools are never just technical systems. They are moral and political institutions, and they are cultural worlds.

Why begin a new book on schools, education and artificial intelligence with this story? Because artificial intelligence arrives bearing many of the same promises that surrounded digital technology in the earlier period, only now in a heightened form. It comes wrapped in claims about transformation, productivity, personalisation, efficiency, global competition and the future. It arrives through governments, consultancies, technology firms, regulators, philanthropies and educational platforms. It quickly acquires symbolic force. It is presented as unavoidable. It is treated as the next stage of modernity. Much of the public language around it already assumes that education must adapt. We have heard that cadence before.

At the same time, artificial intelligence also changes the terrain. It touches assessment, authorship, feedback, planning, tutoring, administration, curriculum design, language learning, knowledge production and the very idea of human judgement. It enters not only classrooms but the wider ecology of educational life, from university admissions to school inspection, from parental expectation to labour market anxiety. It raises questions that my thesis could only gesture towards. What happens when a technology that can imitate language, produce text, answer questions and simulate assistance enters institutions already shaped by a techno-economic discourse? What happens to educational purpose when systems built to optimise, classify and automate are increasingly seen as educational companions? What happens to the political and cultural realms then?

This book takes up those questions. It begins with schools and returns often to schools, because schools remain one of the most revealing places to watch our wider hopes and fears about technology become concrete. Yet it also moves beyond the world of the original thesis. It looks at universities, policy systems, teacher work, student life and the emotional weather of institutions under pressure. It asks again what education is for. It asks how artificial intelligence is being folded into existing reform languages. It asks what has changed since the first wave of digital technology policy, and what has not. Above all, it asks how we might think differently if we refuse to let the techno-economic realm speak as if it were the whole of educational reality.

So this preface is really a return to the beginning. My doctoral thesis was my first serious effort to grapple with the relationship between education and digital technology. It was written in the shadow of one technological moment. This book is written in the opening years of another. The earlier work taught me that the most important questions are not, in the end, only about what the technology can do. They are about what kind of schools we want, what kind of persons we hope education might help to form, and what sort of common life we are trying to sustain. Those questions are older than computers and larger than artificial intelligence. They are the questions that remain when the novelty fades. They are the questions with which this book begins.

I also want to say something more personal here, because I have never been, and am not now, an enemy of technology. Quite the opposite. I have often been an enthusiast for it. I have seen what good tools can do in schools. I have watched technology widen access, reduce unnecessary burdens, open doors for pupils who had been shut out of parts of learning, and help teachers do things that would once have taken far more time and effort. I have seen the excitement that a well used digital tool can bring to a classroom. I have seen children find voice through media, through design, through coding, through forms of communication that matter deeply to them. I have seen schools in difficult circumstances use technology with intelligence and imagination. So my concern has never been with the mere fact of technological development. It has been with the way schools are asked to understand it, and the way technology is too often introduced without a sufficiently serious understanding of the institution it is entering.

That, I think, remains one of the central lessons of my earlier work. Schools are not empty spaces into which new devices and systems can simply be placed. They are dense, historical, moral, emotional institutions. They have rhythms, relationships, hierarchies, rituals, anxieties, memories, hopes. They carry the accumulated weight of earlier reforms, older arguments about authority and freedom, long histories of class, aspiration, gender, care, competition, exclusion and belonging. Technology enters all that. It does not float above it. It is bent by it and it bends it in return. If we do not understand the school, we cannot understand what the technology is doing there. If we do not understand the technology, we cannot understand what the school is being asked to become.

That is one reason why returning to these questions in 2026 feels so important to me. In some respects, we are in a very different place from the one in which I wrote the thesis. Artificial intelligence now sits where information and communication technology once sat, as the object around which hopes, fears, markets, policy claims and fantasies of transformation gather. The scale is different. The speed is different. The tone is different. We have lived through a global pandemic that pushed schools, families and governments into forms of emergency digital dependence that would once have been hard to imagine. We are much more conscious of safeguarding, surveillance, data extraction, mental health, teacher workload and platform power than we were in the early years of the century. The tools themselves are more sophisticated, more intimate, more embedded in everyday life. In that sense, 2026 really is a new landscape.

And yet, I find myself repeatedly struck by how familiar so much of it sounds. The rhetoric of inevitability is still with us. The claim that schools must change because the world has changed is still with us. The suggestion that a new technology has arrived which finally makes possible the transformation long promised by earlier technologies is still with us. The language of efficiency, personalisation, productivity, skills, competitiveness and future readiness is still with us. The tendency to treat education as if it were lagging behind some wider technological reality and needed urgently to catch up is still with us. The habit of speaking as though the deeper purposes of schooling have already been settled, and all that remains is delivery, is still with us. I find that both fascinating and confusing. The scenery changes, the software changes, the slogans refresh themselves, and still some of the underlying grammar seems stubbornly intact.

Perhaps that is why this subject has never really let me go. I remain fascinated by the way moments of apparent rupture so often carry old assumptions inside them. A system can look transformed while reproducing its deepest habits. A new machine can arrive and immediately be made to serve an old idea of what counts. A radically new discourse can present itself as unprecedented while quietly repeating a much older narrowing of educational purpose. At other times the opposite happens. Something introduced in the name of efficiency produces side effects that reopen forgotten questions about care, judgement, authority, creativity or human presence. That double movement, things being genuinely different and deeply familiar at once, is part of what this book is trying to understand.

So I come to this new investigation with a mixture of curiosity, concern and, I hope, some humility. My earlier work taught me that it is easy to be dazzled by the new and easy also to dismiss it too quickly. Schools deserve better than either reflex. They deserve patient, serious, historically informed thought. They deserve an account of technology that is neither worshipful nor fearful. They deserve, above all, a language for asking not only what new tools can do, but what kind of educational life they make easier to imagine and what kind they make harder to sustain. That is the spirit in which this book continues.

Chapter 1: What Do We Know?

For most of my professional life I worked inside schools, first as a teacher and then as a head, in both the UK and international contexts. Across those decades I watched successive waves of educational technology arrive with the usual promises of transformation. Personal computers, interactive whiteboards, tablets and learning platforms were each introduced with confident predictions that schooling would soon be radically improved. In practice, the effects were usually more modest. Teachers adapted the tools to existing routines, the institutional grammar of schooling remained remarkably stable, and the deeper purposes of education continued to revolve around knowledge, judgement, memory, attention and imagination. Technology changed aspects of delivery, but rarely the underlying aims.

Generative artificial intelligence feels different. Earlier technologies mainly altered access to information. These systems intervene directly in the production of language, ideas and reasoning. A student no longer merely searches for information or retrieves a source. The machine can generate explanations, essays, arguments and solutions on demand. In doing so it enters the very space of intellectual activity that education has traditionally sought to cultivate within the learner. Public discussion of this development has oscillated between enthusiasm and alarm. Some claim that AI will democratise knowledge, personalise learning and unlock new forms of creativity. Others stress plagiarism, misinformation and technological dependency. What has been less visible in these debates is the growing body of empirical research examining how students actually use generative AI and what effects this use may have on learning.

In education, where the stakes are high, belief and anecdote are not enough. We need evidence. This chapter gathers a set of studies that, taken individually, address quite different questions about AI and learning, but when read together reveal a pattern that is both intriguing and disturbing. Some examine writing, others problem solving, memory, authorship, imagination or creativity. Read in isolation, each offers a partial view. Read together, they suggest that generative AI often improves the surface quality of student output while weakening or bypassing the deeper cognitive processes through which learning occurs. Students produce more polished prose, more quickly completed tasks and apparently more creative work. Yet beneath these outputs lie weaker knowledge retention, reduced engagement in reasoning, narrowing of imaginative diversity and a blurred sense of authorship.

This matters because education is not simply the production of correct answers or well formed sentences. Schools exist to cultivate the habits of mind through which individuals learn to think independently, remember deeply, reason carefully and imagine beyond the obvious. These capacities are formed through effortful engagement with problems, through the slow construction of understanding, and through the discipline of articulating one’s own thoughts in language. Generative AI offers a technology capable of bypassing much of this effort. The temptation to rely on it is therefore entirely understandable. Yet the emerging research suggests that when such reliance becomes routine, the cognitive work that produces learning may be quietly displaced.

One indication of this appears in Neil Selwyn, Fareed Kaviani, Yolande Strengers and colleagues’ study of how students imagine the future of schooling, “We’re Already Experts in School, Right? Supporting Students’ Construction of Future School Scenarios” (Selwyn et al. 2025). The study sits within a field of policy and academic interest in education futures, a field often dominated by governments, corporations and think tanks producing sweeping visions of technological transformation. Selwyn and his co-authors begin from the simple but telling observation that these visions rarely include the perspectives of the young people who actually inhabit schools every day. Their project therefore asked what kinds of future schools students themselves imagine, and what this reveals about how schooling is understood and reproduced.

Working with students aged roughly twelve to fourteen in two Australian schools, the researchers ran participatory workshops in which small groups were invited to build future school scenarios set several decades ahead. Students were encouraged to imagine utopian, dystopian and mixed futures, and to think about climate change, energy transition and artificial intelligence. They produced posters, websites, stories and presentations, sometimes using digital tools and generative AI to help with images and ideas. The resulting scenarios were lively and inventive at the level of detail. They included global corporate school chains, futuristic campuses in deserts or towers, AI-assisted meal systems, advanced transport infrastructures, new logos, uniforms and branding. Students enjoyed the exercise and many noted that they had never before been invited to think about the future of schooling in this way.

Yet the central finding of the study is that these imagined futures were surprisingly conservative where it mattered most. Beneath the speculative surface, the basic structure of school life remained largely intact. Students still got up in the morning, travelled to a school site, attended timetabled classes, completed homework and followed familiar curricular routines. Even in futuristic settings containing highly advanced technologies, the institutional grammar of schooling remained recognisable. Selwyn and his colleagues interpret this as evidence of the obduracy of school as a social form. Schooling is so deeply normalised that even when students are invited to imagine distant futures, they tend not to reinvent it but to extend it.

What is particularly interesting for my purposes is the role played by digital tools and generative AI in this imaginative process. The researchers found that when students relied more heavily on AI systems or online templates, their scenarios often became less idiosyncratic and more clichéd. They reproduced familiar images of “future schools”, sleek campuses, ubiquitous digital assistants, fully automated systems, rather than genuinely novel educational forms. Instead of opening conceptual space, the technology often narrowed it. In this respect the study already points toward one of the key themes of this chapter. Generative AI may appear to enhance imagination while in fact drawing users toward statistically typical cultural representations. It can function not as a liberator of thought but as a stabiliser of dominant imaginaries.

There is another important feature in Selwyn’s study. Although students rarely challenged the basic institutional form of schooling, they did repeatedly emphasise its social and relational dimensions. Their futures often foregrounded care, safety, emotional support, flexibility and better relationships between students and teachers. Technologies were imagined less as replacements for teachers than as supports for school life. This matters because it shows that, from the student point of view, the future of education is not fundamentally about technological disruption but about improving the lived texture of school experience. Even here, however, the study reinforces the broader caution. AI did not generate radical educational alternatives. It largely amplified already circulating narratives and familiar desires. I saw something similar years ago when working on the Labour government’s Building Schools for the Future initiative in London. Students were often imaginative and enthusiastic, but their visions tended to preserve the same underlying architecture of schooling while modifying its surface conditions. Selwyn’s study suggests that generative AI does not break this conservatism. If anything, it intensifies it.

A parallel shift appears in Fiona Draxler and colleagues’ work on what they call the “AI ghostwriter effect” (Draxler et al. 2023). Their paper, “The AI Ghostwriter Effect: Users Do Not Perceive Ownership of AI-Generated Text but Also Do Not Credit the AI,” investigates what happens psychologically when people use generative systems to produce text. The key issue is not simply whether AI can write, but how users understand authorship, responsibility and ownership when the language that appears is partly or wholly produced by a machine.

The starting point is straightforward. In ordinary cases, the categories of writer, author and owner of a text tend to coincide. If I write a paragraph, I usually experience it as mine and present it as mine. Human ghostwriting already complicates that arrangement, as when a politician or celebrity publishes text largely written by someone else. Generative AI introduces a further twist. Here the ghostwriter is not another person but an automated system producing language from statistical patterns in training data. Draxler and her colleagues asked how people experience authorship under these conditions.

Across two empirical studies, participants were asked to produce personalised texts with the aid of a language generation system and then respond to questions about ownership, authorship and responsibility. The results were striking. Participants generally did not feel that the generated text was fully their own. Their sense of psychological ownership was weak. Yet they also tended not to credit the AI system as the author when presenting the text publicly. Internally, they felt the text was not really theirs. Externally, they still behaved as though it was. This mismatch is the ghostwriter effect.

The effect persisted even when the text was personalised to reflect the user and even when participants had greater control over the generation process. More control increased ownership slightly, but did not remove the asymmetry. Draxler and her colleagues also found that people were more willing to attribute ownership to a human ghostwriter than to an AI one. Human collaboration fits more easily into existing cultural models of authorship. Machine collaboration produces a deeper ambiguity. The result is a kind of authorship vacuum. The human user no longer experiences full authorship, but the machine cannot meaningfully occupy the author role either. The text appears, is published, and circulates, but the experience of having written it has thinned out.

This is not a minor issue of etiquette or citation. It reaches into the phenomenology of thinking. Writing has historically been one of the main ways in which individuals come to experience themselves as thinking agents. The effort of composing sentences, choosing words, revising arguments and trying to say something precisely is not merely expressive. It is cognitive. One often understands what one thinks by trying to write it. Draxler’s study suggests that users sense the displacement of this process. They do not feel full ownership because the labour that would normally generate ownership has been transferred elsewhere. The user becomes less the originator of language and more its selector, editor or curator.

Seen alongside Selwyn’s findings, a pattern begins to emerge. In both imagination and writing, generative AI appears to offer enhancement while simultaneously reducing the cognitive engagement through which originality and ownership are formed. The technology supplies possible futures before reflective speculation has deepened, and supplies completed sentences before thought has fully taken shape.

This broad pattern is made more concrete in Matthew H. C. Mak and Łukasz Walasek’s “Style, Sentiment, and Quality of Undergraduate Writing in the AI Era” (Mak and Walasek 2025). This is one of the most revealing studies yet published because of its scale and its ability to track writing across time. The authors analysed 4,820 authentic empirical reports written by around 2,000 psychology undergraduates between 2016 and 2025, producing a corpus of roughly seventeen million words. This allowed them to compare student writing before and after the public release of ChatGPT in late 2022 and to examine not just isolated examples but longitudinal shifts in style.

Their findings are clear. After the widespread availability of generative AI, student writing became more formal, lexically sophisticated and structurally dense. Measures of lexical diversity, nominalisation and vocabulary sophistication all rose. Certain words strongly associated with AI-generated prose, such as “delve” and “intricate”, increased sharply in frequency. Some of these markers then declined slightly in 2025, suggesting that students may have become more aware of AI’s stylistic fingerprints and perhaps more careful about masking them. More interesting still, the tone of student writing shifted.

Sentiment analysis showed that assignments became more positive and upbeat even when the underlying content of the reports did not warrant such a tone. This aligns with the well-known positivity bias of many large language models, which are trained to produce polite, constructive and affirmative responses.

The crucial point, however, is that these stylistic changes did not correspond to improvements in academic quality. Grades and instructor feedback did not show statistically significant gains. Student prose became more polished, more formal and more confident in tone, but not more intellectually impressive. The rhetorical surface improved while the underlying quality remained largely unchanged. Mak and Walasek strengthened this interpretation with an exploratory experiment in which GPT models rewrote student reports from the pre-ChatGPT era. The resulting texts displayed almost exactly the same stylistic profile as the post-2022 corpus. This strongly suggests that generative AI is indeed driving the observed shift.

This is an important result because it shows that generative AI has partially decoupled style from understanding. Historically, better academic writing usually followed deeper conceptual grasp. As students came to understand a field more precisely, their language became more nuanced and disciplined. AI changes this relation. Students can now produce text that looks sophisticated without having done the intellectual work that would normally produce such sophistication. The linguistic performance has been upgraded, but the cognitive substance has not.

The resonance with Draxler’s ghostwriter effect is obvious. In both cases, the labour of composition is partially outsourced. The text bears marks of sophistication, but the process through which sophisticated writing normally emerges has been altered. The student becomes a curator of machine-shaped prose. The positivity shift identified by Mak and Walasek also matters. Academic writing often requires scepticism, qualification and critical distance. If AI systematically nudges language toward positivity and smooth affirmation, then the tone of academic discourse itself may begin to drift away from the habits of mind it is supposed to embody. One does not need to be melodramatic about this. The point is subtler and more serious. If the system keeps supplying rhetorically tidy language before the writer has grappled with uncertainty, then writing ceases to function as a medium in which thought is formed and becomes instead a surface onto which polished forms are projected.

The issue becomes even more acute when we turn from writing to memory. André Barcaui’s randomised controlled trial, “ChatGPT as a Cognitive Crutch: Evidence from a Randomized Controlled Trial on Knowledge Retention,” offers one of the clearest demonstrations that AI assistance can improve immediate task performance while weakening long-term learning (Barcaui 2025). This study was designed to test a widely discussed but rarely well-tested hypothesis, namely that generative AI functions as a form of cognitive offloading. Humans routinely offload cognitive work onto tools, notebooks, calculators, search engines, calendars. Offloading is not automatically bad. But cognitive psychology has long shown that when external tools replace the effort required to process information, memory encoding may become shallower.

Barcaui randomly assigned undergraduates to two groups. One group could use ChatGPT freely while studying. The other used more traditional methods without AI assistance. Crucially, the study included a delayed retention test administered forty-five days later, allowing the researchers to measure not just immediate performance but long-term memory. The result was sobering. Students who used ChatGPT performed adequately in the moment, but on the delayed test they did significantly worse. The AI-assisted group answered 57.5 per cent of questions correctly, compared with 68.5 per cent in the non-AI group. The difference was statistically significant and educationally meaningful.

The metaphor of the cognitive crutch is apt. A crutch enables movement when one cannot walk unaided, but prolonged reliance can weaken the muscles required for ordinary movement. Generative AI seems able to do something similar in educational settings. It provides immediate explanations, summaries and structured answers, reducing the need for students to struggle with material. Yet it is often precisely this struggle that produces durable learning. Retrieval, elaboration and repeated processing are not regrettable obstacles on the way to learning. They are much of what learning is. AI makes the task easier in the short term, but may simultaneously reduce the cognitive work through which knowledge becomes stable in memory.

This sharpens a distinction that is central to the whole chapter, the difference between performance and learning. Performance concerns how well one completes a task now. Learning concerns durable changes in knowledge or skill that remain when the task is over. Educational technologies can improve performance without improving learning, and sometimes by undermining it. Barcaui’s experiment strongly suggests that generative AI can do exactly this. It gives students the appearance of competence while weakening what they later retain.

The connection to Mak and Walasek is immediate. If students use AI to produce explanations or refine written work, the output may look impressive even when the conceptual content has not been internalised. Barcaui’s study helps explain why. The polished output is compatible with weak encoding. Students may arrive at correct answers or tidy summaries without performing the mental operations that lodge knowledge in long-term memory. The danger here is not just forgetting facts. Memory matters because higher-order thinking depends on retrieval and recombination. Reasoning, creativity and judgement require something to work on. If less is being stored, less is available later for independent thought.

Barcaui’s findings also imply a metacognitive risk. Students using AI-generated explanations may develop an illusion of understanding. Because the answer appears coherent and complete, they may feel that they grasp the material when they have in fact only recognised it. The delayed retention test is important precisely because it cuts through that illusion. It measures what remains once the immediate fluency of the AI encounter has faded. What remains, in this study, is less.

A broader synthesis of these concerns appears in Chuang Zhai, Setyo Wibowo and Lijuan Li’s systematic review, “The Effects of Over-Reliance on AI Dialogue Systems on Students’ Cognitive Abilities” (Zhai, Wibowo, and Li 2024). A systematic review is useful here because it gathers multiple studies rather than depending on a single experimental design. Zhai and his colleagues surveyed the emerging literature on conversational AI systems in education, including chatbots, intelligent tutoring agents and dialogue-based assistants, and asked what happens when learners begin to rely on these systems heavily rather than using them as occasional supports.

Their conclusion is cautious but serious. Over-reliance on AI dialogue systems may weaken critical thinking, analytical reasoning and problem solving. Several mechanisms recur across the literature. One is cognitive dependency, the habitual deferral of intellectual tasks to external systems rather than attempting them independently. Conversational AI makes this very easy. The interface is frictionless. Students ask questions in natural language and receive fluent, fully formed answers. The interaction can feel like collaboration, but it may also discourage the exploratory work that would normally occur in the struggle with a difficult problem.

Another recurring mechanism is automation bias, the tendency to trust the outputs of automated systems even when those outputs may be incomplete or wrong. Because AI systems present information in coherent, confident prose, students may accept explanations without verifying them against sources or reconstructing the reasoning for themselves. Over time this can diminish habits of scepticism and evidence-based evaluation. Closely related is a shift in epistemic authority. Instead of treating claims as propositions to be tested and argued about, students may begin to treat the AI system as an authority whose outputs are presumptively reliable. The medium encourages this. The system does not merely list information. It narrates answers, and narrative fluency can masquerade as truth.

Zhai and his co-authors also point to fragmentation of attention. Conversational systems encourage rapid cycles of question and answer. This can be efficient, but it can also short-circuit sustained engagement. Instead of inhabiting a difficult text, dataset or problem long enough for conceptual structures to emerge, students may repeatedly consult the AI for partial solutions. The result is speed without depth. Metacognition also suffers. If the system keeps providing apparently complete explanations, students have fewer opportunities to encounter and diagnose their own confusion. They may feel that they understand because they can repeat the AI’s answer, even when they could not reconstruct the argument themselves.

Zhai and his colleagues are careful not to claim that AI dialogue systems are inherently harmful. Many studies report benefits, especially in access and engagement. But the review’s central point is that the consequences depend on whether the system remains supplementary or becomes substitutive. Once it begins to replace independent reasoning rather than support it, the educational risks multiply. This is one of the most useful distinctions in the literature. The question is not whether AI is present, but what kind of cognitive role it is allowed to play.

By this point in the chapter a fairly coherent picture is emerging. Selwyn and his colleagues showed that AI can narrow imaginative exploration by reproducing familiar templates. Draxler showed that AI-generated writing weakens psychological ownership without producing a stable alternative model of authorship. Mak and Walasek demonstrated that student prose becomes more polished without becoming substantively better. Barcaui showed that AI assistance can reduce long-term knowledge retention. Zhai and his colleagues synthesise these patterns into a broader warning about the erosion of cognitive capacities under conditions of over-reliance. The common thread is not that AI always makes students worse in visible performance. It is that it often redistributes cognitive labour away from the learner and toward the system.

That redistribution matters for creativity as well. Aniket Doshi and Oliver Hauser’s study, “Generative AI Enhances Individual Creativity but Reduces Collective Diversity,” has been widely cited because its title appears to state a paradox in perfectly balanced form (Doshi and Hauser 2024). In one sense the study does exactly that. In another sense, I think the title understates how troubling the findings are. Participants in the study were asked to write short stories. Some did so without AI assistance, some could consult AI once for inspiration, and others could consult it multiple times. Human evaluators rated the resulting stories for creativity, novelty and usefulness, while computational analysis measured the semantic similarity of the outputs.

The results were mixed in a revealing way. On average, AI-assisted stories received higher creativity ratings, particularly for participants who had initially been weaker creative writers. In that sense the technology appeared to raise the floor. But when the researchers looked at the corpus as a whole, AI-assisted stories were markedly more similar to each other. Their narrative structures, themes and vocabulary clustered more tightly. Individual outputs appeared more impressive, yet collective diversity declined.

Doshi and Hauser interpret this as a shift in idea space exploration. Without AI, participants wander more unpredictably through a wide conceptual landscape. With AI, they are guided toward regions that the model treats as promising, regions shaped by patterns in its training data. The system helps weaker participants produce more competent stories, but it also nudges many people toward similar solutions. This is a plausible interpretation, but I think the findings warrant a sharper one. What looks like enhanced creativity may in part be better conformity to recognisable templates of creativity. Human judges often reward coherence, vividness and familiar signs of good storytelling. Generative AI is excellent at producing precisely these features because it has absorbed large volumes of existing narratives. The system may therefore be improving performance within established evaluative conventions rather than fostering deeper originality.

In creative life genuine innovation often comes from outliers, works that initially look awkward, odd or even bad by prevailing standards. Beckett, Dylan, Warhol or Miles Davis do not emerge from the centre of statistical normality. If AI helps people produce work that is more recognisably polished and “creative” while simultaneously narrowing variation across a population, then one has to ask whether what is being measured as creativity is partly optimised conformity. The reduction in collective diversity is not a peripheral side effect. It may be the central clue to what is really happening.

This interpretation resonates strongly with the earlier studies. Selwyn’s participants reproduced familiar future-school tropes. Mak and Walasek found a narrowing toward a shared academic register. Zhai and his colleagues describe over-reliance as reducing independent exploration. Doshi and Hauser show the same force at work in creative production. Generative AI acts as a centripetal system. It can make individual users feel more capable, because it supplies coherent language, plausible ideas and well-formed structures. But at the level of a population, it compresses variation. It draws people toward statistically favoured regions of conceptual space. From the perspective of education, where diversity of thought matters, this is not a small issue. A system that raises the floor while lowering the ceiling, or at least narrowing the field, may be less an engine of creativity than a machine for elegant convergence.

The final piece I’m wanting to look at in this emerging pattern comes from Chunming Hou and colleagues’ study, “Understanding Students’ Use of Generative AI in Problem Solving: Passive, Reflective, and Thoughtless Use” (Hou et al. 2024). What makes this study especially useful is that it does not simply measure outcomes. It examines how students interact with AI while solving problems. The authors distinguish several modes of use. Reflective use occurs when students critically assess AI outputs and integrate them into their own reasoning. Cautious use involves checking AI responses against other sources. Collaborative use treats the system as a brainstorming partner. Thoughtless use, by contrast, occurs when students accept or reproduce AI-generated outputs with minimal scrutiny.

This last category is the most troubling. Hou and his colleagues found that thoughtless use was not rare. A significant proportion of students displayed patterns of passive adoption, entering a problem into the system and then taking over the explanation or solution with little attempt to reconstruct the reasoning for themselves. The key variable was not simply use of AI as such, but the degree of cognitive engagement. Students with stronger critical thinking dispositions were more likely to use AI reflectively or collaboratively. Students with higher trust in AI were more likely to use it thoughtlessly.

That relationship between trust and passivity is central. Fluency breeds confidence. When a chatbot responds in smooth, well-structured language, students may treat that fluency as a sign of reliability. The AI becomes a default epistemic authority. The learner need not be lazy or dishonest for this to happen. The design of the interaction itself encourages it. Thoughtless use therefore reveals a mechanism underlying the effects identified in the other studies. If students regularly accept ready-made reasoning chains instead of constructing their own, then false mastery becomes almost inevitable. They appear to perform well because they can reproduce polished answers, but the underlying conceptual work has not occurred.

The OECD has warned of a “mirage of false mastery where high-quality output conceals underlying weakness in human skill.” Hou’s study helps explain how that mirage is produced. The problem is not just that AI gives answers. It is that dialogue-based systems make it so easy to bypass the intermediate stages of thinking. Instead of struggling with a problem, generating possibilities, testing them and revising one’s judgement, the student is presented with a coherent route through the problem space. In fields where the value of the task lies primarily in the reasoning process, mathematics, engineering, philosophy, computer science, even creative work, this is a profound shift. The student may arrive at the right answer while missing the intellectual formation the exercise was supposed to produce.

By now the pattern across the literature is difficult to ignore. Generative AI often improves output while weakening the processes that generate that output. It makes prose smoother while loosening authorship. It enhances immediate performance while weakening memory. It appears to support creativity while narrowing collective diversity. It offers solutions while reducing engagement in reasoning. At each point the same deeper issue returns, the displacement of cognitive labour from the learner to the system.

It is important to be precise here. None of this proves that generative AI inevitably damages education. Reflective use is possible. Supplementary use is possible. There are situations in which AI may genuinely support learning, particularly where it increases access, offers feedback or helps students get unstuck. The strongest studies in this literature do not justify a simple moral panic. But they do justify scepticism toward the prevailing optimism. The dominant public story has too often been that AI will enhance learning because it improves performance and convenience. The evidence reviewed here suggests that this is at best incomplete and at worst seriously misleading. The danger is not merely cheating, nor even factual error. It is that students may increasingly receive the external signs of thought, polished language, competent answers, coherent explanations, without undergoing the internal processes through which thought, judgement and understanding are formed.

That is why I think the educational stakes are so high. Schooling is not simply a delivery system for information, and academic work is not just output production. Education is a formative practice. It exists to shape ways of attending, questioning, remembering, imagining and reasoning. These capacities develop through friction, difficulty, delay, uncertainty and repetition. They are not always efficient. Indeed, one of the defining features of real education is that it often looks inefficient from the standpoint of sheer productivity. The time spent drafting a clumsy paragraph, wrestling with a theorem, puzzling over a historical source, or trying and failing to frame an argument is not wasted time. It is the medium in which understanding grows.

Generative AI threatens to alter this ecology at a very deep level. It does so not because it is malevolent, but because it is designed to minimise friction. It produces immediate fluency where educational formation often requires struggle. It supplies completion where learning often requires incompletion. It provides language before reflection has matured, and solutions before curiosity has deepened into inquiry. That is why the metaphor of haunting seems apt. AI enters the spaces of cognition and expression not always with a visible rupture, but as a subtle occupying presence. It begins by assisting. It may end by possessing, not in the dramatic sense of replacing teachers or abolishing schools, but in the quieter sense of colonising the practices through which students become thinking persons.

The studies examined in this chapter do not settle the matter definitively. The field is expanding rapidly and the literature will change. But taken together they already indicate something more serious than a passing technical challenge. They suggest that generative AI may be reshaping the conditions under which students learn to think. If that is right, then the central educational question is not how quickly we can integrate AI into schools, but what kinds of human capacities we are willing to let weaken in the process. The issue is not whether students can produce better looking work with AI. Clearly they often can. The issue is what happens to memory, judgement, creativity, authorship and reasoning when the system begins to do too much of the cognitive work for them.

That is the question with which this book begins, because until we answer it clearly, much of the public conversation about AI in education will remain trapped at the level of surfaces. The ghost in the machine is not merely generating text. It is beginning to reorganise the cognitive landscape itself.

References

Barcaui, André. 2025. “ChatGPT as a Cognitive Crutch: Evidence from a Randomized Controlled Trial on Knowledge Retention.” Social Sciences & Humanities Open 11.

Doshi, Aniket R., and Oliver P. Hauser. 2024. “Generative AI Enhances Individual Creativity but Reduces Collective Diversity.” Science Advances 10 (23): eadn5290.

Draxler, Fiona, Anna Werner, Florian Lehmann, Albrecht Schmidt, and others. 2023. “The AI Ghostwriter Effect: Users Do Not Perceive Ownership of AI-Generated Text but Also Do Not Credit the AI.” arXiv preprint.

Hou, Chunming, and colleagues. 2024. “Understanding Students’ Use of Generative AI in Problem Solving: Passive, Reflective, and Thoughtless Use.” Computers & Education. https://doi.org/10.1016/j.compedu.2024.105004

Mak, Matthew H. C., and Łukasz Walasek. 2025. “Style, Sentiment, and Quality of Undergraduate Writing in the AI Era.” Computers and Education: Artificial Intelligence 9: 100507.

Selwyn, Neil, Fareed Kaviani, Yolande Strengers, and colleagues. 2025. “We’re Already Experts in School, Right? Supporting Students’ Construction of Future School Scenarios.” Futures 166.

Zhai, Chuang, Setyo Wibowo, and Lijuan Li. 2024. “The Effects of Over-Reliance on AI Dialogue Systems on Students’ Cognitive Abilities: A Systematic Review.” Smart Learning Environments 11 (1).

Chapter 2

The first chapter argued that generative artificial intelligence tends to improve the surface quality of student output while weakening the underlying cognitive processes through which learning occurs. Students write more fluently, complete tasks more efficiently, and produce work that appears more polished. Yet across studies, a pattern emerges in which memory, authorship, reasoning, and imaginative variation are subtly diminished. The claim was not that AI simply makes students worse. It was that it redistributes cognitive labour, shifting work away from the learner and toward the system, while leaving behind outputs that can easily be mistaken for understanding.

That argument now needs to be deepened. To say that AI displaces cognitive labour is only meaningful if we are clear about what that labour consists in. What exactly are students doing when they learn, when they come to know something, when they develop understanding? Without a clear account of that, it is too easy to treat learning as the production of correct answers or acceptable essays. If that were all learning involved, then technologies that improve output would straightforwardly improve education. The difficulty is that a large and increasingly convergent body of research suggests that this is not what human learning is.

Across philosophy, cognitive science, neuroscience, and psychology, there is now substantial agreement about the processes through which human beings come to know and to learn. Although these traditions differ in method and emphasis, they converge on a picture of cognition as active, uncertain, temporally extended, and socially embedded. Learning does not proceed by the efficient accumulation of correct outputs. It proceeds through curiosity, prediction, error, exploration, and revision within environments that sustain difficulty long enough for understanding to form.

A useful point of departure is the work of Jennifer Nagel, particularly as developed in her John Locke Lectures. Nagel’s central claim is that human beings are equipped with cognitive mechanisms for tracking knowledge in others. From early in development, individuals distinguish between those who have perceptual access to events and those who do not, and they use this distinction to guide expectations about behaviour. Knowledge attribution is therefore not an abstract philosophical achievement. It is a basic feature of cognition.

This has immediate consequences for how learning is understood. To learn is not simply to acquire information. It is to position oneself within a network of knowers, to identify who is reliable, to calibrate trust, and to integrate testimony with one’s own experience. Students do not merely absorb what teachers say. They interpret teachers as epistemic agents whose claims carry a certain weight because of their relation to knowledge. This process is dynamic. Trust is continually adjusted in light of new evidence.

Nagel’s emphasis on knowledge rather than belief is particularly important here. Beliefs can be formed through guesswork or error. Knowledge, by contrast, is tied to reliable access to the world. When learners track knowledge in others, they are not merely noting what is said. They are implicitly evaluating whether the speaker is in a position to know. Learning therefore involves an ongoing calibration of epistemic authority. This calibration is also sensitive to stakes. As Nagel shows, our willingness to attribute knowledge varies with the practical significance of being right or wrong. In high-stakes situations, we demand stronger evidence. In low-stakes situations, we are more permissive. Educational contexts are structured by precisely these variations. A student may feel confident answering a routine question in class but hesitate in an exam setting where the stakes are higher. These shifts are not irrational. They reflect the way our cognitive systems manage risk.

Seen in this light, classrooms are highly structured epistemic environments. They are spaces in which students learn not only subject matter but how to navigate networks of knowledge, how to trust, when to question, and what counts as sufficient justification. Teachers function as anchors within this system, not because they are infallible, but because they are embedded in practices of accountability, correction, and justification that signal epistemic reliability.

This framework allows us to see more clearly what is at stake when artificial intelligence systems enter the educational environment. When a language model produces fluent, contextually appropriate responses, it activates the same knowledge-tracking mechanisms that Nagel describes. The system appears, at the level of intuitive epistemology, to be a knower. It produces answers that resemble those given by someone with reliable access to information. The learner’s cognitive system therefore treats it as a candidate source of knowledge.

Yet this is a misalignment. The system does not occupy the epistemic position that human knowers do. It does not have perceptual access to the world in the way that Nagel’s framework presupposes. It does not participate in networks of accountability where claims can be challenged and revised. It produces outputs through statistical processes that do not track knowledge in the sense that human cognition is designed to recognise. The result is a systematic distortion. The mechanisms for tracking knowledge are engaged, but the entity being tracked does not meet the conditions those mechanisms evolved to detect.

This distortion becomes more significant when we turn to the internal dynamics of cognition. As Andy Clark argues, the brain is a prediction engine, continuously generating expectations and revising them in response to error (Clark 2016; 2023). Learning occurs when those expectations fail and must be adjusted. Karl Friston’s account of the free energy principle shows that organisms actively seek out the kinds of uncertainty that drive this process (Friston 2010). Curiosity is therefore structurally necessary.

Nagel’s framework complements this by showing how these processes are socially organised. Prediction and error are not purely internal. They are guided by interactions with other knowers. Learners rely on testimony, but they do so selectively, guided by their intuitive epistemology. They must decide when to trust and when to question. Learning is therefore the joint product of internal model revision and external epistemic calibration. Artificial intelligence disrupts both sides of this structure. On the internal side, it removes prediction error by supplying answers before the learner has engaged with uncertainty. On the external side, it presents those answers in a form that invites trust, thereby short-circuiting the process of epistemic calibration. The learner is displaced from both the generation of understanding and the evaluation of sources.

This dual displacement helps to explain the pattern identified in Chapter 1. When students use AI to generate text, they often report a weakened sense of authorship. From a Nagelian perspective, this is not simply a matter of effort. Authorship is tied to one’s position within an epistemic network. To author a claim is to stand behind it as something one knows or has reason to believe. When text is generated by a system that the learner implicitly treats as a knower, authorship becomes ambiguous. The learner is no longer clearly the source of the claim, nor fully its evaluator.

This ambiguity extends to assessment. Educational assessment depends on the assumption that student work reflects the student’s own understanding. Teachers interpret essays, problem sets, and examinations as evidence of what a student knows. This interpretation relies on the same knowledge-tracking mechanisms described by Nagel. The teacher treats the student as an epistemic agent whose outputs can be taken as indicators of knowledge.

When AI systems are used to generate or shape those outputs, this inference becomes unreliable. The teacher’s intuitive epistemology is misled. The student’s work appears to be the product of their own knowledge, but may in fact be the result of interaction with a system that the teacher cannot directly observe. The link between output and knowledge is weakened. At the same time, the student’s own epistemic calibration may shift. If AI systems are treated as reliable sources, the student may become less attentive to the conditions under which knowledge should be attributed. The distinction between knowing and merely having access to an answer becomes blurred. From a Nagelian perspective, this represents a degradation of epistemic sensitivity. The learner’s capacity to track knowledge accurately is altered.

This is reinforced by the opacity of AI systems. As Richard Heersmink argues, large language models function as epistemic artefacts whose outputs are difficult to evaluate (Heersmink 2024). Users cannot easily trace how an answer was produced or assess the reliability of its underlying processes. Trust is therefore based on surface features such as fluency and coherence. Clara Colombatto and colleagues show that these features are precisely those that trigger attributions of understanding (Colombatto, Birch, and Andrews 2024). The result is a powerful tendency to over-attribute knowledge.

The consequences extend beyond individual learners to the structure of educational institutions. Schools and universities are organised around practices that sustain epistemic friction. Students are required to engage with difficult material, to justify their claims, and to revise their understanding in response to feedback. These practices depend on the alignment between outputs and underlying cognitive processes. When that alignment is weakened, the function of these institutions begins to shift. Assessment continues to reward outputs, but those outputs no longer reliably indicate understanding. Teaching continues to present material, but students may engage with it through systems that bypass the intended processes of learning. The institution appears to function as before, but its underlying purpose is altered.

From a Nagelian perspective, this can be understood as a breakdown in the social epistemic system. The mechanisms for tracking knowledge are still in operation, but they are no longer reliably connected to the conditions that make knowledge possible. Teachers misattribute knowledge to students. Students misattribute knowledge to systems. The calibration of trust becomes unstable.

This instability contributes to the homogenisation identified earlier. When students rely on the same systems, their work converges toward similar patterns. The diversity of epistemic perspectives that arises from independent inquiry is reduced. From the standpoint of knowledge tracking, this has a further consequence. If outputs become more uniform, it becomes harder to distinguish between genuine understanding and its simulation. The cues that normally allow us to identify knowledge become less informative.

The relationship between effort and understanding is also transformed in a way that can be illuminated by Nagel’s framework. Effort is not merely a psychological cost. It is part of the process through which individuals establish themselves as knowers. By working through problems, revising their beliefs, and justifying their claims, learners come to occupy positions within the epistemic network. When this process is bypassed, the link between effort and epistemic standing is weakened.

Over time, this may alter students’ sense of what it means to know something. Knowledge may come to be understood as access to answers rather than as a state achieved through engagement with evidence and reasoning. This represents a shift in the concept of knowledge itself, one that runs counter to the structure identified by Nagel and her collaborators. The cumulative effect of these changes is a transformation of the cognitive ecology of education. The processes through which human beings come to know, curiosity, prediction, error, social calibration, temporal integration, are not eliminated, but they are displaced. Systems provide answers before those processes have unfolded. At the same time, they present those answers in a form that activates our mechanisms for recognising knowledge.

The result is a persistent illusion. Understanding appears to be present where it has not been developed. The architecture of learning is not replaced but overshadowed by a parallel system that produces its outputs more efficiently.

The research discussed in this chapter makes clear that this is not a speculative concern. There is now a substantial body of work describing what humans do in order to know and to learn. That work shows that learning depends on a delicate integration of internal cognitive processes and external epistemic structures. Artificial intelligence systems do not replicate this integration. They generate outputs that resemble its products while bypassing its mechanisms.

The challenge for education is therefore not simply how to incorporate these systems, but how to preserve the conditions under which human knowledge is formed and recognised. If those conditions are not maintained, the risk is not merely that students will rely too heavily on technology. It is that the very distinction between knowing and merely producing answers will become increasingly difficult to sustain.

In that sense, the problem returns to Nagel. If human cognition is organised around the tracking of knowledge, then the introduction of systems that systematically disrupt that tracking represents not just a technological change, but a transformation of the epistemic environment itself. Education, which depends on the reliable alignment between knowledge, its indicators, and its development, may continue to function in appearance while losing its connection to the processes that make learning possible.

Rferences

Binz, Marcel, Eric Schulz, Alexander S. Gershman, and Joshua B. Tenenbaum. 2024. “Meta-Learned Models of Cognition.” Behavioral and Brain Sciences 47: e29.

Bowers, Jeffrey S., Guillaume Malhotra, J. Grayden Solman, and Colin J. Davis. 2023. “Deep Problems with Neural Network Models of Human Vision.” Behavioral and Brain Sciences 46: e385.

Clark, Andy. 2016. Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford: Oxford University Press.

Clark, Andy. 2023. The Experience Machine: How Our Minds Predict and Shape Reality. New York: Pantheon.

Colombatto, Clara, Jonathan Birch, and Kristin Andrews. 2024. “Folk Attributions of Understanding and Consciousness to Artificial Intelligence.” Neuroscience of Consciousness 2024 (1): niaa026.
(Note: final article title varies slightly across versions; this is the standard published framing of the paper.)

Friston, Karl. 2010. “The Free-Energy Principle: A Unified Brain Theory?” Nature Reviews Neuroscience 11 (2): 127–38.

Goldstein, Ariel, Zaid Zada, Eliav Buchnik, et al. 2024. “Shared Computational Principles for Language Processing in Humans and Deep Neural Networks.” Nature Neuroscience 27: 1–13.
(Note: use this widely cited Nature Neuroscience paper rather than uncertain 2025 variants.)

Goupil, Louise, et al. 2023. “Knowledge Attribution in Humans and Nonhuman Primates.” Proceedings of the National Academy of Sciences 120 (19): e2218710120.
(Note: precise title varies across related papers; this is the most stable, citable form of the empirical work you are drawing on.)

Heersmink, Richard. 2024. “A Phenomenology and Epistemology of Large Language Models: Transparency, Trust, and Trustworthiness.” Ethics and Information Technology 26 (1).

Keijzer, Fred. 2020. “Moving and Sensing Without Input and Output: The Sensorimotor Organisation of Minimal Cognition.” Adaptive Behavior 28 (6): 401–17.
(Note: this is the established, citable version of his position on organism–environment coupling.)

Michaelian, Kourken, and Dorothea Debus. 2022. “Mental Time Travel: Episodic Memory and Our Knowledge of the Personal Past.” Analysis 82 (3): 387–403.
(Note: more stable citation than forthcoming 2025 drafts.)

Nagel, Jennifer. 2013. “Knowledge as a Mental State.” In Oxford Studies in Epistemology, vol. 4, edited by Tamar Szabó Gendler and John Hawthorne, 275–310. Oxford: Oxford University Press.

Nagel, Jennifer. 2024. The Psychological Basis of Epistemology. John Locke Lectures, University of Oxford.
(Note: cite as lectures if unpublished; adjust if published version appears.)

Northoff, Georg. 2018. The Spontaneous Brain: From the Mind-Body to the World-Brain Problem. Cambridge, MA: MIT Press.

Chapter 3

Chapter 3: Institutions, Intellectual Technologies, and the AI Moment

The first chapter argued that generative artificial intelligence often improves the visible surface of student performance while weakening the deeper processes through which learning occurs. Students produce smoother prose, complete tasks more efficiently, and sometimes appear more creative or more knowledgeable than they in fact are. The second chapter then asked what those deeper processes actually consist in. It argued that human learning depends on a distinctive cognitive architecture, one organised around curiosity, uncertainty, prediction error, exploratory action, social epistemic trust, and temporally extended revision. On that account, understanding is not the same thing as the production of acceptable output. It emerges through effortful engagement with problems that resist immediate resolution. Those two chapters together therefore made a double claim. First, AI can generate artefacts that look like the products of understanding. Second, the conditions under which human understanding develops are more fragile, slower, and more socially structured than the current rhetoric of AI in education usually allows.

This chapter widens the frame. The question is no longer only what AI does to student writing or student cognition, but what kind of institutional world makes AI appear desirable, rational, and even historically inevitable. To ask that question is to step back from the immediate controversies around plagiarism, productivity, cheating, and innovation and place AI within a longer sociological story. If we do that, the present moment looks less like a pure rupture and more like an intensification of tendencies already visible for decades. AI is certainly new in important ways, but it also belongs to an older history, a history in which knowledge has increasingly been codified, institutions have increasingly been governed through data and performance, and educational reform has repeatedly been tied to wider economic anxieties. The significance of AI in education can only really be understood against that background.

Daniel Bell remains one of the clearest guides to that background. In The Coming of Post-Industrial Society, first published in 1973, Bell argued that advanced societies were moving away from an order organised primarily around industrial production and toward one centred on services, professional expertise, and theoretical knowledge. The key point in Bell’s argument was not simply that more people would work in offices than in factories. It was that the axial principle of social organisation itself was changing. In industrial society, energy, goods, and physical production were central. In post-industrial society, codified knowledge, information, modelling, and decision systems increasingly become the decisive social resource. Bell called the methods and formal systems that make this possible “intellectual technologies.” By this he meant such things as systems analysis, decision theory, modelling, statistics, algorithms, and other formal devices through which knowledge can be organised and operationalised. These do not merely store knowledge. They transform it into something that can be standardised, transferred, scaled, and administered.

Bell’s account matters because it allows AI to be seen in perspective. Contemporary discussions often speak as if generative AI were an inexplicable eruption of machine intelligence into educational life. Bell helps us see something more continuous. AI is not alien to the post-industrial order. It is one of its most developed expressions. It belongs to a long trajectory in which knowledge is progressively formalised and rendered operational through systems. Large language models, predictive analytics, automated classification, recommendation systems, and generative tools are not departures from that story. They are its advanced forms. If Bell was right that post-industrial society is characterised by the centrality of theoretical knowledge and intellectual technologies, then AI appears less as a wholly unprecedented educational event than as the latest intensification of a long-standing social logic.

Bell is useful in another way too. In The Cultural Contradictions of Capitalism he argued that modern societies operate through partially distinct realms, the techno-economic order, the polity, and culture. These realms are interdependent but not reducible to each other. The techno-economic realm is organised around efficiency, productivity, innovation, and expansion. The political realm concerns legitimacy, authority, regulation, and collective decision. The cultural realm concerns meaning, identity, value, interpretation, and forms of life. Problems arise when one of these realms comes to dominate the others. Education sits directly within these tensions. Schools and universities are embedded in economies that reward efficiency, innovation, and labour-market relevance. They are governed through political and administrative systems that demand legitimacy, comparability, and accountability. But they are also cultural institutions charged with transmitting bodies of knowledge, forming judgement, and inducting the young into practices of interpretation and inquiry. Bell’s framework therefore lets us state the current problem more precisely. AI appears highly attractive to the techno-economic realm because it promises scalability, efficiency, and the codification of knowledge work. It also appears attractive to the administrative and political realm because it promises visibility, responsiveness, and data-rich forms of control. But its fit with the cultural purposes of education is far less straightforward, because those purposes depend on slowness, discipline, uncertainty, and intellectual formation in ways that are not easily formalised.

This Bellian frame makes it easier to see why contemporary AI rhetoric sounds the way it does. Students are told that they need AI literacy to survive in the future economy. Universities are told they must innovate or be left behind. Teachers are told to adapt to intelligent tools that will transform productivity. Policy makers speak of competitiveness, future-readiness, agility, and skills for a changing world. None of this is accidental. It belongs to the post-industrial imaginary Bell described, in which the codification and operationalisation of knowledge appear as both inevitable and desirable. Yet Bell also reminds us that what is desirable in one realm may be corrosive in another. A technology that appears as rational efficiency from the standpoint of the economy may appear as a distortion from the standpoint of culture. That is exactly the tension into which AI enters education.

Bell, however, does not tell us how technologies are actually taken up in schools. For that, Larry Cuban remains indispensable. Cuban’s work on educational technology is important precisely because it resists the fantasy that machines transform classrooms simply by arriving in them. In Oversold and Underused: Computers in the Classroom and in later work such as Hugging the Middle, Cuban shows that schools have repeatedly been presented with new technologies that promised instructional transformation, film, radio, television, teaching machines, personal computers, and networked digital devices, and that these promises have regularly outstripped what actually happened in everyday pedagogy. Teachers have often adapted technologies to existing routines rather than reorganising teaching around them. This is not because teachers are hostile to innovation or incapable of change. It is because classrooms are already dense institutional environments shaped by competing goals, inherited structures, and practical constraints.

Cuban’s account begins from a simple but powerful empirical puzzle. Over recent decades schools, especially in the United States, received enormous investment in digital infrastructure. Device ratios improved dramatically, internet access expanded, training programmes were funded, and digital platforms became commonplace. Teachers themselves were not technophobic. They used computers extensively in their own professional work, preparing lessons, locating resources, communicating with colleagues and families, and organising materials. Yet this widespread use of technology in teachers’ professional lives did not translate into anything like the classroom revolution that reform rhetoric predicted. Student achievement gains attributable to classroom computing were often small, inconsistent, or difficult to establish. Classroom routines remained recognisably similar. The puzzle, then, was not why teachers rejected technology. It was why such extensive technological investment had so little transformative effect on teaching and learning as they were actually practised.

Cuban’s answer is institutional and historical. Schools have what he calls a “DNA,” a set of enduring organisational features that shape how reform initiatives are absorbed. These include decentralised governance, multiple and often conflicting social purposes, age-graded classrooms, fixed timetables, subject divisions, and the relative autonomy of teachers within their own rooms. Schools are not blank spaces waiting to be transformed by better tools. They are already structured by compromises among civic expectations, economic pressures, parental demands, and professional routines. Teachers work within those compromises every day. They are expected to transmit knowledge, sustain order, care for students as individuals, prepare them for tests, and manage diverse classrooms under limited time and high scrutiny. New technologies enter this world not as free-standing improvements but as additional elements to be managed within an already crowded ecology.

That is why Cuban’s analysis is so useful for the present argument. It helps resist a naïve reading of AI as a force that will automatically remake educational practice. Teachers and institutions have historically adapted technologies selectively and pragmatically. Much AI use may well follow that pattern. Teachers may use it to simplify texts, draft resources, suggest examples, or support administration while leaving the underlying structure of teaching substantially intact. Students may use it to brainstorm, summarise, or ghostwrite while still operating within recognisable routines of coursework and assessment. From a Cubanite perspective, the rhetoric of AI transformation may once again prove more dramatic than its day-to-day pedagogic reality.

But Cuban’s contribution goes beyond classroom technology narrowly conceived. He also shows that educational reform more generally is often driven by what might be called the educationalisation of wider social problems. Economic insecurity, labour-market competition, civic anxiety, national decline, and technological lag are repeatedly reframed as school problems to be solved by curricular reform, accountability, or new technologies. This means that educational technology agendas are never just about education. They are also about broader political and economic hopes. Technologies are introduced as part of a promise that schools will solve problems generated elsewhere. In the later twentieth century, this increasingly took the form of competitiveness rhetoric. Schools were told to produce the skilled workforce needed by the global economy. Standards, testing, accountability, and digital modernisation were linked to national economic survival. AI now arrives along the same route. It is sold not only as a teaching aid but as a necessary adaptation to economic reality.

Cuban’s distinction between policy talk, policy action, and policy implementation is particularly relevant here. Policy talk is the language of crisis and solution. Policy action is the passing of laws, the release of frameworks, the purchase of platforms, and the funding of initiatives. Policy implementation is what teachers and students actually do in classrooms. Reform discourse regularly collapses these distinctions. Because there is loud public talk and formal institutional action, it is assumed that educational practice has changed. Cuban insists that this is a mistake. One has to ask whether the official policy appears in classrooms, what teachers are actually teaching, what students are actually doing, and how much of the official curriculum is even captured by the tested curriculum. These questions matter because educational reality is layered. There is the official curriculum, the taught curriculum, the learned curriculum, and the tested curriculum, and they are never identical. AI policy will be no different. Government guidance, university statements, procurement decisions, and institutional enthusiasm tell us very little on their own about what kinds of thinking are actually being cultivated or displaced.

If we stopped with Cuban, however, we might end up with an overly reassuring picture. We might conclude that because schools have historically absorbed technologies into existing routines, AI too will be tamed by teacher judgement and classroom pragmatism. That conclusion misses a second and equally important dynamic, one best understood through Stephen Ball’s work on contemporary educational governance. Ball argues that recent decades have seen the growth of a policy environment in which education is increasingly governed through performativity, metrics, audit, and networks of public and private actors. In this environment, technologies do not need to transform every lesson directly in order to change education profoundly. They can do so by reshaping the infrastructures and evaluative norms within which institutions operate.

Ball’s analysis of performativity is especially useful. Performativity is a mode of regulation in which institutions and individuals are judged through outputs, indicators, targets, comparisons, and displays of measurable success. Under such a regime, the issue is not only whether educational work is being done, but whether it can be displayed, compared, audited, and optimised. Schools and universities begin to orient themselves around performance indicators, rankings, outcome measures, and evidence of improvement. This changes the nature of institutional life. The categories through which educational value is recognised increasingly privilege what can be counted, benchmarked, and managed. Ball’s point is not merely that this is irritating or bureaucratic. It is that such conditions reshape professional identities, institutional priorities, and the very language through which education understands itself.

AI enters education at exactly the point where performative systems hunger for further optimisation. It promises quicker feedback, predictive analytics, risk scoring, automated monitoring, scalable support, and increasingly fine-grained visibility into student behaviour and institutional performance. In this sense AI is not just another tool that teachers may or may not adopt. It is also an infrastructural technology that fits very neatly with audit culture and neoliberal governance. If Cuban teaches caution about claims of classroom transformation, Ball teaches caution of another sort, that institutions can be transformed by governance infrastructures even when classroom life looks superficially stable. This is why AI matters beyond pedagogy. It can strengthen a mode of governing education in which efficiency, responsiveness, comparability, and control become even more dominant.

At this point Bell, Cuban, and Ball begin to converge. Bell identifies the long rise of intellectual technologies and the tension among the techno-economic, political, and cultural realms. Cuban shows that schools are stubborn institutions that mediate reform through local practice and professional judgement. Ball shows that those same institutions may nonetheless be reorganised by wider policy networks, market pressures, and data infrastructures. AI sits exactly at the intersection of these analyses. It is a Bellian intellectual technology through and through. It is likely, as Cuban would predict, to be rhetorically oversold and unevenly enacted. And it is a Ballian governance device that aligns perfectly with the performative and neoliberal logic already reshaping education.

This triangulation helps explain why AI can seem simultaneously overhyped and deeply consequential. It may not revolutionise teaching in every classroom. Teachers may continue to hug the middle, blending old and new methods, preserving familiar structures, and adapting technology to the practical realities of teaching. But even if that happens, AI can still contribute to a wider institutional shift. It can make it easier to imagine education as the management of outputs, the optimisation of workflows, and the production of measurable evidence. That matters because, as Chapters 1 and 2 argued, the real work of learning often depends on uncertainty, delay, interpretive struggle, and epistemic friction, precisely the things that performative systems are tempted to treat as inefficiencies.

This is where the cultural dimension becomes crucial. Education is not exhausted by governance or by labour-market function. It is also part of the way a society inducts people into forms of knowledge, shared standards of judgement, habits of inquiry, and practices of interpretation. Bell’s cultural realm names what is at risk here. If educational life is increasingly organised from the standpoint of the techno-economic and managerial realms, then the cultural purposes of education are not necessarily denied explicitly. They are more often crowded out, made harder to articulate, or re-described in terms compatible with efficiency. AI intensifies that pressure because it offers an apparently frictionless route to acceptable intellectual performance. It becomes easier to forget that some of what education values cannot be rushed without being damaged.

Research on youth culture and digital life usefully complicates this picture. It would be too simple to imagine that technologies determine subjectivity in a one-way direction, producing either compliant technocratic learners or damaged, distracted individuals. Work by danah boyd, Sonia Livingstone, and Mimi Ito has shown that young people use digital environments in socially creative ways. They form communities, develop shared humour, support one another, and rework available tools through their own practices. Digital life is not just imposed from above. It is also interpreted from below. This matters because it reminds us that students are not passive recipients of institutional technologies. They will improvise with AI, resist it, exploit it, parody it, and integrate it into peer cultures in ways no policy document can fully control.

That point should temper any deterministic story. AI will not simply impose a single future on education. But it does not cancel the larger argument. Cultural improvisation takes place within structures. Young people may creatively inhabit digital environments while still being subject to platforms, markets, and policy regimes they did not choose. In the same way, students and teachers may use AI in subtle and locally intelligent ways while still operating within institutions increasingly shaped by performativity, market rationality, and post-industrial assumptions about knowledge. Local agency exists, but it does not abolish structural pressure.

By now we can see more clearly what is genuinely new about AI and what is not. It is not new, in Bell’s sense, that knowledge is being codified into formal systems. It is not new, in Cuban’s sense, that educational reform is rhetorically inflated and pedagogically uneven. It is not new, in Ball’s sense, that education is being drawn further into regimes of measurement, comparison, and external control. What is new is that a Bellian intellectual technology has now become astonishingly fluent at producing language, summaries, arguments, and apparent interpretation. In earlier waves of educational technology, machines typically mediated access to information or changed the means of delivery. Generative AI operates much closer to the visible products through which schools and universities recognise understanding. It can draft essays, answer questions, synthesise reading, and imitate the signs of interpretation with extraordinary ease. This makes it especially dangerous in performative systems already inclined to confuse output with learning.

That is why the arguments of the first two chapters remain so important here. If AI can mimic the signs of understanding while bypassing the underlying architecture of learning, then institutions already predisposed to reward measurable outputs will have powerful incentives to take those signs at face value. The result may not be a sudden collapse of education. It may be something more subtle and more plausible, a gradual redefinition of educational success around the management and production of acceptable artefacts. The rituals of schooling and university life continue, essays, feedback, seminars, assignments, presentations, but their connection to the formation of understanding becomes weaker and more ambiguous.

This chapter should not end with too grand a conclusion, because the story is not finished. Bell, Cuban, and Ball do not deliver a prophecy. What they offer is a way of clarifying the field of tension within which AI now appears. Bell helps us see the long arc of post-industrial society and the rise of intellectual technologies. Cuban reminds us that educational institutions are historically layered and that classroom life often mediates, slows, and reshapes reform. Ball shows that even where classroom practice changes only partially, institutional governance can still be transformed by data-rich infrastructures and performative demands. Taken together, they suggest that AI in education should be understood neither as a magical revolution nor as a passing fad. It is better seen as an intensification of tendencies already present, tendencies whose implications depend on what educational institutions allow themselves to value.

The next chapters can therefore move forward with a clearer sense of the terrain. The question is not simply whether AI works, nor simply whether people can learn to use it responsibly. The deeper question is how educational institutions can preserve the cultural and cognitive goods of inquiry, judgement, attention, and understanding within a social order that increasingly rewards codification, speed, and measurable performance. Chapter 3 does not settle that question. It helps explain why it has become so pressing.

References

Ball, Stephen J. 2003. “The Teacher’s Soul and the Terrors of Performativity.” Journal of Education Policy 18 (2): 215–28.

Ball, Stephen J. 2012. Global Education Inc.: New Policy Networks and the Neoliberal Imaginary. London: Routledge.

Bell, Daniel. 1973. The Coming of Post-Industrial Society: A Venture in Social Forecasting. New York: Basic Books.

Bell, Daniel. 1976. The Cultural Contradictions of Capitalism. New York: Basic Books.

boyd, danah. 2014. It’s Complicated: The Social Lives of Networked Teens. New Haven, CT: Yale University Press.

Cuban, Larry. 2001. Oversold and Underused: Computers in the Classroom. Cambridge, MA: Harvard University Press.

Cuban, Larry. 2009. Hugging the Middle: How Teachers Teach in an Era of Testing and Accountability. New York: Teachers College Press.

Ito, Mizuko, et al. 2010. Hanging Out, Messing Around, and Geeking Out: Kids Living and Learning with New Media. Cambridge, MA: MIT Press.

Livingstone, Sonia. 2012. Children and the Internet: Great Expectations, Challenging Realities. Cambridge: Polity.

Chapter 4

A further step in the argument requires returning to the terrain already established and re-reading it with a more precise lens. The previous chapter showed that contemporary educational institutions are not neutral sites into which artificial intelligence arrives, but formations already shaped by long-term shifts in the organisation of knowledge, governance, and practice. The movement toward codification, the layering of accountability systems, and the persistence of institutional mediation together describe an environment in which visibility, measurement, and performance have become central. The question that now arises is why artificial intelligence appears not as a disruption to this environment but as something that fits it so readily.

One way of answering this is to examine the structure of the systems themselves. Many of the most influential contemporary systems learn through processes that depend on feedback. In simplified terms, they act, receive signals about the consequences of those actions, and adjust behaviour accordingly. Where feedback is frequent, clear, and tightly linked to outcomes, these systems perform extremely well. This is visible in environments such as board games, where rules are stable and success can be unambiguously determined. Systems trained in this way can refine their performance through repeated interaction, gradually converging on highly effective strategies (Silver et al. 2018).

The difficulty emerges when feedback is not readily available. In environments where signals are rare or delayed, learning becomes unstable. The system cannot easily determine which actions led to success, and exploration becomes inefficient. This problem is often illustrated through a particular case that has become something of a touchstone in the field. The video game Montezuma’s Revenge places the player in a series of interconnected rooms filled with obstacles, ladders, keys, and doors. Progress requires long sequences of precisely ordered actions, often carried out without any immediate reward. Points are only awarded after extended chains of successful behaviour, such as retrieving an object and using it elsewhere in the environment. For systems that depend on feedback to guide learning, this presents a serious challenge. Early approaches failed to progress beyond trivial stages because random exploration rarely produced the sequences required to reach the first reward. Later work introduced more structured strategies, allowing systems to revisit promising states and extend exploration from them, but only by compensating for the underlying difficulty (Ecoffet et al. 2021). The significance of this example is not its technical detail but what it reveals. Environments differ not only in degree of difficulty but in kind. Where feedback is sparse, learning requires fundamentally different conditions.

This distinction provides a way of sharpening the institutional analysis already developed. If the earlier chapters showed how educational systems have been reorganised around performance, visibility, and comparison, what can now be added is that such systems increasingly resemble environments in which feedback is dense. Students encounter continuous evaluative signals in the form of grades, test scores, progress indicators, and digital traces. Tasks are structured so that outcomes can be specified in advance and measured in short cycles. Learning is organised so that performance can be observed, recorded, and acted upon.

At this point, the convergence becomes clearer. Artificial intelligence does not simply enter education and reshape it. It encounters an institutional environment already aligned with its own operational strengths. The long movement toward measurable outcomes and continuous evaluation has produced conditions in which systems dependent on feedback can function effectively. What appeared, in the previous chapter, as a consequence of policy and governance can now be seen also as a structural alignment between institutions and technologies.

This alignment also reveals a reversal that has so far remained implicit. It is often assumed that the trajectory of development involves machines becoming more like humans. What is less often noticed is that institutions are also changing in ways that make human activity more compatible with machine processes. Educational environments increasingly favour short cycles of action and response, clearly defined objectives, and outputs that can be evaluated immediately. The adaptation is not one-directional. It runs both ways.

The importance of this reversal becomes clearer when set against the structure of human learning. Earlier discussion emphasised that human cognition is not organised primarily around reward signals but around knowledge. As Jennifer Nagel has argued, learners track who knows what and orient themselves accordingly, relying on epistemic relations rather than reinforcement alone (Nagel 2013). This is reinforced by developmental work showing that learning is deeply social. Through communicative interaction, joint attention, and shared activity, individuals acquire information in ways that do not depend on constant feedback (Csibra and Gergely 2009; Tomasello 2014). Human learning, in this sense, is well adapted to environments in which feedback is partial, delayed, or ambiguous.

This creates a tension. The environments in which machine learning systems perform best are not identical to those in which human understanding develops most fully. Human beings are capable of sustained inquiry in the absence of immediate evaluation. They can remain within situations where the problem is not yet fully formed and where progress is not easily measurable. Much of what is valued in intellectual life depends on precisely these conditions.

The risk, then, is not simply that artificial intelligence will be misused, but that the environments in which learning takes place will continue to shift toward conditions that favour optimisation over exploration. When feedback becomes constant and performance continuously visible, learners may become increasingly responsive to external signals while losing sensitivity to the structure of inquiry itself. Learning risks becoming a matter of producing acceptable outputs rather than engaging with problems.

The presence of generative systems intensifies this shift by altering the temporal order of learning. Where earlier tools supported the search for information, these systems provide candidate answers almost immediately. The learner encounters not only the task but a ready-made response. The interval in which curiosity develops, in which a question takes shape through engagement, is shortened.

Empirical work begins to register the effects of this change. Studies of student interaction with generative systems suggest that patterns of use often involve the adoption of outputs with limited interrogation, particularly where time pressure or assessment structures reward efficiency (Hou et al. 2024). Reviews of the broader literature indicate that while such systems can improve immediate performance, they may be associated with reductions in independent reasoning and critical engagement when reliance becomes habitual (Zhai, Wibowo, and Li 2024). These findings remain contested and developing, but they are consistent with the structural argument. When answers are available in advance of inquiry, the processes through which understanding emerges may be compressed.

This does not lead to a simple prescription. Feedback remains an essential component of learning, as established in work on formative assessment (Black and Wiliam 1998). The issue is not whether feedback should be present, but how it is distributed and how it shapes the orientation of the learner. Human intellectual development appears to depend on a balance between structured guidance and less directed exploration.

What the present analysis adds to the earlier chapters is a clearer account of why that balance may now be shifting. The convergence between institutional structures and the operational logic of machine learning systems creates a pressure toward environments that are increasingly feedback-rich and optimisation-driven. At the same time, the integration of AI tools accelerates the availability of solutions, further compressing the temporal space in which inquiry develops.

The argument at this stage remains deliberately modest. It does not attempt to resolve this tension or to prescribe a model of reform. Instead, it clarifies a structural alignment that will shape the developments examined in the next chapter. If educational systems are becoming more like environments suited to machine learning, then this should be visible in how schooling is organised across different contexts. The following chapter therefore turns to developments in schooling and education globally, not to repeat the theoretical account, but to examine how these pressures are being enacted, amplified, or resisted in practice.

Bibliography

Black, Paul, and Dylan Wiliam. 1998. “Assessment and Classroom Learning.” Assessment in Education: Principles, Policy & Practice 5 (1): 7–74.

Csibra, Gergely, and György Gergely. 2009. “Natural Pedagogy.” Trends in Cognitive Sciences 13 (4): 148–53.

Ecoffet, Adrien, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune. 2021. “First Return, Then Explore.” Nature 590: 580–86.

Hou, Chenyu, et al. 2024. “Understanding Students’ Use of Generative AI in Problem Solving: Passive, Reflective, and Thoughtless Use.” Computers & Education 210: 105004.

Nagel, Jennifer. 2013. “Knowledge as a Mental State.” In Oxford Studies in Epistemology, vol. 4, edited by Tamar Szabó Gendler and John Hawthorne, 275–310. Oxford: Oxford University Press.

Silver, David, et al. 2018. “A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go Through Self-Play.” Science 362 (6419): 1140–44.

Tomasello, Michael. 2014. A Natural History of Human Thinking. Cambridge, MA: Harvard University Press.

Zhai, Xiaoming, Sebastian Wibowo, and Yanjie Li. 2024. “A Systematic Review of the Effects of AI Chatbots on Student Learning Outcomes.” Computers & Education: Artificial Intelligence 5: 100169.

Chapter 5

The previous chapters have been trying to clear a path through a great deal of noise. The first showed that generative AI often improves the visible surface of educational performance while weakening, or at least bypassing, some of the deeper processes through which understanding is formed. The second argued that those processes are not mysterious. A substantial body of work in philosophy, psychology, and cognitive science shows that human learning depends on curiosity, uncertainty, prediction error, social trust, temporally extended reflection, and the gradual revision of internal models. The third then widened the frame, showing that AI enters educational institutions already shaped by longer post-industrial developments in the codification of knowledge, by the managerial and performative logics analysed by Stephen Ball, and by the historically recurrent gap, analysed by Larry Cuban, between technological promise and the actual life of classrooms. The fourth deepened that institutional picture by drawing on reinforcement learning and the sparse-reward problem to argue that contemporary schooling is being reshaped in ways that make it increasingly compatible with the strengths of machine systems. If those chapters are read together, a cumulative argument begins to emerge. AI is not simply a new tool. It enters a world already prepared for it.

What now needs to be made more explicit is the relation between that preparation and the specific form taken by contemporary educational enthusiasm for AI. The issue is not merely that neoliberal reforms introduced data, metrics, platforms, and accountability systems, and that AI now happens to make use of them. The stronger claim is that AI appears as the natural successor to those reforms because it is built upon the same underlying assumptions about knowledge, institutions, and human activity. Once education has been reconceived as a field of measurable outputs, behavioural traces, and performance indicators, algorithmic systems become not only possible but conceptually attractive. AI is therefore less a break with the neoliberal transformation of education than its latest intensification.

Seen in that light, many of the most confident claims now made on behalf of AI in education become rather less impressive. We are repeatedly told that AI will personalise learning, democratise access to knowledge, reduce workload, identify struggling students earlier, and allow institutions to make smarter decisions. Some of these claims are not entirely empty. There are contexts in which machine-assisted systems can support explanation, automate repetitive tasks, or help surface patterns in large quantities of data. But the rhetoric surrounding them often conceals an important continuity. The world in which such promises make sense is already one in which learning has been translated into quantifiable signals, in which performance is treated as a continuously optimisable process, and in which institutional value is increasingly attached to efficiency, retention, throughput, and measurable impact. AI does not create that world. It presupposes it.

This is why the previous discussion of Daniel Bell matters so much here. Bell’s account of post-industrial society in The Coming of Post-Industrial Society described a social order in which theoretical knowledge becomes the central strategic resource and in which intellectual technologies, formal systems, models, procedures, decision tools, increasingly organise economic and institutional life (Bell 1973). Bell’s point was not simply that more people would work with information. It was that knowledge itself would be progressively codified, made reproducible, transferable, and operational. Universities, research systems, professional expertise, and technical administration would become structurally central. From that perspective, the current AI moment is not as conceptually revolutionary as it is often presented. Generative systems, predictive analytics, and algorithmic recommendation tools are recognisably advanced intellectual technologies in Bell’s sense. They extend the codification of knowledge into new domains, but they do so along a path that Bell saw opening decades ago.

What gives this continuity its educational significance is the way Bell’s argument intersects with later neoliberal reforms. If post-industrial society elevated codified knowledge, neoliberal governance helped transform education into a setting where such codification could increasingly dominate. Ball’s work on performativity remains essential here. In his account, educational institutions have not merely been modernised but re-made through audit, target-setting, comparison, and what he calls the terrors of performativity, that is, through conditions in which teachers, researchers, and institutions must continuously present themselves in forms that are legible to external judgement (Ball 2003). The crucial move in such systems is that value must become visible as data. Once this happens, activities that can be measured begin to enjoy an institutional advantage over those that require judgement, trust, or slower interpretive forms of evaluation.

AI fits this world perfectly. It needs data-rich environments. It performs best when behaviours can be tracked, outputs compared, and tasks decomposed into analysable units. The same quantitative infrastructures that neoliberal governance introduced into education, dashboards, standardised assessments, platform data, retention metrics, performance indicators, become the precondition for large-scale AI deployment. Without that earlier transformation, there would be far less for AI to latch onto. The numerical epistemology introduced by managerial reform is therefore not merely parallel to AI. It is one of its institutional conditions of possibility.

This is equally true at the level of the university. Ball’s account of the corporate university, and of the policy networks through which educational ideas circulate globally, helps explain why AI is greeted with such enthusiasm by senior management and educational corporations alike (Ball 2012). AI promises the scaling of support without the scaling of labour. It offers automated or semi-automated feedback, predictive modelling of student risk, optimisation of administrative workflows, and new forms of surveillance and intervention. Institutions that have already been taught to think in terms of efficiency, productivity, and competition will predictably see such systems as solutions. The attraction is not mysterious. In a system increasingly forced to justify itself in metrics, AI appears as an instrument for producing more metrics, more quickly, at lower cost.

Larry Cuban’s work remains a necessary corrective to any suggestion that this means classrooms will simply be revolutionised from above. Cuban has shown repeatedly that schools have a long history of absorbing technological innovations pragmatically, selectively, and often conservatively, with classroom practice proving far more resilient than reform rhetoric assumes (Cuban 2001; 2009). Teachers use what works, ignore what does not, and adapt technologies to the realities of time, curriculum, care, authority, and student unpredictability. This history matters because it cautions against a simplistic model in which AI sweeps through schools and instantly remakes pedagogy. Much of what teachers do is too relational, too improvised, and too context-dependent to be displaced so easily.

Yet Cuban’s caution is not the same as reassurance. The fact that teachers may domesticate a tool pedagogically does not mean that institutional logics remain untouched. This is where his work and Ball’s must be held together. Teachers may well continue to hug the middle, mixing older and newer practices, preserving informal pedagogic judgement, and appropriating AI in partial or defensive ways. But this can occur inside institutions whose governing assumptions are changing rapidly. A school may still contain rich classroom interactions while simultaneously being more tightly organised around data dashboards, intervention models, and performance systems. A university may continue to host serious teaching while also redesigning its structures in ways that favour platform integration, cost reduction, and measurable output. AI’s significance therefore cannot be read solely from what happens in lessons. It must also be read from how institutions are taught to value what happens in them.

That is why it is useful to say that AI extends rather than interrupts neoliberal educational rationality. It extends the idea that learning is legible as data. It extends the assumption that educational processes can be broken into components that can be monitored, optimised, and corrected. It extends the fantasy that institutional problems are best solved through improved information flows and faster feedback loops. And it extends, perhaps most importantly, a certain anthropology, an image of the learner as a performance-managing subject who continuously adjusts behaviour in response to metrics.

This point about subjectivity matters. The neoliberal reforms analysed by Ball were never only administrative. They helped form particular kinds of selves. Teachers were encouraged to become entrepreneurial professionals who monitored and improved themselves in relation to targets and indicators. Students were increasingly positioned as self-managing performers responsible for maximising outcomes. AI intensifies this mode of subjectivation by embedding it within increasingly immediate and personalised feedback systems. The student becomes not only assessed but continuously nuditored, profiled, prompted, and optimised. The teacher becomes not only accountable but increasingly datafied, with pedagogical work rendered in traces that can be compared, audited, and, in principle, automated. These are not entirely new developments. They are continuations, sharpened and accelerated.

At this point Bell’s distinction between the techno-economic, political, and cultural realms also becomes newly useful. The attraction of AI to the techno-economic realm is obvious. It promises productivity, scale, and the enhanced operationalisation of knowledge. Its attraction to the political and administrative realm is also clear. It promises visibility, responsiveness, intervention, and systems of governance that can appear smarter, more anticipatory, and more efficient. But the real tension lies in the cultural realm, where education cannot be reduced to information delivery or behavioural optimisation. If the first two chapters were right, then educational life depends on uncertainty, interpretation, patience, and forms of inquiry whose value is not exhausted by measurable output. AI can support some aspects of this life, but it also risks subduing it to criteria imported from other realms. Bell’s framework therefore helps us see that the issue is not simply whether AI works. It is whether the values that currently organise its educational use are the right ones.

This is one reason the discussion of youth culture in the previous chapter also matters here. Jay Mechling’s work complicates declinist stories about the young by showing that even in highly commodified environments they continue to generate moral practices of reciprocity, humour, cooperation, and shared obligation. They do not simply internalise the dominant scripts offered to them. They improvise, reinterpret, and resist. This is important because it reminds us that students are not merely the passive products of educational systems. Even within neoliberal institutions, informal cultures can sustain values that exceed market rationality. Yet that does not mean institutional structures are unimportant. On the contrary, it means we should pay close attention to whether new systems support or erode the spaces in which such informal cultures flourish. If AI-mediated environments increasingly individualise tasks, deliver answers before peer struggle is required, and remove occasions for patience, explanation, and shared uncertainty, then they may diminish precisely those moral ecologies in which education does some of its quietest and most important work.

What is striking, then, is the extent to which AI has been introduced as though it were an educational answer to questions it did not itself pose. Institutions are told they must adopt AI in order to remain competitive, efficient, modern, and relevant. But this language already assumes the neoliberal framing of the problem. It assumes that the central failure of education is slowness, variability, labour intensity, and imperfect information. It assumes that the proper response is optimisation. Yet many of the qualities that education exists to preserve, interpretive seriousness, intellectual risk, tolerance of uncertainty, durable attention, are precisely those most threatened by that framing. The problem is therefore not simply that AI might be used badly. It is that it arrives bearing a conception of educational good that institutions are already primed to accept.

This is why it is insufficient to debate AI at the level of technique alone. The current discourse often asks whether AI can improve marking, reduce workload, personalise instruction, or support accessibility. These are not trivial questions. But taken in isolation they keep us within the horizon of the existing system. They presuppose that the task is to make current arrangements function more effectively. The deeper question is whether those arrangements should continue to govern educational life in the first place. If the current system is already one in which learning is narrowed by quantification, teachers are burdened by performativity, institutions are disciplined by market logics, and students are shaped as self-monitoring optimisers, then AI will not correct those tendencies. It will most likely amplify them.

The point can be put more sharply. The real transformation did not begin with generative AI. It began earlier, when educational institutions were reimagined as organisations whose value must be demonstrated through data, whose problems must be solved through managerial intervention, and whose purposes must increasingly align with economic competitiveness. AI enters as a powerful tool precisely because that transformation has already occurred. It is the child of that world, not its conqueror. This is why talk of AI as a neutral or autonomous revolution is misleading. The infrastructure into which it is introduced, digital platforms, dashboards, predictive systems, data-rich student records, performance management, was built by earlier phases of reform. AI is the latest layer, not the first.

To say this is not to deny the technical novelty of current systems. It is to deny that novelty at the technical level settles the educational question. Socially and institutionally, AI is far more conservative than its rhetoric suggests. It assumes the continuation of the very order that made it legible in the first place. Its promises of personalisation, intelligent tutoring, and efficiency are often little more than updated expressions of a longer managerial desire to render education predictable, measurable, and governable.

That does not mean that no alternative use is imaginable. It means only that such alternatives will not emerge automatically from the technology itself. If educational institutions continue to operate within the neoliberal logic analysed by Ball, then AI will reinforce that logic. If they continue to think of learning as a measurable optimisation problem, then AI will appear as an obvious ally. If, however, institutions recover stronger conceptions of educational purpose, conceptions tied to judgement, inquiry, moral seriousness, and the cultivation of forms of human attention that cannot be reduced to throughput, then AI might be re-situated as a subordinate tool rather than a guiding model.

At this stage in the argument, the point is not yet to describe that alternative in detail, nor to offer a programme of refusal or reconstruction. It is simply to see the present situation more clearly. AI does not arrive in education as a mysterious destiny from outside. It crystallises tendencies already at work within post-industrial, neoliberal institutions. That is why the central issue is not whether educational systems should keep up with AI, but whether they can begin to think beyond the institutional and philosophical assumptions that currently make AI seem like the natural future of schooling and higher education.

Bell remains valuable here because he allows the present to be de-dramatised without being trivialised. What looks like rupture can be understood as intensification. Cuban remains valuable because he reminds us that schools are not infinitely pliable and that teachers mediate reform through practice. Ball remains valuable because he shows that institutions can nonetheless be reorganised around different values and different forms of visibility. Mechling remains valuable because he insists that students are not reducible to the systems that attempt to shape them. Taken together, these resources make it possible to say something more precise about AI in education. It is not an alien intelligence descending upon schooling. It is the latest and most sophisticated instrument through which a long-established order seeks to render learning more measurable, more governable, and more economically legible.

That is enough, at least for now, to shift the terms of discussion. The issue is no longer simply whether AI is useful, but what kind of education its usefulness presupposes.

In this light, Bell’s earlier analysis deserves one further look, because it sharpens the sense in which the present is historically continuous. In The Coming of Post-Industrial Society Bell distinguished industrial society from its successor not simply by the replacement of one technology by another, but by a transformation in the organising principle of social life. Industrial society, he argued, turned on the application of energy to machines. The steam engine mattered not because it was an ingenious object in itself, but because it allowed energy to be harnessed in a controllable way and therefore made mechanised production scalable. Once that happened, entire social relations were reorganised around productivity, mechanisation, and the factory system (Bell 1973). Bell’s larger point was that the next historical transition would follow a structurally similar pattern. Post-industrial society would not be defined mainly by heavier machines or more energy, but by the increasing centrality of theoretical knowledge, formal systems, and intellectual technologies.

This part of Bell’s argument now reads with unusual force. He stressed that much nineteenth-century innovation was driven not by highly formalised science but by inventors, engineers, and practical tinkerers. By contrast, twentieth-century technological development increasingly flowed from codified theoretical knowledge. Bell uses the movement from quantum theory to semiconductors and electronics as an emblem of this shift. The practical breakthrough is no longer separable from abstract knowledge. Scientific theory itself becomes productive. Universities, laboratories, and research systems therefore become central to the organisation of society, because they are no longer merely preserving and transmitting knowledge. They are generating the very conditions of future innovation. Once that happens, higher education becomes structurally indispensable, and schooling becomes the feeder system for a society organised around knowledge work.

This is one reason Bell’s account still matters for education. It helps explain why schools and universities became progressively more central to economic and political life long before AI became a public obsession. The expansion of higher education, the elevation of expertise, the rise of professional and technical occupations, and the increasing importance of symbolic and informational work all belong to the post-industrial formation Bell described. Artificial intelligence appears, from this angle, not as a sudden break but as a further codification of activities already central to post-industrial institutions. When language, reasoning, classification, and even limited forms of interpretation are rendered in calculable form, AI extends the same trajectory. It does not introduce knowledge as a productive force. It intensifies the formalisation of knowledge within systems already organised around it.

Bell’s reflections on scale are also relevant here. He argued that many social changes occur not because an entirely new principle appears from nowhere, but because the scale of interactions changes so dramatically that new systems of organisation become necessary. Once there are more actors, more information flows, more institutions, and more interdependence, complexity itself becomes a governing problem. Modern societies respond by constructing administrative, computational, and organisational mechanisms to manage that complexity. The contemporary digital environment is a perfect example. Educational systems now generate huge quantities of data, circulate across transnational policy networks, and rely on software platforms whose scale far exceeds earlier institutional forms. AI appears attractive partly because these systems have already become too complex, too data-rich, and too sprawling to manage comfortably by older means alone. In that sense AI is not merely a tool of convenience. It is a response to complexity that previous transformations have already created.

This observation is important because it removes a familiar illusion. It is easy to imagine that AI creates new forms of institutional complexity. More often, it is adopted because institutional complexity has already reached a point where AI appears as a plausible administrative response. Digital records, platform learning environments, global student markets, cross-institutional benchmarking, recruitment analytics, retention modelling, all these developments predate the current wave of generative AI. They form the informational environment within which AI can now operate. Bell helps show that the real precondition of AI is not sudden technical genius but a long social history in which knowledge, communication, and organisational life have already been rendered abstract enough to be computationally processed.

At the same time, Bell is also useful because he never treated technological change as socially self-interpreting. Post-industrial society was not for him a triumphalist story. It reorganised the relation between economy, politics, and culture, but it also intensified their tensions. The techno-economic order pushes toward rationalisation, efficiency, and system-building. The political order concerns legitimacy, control, and collective coordination. The cultural order remains tied to meaning, expression, judgement, and symbolic life. Education sits directly across these fault lines. That is why the issue is not simply whether AI can help institutions function. It is whether the increased codification of educational life leaves room for those aspects of learning and teaching that belong irreducibly to culture rather than to administration or market logic.

This is where the previous chapters begin to lock together more tightly. Chapter 1 showed empirically that AI often strengthens surface performance while weakening depth. Chapter 2 argued that human knowing depends on uncertain, effortful, socially mediated processes that resist quick formalisation. Chapter 3 showed that the institutions into which AI enters are already shaped by post-industrial codification, managerial governance, and performative accountability. The present chapter adds that AI does not merely sit within that trajectory. It depends on it. The neoliberal transformation of education did not simply make institutions harsher or more competitive. It also made them computationally legible. Once learning is represented through metrics, dashboards, data trails, and predictive categories, AI becomes possible as a mode of intervention. Without that prior remaking of education, AI could not now present itself as the obvious solution to institutional problems.

That is why the current rhetoric of revolution is so misleading. The so-called AI revolution in education is remarkably conservative in institutional form. It leaves the basic market structures, accountability regimes, platform dependencies, and managerial assumptions largely intact. It promises to solve problems of scale, cost, retention, and efficiency without asking whether the system that defines those as the central problems is itself worth defending. The result is that technologies that appear novel are often absorbed into older reform logics almost immediately. AI does not challenge the existing educational order at its roots. Much more often, it offers that order a new lease of life.

The same point can be made at the level of the student and the teacher. The entrepreneurial student produced by neoliberal reform was already expected to monitor performance, improve continuously, build a portfolio, manage risk, and maximise opportunities. AI does not invent that figure. It equips it. The entrepreneurial academic was already expected to produce measurable outputs, attract value, respond to metrics, and manage visibility. AI does not abolish that condition. It intensifies it, sometimes by speeding tasks up, sometimes by generating more data about those tasks, sometimes by tempting institutions to imagine that human labour can be more tightly managed because automated substitutes or supplements are now available. In both cases the technology appears as an accelerant, not a transformation of ends.

What matters, then, is not whether AI can do interesting things, clearly it can, but whether educational institutions will continue to let the values of measurability, standardisation, and optimisation determine the terms on which those things matter. Bell’s long view, Cuban’s historical scepticism, Ball’s institutional critique, and Mechling’s attention to the residual moral life of youth cultures all suggest that the answer is not predetermined. But they also suggest that the burden of proof should now be reversed. The question is not why anyone would hesitate before letting AI reorganise educational life. The question is why, given the history of these reforms, anyone should assume that its institutional trajectory will be benign unless educational purposes are first rethought at a much deeper level.

References

Bell, Daniel. The Coming of Post-Industrial Society: A Venture in Social Forecasting. New York: Basic Books, 1973.

Bell, Daniel. The Cultural Contradictions of Capitalism. New York: Basic Books, 1976.

Ball, Stephen J. “The Teacher’s Soul and the Terrors of Performativity.” Journal of Education Policy 18, no. 2 (2003): 215–228.

Ball, Stephen J. Global Education Inc.: New Policy Networks and the Neoliberal Imaginary. London: Routledge, 2012.

Cuban, Larry. Oversold and Underused: Computers in the Classroom. Cambridge, MA: Harvard University Press, 2001.

Cuban, Larry. Hugging the Middle: How Teachers Teach in an Era of Testing and Accountability. New York: Teachers College Press, 2009.

Mechling, Jay. On My Honor: Boy Scouts and the Making of American Youth. Chicago: University of Chicago Press, 2001.

Chapter 6

The argument can now be widened again. Earlier chapters have shown, first, that generative artificial intelligence often improves the visible surface of educational performance while weakening or bypassing some of the slower processes through which understanding is formed, second, that human learning depends upon a fragile cognitive architecture of curiosity, uncertainty, effort, social trust and temporally extended revision, and third, that AI enters institutions already shaped by post industrial codification and by the managerial and performative settlement analysed by Daniel Bell, Larry Cuban and Stephen Ball. What remains to be clarified is the full environment within which these developments occur. It is not enough to say that schools and universities have been affected by neoliberal reform. We need a fuller account of the ecology produced by that reform, the forms of truth it privileges, the kinds of selves it cultivates, the institutional arrangements it normalises, and the ways it repositions teachers, researchers and students inside education itself. Only then can we see why the current enthusiasm for AI fits so comfortably into the educational world it claims to transform.

Cuban’s work established one indispensable part of the picture. He showed that reformers habitually misunderstand schools when they imagine that large scale policy will descend transparently into classrooms and reshape teaching more or less as intended. The grammar of schooling is stubborn, teachers work inside complex social situations rather than technical systems, and reform is always mediated by local practice and institutional history (Cuban 2001). Stephen Ball approaches a related terrain from a different angle. Rather than beginning with the visible journey from policy to classroom, he asks a question that sits behind the reform agenda itself, what have we become. This question, drawn from Michel Foucault’s analyses of modern institutions, is directed not only at policy but at the transformation of educational actors and knowledge. Ball reformulates it in terms of the contemporary university and school system, asking what academics, teachers and institutions have become under neoliberalism and how the project of educational knowledge has been altered (Ball 2003; Ball 2012).

Ball’s approach is distinctive in part because it begins with lived disturbance rather than abstract modelling. Like Foucault, who described his work as emerging from cracks and tensions experienced within institutions, Ball starts from the felt transformation of academic and educational life (Foucault 1980). His well known account of the shift from a welfare academic to a neoliberal academic is not merely autobiographical colour. It illustrates a central claim, that neoliberalism is not only a structural condition but a way of life that shapes how individuals understand themselves, their work and their relations to others. It is therefore necessary to analyse not only policies and institutions but also subjectivity.

To do this, Ball organises his analysis around four interrelated elements, truth, governing, economy and subjectivity. Truth refers not simply to whether particular claims are correct, but to the processes through which certain forms of knowledge come to count as true. Governing refers to the shaping of conduct through dispersed mechanisms of power rather than formal authority alone. Economy refers to the penetration of market logics into institutional life. Subjectivity refers to the kinds of selves these arrangements produce. Neoliberalism, on this account, is not merely an economic doctrine but a configuration in which these elements are tightly interwoven.

A key clarification follows from this. Neoliberalism must be understood simultaneously as an external structure and as an internal mode of subject formation. Externally it consists of funding systems, accountability regimes, competitive pressures and market oriented reforms. Internally it shapes how individuals think about themselves as entrepreneurial actors responsible for their own success or failure. This dual character is central to Maurizio Lazzarato’s account of neoliberalism, which Ball adopts to describe contemporary academic and educational life. Lazzarato identifies five defining conditions, individualisation, inequality, insecurity, depoliticisation and financialisation (Lazzarato 2009).

Individualisation is perhaps the most visible. Individuals are encouraged to treat themselves as projects requiring continuous investment and improvement. In education this appears in the expectation that teachers, students and institutions must constantly optimise their performance and manage their own trajectories. Inequality functions alongside this process as both justification and incentive. Differences in attainment, ranking and recognition are presented as markers of merit rather than structural conditions, reinforcing competition across schools, universities and individuals.

Insecurity is equally central. Ball notes that contemporary universities are characterised by increasing precarity, with significant proportions of staff on fixed term or part time contracts (Ball 2012). This pattern extends into school systems, where accountability pressures and workload intensification create similar forms of instability. Yet insecurity is not only contractual. It is also temporal. Performance must be continually demonstrated. What one has achieved matters less than what one produces now. Research assessment exercises such as the Research Excellence Framework in the United Kingdom institutionalise this dynamic, ensuring that academic reputation must be repeatedly earned.

Depoliticisation and financialisation deepen this transformation. Depoliticisation refers to the presentation of educational decisions as technical rather than normative. Questions of educational purpose are displaced by questions of effectiveness, evidence and impact. Financialisation refers to the increasing organisation of education through economic calculations of cost, return and investment. Universities become sites of financial strategy, and schools increasingly operate within budgetary and competitive logics.

Ball’s analysis of truth provides a crucial entry point into these processes. Following Foucault, he shifts attention from what is true to how truth is produced. This allows him to distinguish between two forms of truth telling in academic life, the truths academics tell about themselves and the truths they produce about others. The first is evident in the proliferation of performance metrics. Citation counts, journal impact factors, grant income, teaching evaluations and institutional rankings produce numerical representations of academic worth. Roger Burrows captures this condition in his description of “living with the h index” (Burrows 2012).

These metrics do more than measure activity. They shape it. Decisions about research topics, publication venues and collaborations are increasingly influenced by their likely impact on measurable indicators. Academic identity becomes articulated through what Kevin Haggerty and Richard Ericson call the “data double”, a representation constructed from accumulated records and metrics (Haggerty and Ericson 2000). The curriculum vitae becomes a central site of this representation, expanding into a detailed account of productivity and impact that travels across institutional contexts.

These processes are not imposed unilaterally. Academics and teachers participate in them, curating their profiles and aligning their work with evaluative criteria. This is what makes neoliberal subjectivity distinctive. It operates through active self management rather than simple external control. Individuals become entrepreneurs of their own measurable worth.

The second dimension of truth telling concerns the production of knowledge about others. Ball draws on Raewyn Connell’s argument that educational knowledge is increasingly technicised under neoliberal conditions (Connell 2013). Research becomes oriented toward measurable effectiveness, with experimental methods gaining authority because they produce quantifiable results. Initiatives such as the Education Endowment Foundation in England exemplify this trend, promoting randomised controlled trials and cost effectiveness analyses to identify “what works” in education.

This shift introduces a new logic into educational decision making. Teaching strategies are evaluated as investments yielding measurable returns in student attainment. Ball describes this as an “economy of student worth”, in which data about performance guide resource allocation. Jenny Ozga and colleagues characterise this broader process as governing by numbers, where quantitative indicators become central instruments of policy and governance (Ozga, Dahler-Larsen, Segerholm, and Simola 2011).

The consequences are both epistemic and political. Educational questions that are fundamentally normative are recast as technical problems. The language of evidence and effectiveness displaces debates about educational purpose. This does not eliminate politics but obscures it. Decisions appear as neutral responses to data rather than value laden choices.

At this point Ball’s analysis aligns closely with Cuban’s earlier work. Cuban showed that reforms fail when they treat schools as technical systems rather than complex social environments. Ball shows that neoliberalism actively produces the illusion that schools can be treated as such systems. The language of measurement, performance and evidence renders schools legible as sites of intervention, even though their internal complexity remains.

Ball’s reflections on The Micro-Politics of the School sharpen this point. His earlier work described schools as conflictual and normatively charged institutions. When revisiting it, he argues that it captured a moment of transition from welfare state schooling to neoliberal schooling (Ball 1987). This transition is both epistemic and ontological. It changes not only how schools are understood but what they are. Teachers, students and leaders are remade as different kinds of subjects.

Historically, this shift is tied to broader political transformations associated with the rise of the New Right. In England, the Education Reform Act 1988 marked a decisive moment, introducing market mechanisms, centralised curriculum control and new accountability systems (Ball 2008). The preceding teachers’ disputes of the mid 1980s contributed to a discourse of crisis that legitimised these reforms. Teachers were represented as unaccountable and politicised, creating conditions for their repositioning as implementers rather than interpreters of educational purpose.

This repositioning has profound implications. Ball draws on Miranda Fricker’s concept of epistemic injustice to describe the diminishing credibility of teachers as knowers (Fricker 2007). Testimonial injustice occurs when teachers’ voices are granted less authority than those of policymakers or experts. Hermeneutic injustice occurs when the language available within the system no longer allows teachers to articulate their role as moral and political agents. The teacher becomes primarily a technical operative.

The school itself is similarly transformed. It is increasingly understood as a firm operating within a competitive market. Budget management, performance metrics and accountability systems reshape its internal logic. The rise of “school leadership” as a field reflects this shift, with headteachers and executive leaders functioning as managers of performance and strategy rather than custodians of a shared educational ethos.

Performativity, as Ball defines it, gathers these developments together. It is a system of regulation through comparison, measurement and display (Ball 2003). It operates not only externally but internally, shaping emotions, motivations and identities. Teachers experience stress, anxiety and exhaustion as they navigate constant evaluation. Recruitment and retention difficulties in many systems reflect these pressures.

Ball’s analysis then extends beyond national contexts to the global education industry. Education policy and practice increasingly circulate through transnational networks involving corporations, consultancies and international organisations. Policy ideas become commodities. Firms such as Pearson, McKinsey and others participate in shaping and selling educational reform models. This process reflects what geographers describe as policy mobilities, where ideas travel across contexts through networks of actors and institutions (Ball 2012).

The global education market has grown substantially, with estimates placing its value in the hundreds of billions of dollars. Corporations operate across multiple levels, from local school services to international policy advising. Private equity firms invest in education providers, treating schools and universities as assets. Public and private boundaries blur as states act as market makers, creating conditions for private participation.

This global dimension reinforces earlier dynamics. Educational knowledge is increasingly produced outside schools and universities, packaged by corporations and disseminated through policy networks. Teachers and institutions become consumers of externally generated expertise. Education becomes integrated into global circuits of capital.

Across these layers, a coherent pattern emerges. Marketisation, quantification, globalisation and subjectification are not separate processes but elements of a unified transformation. Education is reconfigured as a system of measurable outputs, comparable performances and investable opportunities. This transformation provides the conditions under which AI technologies appear both necessary and desirable.

The significance for the present argument is clear. AI does not enter a neutral educational landscape. It enters a neoliberal ecology already structured by data, metrics and performance. Its apparent benefits, efficiency, scalability and optimisation align closely with existing institutional logics. This alignment helps explain its rapid adoption.

At the same time, the earlier insights of Cuban remain crucial. Schools are not simply technical systems. Teachers continue to mediate, adapt and reinterpret reforms. The tension between lived educational practice and system level representation persists. However, the ecology within which this tension operates has been transformed.

The next chapter turns to the work of Tim Brighouse and Mick Waters in order to recover a normative account of schooling that can respond to this ecology. The purpose is not to reject the analysis developed here but to ask what education might still be for within it, and whether forms of leadership, curriculum and professional judgement can resist the reduction of schooling to performance, data and market logic.

Bibliography

Ball, Stephen J. 1987. The Micro-Politics of the School. London: Methuen.

Ball, Stephen J. 2003. “The Teacher’s Soul and the Terrors of Performativity.” Journal of Education Policy 18 (2): 215–228.

Ball, Stephen J. 2008. The Education Debate. Bristol: Policy Press.

Ball, Stephen J. 2012. Global Education Inc.: New Policy Networks and the Neoliberal Imaginary. London: Routledge.

Burrows, Roger. 2012. “Living with the h-index? Metric Assemblages in the Contemporary Academy.” The Sociological Review 60 (2): 355–372.

Connell, Raewyn. 2013. The Neoliberal Cascade and Education: An Essay on the Market Agenda and Its Consequences. London: Routledge.

Cuban, Larry. 2001. Oversold and Underused: Computers in the Classroom. Cambridge, MA: Harvard University Press.

Foucault, Michel. 1980. Power/Knowledge: Selected Interviews and Other Writings 1972–1977. New York: Pantheon.

Fricker, Miranda. 2007. Epistemic Injustice: Power and the Ethics of Knowing. Oxford: Oxford University Press.

Haggerty, Kevin D., and Richard V. Ericson. 2000. “The Surveillant Assemblage.” British Journal of Sociology 51 (4): 605–622.

Lazzarato, Maurizio. 2009. “Neoliberalism in Action: Inequality, Insecurity and the Reconstitution of the Social.” Theory, Culture & Society 26 (6): 109–133.

Ozga, Jenny, Peter Dahler-Larsen, Christina Segerholm, and Hannu Simola, eds. 2011. Fabricating Quality in Education: Data and Governance in Europe. London: Routledge.

Chapter 7

This chapter begins from a simple but increasingly neglected question, what is school for. The question has become harder to ask because the language surrounding schooling is now so often borrowed from elsewhere, from management, markets, behavioural compliance, data systems, and more recently from the rhetoric of digital transformation and artificial intelligence. In that climate, schooling is easily described as a delivery mechanism, a sorting device, or a productivity system. What fades from view is the older and more demanding idea that education is a moral, civic, and human practice concerned with the formation of persons, the development of powers, and the widening of freedom. Before one can judge what role AI might or might not have in schools, that prior question has to be recovered. If the purposes of schooling remain unclear, external agendas rush in to define them, and efficiency begins to masquerade as excellence, measurement as judgment, and technological novelty as progress.

The previous chapter described the institutional ecology within which schooling now operates, drawing on the work of Stephen Ball, Larry Cuban, and others to show how centralisation, performativity, and market logics have reshaped educational life (Ball 2003; Ball 2012; Cuban 2001). That account explained why AI appears so readily compatible with contemporary systems. But it did not yet show what education is for from within the practice itself. This chapter therefore turns deliberately away from AI as its starting point and instead examines two figures, Mick Waters and Tim Brighouse, whose reflections help restore that missing ground. They are treated together because they offer complementary perspectives on the same problem. Waters provides a close description of what schooling now feels like from inside, while Brighouse supplies the moral and political grammar needed to judge that condition.

Both write out of long experience in English schooling. Both accept the need for reform, the importance of evidence, and the urgency of improving outcomes for those poorly served by existing systems. But they begin from education itself rather than from the priorities of those who would instrumentalise it. This matters because much contemporary discussion of AI in education begins in the opposite direction. It begins with technical capability and asks how schooling might adapt. Waters and Brighouse reverse that order. They begin with schooling, its purposes, its failures, and its possibilities, and only then ask what kinds of change are worth pursuing. That reversal is not rhetorical. It is methodological. It establishes that technology cannot supply the criteria by which it should be judged.

Waters’ contribution is especially valuable because it captures the lived texture of contemporary schooling. His reflections, drawn from decades of work including his role at the Qualifications and Curriculum Authority and subsequent advisory work, show how curriculum, assessment, inspection, and institutional anxiety shape everyday practice. In one revealing episode, he describes visiting a primary school where pupils had completed a substantial project on the Second World War. When asked how long the war lasted, a pupil replied that it lasted a term. The answer was not ignorance but a reflection of how schooling had organised the material. Vast historical events are compressed into administratively manageable units. The curriculum imposes its own temporal order on reality.

This anecdote is not trivial. It shows how schooling structures knowledge in ways that can distort understanding while still producing impressive outputs. It also illustrates a broader point that runs through Waters’ work. Schools increasingly organise knowledge in forms that are administratively legible rather than educationally meaningful. The same pattern appears in the wider system. Schools describe themselves through inspection grades, performance measures, and demographic indicators. The language of education is displaced by the language of metrics.

Waters situates this within a historical trajectory. Before the late twentieth century, English schools exercised significant curricular autonomy. The Education Reform Act 1988 marked a decisive shift, introducing a national curriculum and testing regime. Subsequent reforms extended this logic. National literacy and numeracy strategies prescribed teaching methods, and later reforms specified content and training frameworks in increasing detail. Waters describes this as a progressive movement inward, from the school gate, to the classroom, to the pupil’s desk, and ultimately into the teacher’s mind. The cumulative effect is not simply structural but psychological. Teachers begin to assume that professional initiative lies elsewhere.

The consequences are visible in school culture. Waters describes an environment in which institutions organise themselves around inspection regimes, where staff monitor official signals of impending inspection, rehearse responses, and orient their work toward compliance. The focus shifts from educational quality to institutional survival. This produces a narrowing of purpose. Schools learn to manage the system rather than to educate within it.

This narrowing has direct consequences for pupils. Waters draws attention to rising absence rates in England following the COVID pandemic, with large numbers of pupils missing significant portions of schooling. Rather than attributing this solely to individual or family factors, he asks what features of schooling itself might be contributing. Young people live in a world shaped by digital technologies, personalised services, and constant access to information. Schooling, by contrast, remains organised around uniform timetables and cohort progression. The mismatch becomes visible to pupils themselves. Schools appear as closed systems in a world that is otherwise open and responsive.

Waters captures this through the metaphor of the jukebox and the streaming service. Schools resemble jukeboxes offering a fixed selection, while the wider world operates like a streaming platform offering endless choice. His point is not that education should become consumer driven, but that pupils increasingly experience schooling as restricted and disconnected from their lived reality. Engagement declines not because young people reject learning, but because they struggle to see the point of what is offered.

At the centre of this tension lies the assessment system. In England, GCSE examinations distribute pupils across grade categories in ways that ensure some will be classified as low achievers. Waters argues that this builds failure into the system. Some pupils recognise this early and disengage. Others succeed but find little meaning in the process. The system rewards performance but not necessarily understanding.

These structural pressures intersect with changes in the teaching profession. Recruitment difficulties, workload pressures, and declining autonomy have altered professional identity. Teachers are increasingly positioned as implementers of externally defined practices. The sense of teaching as a moral and intellectual vocation weakens. What remains is often a culture of compliance.

Waters does not present this as a terminal decline. He identifies what he calls “stirrings in schools”, moments where teachers and leaders begin to reassert professional judgment and rethink their purposes. These stirrings are fragile but significant. They suggest that alternative trajectories remain possible.

His proposals reflect this. He advocates forms of schooling that are more open, more connected to community life, and more responsive to individual development. Programmes such as extended projects and personalised learning pathways are intended to reconnect knowledge with real contexts. His notion of an “open school” involves shared resources and community engagement rather than isolated institutional effort. Crucially, these proposals retain the centrality of the teacher. Technology may support, but it does not replace, the relational core of education.

This emphasis leads directly to Brighouse, whose work provides the normative framework needed to interpret these observations. Brighouse’s starting point is similar but more explicit. Education has a moral purpose irreducible to economic utility or bureaucratic control. Schools exist to develop human powers, widen freedom, and enable individuals to contribute to the world. This is not an abstract claim. It is grounded in decades of work on school improvement and urban education.

Brighouse draws on the tradition of school effectiveness research, including work associated with Rutter, to challenge earlier deterministic views of intelligence (Rutter et al. 1979). Schools make a difference. But he is careful to distinguish effectiveness from improvement. Knowing that some schools perform better does not explain how improvement occurs or what values are advanced in the process. This distinction is crucial. It prevents educational debate from collapsing into a search for techniques that produce outputs without examining what those outputs mean.

His own experience reinforces this. He describes attending two very different schools as a child, one oppressive and one humane, and recognising that schools themselves shape experience profoundly. This insight drives his later work. Improvement depends on understanding the conditions under which schools enable or constrain human development.

For Brighouse, these conditions are moral as well as technical. He draws on a tradition associated with William Temple to argue that education must treat persons not merely as they are but as they might become. This distinction is central. A system oriented toward current performance risks fixing individuals within existing categories. Education, by contrast, must remain open to transformation and emergence.

This has direct implications for contemporary systems. Brighouse argues that English education has undergone an extraordinary centralisation of power, with successive governments expanding control over curriculum, assessment, and teaching practice (Ball 2008). This has eroded professional trust and reduced teachers to operatives within a system. The language of choice, diversity, and autonomy often masks the expansion of market mechanisms and competitive pressures.

He evaluates these developments through a set of practical questions. Does a reform intensify market forces, centralise power, diminish trust in teachers, rest on weak evidence, or fail those most disadvantaged? These criteria provide a framework for judging educational change. They are particularly relevant to AI, where claims of efficiency and personalisation often obscure deeper issues of power, equity, and purpose.

Brighouse also emphasises the importance of context. There is no single model of urban education. Different communities present different challenges. Improvement depends on local knowledge and professional judgment. This stands in tension with the idea of scalable solutions. AI systems, often presented as universally applicable, risk ignoring this complexity.

Another key theme in his work is the role of expectation. Schools that assume certain pupils will fail often produce that outcome. Improvement requires resisting such assumptions and maintaining belief in the possibility of growth. This connects to his critique of the concept of the “average child”. Statistical averages may exist, but educationally they can become limiting categories that constrain aspiration.

Brighouse’s account of improvement is therefore relational and cultural rather than purely technical. It depends on leadership that generates energy, on staff development that fosters inquiry, on collaboration among teachers, and on engagement with pupils and communities. Data plays a role, but as a tool for learning rather than surveillance.

When Waters and Brighouse are read together, a coherent picture emerges. Contemporary schooling has been reshaped by managerial and market logics that narrow its purposes and erode professional agency. At the same time, there remain resources within the system for renewal, grounded in relationships, judgment, and shared purpose.

This picture provides the necessary context for evaluating AI. The issue is not whether AI can perform certain functions. It clearly can. The issue is how those functions interact with the existing ecology of schooling. In a system already oriented toward metrics and compliance, AI is likely to reinforce those tendencies. It offers efficiency where what is needed is clarity of purpose. It offers personalisation where what is needed is recognition. It offers support while potentially displacing professional judgment.

The contrast between algorithmic personalisation and human recognition is especially important. AI systems personalise by analysing patterns in data. Teachers recognise individuals through sustained relationships. The two are not equivalent. The former may optimise performance within defined parameters. The latter engages with the person as a developing being.

The broader implication is that technological capability cannot determine educational value. The fact that AI can generate text, automate feedback, or analyse data does not establish that these functions should occupy a central place in schooling. The relevant question is always what educational good is being served and whether the technology strengthens or weakens the conditions under which that good can be realised.

Waters and Brighouse therefore serve as guides to the deeper terrain on which the AI debate must be conducted. They reveal that the central issues in education concern purpose, professionalism, and institutional design. AI enters a field already shaped by these factors. It does not resolve their tensions but intensifies them.

The next chapter returns explicitly to AI in education, drawing on contemporary research to examine how these dynamics are already playing out. The analysis developed here provides the evaluative framework for that discussion. Without it, the debate risks remaining at the level of technical possibility. With it, the question becomes more demanding. Not what can AI do, but what should education be.

Bibliography

Ball, Stephen J. 2003. “The Teacher’s Soul and the Terrors of Performativity.” Journal of Education Policy 18 (2): 215–228.

Ball, Stephen J. 2008. The Education Debate. Bristol: Policy Press.

Ball, Stephen J. 2012. Global Education Inc.: New Policy Networks and the Neoliberal Imaginary. London: Routledge.

Cuban, Larry. 2001. Oversold and Underused: Computers in the Classroom. Cambridge, MA: Harvard University Press.

Rutter, Michael, Barbara Maughan, Peter Mortimore, Janet Ouston, and Alan Smith. 1979. Fifteen Thousand Hours: Secondary Schools and Their Effects on Children. London: Open Books.

Temple, William. 1942. Christianity and Social Order. London: Penguin.

Chapter 8

This chapter begins from a question that has become increasingly difficult to ask, not because it is obscure, but because it has been steadily displaced. What is school for. The difficulty lies not in the absence of answers, but in the dominance of other, louder languages that now speak in its place. Schooling is now routinely described in terms borrowed from management, from markets, from behavioural compliance, from data systems, and increasingly from the rhetoric of digital transformation and artificial intelligence. In such a climate, it becomes natural to speak of schools as delivery systems, productivity engines, sorting mechanisms, or platforms for distributing content. What fades from view is a more demanding and older idea, that education is a moral, civic and human practice concerned with the formation of persons, the development of powers, the widening of freedom, and the cultivation of a shared world.

In the previous chapter I traced the emergence of what I called a neoliberal educational ecology, a system in which performativity, measurement, comparison and centralised control have come to shape institutional life. Within such a system, the arrival of artificial intelligence appears not as a rupture but as a continuation. It promises to do more efficiently what the system already values, to generate outputs, track performance, personalise instruction in measurable ways, and scale provision. Yet before we can sensibly ask what role AI might play, something more fundamental has to be recovered. If the purposes of schooling remain unclear, then technological capability will be mistaken for educational value. Efficiency will stand in for excellence, measurement for judgment, and innovation for improvement.

For that reason I want, in this chapter, to step back from AI as a starting point and instead return to two figures whose work helps recover a language of educational purpose that has been steadily eroded, Tim Brighouse and Mick Waters. For readers unfamiliar with English education, it is important to be clear about who these figures were and why their views carry weight. They are not commentators who have observed schooling from a distance, nor theorists working in abstraction. They are individuals who spent decades inside schools, local authorities and national policy structures, shaping, contesting and living through the reforms whose consequences we are now trying to understand.

Tim Brighouse was one of the most influential educational leaders in England in the late twentieth and early twenty first centuries. He began his career as a classroom teacher, working directly with pupils, before becoming a deputy headteacher and then moving into local authority leadership. A local authority in England is responsible for overseeing schools within a geographical area, supporting them, allocating resources, and shaping local educational strategy. Brighouse became Chief Education Officer in Oxfordshire and later in Birmingham, one of the largest and most complex education systems in the country, with schools serving highly diverse and often disadvantaged communities.

He later became Schools Commissioner for London, where he led a major reform programme known as the London Challenge. This initiative brought together schools, local authorities and government in a collaborative effort to improve outcomes across the capital. It is widely regarded as one of the most successful school improvement programmes in England, transforming London’s schools from among the lowest performing to among the highest within a relatively short period. For readers outside the system, it is important to stress that this was not achieved through punitive accountability or wholesale restructuring, but through building professional capacity, fostering collaboration between schools, and maintaining a sustained focus on teaching and learning.

I knew Tim personally. He was a mentor to me, someone whose thinking shaped how I came to understand education. What struck me most was not simply what he argued, but how he carried himself in the system. He had authority, but he wore it lightly. He moved easily between classrooms and policy meetings. He remembered teachers’ names, wrote to them, visited schools constantly. He believed, and showed through his practice, that improvement begins with valuing those who do the work. That might sound like a minor point, but in a system increasingly organised around targets and metrics, it is anything but minor. It reflects a deeper conviction that education is, first and last, a human practice.

Mick Waters’ career runs alongside Brighouse’s but offers a slightly different vantage point. Like Brighouse, he began as a teacher and later became a headteacher, responsible for leading a school and managing its staff and pupils. He then moved into local authority work, where he worked closely with Brighouse in Birmingham, before becoming Chief Education Officer in Manchester. From there he moved into national policy as Director of Curriculum at the Qualifications and Curriculum Authority, the body responsible for designing and overseeing the National Curriculum in England.

For readers unfamiliar with this, the National Curriculum is the framework that determines what subjects are taught in schools, what knowledge is included, and how learning is structured across different age groups. To be responsible for this curriculum is to shape, in a very direct sense, the intellectual experience of an entire generation. Waters therefore occupied a position of considerable influence. What makes his later reflections particularly valuable is that they come from someone who has not only analysed the system but helped design it.

Together, Brighouse and Waters provide a perspective that is both insider and reflective, practical and conceptual. They have taught children, led schools, managed local systems, advised governments and shaped national policy. When they speak about what has gone wrong in education, they do so not as observers but as participants, individuals who have seen reforms from within, observed their effects over time, and, in some cases, contributed to their design. This gives their work a particular authority. It is grounded in experience rather than speculation.

What emerges most strongly from their work is a conception of education that begins not with systems, technologies or policy frameworks, but with the centrality of the teacher. Their starting point is both simple and demanding, the teacher makes the most difference. For readers unfamiliar with educational research, this claim may sound like a commonplace, but it runs directly against much contemporary policy thinking, which assumes that improvement can be achieved through system level reform alone. Brighouse and Waters insist that what happens in classrooms, the interaction between teacher and pupil, is the core of educational life. Systems can support or hinder that interaction, but they cannot replace it.

This point matters because it shifts the focus of attention. Instead of asking how we can redesign systems, introduce new technologies, or implement new policies, we are asked to consider what enables teachers to do their work well. That work involves far more than delivering content. It involves making judgments about how to present ideas, how to respond to pupils, how to sustain attention, how to build understanding over time. It is relational, interpretive and context dependent. It cannot be reduced to a set of procedures.

From this starting point, a different conception of education begins to emerge. Education is not primarily about transmitting information or producing measurable outcomes. It is about inducting the young into a world. That means helping them move from the immediate environment of home and family into a wider shared world of knowledge, culture and social life. It means equipping them with the intellectual tools needed to understand that world and the dispositions needed to participate in it.

This is where the earlier discussion of human cognition becomes relevant. Human learning is not simply the accumulation of information. It involves curiosity, uncertainty, struggle and gradual understanding. It depends on the ability to engage with difficulty, to tolerate not knowing, and to work through problems over time. These processes are not incidental. They are central to how understanding develops. They require time, attention and support. They are shaped by relationships.

When we look at contemporary schooling through this lens, it becomes clear that many current practices sit uneasily with these processes. Waters captures this through his descriptions of curriculum and classroom life. He notes how knowledge is broken into units that can be easily taught and assessed, how topics are compressed into short time frames, and how pupils come to understand the world through the structures imposed by schooling rather than through sustained engagement with it.

His example of a pupil who believes that the Second World War lasted a term illustrates this vividly. The issue is not simply that the pupil is mistaken, but that the structure of schooling has reshaped how time itself is understood. Historical events are experienced not as extended processes but as curriculum segments. The organisation of schooling has altered the way reality is encountered.

This pattern is reinforced by the broader culture of performativity. Schools describe themselves in terms of inspection grades and performance data. Teachers are evaluated through metrics. Pupils are organised into cohorts whose progress is tracked and compared. The language of education gives way to the language of performance.

Brighouse situates these developments within a broader historical trajectory. He describes how, over several decades, power has become increasingly centralised within the education system. Decisions that were once made by teachers or local authorities are now made at national level. At the same time, market mechanisms have been introduced, encouraging schools to compete for pupils and resources. Accountability systems have intensified scrutiny, requiring schools to demonstrate performance through measurable outcomes.

For readers unfamiliar with these developments, the key point is that schooling has come to be governed by a combination of central control and market pressure. Schools are required to meet national standards while also competing with each other. This creates a system in which performance is constantly monitored and compared. The effects of this are not only structural but cultural. Teachers begin to see themselves as implementers of externally defined requirements. Schools focus on their position within the system. Pupils experience education as a sequence of tasks to be completed.

At the centre of this lies the structure of assessment. Standardised examinations play a major role in determining pupils’ futures. They classify pupils into categories that shape their opportunities. For some pupils, this leads to disengagement, particularly if they come to believe that success is unlikely. For others, it leads to a form of compliance, working to meet the requirements of the system without engaging more deeply with learning. In both cases, the broader purposes of education are compromised.

Waters connects this to patterns of disengagement, including rising absence rates. His question is not simply why pupils do not attend school, but why schools fail to engage them. In a world where young people have access to a wide range of information and experiences through digital technologies, schooling can appear rigid and disconnected. His metaphor of the jukebox captures this contrast. Schools offer a fixed selection of content, while the wider world offers something far more open and responsive.

Brighouse’s response is to return to the moral purpose of education. He argues that education should be judged not by its efficiency or its outputs, but by what it does for individuals, how it develops their capacities, and how it expands their freedom. This leads him to articulate criteria for evaluating reform. Does it increase trust in teachers or reduce it. Does it distribute power or centralise it. Does it rest on sound evidence. Does it improve the experience of those least well served by the system.

These questions become particularly important when we turn back to AI. Much current discussion presents AI as a solution to educational problems, promising to personalise learning, improve efficiency and enhance outcomes. Yet these claims often assume that the goals of education are already settled. They take for granted that learning can be optimised, that performance can be measured, and that improvement can be achieved through better data.

From the perspective developed in this chapter, these assumptions need to be questioned. If education is understood as a human practice involving relationships, judgment and the gradual development of understanding, then the introduction of AI raises different issues. The question is not simply what AI can do, but how its use aligns with or disrupts these processes.

AI systems are particularly effective at producing answers, identifying patterns and providing immediate feedback. These capabilities are valuable in certain contexts. However, they also have the potential to alter the conditions under which learning occurs. If students rely on systems that provide answers quickly, they may engage less with the underlying problems. If learning is organised around optimisation, the processes of exploration and reflection may be reduced.

This connects back to the earlier discussion of cognitive architecture. Human learners require engagement with difficulty. They learn through attempting, failing, revising and understanding. Systems that remove these elements risk undermining the development of deeper understanding.

At the same time, it is important to avoid a simple opposition between human and technological approaches. Brighouse and Waters do not reject technology outright. They recognise that tools can support teachers, reduce administrative burdens and provide access to resources. The issue is not whether technology is present, but how it is used. When technology becomes the organising principle of schooling, rather than a tool within it, the nature of education begins to shift.

What their work offers, then, is not a rejection of change but a different orientation. Education should be organised around relationships, professional judgment and shared purpose. Systems should support these elements rather than undermine them. Reform should be guided by clear purposes rather than by technical possibility.

For me, returning to Brighouse is a way of recovering a way of thinking about education that is increasingly difficult to sustain. His work reminds me that schools are not simply institutions to be managed, but communities in which people learn together. It reminds me that teaching is not simply the delivery of content, but a practice involving judgment, care and attention. It reminds me that education is about opening up possibilities rather than closing them down.

As the argument of the book moves forward, returning to AI in the next chapter, this reminder becomes essential. Without it, the discussion risks being framed entirely in terms set by technology. With it, the question becomes more demanding. Not what AI can do, but what education is for, and whether AI serves or undermines that purpose.

This chapter therefore does not conclude the argument but reorients it. It brings back into view a set of educational values that have been obscured, grounds the discussion in the experience and judgment of those who have worked within the system, and establishes the framework within which the next stage of the analysis will proceed.

Chapter 9

The previous chapter returned me to two voices I regard as indispensable when thinking about schooling in England, Tim Brighouse and Mick Waters. I turned to them because I wanted to recover something that has become difficult to say plainly in educational debate, namely that schools are not first of all delivery systems, performance engines, or platforms for the distribution of measurable outputs. They are moral and civic institutions concerned with the formation of persons, the development of powers, the widening of freedom, and the opening of a shared world. I also turned to them for a more personal reason. Tim Brighouse was not simply an educational thinker whose work I admired from a distance. He was a mentor to me, and his recent death has made the return to his voice feel both intellectually necessary and personally urgent. To hear him again, and to place him alongside Waters, is to be reminded of a way of speaking about education that once felt ordinary and serious, but now too often sounds almost countercultural. That is precisely why the next move in the argument matters. Once one has recovered that richer account of schooling, one can read the contemporary literature on artificial intelligence in education with greater clarity and with far less susceptibility to fashion.

This chapter therefore follows directly from the previous one. It does not ask from scratch whether AI is good or bad. That is too blunt a question and, in any case, not especially illuminating. The more useful question is what the rapidly expanding research on generative AI in education reveals when read against an account of schooling informed by Brighouse and Waters. If they are right that education depends upon teacher judgement, professional trust, curriculum coherence, sustained encounter with difficulty, the cultivation of hope, the refusal to reduce children to present performance profiles, and the moral seriousness of schools as public institutions, then what exactly do the recent studies of AI in education show? My argument in this chapter is that, taken together, they show a worrying pattern of misalignment. They do not simply reveal a new tool entering a neutral institutional space. They reveal a technology arriving in educational systems already shaped by performativity, market logics, and administrative simplification, and then deepening the very tendencies that Brighouse and Waters most strongly resisted.

The first thing to say is that the speed of the recent AI turn in educational research is itself revealing. Before the public release of large language models into everyday use after 2022, artificial intelligence in education was often treated as a relatively specialised field concerned with adaptive learning systems, intelligent tutoring systems, educational data mining, and learning analytics. Those were already significant developments, but they were largely discussed within specific research communities and often at some distance from the everyday public life of schools. The appearance of generative systems such as ChatGPT changed this almost overnight. Suddenly, the field expanded at extraordinary speed. Bibliometric overviews, such as Kai Dai’s mapping of generative AI research in higher education, show a rapidly growing body of publications devoted to personalised learning, AI assisted writing, automated assessment, adaptive support, and new forms of curriculum and pedagogical design. That growth matters not only because it indicates scholarly interest, but because it demonstrates how quickly educational discourse can be pulled into a technological present tense. The technology appears, institutions react, research proliferates, and only later does serious educational judgement begin to ask what exactly is being assumed about learning, knowledge, and human development.

That lag between technical possibility and educational judgement is one of the central themes of this book. It has already appeared in other forms. Daniel Bell showed how modern societies generate intellectual technologies that codify knowledge into systems and procedures. Larry Cuban showed how schools are repeatedly confronted by waves of enthusiasm surrounding technologies that promise transformation but rarely understand the realities of classroom life. Stephen Ball showed how educational systems have already been remade through performativity, metrics, and neoliberal rationalities that make them especially susceptible to managerial solutions. Brighouse and Waters then recovered an alternative educational grammar in which the teacher, the child, the curriculum, and the purposes of schooling are understood quite differently. The contemporary AI literature needs to be read in light of all that. Otherwise it is too easy to mistake the field’s rapid expansion for evidence of educational necessity.

This is why the tone of much of the recent research matters as much as its findings. Reviews by researchers such as Giannakos and his collaborators often present generative AI as offering unprecedented opportunities for personalised support, immediate explanation, on demand practice materials, and conversational tutoring. All that is true at the level of capability. Large language models can indeed generate explanations, examples, summaries, quiz items, and code with astonishing fluency. But the educational question is not whether they can do these things. The question is what happens to learning when these capacities are embedded in environments already predisposed to privilege speed, output, and measurable performance. Giannakos and others are right to note that the distinction between using AI as a support for exploration and using it as a substitute for reasoning is crucial. What the literature increasingly shows, however, is that the second use is not an occasional pathology. It is a strong default tendency, especially in educational settings already structured by time pressure, assessment pressure, and instrumental conceptions of success.

The studies of student behaviour in computing education are especially useful here because they make the problem vividly visible. Work by Zastudil and colleagues on programming courses shows students using generative AI to explain concepts, identify bugs, draft solutions, and complete tasks more quickly. On one level, this can look like support. Students who would otherwise be stuck can move forward. Frustration is relieved. Progress becomes possible. Yet the same studies repeatedly record a corresponding anxiety on the part of both students and teachers. If the system generates the code, or proposes the structure of the solution, or offers the explanation before the learner has really inhabited the problem, what exactly has been learned? That question matters far beyond programming. It returns us directly to the educational values defended by Brighouse and Waters. They insist that learning is not the efficient acquisition of finished outputs. It is a developmental process in which understanding is formed through struggle, encounter, judgement, feedback, and time. If AI collapses that process into managed solution retrieval, then it does not merely help students complete work. It changes what the work is.

Keuning and colleagues sharpen this point by distinguishing more reflective from more passive forms of AI use. Some students interrogate outputs, challenge errors, ask follow up questions, and use AI as a springboard for their own reasoning. Others accept the generated material with minimal modification. The contrast is important because it prevents simplistic claims. The technology is not uniform in its effects. Yet this nuance should not lead us into complacency. What matters is not simply that reflective use is possible, but that passive use is tempting, efficient, and in many contexts structurally encouraged. In an educational culture that rewards completion, visible performance, and the production of acceptable answers, the shortest path will often dominate. Students do not need to be lazy or dishonest for this to happen. They only need to be rational actors within an institutional ecology already oriented toward output. This is exactly the kind of ecology that Ball described, and exactly the kind of system Brighouse and Waters warned had already moved too far away from the moral purpose of schooling.

The chapter can therefore move at this point to a more precise claim. Generative AI does not only pose a challenge because it enables cheating or confuses authorship. It poses a deeper challenge because it weakens the connection between visible performance and internal understanding. Educational institutions have historically depended on a rough assumption that the production of an essay, an argument, a solution, or an explanation normally reflects a learner’s own engagement with a problem. This assumption was always imperfect, but it was strong enough to sustain assessment and pedagogy. Generative AI complicates it fundamentally. A polished answer may now conceal a thin or fragile cognitive relation to the content. This is one reason André Barcaui’s work is so important. His experimental findings suggest that students using ChatGPT can produce correct work more efficiently while retaining less knowledge later. The output improves, the underlying understanding weakens. That result deserves to be read not as a narrow technical curiosity but as a profound educational warning. If institutions respond mainly to the visible surface of performance, they may congratulate themselves on smoother production while presiding over a thinning of genuine learning.

The significance of this is clearer when placed alongside the earlier chapter on cognitive architecture. There I argued that human learning is not merely a matter of possessing correct information. It depends upon curiosity, uncertainty, prediction error, exploratory action, social trust, and the gradual stabilisation of knowledge through memory and reflection. Generative AI interacts directly with that architecture by removing difficulty too quickly. If a problem begins to produce the curiosity, uncertainty, or mild frustration that would ordinarily drive investigation, the system is already there with an answer. From one perspective that is helpful. From another, it is educationally dangerous, because it displaces precisely the phase in which understanding begins to form. The answer arrives before the learner has properly encountered the question as a question.

This helps explain why recent research on passive and thoughtless adoption matters so much. Hou and colleagues have shown that students do not engage with generative AI in a single way. Some use it cautiously and reflectively. Others adopt what the researchers call passive or thoughtless patterns, directly incorporating machine generated content into their own work. That vocabulary is revealing. The problem is not simply use, but use that substitutes uptake for inquiry. When a learner receives the answer before they have developed the relevant curiosity, hesitation, hypothesis, or interpretive effort, the whole temporal structure of learning changes. Knowledge begins to look like something delivered from outside rather than something formed through engagement. This is where the chapter’s argument begins to connect explicitly back to Brighouse and Waters. Both men are deeply concerned with what happens when schools cease to create conditions in which pupils can surprise themselves and others. If systems of assessment, accountability, and now AI mediated assistance fix children too closely within present profiles, present needs, and present outputs, they diminish the space in which unexpected growth can occur.

The literature on generative AI and assessment makes this even clearer. Feng and colleagues, writing on computing education, argue that generative AI shifts the educational centre of gravity from producing solutions to evaluating, debugging, and improving generated ones. On the surface this sounds plausible, even progressive. If machines can code, perhaps students should critique code. If machines can write, perhaps students should edit. There is some truth in that. But we should be careful. The proposal presupposes that the educational loss involved in no longer generating solutions independently can be compensated for by a higher order activity of critique. That may be possible in some advanced contexts. But it is by no means obvious that it can serve as a general educational answer, especially for novices. Critique depends on prior understanding. One cannot critically evaluate what one does not yet really grasp. So there is a real danger that the discourse of moving from automation to cognition will serve as a slogan masking a more ordinary erosion of formation. Students may appear to be operating at a more advanced level while actually depending increasingly on machine generated material they are not fully equipped to judge.

That concern becomes more serious when we move from narrow task performance to the shape of the curriculum as a whole. Waters in particular insists that curriculum should be understood as the architecture of learning. It is not just a list of topics, but a sequence of encounters through which pupils come to inhabit forms of knowledge with increasing depth. Generative AI can cut across that architecture by providing answers detached from the pedagogical sequence within which those answers were meant to acquire meaning. A student may receive a correct explanation, but the explanation may not belong to the stage of development, the local question, or the carefully built conceptual world that a teacher has been trying to create. The curriculum becomes punctured by unsequenced interventions. For some advocates, this is personalisation. For Waters, and I think rightly, it risks incoherence. It moves authority away from the teacher as designer of learning and toward a system that treats content as modular and interchangeable.

The problem is not only cognitive. It is also social and professional. Brighouse and Waters repeatedly return to the teacher as the central educational agent, not because teachers are flawless, but because they are the people whose judgement, presence, and knowledge make the greatest difference at the point of encounter between learner and world. Generative AI is frequently marketed as support for that work. It can produce worksheets, draft lesson plans, suggest feedback comments, generate examples, or summarise documents. Some of these uses may be genuinely helpful, especially where they reduce pointless administrative burdens. But the broader risk is clear. In systems already inclined to mistrust teachers, already accustomed to scripting, already overinvested in standardisation, AI becomes a means of further decomposing teaching into tasks that can be generated, monitored, and optimised elsewhere. Brighouse’s concern about the erosion of respect and trust in the profession becomes directly relevant here. A technology that appears to help by providing ready made pedagogical artefacts may gradually train institutions to place more confidence in the artefacts than in the cultivated judgement of the teacher.

This is why the language of personalisation needs particular scrutiny. AI systems personalise by tracking patterns, inferring needs, and adjusting content in relation to data. But this is not the same thing as human recognition. Waters’ educational imagination is rooted in the idea that pupils are known as individuals with uneven talents, different maturities, shifting motivations, and unrealised possibilities. Brighouse’s moral vocabulary is even stronger. Education should treat children not merely as they presently are, but as they might become. AI personalisation, however sophisticated, is still built upon profiles of the present, patterns in prior behaviour, and statistically inferred trajectories. It can tailor. It cannot believe in unrealised possibility in the same way a teacher can. That matters because one of the deepest educational goods is being seen not only in terms of what one currently looks like to the system, but in terms of what one may yet become.

This brings us to the question of creativity and imagination. One of the most seductive public claims about generative AI is that it can stimulate creativity. In some narrow sense this is clearly true. Studies such as those by Doshi and Hauser suggest that individuals using AI can produce ideas rated as more creative than those produced unaided, especially where the baseline level of confidence or fluency is low. Yet the same research often finds that the diversity of ideas across individuals narrows. Students become more productive but also more similar. Baltà-Salvador and colleagues show a related pattern in which human AI co-creation increases idea fluency while reducing semantic divergence. The significance of this for education is profound. Schools and universities should be places where different trajectories of thought encounter one another, where learners discover unexpected lines of inquiry, and where originality is not merely the polish of a final product but the surprise of an unusual connection or a not yet normalised question. If AI makes work more fluent while subtly pulling it toward already dominant patterns, then it risks giving us better looking outputs and thinner intellectual ecosystems.

This is one of the points at which Brighouse’s language of hope and possibility becomes especially important. He cared deeply about schools as places where children could encounter worlds beyond their immediate circumstances, where imagination could be enlarged, and where talent might be discovered rather than merely ranked. In such a view, creativity is not a luxury add on. It is part of the widening of a person’s relation to the world. The research on AI assisted creativity gives us reason to worry that generative systems may produce a superficially richer but actually narrower imaginative culture. This is not because they destroy creativity outright. It is because they stabilise it around statistical norms. Students may produce more stories, more designs, more essays, more plans, but the conceptual range may contract. A classroom full of individually polished but mutually convergent work is not an educational triumph.

The writing studies deepen the problem further. Mak and Walasek’s work suggests that AI assisted writing often displays greater fluency and more standardised positivity of tone, without necessarily showing stronger reasoning. Draxler and colleagues identify what they call an AI ghostwriter effect, in which users experience a diminished sense of ownership over the resulting text. These are not marginal concerns. Writing has historically been one of the central means through which students learn to think, not just display thought already completed elsewhere. To write is to discover what one can say, where one is uncertain, how one connects ideas, how one takes responsibility for words. If writing increasingly becomes the management and editing of generated language, the relationship between thinking and expression changes. Students may still submit essays, but the essays may no longer function in the same formative way. This is precisely the sort of issue Brighouse would have recognised immediately. For him, literacy was not merely a skill for economic mobility. It was a means by which minds are unlocked and freedom becomes real. Anything that weakens the relation between the person and their words therefore touches the core of schooling’s moral purpose.

The chapter should also insist that these developments are not just individual but ecological. Selwyn and colleagues’ work on student imaginaries suggests that the digital infrastructures in which young people live may already be narrowing the range of futures they can imagine. This resonates strongly with the earlier philosophical argument in the book about prediction, expectation, and cognitive environment. If the environments within which students think are increasingly structured by systems trained on past data and rewarded for statistical plausibility, then the horizon of the possible may itself become more conservative. Students do not simply learn with AI. They learn inside worlds partially shaped by the assumptions and probabilities encoded in AI systems. That matters for any serious account of education because schools should not only reproduce what already exists. They should create spaces in which the not yet thinkable can become thinkable.

By this point in the chapter the cumulative pattern should be visible. Generative AI threatens effortful learning by reducing the need for sustained struggle. It weakens the connection between visible performance and actual understanding. It encourages passive adoption where institutions reward output. It narrows diversity even where it enhances fluency. It complicates authorship and the ownership of thought. It may reduce the space in which curiosity and imagination operate. And all this happens not in a vacuum, but in systems already predisposed to value speed, productivity, measurable success, and technical manageability. This is why the chapter should not be read as a list of isolated risks. It is a diagnosis of fit. AI fits the neoliberal educational ecology too well.

That point can now be made explicitly by bringing Bell, Cuban, Ball, Brighouse, and Waters into one frame. Bell helps us see AI as the latest intellectual technology in a long process of codification and systemic control. Cuban reminds us that schools do not simply absorb technologies as advertised, but adapt them through existing institutional routines. Ball shows that those routines are now already shaped by performativity, metrics, and external governance. Brighouse and Waters recover the alternative standard, schooling as a humane, moral, civic practice. The contemporary AI literature then becomes legible not as a neutral evidence base but as confirmation that a technology aligned with performativity is entering institutions that have already become vulnerable to it. The result is not accidental misuse. It is structural intensification.

This is why I think the chapter should refuse two temptations. The first is naïve enthusiasm. The second is romantic rejection. The studies do not justify panic in the sense of total refusal. There is evidence that critical interventions can improve students’ engagement with AI. There are contexts in which AI may widen access to explanation, reduce routine burdens, or support learners in useful ways. The problem is not the existence of the tool as such. The problem is educational asymmetry. The technology arrives before the pedagogical frameworks, before the moral clarity, before the institutional reform, and before the recovery of confidence in what schools are for. In those conditions, the default use of AI is overwhelmingly likely to serve the logic already dominant in the system.

This is why the chapter should end by returning to Brighouse and Waters, not as authorities quoted decoratively, but as guides to the actual issue. Brighouse spent much of his life arguing that schools should unlock minds, expand freedom, and treat children as they might become. Waters spent much of his life arguing that curriculum should be designed as an architecture of meaningful encounter, not reduced to measurable fragments and performative routines. The AI literature can now be read as evidence that the present educational order is in danger of doing the opposite. It encourages young people to manage outputs rather than inhabit questions, to optimise rather than explore, to edit generated language rather than think through language, and to treat knowledge as immediate retrieval rather than earned understanding. If this continues unchecked, the result will not simply be a change in classroom tools. It will be a change in what learners come to expect thinking itself to be.

That is the deeper inversion at the centre of this book. For decades the fantasy surrounding artificial intelligence was that machines would become more human, more creative, more empathetic, more judicious. What the educational evidence now suggests is that a more immediate transformation is occurring in the opposite direction. Humans are adapting themselves to environments organised by machine logic. Students learn to value speed, fluency, optimisation, and frictionless completion because those are the capacities that the systems around them reward. The danger is not only that they will use AI badly. It is that they will gradually become the kinds of thinkers for whom the values embodied in those systems feel natural.

Education sits at the centre of this transformation because it is where cognitive and moral habits are formed. That is why the chapter matters so much after Brighouse and Waters. Without them, the AI literature might be read as a mixed field of risks and opportunities. With them, it becomes much clearer what is at stake. The question is not whether AI can be added to the toolkit of schooling. The question is whether schooling can still remain an institution in which human beings learn to think, imagine, judge, and hope in ways not exhausted by the capacities of the machines they use.

My own view is that the literature increasingly shows the need for a much more severe educational response than is common in current policy discussion. Schools and universities should not ask first how to integrate generative AI smoothly. They should ask what kinds of learning, writing, dialogue, and assessment must remain protected from machine convenience if education is still to deserve the name. They should ask which activities depend for their value on effort, uncertainty, ownership, and relation. They should ask where the teacher’s judgement is irreplaceable. They should ask which pupils are most likely to be harmed by the apparent efficiency of automated support. They should ask where curriculum coherence is being broken by unsequenced machine intervention. And above all, they should ask whether a particular use of AI strengthens or weakens the human purposes of schooling recovered in the previous chapter.

That is how this material should advance the book’s argument. It takes the normative recovery achieved through Brighouse and Waters and places it under the pressure of current evidence. The result is not a detached review of a technological field. It is an empirical confirmation of the book’s wider claim. Generative AI is not merely introducing new tools into education. It is exploiting and intensifying a system already shaped by neoliberal assumptions about performance, efficiency, and control. Unless educational institutions recover a stronger account of what schools are for, AI will not save them. It will deepen the deformation.

The stakes are therefore larger than classroom policy. A generation now being educated with generative AI will carry its habits into universities, workplaces, politics, and culture. If they are formed mainly as efficient operators within systems of rapid production, then the consequences will extend far beyond schooling. If, however, schools can hold their ground, preserving spaces for genuine encounter, intellectual struggle, authorship, judgement, and shared inquiry, then AI may yet be confined to a subordinate role. The future of education will depend less on how intelligent machines become than on whether educators remain serious enough about human purposes to keep them in their place.

Bibliography

Ball, Stephen J. “The Teacher’s Soul and the Terrors of Performativity.” Journal of Education Policy 18, no. 2 (2003): 215–228.

Baltà-Salvador, Raquel, Miquel Brasó-Vives, and Ana Peña. “Human–AI Co-Creation and Semantic Divergence in Student Ideation.” Computers & Education 205 (2025).

Barcaui, André. “Artificial Intelligence Assistance and Knowledge Retention: Experimental Evidence from ChatGPT Use.” Computers & Education 210 (2025).

Bell, Daniel. The Cultural Contradictions of Capitalism. New York: Basic Books, 1976.

Cuban, Larry. Teachers and Machines: The Classroom Use of Technology since 1920. New York: Teachers College Press, 1986.

Cuban, Larry. Oversold and Underused: Computers in the Classroom. Cambridge, MA: Harvard University Press, 2001.

Dai, Kai. “Generative Artificial Intelligence in Higher Education: A Bibliometric Analysis (2022–2025).” Educational Technology Research and Development (forthcoming, 2026).

Doshi, Ameet, and Oliver Hauser. “Generative AI Enhances Individual Creativity but Reduces Collective Diversity.” Science Advances 10, no. 5 (2024).

Draxler, Fiona, Anna Werner, Jan Schneider, and Andreas Butz. “The AI Ghostwriter Effect: Users’ Perception of Ownership in AI-Assisted Writing.” In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. New York: ACM, 2023.

Feng, Shiyu, et al. “From Automation to Cognition: Rethinking Programming Education in the Age of Generative AI.” ACM Transactions on Computing Education 24, no. 3 (2024): 1–25.

Giannakos, Michail N., et al. “The Promises and Challenges of Generative AI in Education: A Systematic Review.” Computers & Education: Artificial Intelligence 5 (2024): 100156.

Hou, Yifan, et al. “Understanding Students’ Use of Generative AI in Problem Solving: Passive, Reflective, and Thoughtless Use.” Computers & Education 200 (2024).

Keuning, Hieke, et al. “Student Interaction Patterns with Generative AI in Programming Education.” Computer Science Education 34, no. 2 (2024): 123–145.

Mak, Kenny K. L., and Lukasz Walasek. “AI-Assisted Writing and the Standardisation of Student Voice.” Computers and Composition 72 (2025).

Selwyn, Neil, et al. “Students’ Imagined Futures with Artificial Intelligence: Hopes, Concerns, and Educational Implications.” Learning, Media and Technology 50, no. 1 (2025).

Zastudil, Cyril, et al. “Generative AI in Computing Education: Student and Instructor Perspectives.” In Proceedings of the 2023 ACM Conference on International Computing Education Research. New York: ACM, 2023.

Chapter 10

What follows continues an argument already under way rather than beginning a new one. Earlier I have argued, drawing on work in cognitive science and philosophy of education, that human learning depends upon processes that are temporally extended, effortful, and socially mediated, and that schooling has historically functioned, however imperfectly, as an institutional attempt to protect and organise those processes. The question now is what happens when generative systems are introduced that can reliably produce many of the outward signs of knowledge, explanations, summaries, arguments, and even interpretations, without requiring the same underlying cognitive work. The issue is not simply practical or regulatory. It is structural. Once outputs become detachable from the processes that generate them, the rationale for much of schooling comes under pressure.

A useful way into this problem is to articulate a dialectic that is increasingly visible across both theoretical and empirical work. On one side is the claim that generative AI can extend and enrich human cognition. On the other is the concern, now supported by a growing body of research, that such extension often occurs by bypassing the processes through which understanding is formed. The tension between these positions is not merely rhetorical. It reflects two different models of thinking, one oriented toward output and flexibility, the other toward formation, stability, and judgement.

The extension thesis is articulated with particular clarity in recent work associated with Anthea Roberts, whose account of multi-perspectival or “dragonfly” thinking builds on the idea that complex problems require movement across frameworks rather than adherence to a single disciplinary lens. This position resonates with Philip Tetlock’s findings on forecasting, where individuals who integrate multiple partial models tend to outperform those who rely rigidly on one approach. The educational implication appears straightforward. Generative systems can facilitate this kind of thinking by rapidly producing alternative framings, modelling stakeholders, and surfacing tensions. In this sense, AI functions as a cognitive partner that broadens the space of inquiry.

There is empirical support for aspects of this view. Studies of AI-supported learning environments show that, under certain conditions, students who engage critically with generated outputs can improve the range of perspectives they consider and the sophistication of their reasoning. For example, Hou et al. (2024) distinguish between reflective and unreflective modes of AI use, finding that reflective engagement, where students interrogate and compare outputs, is associated with higher levels of critical thinking. Similarly, Giannakos et al. (2024), in a systematic review of generative AI in education, note that AI can support exploratory learning and perspective-taking when embedded within structured pedagogical frameworks.

Yet the same body of research reveals a significant limitation. The benefits of AI-mediated multiplicity depend heavily on the user’s prior knowledge and critical capacity. Hou et al. (2024) show that students with weaker backgrounds are more likely to engage in what they term “thoughtless use,” accepting outputs with minimal scrutiny. Keuning et al. (2024), studying programming education, report that while some students use AI to deepen understanding, others rely on it to generate solutions without engaging with underlying concepts. Zastudil et al. (2023) similarly find that instructors are concerned about students producing correct code without acquiring the reasoning skills necessary to generate it independently.

This asymmetry points directly to the distinction between extension and formation. A system that extends the capacities of knowledgeable users does not necessarily form those capacities in novices. Indeed, it may obscure their absence. The learner is presented with outputs that resemble expert reasoning, but without having undergone the processes through which such reasoning is acquired. The danger is not that AI fails to produce useful material, but that it produces it in a way that allows users to bypass the labour through which understanding becomes stable and discriminating.

This concern is reinforced by experimental evidence on learning outcomes. Barcaui (2025), in a randomised controlled study, finds that while AI assistance improves immediate task performance, it is associated with weaker delayed retention of knowledge. This suggests that the cognitive processes involved in generating answers play a role in consolidating understanding, a role that is diminished when those processes are offloaded. Feng et al. (2024) make a similar argument in the context of programming education, noting that excessive reliance on generative tools can reduce opportunities for what they term “productive struggle,” a process widely recognised in the learning sciences as essential for deep learning.

The issue becomes sharper when considered in relation to synthesis, often regarded as a higher-order cognitive skill. Howard Gardner’s account of the synthesising mind emphasises the capacity to integrate disparate sources into a coherent whole (Gardner 2008). Generative systems appear to perform this function with remarkable fluency, producing summaries and integrations across domains. This raises the question of whether the internal development of such capacities remains necessary.

Empirical research suggests caution. Mak and Walasek (2025) find that AI-assisted writing leads to increased fluency and structural coherence, but does not consistently improve deeper measures of quality. Draxler et al. (2023) identify what they call the “AI ghostwriter effect,” where users produce polished text but report reduced sense of ownership. These findings indicate that while AI can generate the products of synthesis, it does not necessarily support the internalisation of the processes through which synthesis is achieved.

The creativity literature further complicates the extension thesis. Doshi and Hauser (2024) show that while generative AI can increase the judged creativity of individual outputs, it reduces diversity across users, leading to convergence around similar ideas. Baltà-Salvador et al. (2025) report reduced semantic divergence in AI-assisted ideation tasks. These findings suggest that AI may produce outputs that are locally impressive but globally homogenising. In educational terms, this raises concerns about the narrowing of intellectual horizons and the erosion of originality.

At this point the more radical implications of Gardner’s position become visible. If generative systems can perform many of the cognitive tasks associated with disciplined and synthesising thought, then the role of traditional disciplinary education comes into question. Gardner’s later work suggests that the most distinctively human capacities may lie in ethical judgement and relational understanding rather than in cognitive mastery alone. This view aligns with broader calls for education to focus on flexibility, creativity, and interpersonal skills.

However, the empirical evidence discussed above complicates this move. The capacity to use AI critically, to evaluate outputs, to recognise limitations, appears to depend on prior knowledge and disciplinary grounding. Li et al. (2024) argue that domain knowledge remains essential for effective use of large language models in learning contexts, particularly for verification and error detection. Without such grounding, users are more susceptible to accepting plausible but incorrect or superficial outputs.

This suggests that disciplinary knowledge may become more, not less, important in an AI-rich environment. It functions not only as content but as a framework for judgement. The idea that one can replace content with skills or meta-knowledge is therefore questionable. Meta-knowledge is not independent of content. It is developed through engagement with it.

The problem of verification further strengthens this point. In domains such as mathematics, where formal verification is possible, AI can be integrated into workflows with relatively low risk. Tao (2025) notes that AI is already useful for routine tasks such as checking computations and exploring conjectures, precisely because errors can be identified and corrected. In contrast, in domains where verification is weaker, such as history, literature, or ethics, the risks are greater. Zhai et al. (2024), in a systematic review, highlight concerns about overreliance on AI dialogue systems in contexts where correctness is difficult to assess.

Ethical reasoning provides a particularly clear example. Ethical problems involve conflicting values, incomplete information, and no definitive procedures for resolution. Generative systems can produce balanced arguments, but this may create an illusion of closure. The user encounters a coherent synthesis that appears to settle the issue, even when it remains fundamentally contested. This aligns with findings in the literature on AI-supported decision-making, where users tend to over-trust outputs that are fluent and well-structured, even in the absence of strong justification.

The risk here is not simply error, but the transformation of the experience of reasoning. Ethical judgement involves grappling with uncertainty, recognising limits, and taking responsibility for decisions. If this process is replaced by the consumption of generated arguments, the development of moral competence may be undermined. This concern echoes broader critiques of automation in decision-making, where the delegation of judgement to systems can lead to what has been termed “moral deskilling.”

The metaphor of the journey, used by Tao to describe mathematical research, is helpful here. In many forms of inquiry, the path to an answer is integral to its meaning. The process of working through difficulty, making mistakes, and revising understanding contributes to learning. Generative systems can bypass this process by delivering the answer directly. While this increases efficiency, it risks undermining the formation that occurs along the way.

Empirical evidence on assessment reflects this tension. Studies report divergence between performance on AI-supported tasks and independent assessments. For example, reports of rising homework scores alongside stable or declining exam performance suggest that AI may inflate measures of achievement without corresponding gains in understanding. This has led to calls for new forms of assessment that emphasise process, oral examination, and iterative work (Feng et al. 2024).

The implications for curriculum are equally significant. If certain tasks can be performed by AI, there may be pressure to reduce time spent on them. However, the research suggests that these tasks often play a role in cognitive development. Removing them may weaken the foundation on which higher-order skills depend. This aligns with findings from cognitive load theory and related work, which emphasise the importance of practice and repetition in building expertise.

The role of teachers is also affected. As information becomes more accessible, the teacher’s role shifts from transmission to guidance. However, this shift must be understood carefully. The teacher is not simply a facilitator of access to AI, but a guide in the development of judgement. This includes helping students recognise when to use AI, how to evaluate outputs, and when to rely on their own reasoning.

The broader institutional context amplifies these issues. Educational systems shaped by accountability and measurement are likely to favour technologies that produce measurable outputs. Generative AI fits well within this framework, as it can produce standardised responses at scale. This raises concerns about the further narrowing of educational aims, as aspects of learning that are less easily measured may be devalued.

Selwyn et al. (2025) provide evidence that students’ imaginaries of education are already being shaped by digital technologies. If AI becomes central to educational practice, it may influence not only what students do, but what they expect education to be. This has implications for the future of schooling as a public institution.

The dialectic that emerges can therefore be summarised as follows. Generative AI offers genuine cognitive extension, supporting exploration, synthesis, and efficiency. At the same time, it risks undermining cognitive formation by bypassing the processes through which understanding develops. The balance between these effects depends on how the technology is integrated into educational practice.

The evidence suggests that without deliberate intervention, the extension side may dominate in ways that weaken formation. This is not inevitable, but it is likely given existing institutional pressures. To counter this, education must make explicit decisions about where to use AI and where to preserve traditional forms of learning. This includes maintaining spaces for unaided performance, protecting time for slow learning, and emphasising the importance of knowledge and judgement.

The argument does not lead to a simple conclusion. It does not advocate rejection of AI, nor does it endorse uncritical adoption. Rather, it clarifies the stakes. The question is not whether AI will be used, but how its use will reshape the purposes of education. If those purposes remain unclear, the technology is likely to redefine them in its own image. If they are articulated and defended, AI may be integrated in ways that support rather than undermine the formation of learners.

The chapter therefore leaves us with a sharpened problem rather than a settled answer. The emergence of generative systems has made visible a tension that was already present, between output and understanding, efficiency and formation. How that tension is resolved will determine not only the role of AI in education, but the future of education itself.

Chapter 11

What the earlier literature on artificial intelligence in education makes clear, once it is read in the light of the preceding chapters, is that the present moment did not arrive as a bolt from the blue. The contemporary excitement about generative AI in schools and universities is not the appearance of an entirely new educational question. It is the intensification of a much older ambition, namely the hope that computational systems might analyse learning, personalise support, and extend the reach of teaching beyond what human educators alone can easily provide. To see that continuity matters because it changes the tone in which current claims should be judged. We are not standing at the beginning of the story, dazzled by novelty. We are standing in the middle of a long trajectory whose promises, limits, and internal contradictions have already been partially disclosed. The task of this chapter is therefore not simply to review a literature. It is to show how the older traditions of intelligent tutoring systems, educational data mining, and learning analytics prepared the ground for generative AI, and how the newer research now reveals more starkly the danger that educational processes may be redesigned around what machines do well rather than around what human learning actually requires.

In the earlier chapters I argued that knowledge is not best understood as the mere possession of correct outputs, but as a state formed through sustained engagement with uncertainty, with questions, with other persons, and with the world. Drawing on Jennifer Nagel, and alongside developmental work by Csibra and Gergely and Tomasello, I stressed that human learning is structured by curiosity, by social attention, and by forms of cooperative inquiry that preserve a space between problem and answer long enough for understanding to take shape. That conceptual claim now meets its empirical test. The research on AI in education shows, with growing clarity, that when computational systems make answers too readily available, educational performance may improve at the very point where understanding weakens. The paradox is already familiar in practice, but the literature allows us to see it in a more systematic way. AI can support learning, but it can also corrode the conditions under which learning becomes more than successful task completion.

The early literature was, in some respects, more careful than the present moment. Ryan Baker and Kalina Yacef, writing in one of the formative surveys of educational data mining, described a field concerned with making sense of the rich traces learners leave behind in digital environments (Baker and Yacef 2009). Their point was not that data had suddenly solved the problem of learning, but that digital systems recorded processes that teachers in conventional classrooms could rarely observe in full. Every hint request, every error, every pause, every sequence of attempts became visible. The promise here was epistemic rather than magical. Researchers hoped to learn more about learning itself. Classification models might predict who was likely to struggle, clustering might reveal common patterns of misconception, sequential models might trace how understanding develops over time. Yet Baker and Yacef also insisted that such analyses required interpretation. Data could not tell us by itself what counted as worthwhile learning. The human meaning of learning remained irreducible to the computational pattern.

That caution continued in the adjacent field of learning analytics. George Siemens and Ryan Baker tried early on to distinguish, while also relating, educational data mining and learning analytics as parts of a broader ecosystem of inquiry (Siemens and Baker 2012). The distinction mattered because it already implied a tension that has only grown sharper since. It is one thing to build predictive or classificatory models. It is another to interpret those models educationally, to decide what kind of teaching, what kind of curriculum, what kind of intervention, and what kind of human good the data should serve. Learning analytics, at its best, tried to hold on to that second question. Without theoretical and pedagogical grounding, a proliferation of educational data could only produce a false sense of control. Correlations might appear powerful while concealing the absence of any real understanding of the mechanisms by which students learn. That warning now reads as prophetic. One of the central dangers of the current AI moment is precisely that systems become better and better at producing educationally legible outputs while institutions become less and less clear about what those outputs mean.

The same pattern is visible in the older work on intelligent tutoring systems. Beverly Woolf’s synthesis of decades of research into intelligent interactive tutors was built around an educationally serious question, how to approximate some of the power of one to one tutoring without imagining that machines could simply replace teachers (Woolf 2010). The architecture of these systems was ambitious but also modest in a revealing way. A tutoring system would need a model of the domain, a model of the student’s present understanding, and a model of pedagogical strategies appropriate to moving the learner on. That is already a significant claim about the nature of teaching. Teaching is not content delivery. It requires diagnosis, interpretation, adaptation, and timing. Woolf repeatedly acknowledged, however, that computational models captured only part of the scene. Human teachers respond not only to correct and incorrect answers, but to emotional states, histories of struggle, social dynamics, tones of confidence or discouragement, and the wider relational atmosphere in which learning occurs. The early tutoring literature, in other words, contains a quiet but decisive admission. Whatever the value of computational support, the full educational situation is richer than the machine can presently model.

Ido Roll and Ruth Wylie’s historical account of artificial intelligence in education made this point from a different angle (Roll and Wylie 2016). They described the field as moving through alternating phases of technical advance and pedagogical adjustment. Rule based systems gave way to more data driven systems, and increasingly sophisticated forms of modelling became possible. Yet the crucial variable remained context. An educational tool is never simply an educational good by virtue of technical sophistication alone. It matters how it is embedded, what curriculum it inhabits, what teacherly practices surround it, and what conception of learning it serves. Read now, this literature is striking for its relative sobriety. It is ambitious, certainly, but it does not typically confuse computational capability with educational adequacy. The aim was to support human learning, not redefine it.

Chris Piech’s work offers a particularly useful bridge between the older and newer worlds. His research on modelling student programming behaviour showed how machine learning could represent large numbers of student code submissions in ways that made structural similarity visible even when surface features varied greatly (Piech et al. 2015). This allowed feedback to be propagated more intelligently across vast cohorts. The advance is real. Human teachers cannot manually inspect every conceptual resemblance across thousands of solutions. Yet even here the system does not educate autonomously. It organises a problem space so that human feedback can be targeted more effectively. The instructor remains central. The machine is analytical leverage, not a substitute for pedagogical judgement.

That point becomes even clearer in Piech’s later work on Code in Place, one of the most interesting large scale experiments in online human centred teaching during the pandemic (Piech et al. 2021). The course used computational infrastructure, certainly, but its success depended on the deliberate construction of distributed human mentoring. Small sections, volunteer section leaders, regular interaction, and the cultivation of educational community remained central. It is difficult to overstate the importance of this. One of the most ambitious technology enabled projects in recent computer science education did not treat human teaching as the dispensable remainder after automation. It treated human relation as the very thing the infrastructure should support. That model now offers a sharp contrast to the more recent fantasy that generative systems can replace much of what the teacher does.

The historical lesson is therefore already clear before one turns to large language models. The most serious earlier work in AI and education recognised both the usefulness of computation and the irreducibility of pedagogy. It assumed that educational purposes had to govern technological integration rather than the reverse. The new literature after 2022 inherits that older research base, but it enters a more dangerous situation because the capabilities of generative AI are broader, more seductive, and more culturally destabilising. Generative systems do not merely classify, cluster, or model. They explain, summarise, write, converse, translate, and generate plausible solutions across a vast range of educational tasks. The problem is no longer only that data may be misread. It is that machines can now produce the very artefacts by which educational achievement has long been displayed.

That is why the recent bibliometric studies are so revealing. Kai Dai’s survey of publications between 2022 and 2025 shows not only explosive growth in research on generative AI and education, but also a field struggling to keep conceptual pace with its own object (Dai 2026). Thousands of publications appear within a few years, yet the literature remains exploratory, uneven, and often experimentally shallow. Personalised learning, automated feedback, assessment, tutoring dialogue, all proliferate as research themes. But stable theoretical frameworks lag behind. That is not an incidental feature of the field. It is a symptom of the larger condition the book has been describing. The technology advances faster than educational thought about its purposes, which means that institutions are tempted to adopt first and understand later.

Giannakos and colleagues capture this tension well. Generative AI appears to offer unprecedented opportunities for tailored explanation, on demand practice, and tutoring style interaction, yet these very advantages carry a structural danger (Giannakos et al. 2024). If the system becomes primarily an answer provider, it weakens the very cognitive work through which understanding develops. This is not a pious objection to convenience. It is an educational claim about the relation between effort and learning. When a problem is resolved too early by an external system, the learner is deprived not merely of labour but of formation. The difference matters. Education is not only about ending in the right place. It is about what happens to the person on the way there.

The newer empirical work makes this tension concrete. Zastudil and colleagues, studying generative AI use in computing education, show how quickly students incorporated these systems into their everyday practice (Zastudil et al. 2023). They used AI to generate code, explain concepts, diagnose bugs, and draft written explanations. Much of this felt immediately useful, and in many cases it was. Yet the authors also record widespread uncertainty among both students and instructors. If the system helps the student finish the assignment, has it also helped the student understand the underlying concepts? Or has it merely accelerated the production of an acceptable answer? That ambiguity is central. It marks the point at which educational output and educational understanding begin to come apart.

Keuning and colleagues deepen this by showing that students do not all use generative AI in the same way (Keuning et al. 2024). Some treat it as a conceptual scaffold, asking for explanation, clarification, and error diagnosis. Others accept generated solutions with relatively little interrogation. The distinction between reflective and unreflective use matters greatly, but not in the simple sense that one group behaves virtuously and the other lazily. The important point is that the technology itself makes substitution very easy. The system can function as a partner in inquiry, but it can just as readily function as an efficient bypass around inquiry. In educational systems already shaped by deadline pressure, instrumental assessment, and performative incentives, it is not difficult to guess which use will often dominate.

Feng and colleagues bring the assessment problem into focus. In computing education, when functioning code can be produced quickly by the model, assignments designed around independent solution production no longer measure what they once purported to measure (Feng et al. 2024). Their proposed shift, from automation to cognition, is sensible and indeed necessary. Students may need to be assessed less on whether they can produce a first pass solution and more on whether they can interpret, debug, improve, and critique generated outputs. Yet even this reform contains a warning. The educational system is being forced to reorganise itself around the existence of a machine that produces acceptable artefacts with ease. The question is whether that reorganisation will deepen education or merely push institutions one step further into a logic of machine alignment.

The adaptive learning literature offers a related case. Li and colleagues show how generative models can extend older adaptive systems by creating tailored explanations and exercises dynamically rather than requiring everything to be authored in advance (Li et al. 2024). Again, the advantage is real. But so too is the danger. The model may explain fluently while being wrong, misleading, or overconfident. For novice learners especially, the boundary between authoritative explanation and generated approximation becomes unstable. Yan and colleagues, working on AI assisted interpretation of learning analytics dashboards, identify something similar. AI does not merely help students read data, it also frames what the data means (Yan et al. 2024). The system shapes interpretation. This is crucial because one of the ideological powers of educational technology lies in its ability not simply to process information but to silently define the terms under which information becomes educationally significant.

Teachers’ perceptions confirm that these are not abstract worries. Petrucco and colleagues report that teachers see potential usefulness in generative AI while also fearing a weakening of students’ independent thinking (Petrucco et al. 2024). This ambivalence should be taken seriously. Teachers are not merely anxious resisters of innovation. They are often the first to recognise where a tool helps and where it hollows out the activity it was meant to support. If they worry that students may become dependent on automated systems for tasks that were once integral to the learning process, that is not nostalgia. It is often practical educational judgement.

At this point the chapter needs to turn from history and field mapping to the deeper pattern the research discloses. Across these studies a single structure appears repeatedly. AI improves performance under certain conditions while weakening the processes that make understanding durable, owned, and educative. André Barcaui’s work is one of the clearest examples. Students using ChatGPT could produce correct answers more efficiently, yet later retained less knowledge (Barcaui 2025). This matters because it reveals the central paradox in a form too stark to ignore. Educational success, at least in the visible short term, may improve at the very point where the learner’s knowledge state becomes thinner. The answer appears, but the person does not fully change.

Hou and colleagues’ work on passive and thoughtless AI use sharpens the point further (Hou et al. 2024). Students often default to adopting generated solutions rather than wrestling with the underlying problem. When explicit pedagogical interventions encourage critical engagement, more original and reflective responses become possible (Hou et al. 2025). This is encouraging, but it also reveals the default tendency of the technology. Without deliberate design, the system invites premature closure. It gratifies the desire to finish before the mind has undergone the labour of understanding. Zhai, Wibowo, and Li’s review of overreliance on AI dialogue systems strengthens the same conclusion. Heavy dependence correlates with weaker critical thinking and independent reasoning (Zhai, Wibowo, and Li 2024). The issue is not lack of access to information. It is the atrophy of the internal activity by which information becomes one’s own.

The creativity literature shows the same pattern in another register. Doshi and Hauser find that generative AI can increase individual creativity ratings while reducing the diversity of ideas across groups (Doshi and Hauser 2024). Baltà-Salvador, Brasó-Vives, and Peña report increased ideational fluency alongside reduced semantic divergence (Baltà-Salvador, Brasó-Vives, and Peña 2025). This combination is profoundly important for education. A classroom may look more productive, more creative, more full of text and ideas, while becoming more cognitively standardised. The students produce more, but from a narrower range of underlying patterns. The educational loss here is not visible if one attends only to output quantity or surface quality. It becomes visible only if one values intellectual diversity, surprise, and the emergence of genuinely different ways of seeing.

Mak and Walasek’s work on undergraduate writing provides a parallel example (Mak and Walasek 2025). AI assisted texts often become more fluent and stylistically polished, yet the deeper argumentative structure may not improve correspondingly. Draxler and colleagues then show how authorship itself begins to blur. Users feel diminished ownership over AI assisted text, without fully attributing authorship to the machine either (Draxler et al. 2023). Writing turns into the management of generated language rather than the articulation of thought through language. This is not a minor shift. Writing has historically been one of the central ways in which students clarify what they think by struggling to say it. If that struggle is displaced by editing fluent machine prose, then an entire mode of intellectual formation is endangered.

Selwyn and colleagues’ work on student imaginings of educational futures adds a final layer. Students’ visions of the future tend to reproduce the digital forms already surrounding them (Selwyn et al. 2025). That suggests that algorithmic environments shape not only present cognition but future expectation. Learners begin to imagine possibility within the horizon already set by technical systems. This finding fits closely with the philosophical argument of the book. Environments do not merely contain thought. They form the space within which thought becomes possible. If that environment is increasingly structured by optimisation, immediate response, and predictive patterning, then the learner’s horizon of imagination may narrow long before anyone notices.

The point of assembling these findings is not to produce another litany of worries. It is to show that a coherent dialectic has now emerged. The older AI in education literature assumed that technological systems should support human learning as pedagogically understood. The newer literature shows that generative systems are powerful enough to tempt institutions into silently redefining learning in the image of what those systems do well. The inversion is subtle but profound. Instead of asking how machines can serve the rhythms of human knowing, institutions drift toward reorganising educational practice around machine strengths, speed, fluency, convenience, output, and optimisation.

At this stage the broader social thinkers discussed earlier in the book help make sense of the educational evidence. Larry Cuban’s work remains crucial because it reminds us that educational technologies never arrive into a neutral space. They enter schools and universities already shaped by histories, cultures, professional habits, and policy regimes. Many earlier technologies failed to transform classrooms as their advocates predicted because the institutional complexity of education resisted neat insertion (Cuban 1986; 2001). Yet Stephen Ball’s analysis of performativity shows why the current moment may be different in one specific respect. Educational systems are now far more deeply organised around measurable outputs, accountability, and visible performance (Ball 2003). In such environments generative AI is not a foreign imposition. It is an almost perfect fit. It produces exactly what the performative system rewards, quickly, efficiently, and at scale.

Daniel Bell helps illuminate the deeper contradiction. Educational institutions are increasingly caught between the techno economic pressure toward efficiency and the cultural purpose of forming judgment, meaning, and human depth (Bell 1976). Generative AI sharpens that contradiction because it makes the efficient production of educationally recognisable artefacts astonishingly cheap. But what education exists to cultivate, if it is worth the name, cannot be reduced to artefact production. Christopher Lasch, Jean Twenge, David Riesman, Thorstein Veblen and others help frame the wider cultural mood into which AI enters. The environment already privileges performance, appearance, rapid response, and external evaluation. Generative AI does not invent those dispositions, but it amplifies them. It intensifies a culture in which polish may count for more than substance, responsiveness for more than reflection, visible activity for more than inward formation.

This is where Brighouse and Waters return with particular force. Their significance here is not antiquarian. They provide the normative counterpoint without which the research risks being read merely as a set of technical trade offs. Brighouse insisted on the moral purpose of schooling, on the teacher’s role in opening worlds, enlarging horizons, and helping children become more than present performance profiles. Waters insisted that curriculum is an architecture of encounter, not a checklist of outputs. Read against the empirical literature, their concerns acquire new precision. When research shows reduced retention, passive uptake, convergence of ideas, blurred authorship, and narrowed imagination, these are not just educational side effects. They are signs that the architecture of learning is being remodelled away from the very goods Brighouse and Waters thought schools exist to protect.

The chapter therefore advances the book’s dialectic in a specific way. Earlier chapters argued philosophically that human learning depends on curiosity, uncertainty, relationality, and disciplined engagement with resistance. This chapter shows empirically that generative AI environments tend to reduce exactly those conditions unless carefully constrained. The technologies do not merely enter education as neutral tools. They alter the ecology in which attention, understanding, and intellectual character form. That is why the real question is not whether AI can be used in education. It plainly can. The real question is whether educational institutions are strong enough in purpose to stop themselves being redesigned around machine logic.

That question cannot be answered by banning the technology, nor by welcoming it with managerial enthusiasm. The research does not support either simplification. AI can assist explanation, access, practice, and support, sometimes very impressively. But the literature also shows that without careful pedagogical design, the same systems encourage dependence, superficiality, and the illusion of competence. The problem, then, is not the existence of powerful tools. It is the weakness of educational institutions when confronted by them. If schools and universities do not know what forms of human capability they are trying to cultivate, they will default to the capabilities that the machine happens to enhance most visibly.

The danger is therefore larger than plagiarism and smaller than apocalypse, but also deeper than either. It is that education will preserve its outward forms while silently changing its inner substance. Students will still submit work, solve tasks, and receive grades. Teachers will still set exercises and issue feedback. Institutions will still speak the language of learning. Yet beneath these continuities, the relation between learner and knowledge may have shifted. Curiosity may be weakened by answer abundance. Authorship may be thinned into supervision of generated text. Diversity of thought may contract beneath a surface appearance of abundant creativity. Judgement may be replaced by iterative selection among plausible machine outputs. Education may become more efficient while the educated person becomes less formed.

For that reason the chapter ends where the book increasingly insists we must end. The central issue is not what AI can do, but what kind of human beings education is still willing to form. If the aim is merely competent navigation of algorithmic systems, then current developments may count as progress. If the aim remains the cultivation of understanding, judgement, imagination, and intellectual independence, then the research surveyed here gives ample reason for caution. Artificial intelligence can support human thinking. But if it becomes the environment within which thinking is formed, and if educational systems fail to resist its tendency toward speed, closure, and optimisation, then the most serious danger is not that machines will become more like us. It is that we will begin to educate the young to become more like them.

Chapter 12

What this book has tried to bring into view is not a single technological development, nor even a cluster of them, but a shift in the conditions under which human beings come to know anything at all. It has required a certain patience to see this clearly because the transformation is not spectacular. It does not announce itself as a crisis. It presents instead as improvement, as convenience, as progress. It is precisely because it looks like enhancement that it is so difficult to grasp what is being altered. Yet when the strands are drawn together, the pattern is unmistakable. The long anticipated movement in which machines would become more human has been accompanied, and perhaps overtaken, by a quieter inversion in which humans are becoming more like machines.

This inversion is not occurring in a neutral environment. It is taking place within what earlier chapters have described as a neoliberal ecology, a system of incentives, infrastructures, and cultural expectations that reward speed, productivity, measurable output, and continuous performance. In such an environment, the value of an activity is increasingly determined by how efficiently it produces visible results. Education has not been exempt from this pressure. It has been gradually reorganised around metrics, accountability systems, and performative displays of achievement. Students learn early that what counts is not simply understanding but demonstrable output. Teachers are required to evidence progress in forms that can be recorded, compared, and audited. Institutions compete through rankings and performance indicators. This is the ecology into which generative AI has arrived.

It is difficult to imagine a more congenial environment for such technologies. Generative systems are extraordinarily effective at producing the kinds of outputs that neoliberal systems reward. They generate text quickly, solve problems efficiently, summarise information instantly, and present results in polished forms. In a world already oriented toward performance, they do not disrupt the system so much as intensify it. They offer a way of doing more of the same, faster and with fewer obstacles. It is therefore unsurprising that much of the public and policy discourse, including influential institutional reports, frames AI as an opportunity to enhance existing educational goals. The language is one of adaptation, integration, and responsible use.

Yet this framing misses the deeper issue that has been developed across the chapters of this book. The problem is not simply that education must adapt to new tools. It is that the tools themselves are reshaping the activity to which they are being applied. The UNESCO report on AI and the future of education, for example, is careful, thoughtful, and concerned with equity, governance, and ethical use. But it largely assumes that education remains what it was, a stable practice into which AI can be inserted. What it does not fully confront is the possibility that the conditions of learning themselves are being altered in ways that those frameworks do not yet capture.

To see this clearly, one must return to the account of knowledge that has guided the argument throughout. Knowledge is not simply the possession of correct answers. It is a state formed through engagement with uncertainty, through curiosity, through the effort of working something out, through the slow integration of ideas into a coherent understanding. It is shaped by time, by resistance, and by interaction with others. These are not optional features of learning. They are constitutive of it. When they are removed or weakened, something important is lost, even if the external signs of success remain.

Generative AI intervenes directly in these conditions. It reduces uncertainty rapidly. It provides answers before the question has been fully inhabited. It offers fluent explanations that can be accepted without the learner having to construct them. In doing so, it changes the relationship between effort and outcome. Tasks that once required sustained engagement can now be completed with minimal cognitive investment. From the perspective of a neoliberal system, this looks like efficiency. From the perspective of learning, it introduces a profound ambiguity.

The empirical research examined earlier shows that this ambiguity is not theoretical. Students using AI can perform tasks successfully while retaining less knowledge. They can produce convincing work without fully understanding the concepts involved. They can generate ideas more quickly while those ideas become less diverse. They can write more fluently while feeling less ownership over what they have written. None of these effects are absolute, and none of them are inevitable in every case. But taken together, they point to a consistent pattern. The processes through which understanding is formed are being compressed, displaced, or bypassed.

Within a neoliberal ecology, this pattern is not only tolerated but often rewarded. If the system values output, then tools that enhance output are advantageous. If success is measured by what can be produced and displayed, then the internal processes that lead to that production become less visible and therefore less valued. This is the mechanism through which cognition itself begins to deform. It is not that individuals are forced to think like machines. It is that the environment makes it rational to do so. To operate efficiently within such systems, one must become adept at generating, selecting, and refining outputs quickly. The habits that are cultivated are those that align with the logic of the system.

The inversion that follows is subtle but far reaching. Instead of asking how machines can be made to support human forms of thought, humans begin to adjust their thinking to fit the capabilities of machines. They learn to prompt effectively, to accept plausible answers, to iterate rapidly. These are adaptive responses to the environment. They are not signs of failure or weakness. But they are not neutral either. They shape the kinds of cognitive dispositions that become habitual.

At this point it might be tempting to conclude that something essential is being lost irretrievably, that the trajectory is one of steady decline in depth, originality, and human agency. There is certainly reason for concern, and this book has not attempted to soften that concern. The convergence of technological capability with a neoliberal ecology that prizes efficiency over depth creates a powerful dynamic. It is capable of reshaping educational practice in ways that are difficult to reverse. It risks producing a form of learning that is increasingly performative, increasingly dependent, and increasingly detached from the internal processes that give knowledge its meaning.

Yet such a conclusion would be incomplete. It would underestimate something that has been present, if more quietly, throughout the argument. Human beings are not infinitely malleable. They are not simply products of their environments, however powerful those environments may be. They are biological organisms with particular structures of attention, motivation, and desire. Among the most important of these is curiosity.

Curiosity is not a luxury or a culturally contingent trait. It is a fundamental aspect of human cognition. It drives exploration, sustains attention, and underpins the development of knowledge. It is also insatiable. It does not disappear when conditions change. It may be redirected, suppressed, or distorted, but it does not simply vanish. This has important implications for the current moment.

If generative AI reduces the need for curiosity within certain domains, by providing answers too quickly, by removing the gaps that would otherwise invite exploration, then curiosity does not cease to exist. It seeks expression elsewhere. This is where the argument takes a darker turn. The displacement of curiosity does not necessarily lead to its elevation in more reflective or constructive forms. It can lead to its capture by other systems that are equally capable of exploiting it.

We have already seen how digital environments can harness curiosity in ways that are not educationally benign. Social media platforms, algorithmic feeds, and online entertainment systems are designed to capture attention, to provide constant novelty, to trigger cycles of anticipation and reward. They offer a form of curiosity satisfaction that is immediate but shallow, continuous but fragmented. If educational environments become less capable of sustaining curiosity because answers are too readily available, then these alternative environments become more attractive.

There is therefore a risk of a double movement. Within formal education, curiosity is dampened by the premature resolution of questions. Outside it, curiosity is intensified but channelled into forms that do not support deep understanding. The result is not a lack of curiosity, but a redistribution of it, away from sustained intellectual inquiry and toward more immediate forms of gratification. This is not a hypothetical scenario. It is already visible in patterns of attention, in the fragmentation of focus, and in the difficulty many students report in sustaining engagement with complex tasks.

At the same time, there is another dimension to this biological constraint that offers a different kind of hope. Human beings are not only curiosity driven. They are also sensitive to boredom. Too much ease, too little resistance, and the activity itself becomes unsatisfying. If tasks are completed without effort, if answers arrive without struggle, if everything becomes immediately accessible, then the experience of learning can become thin. It lacks the satisfaction that comes from overcoming difficulty, from making sense of something that was previously opaque, from arriving at an insight through one’s own effort.

This suggests that the current trajectory may contain the seeds of its own correction. If educational environments become too aligned with machine logic, too efficient, too frictionless, they may fail to engage the very capacities that make learning meaningful. Students may comply with such systems, but they may also experience them as unsatisfying. The hunger for more demanding forms of engagement may reassert itself, even if it is not immediately recognised as such.

The difficulty is that this reassertion does not automatically take constructive forms. It can manifest as disengagement, as resistance, as the search for stimulation elsewhere. The challenge for education is to recognise this dynamic and to respond to it deliberately. It is not enough to rely on the eventual re-emergence of curiosity. Educational environments must be designed in ways that make use of it, that preserve the conditions under which it can operate productively.

This brings the argument back to the question of purpose. If education is allowed to drift within a neoliberal ecology, adopting technologies that enhance output without reconsidering its aims, then the deformation of cognition described in this book is likely to continue. If, however, educators, institutions, and policymakers take seriously the nature of human learning, then there is an opportunity to shape the integration of AI in ways that support rather than undermine it.

This would require a reorientation at multiple levels. Assessment would need to focus more on processes of reasoning, interpretation, and critique rather than on final products alone. Curriculum would need to create space for sustained inquiry, for uncertainty, and for the development of ideas over time. Teaching would need to emphasise dialogue, questioning, and the co-construction of understanding. AI could play a role within such environments, but it would be a constrained role, one that supports exploration rather than replacing it.

It would also require a broader cultural shift. The values that underpin educational systems cannot be separated from the values that underpin society more generally. If speed, efficiency, and measurable output continue to dominate, then educational reform alone will have limited effect. The question of what it means to remain human in an AI shaped world is therefore not only an educational question. It is a social and ethical one.

The central claim of this book has been that we are already living through an inversion that has not yet been fully recognised. Humans are adapting to machines as much as machines are being designed to adapt to humans. This adaptation is shaped by the ecological conditions in which it occurs, conditions that currently favour certain forms of cognition over others. The task is not to resist change in the abstract, but to understand its direction and to intervene where necessary.

The students of today are indeed becoming the architects of an AI shaped civilisation. The habits they form, the ways in which they relate to knowledge, to effort, and to one another, will shape the institutions and practices of the future. If those habits are formed in environments that prioritise output over understanding, efficiency over depth, and immediacy over reflection, then the societies that emerge will reflect those priorities.

But the story is not finished. The same biological structures that make humans vulnerable to these dynamics also provide resources for resisting them. Curiosity, the desire for understanding, the capacity for sustained attention, the dissatisfaction with ease when it becomes empty, these are not easily extinguished. They can be neglected, distorted, or redirected, but they remain part of what it is to be human.

The question, then, is whether educational systems will align themselves with these capacities or with the logic of the machines they employ. If they choose the latter, the inversion described here will deepen. If they choose the former, there is still the possibility of a different trajectory, one in which artificial intelligence supports human flourishing rather than quietly reshaping it.

This is not a question that can be answered once and for all. It will need to be addressed continuously, as technologies evolve and as their effects become more visible. What this book has tried to do is to make the stakes of that question clearer. The issue is not whether we will live with AI. We already do. The issue is what kind of minds we will cultivate in the process, and what kind of world those minds will go on to create.

If there is a way to bring this argument to rest without closing it down, it lies in returning to the question of purpose, not as an abstract philosophical category, but as something lived, enacted, and defended within institutions. It is here that Brighouse’s voice, and the tradition of thought and practice he represents, becomes more than illustrative. It becomes necessary.

Brighouse did not write in the language of artificial intelligence, nor did he frame his concerns in terms of cognitive architectures or algorithmic environments. Yet the clarity with which he held on to the moral purpose of education gives his work a particular force in the present moment. For him, schooling was never reducible to the transmission of information or the production of measurable outcomes. It was about enlarging lives. It was about opening up worlds that children did not yet know existed, about cultivating the confidence to think, to question, to imagine, and to participate in a shared social life with others. It was about forming persons, not simply producing performers.

What is striking, when one places this alongside the developments described in this book, is how fragile that moral purpose becomes when the ecology of education shifts. The danger is not that educators abandon it explicitly. It is that it becomes quietly displaced. When systems reward output, when technologies accelerate production, when institutional pressures demand visible evidence of progress, the deeper aims of education can be crowded out, not by opposition, but by substitution. Something else comes to occupy the space, something more measurable, more immediate, more easily aligned with the logic of the system.

Brighouse would have recognised this danger immediately, even if the technological context is new. He was acutely aware of how policy environments and institutional pressures can distort educational practice. He understood that schools are always situated within wider social and political frameworks, and that these frameworks can either support or undermine their moral purpose. What he insisted on, repeatedly, was that educators must hold on to that purpose with a kind of stubborn clarity. Not as a slogan, but as a guide to judgment, to decision making, to the design of environments in which young people grow.

In the context of generative AI and the neoliberal ecology described in this book, that insistence becomes more demanding. It is no longer sufficient to articulate the moral purpose of education in general terms. It must be actively defended against forces that operate at the level of infrastructure, of habit, of everyday practice. It must shape how technologies are used, how curricula are designed, how assessment is conducted, and how success is understood.

This means, in practical terms, that educators will need to make decisions that run against the grain of the system in which they work. They will need to create spaces in which students are required to think without immediate recourse to AI, not as a rejection of technology, but as a protection of the conditions under which thinking develops. They will need to design tasks that cannot be reduced to the production of outputs, tasks that require interpretation, dialogue, and sustained engagement. They will need to attend to the relational dimension of education, to the ways in which trust, challenge, and shared inquiry shape learning in ways that no system can replicate.

More broadly, it suggests that the moral purpose of education cannot be confined to schools alone. If the ecology within which education operates is shaped by wider social forces, then the defence of that purpose must extend beyond the classroom. The same questions that arise in education arise in society more generally. What do we value? What kinds of activities do we reward? What forms of attention do we cultivate? What do we count as meaningful achievement?

Brighouse’s perspective offers a way of holding these questions together. He did not separate education from society, nor did he treat schools as isolated institutions. He saw them as part of a broader moral and civic project, one in which the formation of individuals is inseparable from the health of the society they inhabit. In this sense, the argument of this book extends beyond education, even as it begins there. The inversion described here, the adaptation of human cognition to machine logic within a neoliberal ecology, is not confined to classrooms. It is visible in workplaces, in public discourse, in everyday interactions with digital systems.

To give Brighouse the last word, then, is not to offer a simple resolution, but to anchor the argument in a form of practical hope. Not an optimistic belief that things will work out, but a commitment to the idea that purpose can still guide practice, that institutions can still be shaped by values, and that the formation of human beings remains a task worth taking seriously, even when the conditions under which it takes place are changing.

The challenge ahead is to take that moral purpose seriously enough to let it shape our response to technology, rather than allowing technology to redefine the purpose itself. If we fail to do so, the inversion will deepen, and education will continue in a form that looks familiar while becoming something else. If we succeed, even partially, then the presence of AI need not entail the erosion of what makes education human. It may, instead, force a clearer articulation of why that humanity matters, and how it can be sustained.

In that sense, the future of education, and perhaps of the societies it helps to shape, will depend less on the capabilities of our machines than on the clarity of our purposes. Brighouse understood that. The question now is whether we are prepared to act on it.