The Infrastructure of Form

My first paper in artificial intelligence was published in 1995. I have spent the three decades since watching the field travel through a sequence of stages, each of which seemed at the time to be where the consequential action was. Some years ago, at a conference dinner in Disneyland, a colleague was describing the current generation of conversational systems with the enthusiasm of someone for whom the work in front of them was the work. The systems were impressive. The deployments were spreading. The capabilities were the question. I found myself unable to follow the conversation forward, because I kept getting stuck on the same observation. In earlier periods of computing, the analytical vocabulary had shifted as the work shifted. Programming was the right unit for a while, then subroutines, then libraries, then modules, then platforms, then ecosystems. Each shift moved the consequential layer up the stack. The conversation I was in seemed not to have done this. The vocabulary was still the vocabulary of individual systems and individual capabilities, while the layer at which the consequential action was occurring had, on every observable measure, moved several layers up. This book is an attempt to describe the layer where the action is now.

The book's central proposition can be stated directly. The frontier of artificial intelligence is no longer the question of what individual models can do. The frontier is the infrastructure inside which judgment forms: the predictive scaffolding through which everyday cognitive work is now performed, the platform infrastructure on which that scaffolding sits, the regulatory and educational institutions that have not yet developed the categories required to govern it, and the developmental sites at which the next generation of citizens is being formed inside the scaffolding rather than alongside it. The risk is not only that sovereign cognition gets captured. It is that future subjects may never be formed outside capture in the first place. The infrastructure is formative, not merely facilitative. It does not assist thought from the outside; it constitutes a substrate inside which thought develops. The political category for this condition does not yet exist. Building the category, identifying the conditions under which the category becomes operable, and naming the institutional response that does not depend on conditions which the book argues are unlikely to be satisfied: this is what the book attempts.

The level of description matters. Most discussion of these questions operates at the level of the individual model: what it can do, what it cannot do, whether it is aligned, whether it has been jailbroken, whether it should be released. The discussion at this level is real and important. The discussion at this level is also not where the structurally consequential action is occurring. The structurally consequential action is occurring at the level of infrastructure: the orchestration stacks that combine models into agentic systems, the verification regimes that determine whether these systems can be trusted, the standards that determine which firms set the terms for the rest, the educational and developmental institutions that determine which cognitive practices the next generation acquires inside the systems and which it acquires outside them. The book describes this level. The book argues that the level deserves the name "the infrastructure of form" because what is at stake at this level is not what gets done with one's mind but what one's mind is.

The dominant analytical vocabulary in contemporary AI policy is the vocabulary of model safety: capabilities, alignment, evaluation, red-teaming, deployment risk, frontier governance. The vocabulary has produced serious and useful work. The International AI Safety Report 2026, chaired by Yoshua Bengio and published in February 2026 with the backing of approximately thirty national governments and three international organisations, is the most authoritative recent synthesis of where this work has reached. Within its specified scope, the report is careful, empirically grounded, and analytically sophisticated. The book does not contest the report's findings within their scope. The book takes a different scope.

The book's argument is that the model-safety vocabulary, taken as the primary apparatus for understanding what artificial intelligence is doing to the societies that are deploying it, is mistaken about the layer at which the consequential action is now occurring. The vocabulary was developed in a period when the relevant questions were about what individual models could do and whether they were safe to release. The questions are still real. They are no longer the questions that determine outcomes. The questions that determine outcomes are about how models are combined into agentic systems whose failures are correlated across the shared base; about who has the authority to verify the claims that systems' developers make about them; about which firms set the technical and contractual standards that the rest of the sector then has to follow; about which children encounter these systems at which ages under which institutional supervision; about which legal categories are available to describe what is being done to a person whose cognitive substrate is being built and owned by an entity whose interests are not aligned with theirs. These are not capability questions. They are infrastructure questions, governance questions, formation questions. The book argues that they are the questions, and that treating them as second-order issues subordinate to capability assessment misdescribes the situation.

The book does not inherit the model-safety vocabulary's terms. It uses different theoretical anchors and different empirical anchors. The theoretical anchors are Foucault on governmentality and the historical construction of subjects, Arendt on natality and the political conditions of developmental beginning, Dewey on education as the formation of habits of social participation, Sennett on craft as the development of competence in resistant material, with Aristotle as the background of practical reason whose displacement the book is tracking. The empirical anchors are documented agentic-system failures, regulatory records of platform governance disputes, the peer-reviewed literature on cognitive offloading and developmental scaffolding, the public records of judicial and administrative proceedings against the major platforms, and the legal and policy documents of the European Union, the United States, and other jurisdictions that have begun to construct regulatory regimes for the sector. The book's adjacent literatures, where its argument has the most to learn from and to contest, are surveillance capitalism (Zuboff), platform studies (Gillespie, van Dijck), the legal construction of informational capitalism (Cohen), the history of regulated industries (Carpenter, Moss), and the developmental psychology of cognitive formation (Bruner, Vygotsky in the original Russian and the contemporary literature that has developed from him). The book does not pretend to settle questions in any of these adjacent literatures. The book draws on each where the borrowing illuminates the argument.

The book's claims are conditional and structural. They are conditional in the sense that they identify conditions under which the claims hold and acknowledge that the conditions may not hold in every case. They are structural in the sense that they identify regularities at the level of how the infrastructure operates rather than at the level of any specific individual outcome. Two clarifications follow.

The book does not predict catastrophe. The chapters that engage what might be called catastrophic possibilities (Chapter 4 on the cognitive shadow, Chapter 9 on subjects before sovereignty, Chapter 10 on generational asymmetry, Chapter 11 on the catastrophic pattern) are careful to specify the conditions under which the possibilities they describe would materialise. Where the conditions are uncertain, the book says so. Where the empirical evidence is contested, the book identifies the contestation and engages it. The book is not in the business of forecasting how bad things might become; it is in the business of identifying the structural features that make certain bad outcomes possible and specifying what would have to be true for those outcomes to be prevented. The reader who finishes the book with the impression that the author has issued specific predictions has misread the argument.

The book also does not engage in what might be called moral panic about technology. The historical record of overstated technology panic is real, well-documented, and important to acknowledge. Chapter 10 engages this record directly under the heading of the strongest-objection structure that every chapter carries. What the book claims is more specific: that the operational features of platform-mediated cognitive infrastructure are genuinely new in respects the chapter develops at length, that the structural argument the book makes does not depend on emotional or rhetorical investment in any specific outcome, and that the historical pattern of overstated panic is itself one of the conditions that makes the current situation structurally difficult to address.

The conditional and structural character of the argument has implications for what would count as falsifying it. Empirical investigation of cognitive-formation effects across a generation could refute the claim that the infrastructure is formative in the developmental sense the book argues. Comparative jurisdictional outcomes across the European Union, the United States, and other regions could refute the claim that the political category of the externally hosted self is the right category for the regulatory work the book argues it would have to do. Longitudinal data on the cohort whose formation has occurred inside the systems could refute the claim that the reversibility of platform-mediated cognitive practice is structurally uncertain. The book invites this kind of empirical engagement. The book does not invite, and does not respond to, the kind of engagement that takes the structural argument as an emotional commitment to be matched with countervailing emotional commitments. The argument is offered as analytical work; the appropriate response is analytical work in return.

The book is in five parts and twelve chapters.

Part I (Chapters 1 and 2) develops the reframe. The argument is that the model-safety vocabulary mistakes the layer at which consequential action is occurring. Chapter 1 takes up coordination: how agentic systems combine models, how the orchestration layer rather than the model layer determines outcomes, what coordination failures look like in practice, what the policy implications of the coordination focus are. Chapter 2 takes up verification: what it would mean to know whether the systems do what they are claimed to do, why current verification practice is theatrical in the sense the chapter develops, what the verification gap implies for the governance of the sector.

Part II (Chapters 3 through 5) develops the cognitive substrate. Chapter 3 takes up adaptive predictive scaffolds: what they are, how they operate, how they differ from earlier cognitive tools, what their developmental implications are. Chapter 4 takes up the cognitive shadow: the effect of long-term reliance on predictive scaffolding on the underlying cognitive practices of those who rely on it, with attention to the technical literature on machine unlearning as a structural parallel to the cognitive case. Chapter 5 takes up invisible lock-in: the structural feature that the cognitive practices developed inside predictive scaffolding are not, in general, reversible at the level of individual practice once the developmental window has closed.

Part III (Chapters 6 through 8) develops the political category. Chapter 6 takes up cognitive sovereigns: the analysis of major platforms as quasi-political entities whose structural position resembles that of the chartered companies of the early modern period, drawing on Philip Stern's historical analysis and on Julie Cohen's contemporary analysis of joint sovereignty under informational capitalism. Chapter 7 takes up the standards war: the structural feature that the firms setting the technical and contractual standards for the sector enjoy a position of advantage that the regulatory tradition has not yet developed categories to address. Chapter 8 takes up stratified subject formation: the legal and empirical record on differential outcomes for different populations under platform-mediated systems, drawing on the United States v. Meta Platforms litigation and the broader empirical literature on algorithmic stratification.

Part IV (Chapters 9 and 10) develops the formation argument. Chapter 9 takes up subjects before sovereignty: the developmental period during which the externally hosted self is constituted, the political and legal implications of this constitution, the engagement with Foucault on subjection and Arendt on natality. Chapter 10 takes up generational asymmetry: the structural feature that the cohort writing policy on these systems is the cohort whose formation predates them, while the cohort whose formation has occurred inside them has the least direct purchase on the policy that will determine whether their formation conditions are protected, modified, or accepted as they are.

Part V (Chapters 11 and 12) develops the closing window. Chapter 11 takes up the catastrophic pattern: the historical regularity that crises in privately operated economic systems have produced state absorption of regulatory functions, and the structural assessment of whether the pattern is available in the contemporary case. Chapter 12 takes up the educational mandate: the affirmative institutional proposal of the book, which argues that schools are the institutional site at which the formation conditions can be modified without depending on the conditions Chapter 11 specifies.

The order of the parts is not arbitrary. The diagnosis (Parts I and II) precedes the political analysis (Part III); the political analysis precedes the formation argument (Part IV); the formation argument precedes the institutional response (Part V). A reader who takes the parts in another order can follow the chapters, but the cumulative force of the argument depends on the order the book presents.

The book has three disciplines that the reader is entitled to know about in advance.

The first discipline is empirical. The book's claims are anchored, where possible, in documented cases, regulatory records, peer-reviewed literature, public reporting, and the legal and policy documents of the relevant jurisdictions. Where the empirical record is contested, the book identifies the contestation and engages it. Where the empirical record is thin, the book says so. The book does not anchor its claims in anecdote, in author intuition, or in projection from cases the author has personally observed; it anchors them in evidence the reader can independently verify. Each chapter carries its own Notes section in which sources are cited in full on first mention. The citations are arranged in citation order rather than alphabetical order so that the reader can move between the argument and the source material without losing place.

The second discipline is theoretical. The book engages four primary classical interlocutors (Foucault, Arendt, Dewey, Sennett) and one background figure (Aristotle) as genuine intellectual presences whose arguments the book is developing rather than as stepping stones for the book's own contributions. The contemporary interlocutors (Cohen, Stern, Pasquale, Zuboff among others) are similarly engaged on their own terms. The book does not claim to have settled the interpretive questions in any of these literatures. The book draws on each where the engagement illuminates the argument and acknowledges where it is borrowing rather than contributing.

The third discipline is rhetorical. The book aims at a clinical-administrative register throughout. It avoids vocabulary of inflated significance, transitional fillers that signal argument-by-tone, and the rhetorical structures that substitute for analytical structure. The discipline is in service of one purpose: to make the structural argument legible to a reader who can disagree with it on the merits. The reader who finishes the book unconvinced is invited to identify the chapter, the section, the specific claim, and the specific evidence at which the disagreement begins. The book has been written to make that work possible.

This is what the book attempts. The argument starts in Chapter 1.