Intelligence as a Social System

How Technology Reshapes Knowledge, Institutions and Power

The five essays preceding this one treated intelligence as a property to be analysed: a concept with a contested extension, a structure of interdependent conditions, a relation between symbols and their referents, a process separable into generation and evaluation, a status conferred through recognition rather than discovered through inspection. Each of these analyses proceeded largely as if the question were adjudicated by argument and evidence alone. It is not. Every claim about what a system understands, creates or deserves is made inside an infrastructure that someone owns, trained on material that someone extracted, evaluated against standards that someone set, and deployed into an economy that redistributes its consequences unevenly. This essay treats intelligence as that infrastructure, rather than as a property some entities have and others lack.

The move is not a departure from philosophy into politics. Michel Foucault’s work, particularly Surveiller et punir, published in 1975, argued that knowledge and power are not external to one another, one neutral and one applied, but constitute a single circuit: what counts as true in a given period is inseparable from the institutional apparatus that produces, verifies and circulates it, and that apparatus is always also an apparatus of power. Foucault’s point was not that truth is arbitrary. It was that the question “is this true” and the question “who has the standing to say so, using what instruments, verified by what institution” are not two questions but one, examined from different angles. Applied here, the question “is this system intelligent” and the question “who built it, who owns its weights, whose data trained it, and who profits from the answer” are likewise a single question wearing two faces.

What “control” actually bundles

Public discussion of artificial intelligence typically asks who controls it as though control were one thing, held or not held by a given company, government or individual. It is at least four distinct levers, owned separately, governed by different rules, and concentrated to different degrees.

The first is compute: the physical hardware and energy required to train and run large models. Dario Amodei and Danny Hernandez documented in a 2018 analysis that the amount of compute used in the largest published AI training runs had grown by a factor of roughly three hundred thousand between 2012 and 2018, doubling on average every three and a half months, a growth rate far exceeding the historical pace of general semiconductor improvement described by Moore’s law. A trend of that shape has a direct economic consequence: the capital required to train a frontier model rises with it, and the number of organisations able to meet that capital requirement falls. Compute is not a neutral input that anyone with sufficient interest can acquire. It is a scarce, capital-intensive resource whose ownership concentrates as the frontier advances, independent of any decision about openness or restriction made at the level of the resulting model.

The second is data, the raw material from which models derive their statistical structure. Shoshana Zuboff’s The Age of Surveillance Capitalism, published in 2019, described the process by which human experience, behaviour and communication are captured as a free input, rendered as behavioural data, and rendered into a proprietary asset used to train systems that are then sold back into the economy those humans occupy. The people who produced the underlying material, by writing, posting, photographing, conversing, ordinarily hold no claim on what is built from it, and in most jurisdictions no comprehensive legal framework requires that they do. Data governance is therefore a second, separate lever, distinct from compute, with its own separate history of concentration: platforms that accumulated large user bases over the preceding two decades hold a data advantage that a well-capitalised new entrant cannot simply purchase, because much of the relevant material was never for sale.

The third is the model itself, the trained weights and the architecture that produced them, together with the decision of whether and how those weights are released. Whether a given system’s weights are held privately, licensed commercially, or published openly determines who can build on it, audit it, or compete with it, and this decision is currently made unilaterally by whichever organisation trained the system, subject to no general legal obligation in most jurisdictions to choose one regime over another.

The fourth lever is the least visible and the most consequential for the questions raised earlier in this series: the authority to say what a system’s output means, whether it should be trusted, and what standard it should be judged against. This is a standard-setting function, and it does not require owning compute, data or weights at all. It is exercised by benchmark designers, by regulatory bodies, by professional associations that decide whether AI-assisted work meets a field’s evidentiary standards, and by the press and academic institutions that certify some claims about a system’s capabilities as credible and dismiss others as hype or alarmism. Pierre Bourdieu’s analysis of symbolic power, developed across La distinction in 1979 and stated most directly in Language and Symbolic Power, the 1991 English collection of his work on the topic, described exactly this kind of authority: the capacity of a field’s dominant institutions to have their categories of judgment accepted as the natural, self-evident measure of value, such that the exercise of power is no longer visible as power at all, but simply as how things are correctly assessed. A benchmark that defines what counts as reasoning, or a professional body that defines what counts as an acceptable use of a generative system in legal or medical practice, is exercising exactly this kind of authority, and the authority is distinct from, and can be held independently of, ownership of the underlying technology.

These four levers, compute, data, model, and standard-setting authority, are typically discussed as though concentrated in the same hands by default. They are not necessarily concentrated together, and distinguishing them is what allows the question “who controls AI” to be answered with any precision, rather than gesturing at an undifferentiated notion of corporate or technological power.

Whose knowledge counts

The third essay in this series established that a system trained on text has learned relations among symbols without thereby acquiring reference to what those symbols are about. A social consequence follows from this that the earlier discussion did not draw out. The corpus a system is trained on is not a neutral sample of human knowledge. It is whatever text happened to be digitised, indexed and accessible at the moment of training, and digitisation itself has a history with a well-documented skew: languages with large numbers of speakers but limited digital publishing infrastructure are underrepresented relative to their speaker populations, academic and journalistic registers are overrepresented relative to oral and vernacular knowledge traditions, and material produced in institutional and commercial contexts with strong incentives to publish online is overrepresented relative to material that circulates primarily offline or within closed communities.

A system trained on this material inherits a specific, historically contingent distribution of whose knowledge was written down, in what register, and made freely available for scraping. When such a system is then deployed as a general-purpose source of answers, it does not merely reproduce a technical artefact. It reproduces, at scale and with an appearance of neutrality that a named human author does not carry, a particular distribution of epistemic authority that was already unequal before any model was trained, and it does so while presenting its output in a uniform, confident register that gives no indication of which parts of its answer rest on well-attested material and which rest on thin or skewed representation in the training data. Bourdieu’s account of symbolic power applies here with unusual precision: a system that appears to speak from nowhere, with no visible institutional position, is in fact speaking from a very particular position, constituted by the specific history of what got digitised and by whom, and the absence of a visible speaker is what allows that position to pass as neutral.

The restructuring of work

Discussion of AI and employment tends toward a binary that the relevant economics literature has long since rejected: jobs will either be automated away or they will not. David Autor, Frank Levy and Richard Murnane’s 2003 paper, “The Skill Content of Recent Technological Change,” established the task-based framework that has since become standard in labour economics, distinguishing routine tasks, which follow explicit, codifiable procedures and are readily automated, from non-routine tasks, which require flexible judgment, and showing that computerisation over the preceding decades had systematically displaced the former while complementing and increasing the value of the latter. Daron Acemoglu and Pascual Restrepo extended this framework in a series of papers through the 2010s, distinguishing a displacement effect, in which automation of a task directly reduces the labour required to perform it, from a reinstatement effect, in which the same technological change creates new tasks in which labour has a comparative advantage, and showing empirically that the net effect on wages and employment depends on the balance between the two, a balance that is not fixed by the technology itself but by the rate and direction of task creation relative to task destruction.

This literature matters here because it replaces the question “will AI take jobs” with a more tractable and more consequential one: which specific tasks within a given occupation is a given system capable of performing at acceptable reliability, who captures the resulting productivity gain, and is the economy generating new non-routine tasks fast enough to reabsorb the labour displaced from routine ones. None of these sub-questions has a technologically determined answer. The distribution of gains from task automation depends on bargaining power, ownership of the automating technology, and policy choices about how productivity gains are taxed and redistributed, all of which are institutional variables rather than technical ones. A given advance in language model capability can, depending entirely on these institutional variables, either concentrate income toward the owners of the automating capital or be broadly distributed through wage growth, shorter hours, or public revenue, and the technology alone settles none of this.

General-purpose technology and its governance problem

Timothy Bresnahan and Manuel Trajtenberg’s 1995 analysis of general purpose technologies identified a class of innovations, historically including the steam engine, electricity and the internet, characterised by pervasiveness across sectors, continued technical improvement over time, and the generation of complementary innovations in the industries that adopt them. Artificial intelligence, on the evidence available through the mid-2020s, fits this classification: it is not a tool confined to one industry but an input that reshapes production processes across sectors as diverse as software development, diagnostic medicine, legal research and logistics.

Bresnahan and Trajtenberg’s analysis also identified the characteristic governance difficulty that accompanies technologies of this class, and it is worth stating precisely because it is a structural point independent of any particular company or product. A general purpose technology’s value depends on complementary investment made by the industries that adopt it, and coordinating that investment across many independent adopters is difficult, which historically produces a lag between the technology’s availability and its full economic diffusion, together with a period in which the small number of entities that control the technology at its source hold disproportionate influence over the terms on which every downstream industry can use it. Electricity generation and distribution was consolidated into regulated utilities precisely because an unregulated general purpose input of this kind produced unacceptable concentrations of leverage over every industry that depended on it. No comparable regulatory settlement yet exists for the compute, data and model layers described above, and the historical pattern gives reason to expect that the absence of one is not a stable equilibrium but a temporary condition that will eventually be resolved, either by regulation or by the further entrenchment of whichever parties hold the leverage in the meantime.

The European Union’s AI Act, referenced briefly in the first essay of this series, is one attempt at this kind of settlement, structured around a graduated classification of systems by risk rather than around the compute, data, model and standard-setting distinction developed here. Other jurisdictions have moved more slowly or along different lines, and the international variation itself is a further source of leverage for whichever organisations can operate across regulatory regimes and route their most consequential activity through the most permissive one available. This is a standard dynamic in the governance of any technology whose production is more mobile than the regulatory authority meant to oversee it, and it has no technical solution.

Intelligence without a subject

The conclusion this essay is building toward is not that concentrated control over AI infrastructure is a problem layered on top of the philosophical questions examined earlier in this series. It is that the two are the same phenomenon examined from different distances.

When the first essay in this series noted that definitions of intelligence are fixed in the middle of a distributive conflict, by participants with material interests in where the lines fall, that observation was not a passing remark. It is the organising fact of this entire domain. A claim that a system reasons, understands, or possesses genuine creativity is never made from nowhere. It is made by an organisation seeking capital and regulatory latitude, by a competitor seeking to discredit a rival’s product, by a labour organisation seeking to protect a profession’s jurisdiction, or by a regulator seeking a tractable category to legislate around, and the claim’s plausibility to a wider public is itself shaped by which institutions have the symbolic authority, in Bourdieu’s sense, to make it stick.

This does not mean the philosophical questions examined in the preceding essays are unreal or reducible to interest. The hard problem of consciousness, the symbol grounding problem, and the structural analysis of personhood as a relational status are genuine problems, not disguises for economic conflict. It means that in the world as it currently operates, answers to these questions are not arrived at through disinterested inquiry alone and then separately applied. They are contested, provisionally settled, and revised inside an institutional and economic structure that has its own stakes in which answer prevails, and that structure, examined through the four levers of compute, data, model and standard-setting authority, is where the philosophical and the political turn out to be a single object rather than two.

The question that opened this essay, who controls the algorithms, is therefore not a separate department of this inquiry, added to round out its scope. It is the same question the first essay asked, who benefits from a given definition of intelligence taking hold, made concrete: benefits through what channel, measured in what currency, secured against what competing claim. Treating intelligence as a social system rather than as a property of minds or machines is not a broadening of the topic. It is a correction to the assumption, implicit whenever the earlier essays are read in isolation, that the question of what intelligence is could ever have been separated from the question of who gets to decide.