Chapter 5 - Political Economy of Generic Models: Sovereignty and Extraction

5.1 AI mediation and the political economy of knowledge

The tools are not neutral. 💡

Where Chapter 4 described how expertise is materialised in European infrastructure advisory through the integrated three-scale system, this chapter examines what happens to that system under the political-economic conditions of contemporary AI mediation.

Two parallel arguments run through the chapter: a sovereignty argument about identity at risk under generic mediation operating at two scales, and an extraction argument about the dark mirror of the SECI model when firm knowledge flows outward into vendor systems.

The two arguments are connected. One Stieglerian mechanism (dis-individuation under generic technical mediation) operates across the scales of that system, tilting individual practitioners toward common defaults, eroding firm-distinctive ways of seeing, and homogenising what is distinctive about European public-interest practice; §5.6 draws the two together in full. 💡

The chapter is the most politically charged in this work. It engages claims about the cultural and ideological character of major AI labs, about the structure of the model-vendor industry, and about the conditions under which firm knowledge is exposed to extraction through AI-mediated workflows. These are claims where the empirical literature is uneven, where the most pointed sources are journalistic rather than peer-reviewed, and where I am articulating analytical commitments that the existing literature on AI in consulting has not engaged.

I will try to treat the politically charged claims with care, distinguishing structural argument from documented empirical practice, hedging where the evidence does not support stronger formulations, and noting where a claim is anecdotal and the argument does not rest on it.

The general claim applies to any advisory practice whose distinctiveness is rooted in non-Anglo-American institutional contexts: the dis-individuation pressure operates wherever advisory work is analytically distinctive and depends on formation pathways that AI mediation now absorbs. European infrastructure advisory is where this pressure is most institutionally visible, for the reasons Chapter 2 (§2.3) gave. The chapter develops the sovereignty and extraction arguments in this domain; their applicability to other advisory contexts is a direct implication of the general mechanism.

The chapter also draws on the defence reference case (§5.5) as an institutional anchor for the sovereignty argument.

The political-economic analysis of this chapter belongs to what Morozov (2025) distinguishes as worldmaking rather than administration 1 . Administration manages existing conditions: it allocates tasks, negotiates terms, governs workflows within the configuration of practices already in place.

Worldmaking shapes what conditions are possible: which forms of expertise can be reproduced, which institutional traditions remain available as alternatives to a market default, which professional practices can be sustained across time.

The productivity framing operates in the administrative mode: it addresses the present configuration of AI tools and advisory tasks without engaging what that configuration is doing to the possibility space of future practice. The sovereignty and extraction arguments are worldmaking arguments. They ask what institutional conditions will or will not exist for European infrastructure advisory to remain a distinctive form of practice in two decades; how AI mediation in the field should be governed within current conditions is the administrative question.

5.2 Sovereignty as identity at risk: the national-cultural scale

The sovereignty argument works across two scales. The national-cultural scale is where most of the existing literature on AI sovereignty operates, but that literature mostly addresses the procedural level (data residency, cloud infrastructure, regulatory compliance) and leaves aside the level central to this argument.

The argument is this: generic large language models are produced within particular cultural contexts, and their training distributions carry the patterns of those contexts into the texture of any work the models mediate. The claim concerns distributional weight: the framings, exemplars, and tacit assumptions that dominate a model’s training distribution are the ones it produces most fluently, and the ones practitioners working through it absorb most readily. It does not require that any particular model encode any particular ideology in any direct way.

The cultural patterns of contemporary general-purpose tools are predominantly Anglo-American, a point widely documented in the literature on cultural bias, training-data composition, and AI ethics. This work adds a reading of what that dominance implies for European infrastructure advisory specifically. 💡

The European public-interest tradition in infrastructure has a distinctive character that Chapter 4 described: a tradition of public-purpose engineering, a regulatory architecture for utilities, an intensifying EU governance frame around climate transition and strategic autonomy, and a third-box exposure 💡 that operates differently in Europe than in markets where infrastructure is more thoroughly commercialised. This tradition does not dominate the training distributions of generic AI. The texture of European utility regulation, the political weight of public consultation processes, the institutional force of climate commitments, the content of European strategic autonomy thinking: these are present in the training distributions, but as minority patterns, competing with the dominant Anglo-American register for the model’s expressive weight.

When European infrastructure advisory work is mediated by generic AI, it passes through a system that produces Anglo-American patterns most fluently, and practitioners absorb those defaults more readily than the minority registers. The texture of the work tilts toward the dominant registers of the model’s training distribution and away from those distinctive to the European public-interest tradition. The tilt is not dramatic on any single engagement, but it accumulates engagement by engagement, practitioner by practitioner, until it settles into the house style. This is the national-cultural scale of sovereignty.

We make no claim that the European tradition is being deliberately attacked, or that AI labs are pursuing a cultural-political project against European public-interest thinking. The political-economic conditions under which generic models are produced (predominantly Anglo-American contexts, data, and firms) generate a distributional dominance, and that dominance exerts dis-individuating pressure on minority traditions through sheer cumulative weight. Tao et al.’s systematic empirical study of five successive GPT model versions across 107 countries confirms that generic LLMs consistently encode cultural values resembling English-speaking, Protestant European contexts as their default outputs, with the pattern persisting across model iterations 2 . For European practitioners the implication is direct: the regulatory and public-interest registers they work in are minority patterns in the training distributions, and the default output they meet is the Anglo-American register.

Hayek’s (1945) analysis of dispersed local knowledge, introduced in Chapter 3 (§3.4), gives the same incapacity an economic-theoretic form. His diagnosis (that ‘central planning based on statistical information by its nature cannot take direct account of these circumstances of time and place’) describes, transferred to this context, why the generic AI model trained on statistical aggregation of advisory discourse cannot carry European regulatory specificities as working knowledge. The circumstances of European utility regulation, the institutional weight of public consultation traditions, the specific content of climate-transition governance are present in training distributions, but as abstracted patterns rather than as the ‘knowledge of the particular circumstances of time and place’ that European infrastructure advisory practice requires 3 .

The argument keeps its claims within what the available evidence supports. There is a literature on the ideological character of major AI labs, particularly on the political views of certain prominent figures in the AI industry 💡 . It mixes journalism and scholarship, and it asks about the influence of particular ideological currents, including the libertarian-adjacent, longtermist, and effective-accelerationist strands sometimes associated with the Thiel network and broader Silicon Valley political economy.

These ideological currents are real and documented, but the claim that they directly determine the cultural patterns of model training distributions is harder to support than the structural claim about distributional dominance. I rely on the structural claim, which is well-supported, and treat the ideological-content claims as supplementary rather than load-bearing. Where the chapter cites work on the ideological character of major AI labs, it does so to describe the conditions under which the models are produced; it does not advance the stronger claim that any particular ideological content is being directly transmitted through the models.

5.3 Sovereignty at the organisational scale: firms losing their distinctive voice

The second scale is the organisational scale, and here this work departs most directly from the existing literature. That literature addresses task and workflow levels; it does not engage organisational-scale sovereignty as a concern. It is one, because the firm’s distinctive way of seeing (what makes one consulting firm meaningfully different from another even when they sell similar services) is organisational individuation in the sense Chapter 4 described.

Madrid, 2025. A Spanish infrastructure advisory firm has built its reputation on its specific reading of the regulatory dynamics between Spanish network operators and the national energy regulator, a reading formed through two decades of post-liberalisation experience, through specific relationships with staff across the regulator’s directorates, through a body of engagement knowledge that the firm’s senior consultants carry as professional identity. An Applied-stage consultant on a gas transmission project produces a section on regulatory risk framing for a client report. It is structurally correct. It is well-written. It covers the right categories in the right order. It reads as if written by any global strategy house advising a generic European national regulatory authority, in framings the model learned from documents written mostly elsewhere. The firm’s managing partner reads it and asks whether it was AI-assisted. The consultant says: partly; he used an LLM to structure the regulatory landscape section and then edited for firm-specific framing. The managing partner does not say this is wrong. He says: the client is paying for our reading of this regulator, not a reading of what a generic NRA would do, and every firm in Madrid now has the same tool. Those are different things. I need to see the difference in the document. 💡 💡

At the organisational scale the same mechanism operates within the consulting profession rather than across cultural traditions. Generic AI tools used uniformly across firms produce outputs that converge toward common defaults: the standard analytical structures, framings, and vocabulary that the model has absorbed from the wider corpus of consulting work. The firm’s distinctive way of seeing 💡 is at risk under generic mediation that produces standardised outputs by default.

The risk is gradual. Under cumulative AI-mediated production the work drifts toward common defaults, and the firm finds it harder to demonstrate what distinguishes it through the artefacts that historically did so. The Foundation-stage consultant working through generic AI mediation produces an output that reads as competent management consulting work, and not as competent work in the firm’s distinctive style. The Applied-stage consultant who reviews the output does so through AI mediation of their own, often using the same generic tools that produced it. The Chartered-stage consultant who eventually signs off is working on a deliverable produced and reviewed through processes increasingly mediated by tools whose defaults every firm shares.

The firm’s way of seeing lives inside this process, reproduced through the very engagements AI mediation is absorbing: the formative engagements at Foundation stage, the supervisory engagements at Applied stage, the senior-tier judgment at Chartered stage. When those pass through generic tools, the firm has fewer channels through which to transmit what is its own, and its individuation under the SECI cycle becomes harder to sustain.

Benjamin’s concept of the aura (the quality of authentic presence that derives from a work’s embeddedness in a specific tradition, its “here and now”) provides vocabulary for what is at stake at the organisational scale. The AI-generated advisory output is a mechanically reproduced piece of advisory work in Benjamin’s sense: detached from the tradition it nominally represents, bearing the generic marks of the training corpus rather than the marks of the European public-interest tradition the work answers to. The firm’s distinctive way of seeing is an auratic property: it derives from the firm’s accumulated history, the sedimented judgments of its practitioners, the conventions of its analytical approach built across decades of sustained engagement with particular European regulatory institutions. The firm’s output is its own (recognisable to a client across multiple engagements, legible to a regulator who has negotiated with its practitioners) through the auratic quality of embeddedness rather than as a brand or style: the traces of a tradition expressed through one firm’s sustained encounter with it. A system trained on the full range of consulting outputs cannot reproduce this aura; the generic training corpus averages across firm distinctions rather than deepening any one of them 4 .

Mechanical reproduction makes competent advisory output available at scale, a real gain for analytical access across the profession. AI democratises competence: adequate output, information in Benjamin’s sense, technically sound and analytically tractable. The “auratic” quality through which counsel derives its authority from the adviser’s formation within a specific tradition cannot be mechanically reproduced. The gain and the loss are real and simultaneous; the argument here is about what the gain cannot substitute for.

This is the rigorous form of what the informal discourse calls the “digital twin” concern 💡 . In the informal version, AI tools learn how a particular practitioner works and replace them. In the rigorous version, the firm’s organisational individuation, the substance of what makes the firm valuable across generations of practitioners, is increasingly mediated through tools whose default outputs are common across firms, and the cumulative weight of generic mediation dis-individuates the firm just as it dis-individuates the national tradition.

The two scales are connected by the same mechanism (§5.6 draws them together): generic tools, because they are generic, gravitate toward distributional averages and dis-individuate at whatever scale they mediate. Part of this work’s contribution is to make this multi-scale character visible.

To be clear: firm distinctiveness is not yet visibly collapsing into common defaults. Major consulting firms continue to compete on differentiation, and their senior practitioners continue to articulate analytical positions of their own. The pressure named here accumulates through AI mediation; it has not yet fully shown itself. Whether it resolves into actual organisational dis-individuation depends on conditions that vary: the conditions of model production (which §5.6 develops), the institutional choices firms make about how to integrate AI mediation into their workflows, and the wider trajectory of European infrastructure advisory examined in Chapter 8.

However, the vendor-convergence pattern is already measurable at the market level. By mid-2026, major consulting and professional-services firms had converged on the same two or three model providers, deploying them to hundreds of thousands of employees. The reported figures are industry-sourced and illustrative rather than load-bearing: Deloitte to over 470,000 staff through Anthropic’s Claude; Accenture to 743,000 through Microsoft Copilot and to additional staff through separate OpenAI and Anthropic agreements; PwC to 30,000 certified practitioners on Claude following a 200,000-seat ChatGPT rollout. An industry analysis observes that these announcements, each sold as an edge over the firm next door, “stacked up, … mostly show a handful of large organizations arriving at the same shortlist of vendors within about seven months of each other” 5 . The firms frame this as organisational differentiation through AI investment. From the sovereignty argument’s perspective it reads as the opposite: each firm’s analytical posture is increasingly mediated by the same two or three systems operating on the same training distributions. At the market level, the dis-individuating pressure this chapter names is already in place rather than projected; its effect on what firms produce is what has not yet fully shown itself 💡 .

5.4 Extraction: the dark mirror of the SECI model

The extraction argument carries the chapter’s sharpest political-economic claims. It extends the SECI model that Chapter 4 introduced and reads it through the dark mirror: the firm’s externalised knowledge no longer recirculates within the firm’s own SECI cycle but flows outward into vendor systems and back into the consulting market through generic capability.

London, 2024. An infrastructure advisory firm is deploying an LLM vendor’s document summarisation tool for regulatory filing review, genuinely useful, commercially defensible, and significantly faster than manual review. A senior partner asks the technology team to check the vendor’s standard enterprise agreement before the firm processes client materials through the tool. The IT procurement lead’s initial response is that the standard enterprise tier includes a training-data clause. The partner escalates: the filings carry the firm’s annotated reading of a decade of regulatory decisions, knowledge the firm would otherwise be teaching a vendor’s model to sell back to its competitors. A three-week legal review follows. The firm ends up on a bespoke agreement that explicitly excludes client content from the vendor’s training distributions, with audit rights and breach notification obligations. The partner’s question cost the firm additional legal fees and a delayed deployment. Three months later, a peer firm in the same market receives a client request to confirm that the firm’s AI tools do not use client content for training. The peer firm does not have a clear answer, while the partner’s firm can point to its bespoke agreement. 💡

The historical assumption underlying the SECI model, and underlying organisational knowledge management more broadly, is that the explicit artefacts produced through externalisation remain within the firm. The methodology document is the firm’s IP (intellectual property); the playbook is the firm’s competitive asset; the case study is the firm’s institutional memory. These artefacts are transmitted across the firm’s generations of practitioners through internalisation, and they are protected from competitors through institutional architecture: employment contracts, confidentiality agreements, internal-only access, and the practical difficulty of reproducing tacit firm context outside the firm.

Contemporary AI mediation opens a pathway through which this assumption can fail. When firms adopt AI-mediated workflows that involve sending working materials (drafts, frameworks, case-relevant documents, internal analyses) through systems operated by external vendors, the firm’s externalised knowledge is no longer purely within the firm’s control. The vendor’s terms of service, data-handling practices, and use of customer content in training subsequent models all bear on whether that knowledge stays within the firm or flows outward. 💡

Practice varies by vendor, by contract, and by deployment configuration. Some major vendors offer enterprise contracts that explicitly exclude customer data from training. Some firms use deployment configurations (on-premises models, vendor-hosted but isolated tenant environments, retrieval-augmented systems that keep firm content within firm boundaries) that mitigate the risk. Other vendors and other configurations expose customer content to training, often without the customer’s full understanding. The structural risk exists, but how far it is realised varies with each of these arrangements. 💡

My claim is structural rather than evidentiary. I do not claim that every firm’s externalised knowledge is currently being extracted into vendor training distributions. I claim that the possibility now exists in a way it did not before, that the broader pattern of which extraction would be an instance is well-documented in adjacent domains 💡 , and that the consequences, if extraction occurs in the consulting domain, are significant enough to warrant treating the question as a strategic concern.

The dark mirror reading of the SECI model clarifies what is at stake. In the original SECI model, the externalised knowledge produced by senior practitioners feeds back into the firm through internalisation, where junior practitioners absorb the explicit artefacts into their tacit competence. The firm’s individuation is reproduced across generations through this cycle. In the dark mirror reading, the externalised knowledge flows outward into vendor systems whose training distributions absorb it, and the absorbed content is recirculated into the consulting market through the generic capabilities of the vendor’s models, capabilities that are then sold to the firm itself, to its competitors, and to its clients as they build in-house capability.

Competitors and clients then get access, through the generic distribution, to knowledge that was distinctively the firm’s. The firm’s historical competitive asset, through which it has individuated itself, is at risk of being absorbed into the common ground that AI mediation operates on. The threat is to the firm as a knowledge-managing collective, whose distinctiveness depends on its externalised knowledge remaining within its own SECI cycle, rather than to any individual consultant.

The mechanism is not confined to advisory, and the profession that has documented it most carefully is the one that produces knowledge for a living: researchers. Ngwenyama, Klein and Rowe’s 2026 study of academic publishing in the European Journal of Information Systems shows a handful of publishers using their research and publishing platforms, now equipped with generative AI, to capture what they call the general intellect of science: not only researchers’ outputs but “the social processes (informal and formal) by which they are developed and curated” 9 . The platforms hold “the cognitive intellectual assets (routines, skills, norms of practice), hard won from years of study and professional academic experience”, and Elsevier, the publisher their case study examines, retains “perpetual irrevocable rights” to use them with technologies “now known or later developed” 9 . That clause is the academic counterpart of the vendor’s standard training clause in the London case above, signed once and valid for every model still to come. The repositories are already used “to train GenAI tools to mimic the problem-solving skills of scientists without their permission”, and the output is “sold as bespoke consulting” 9 . The loop closes on consulting itself: one expert community’s externalised knowledge, captured into infrastructure it does not own, returns to the market as another community’s product. The structure is the dark mirror of SECI, observed where the extraction channel is easier to document than in consulting, and the authors’ conclusion carries over: the remedy is collective rather than firm by firm. 💡

The most pointed claim in this section requires care: that data brokers systematically purchase Slack archives, email and chat logs from failed startups to feed into LLM training. I have heard it in informal professional discussion but cannot fully ground it in peer-reviewed sources. The bankrupt-asset data sale market is documented, including in FTC interventions in specific cases, and the data-broker pipelines feeding model training are documented in the broader surveillance and platform-governance literature. But the specific claim about consulting firm working materials moving through these channels is anecdotal in the empirical record I have access to. I treat it as illustrative of a risk demonstrated by analogous cases in adjacent domains rather than as documented current practice in consulting itself. The argument depends on the pattern being real and the consequences being significant, both of which are well-supported; it does not depend on every anecdote about specific extraction pathways being verifiable.

5.5 The defence reference case

European policymakers already treat AI sovereignty as a strategic concern in defence, where the logic of sovereignty has been institutionally recognised for decades. The defence case is a reference point rather than a model to copy: evidence that the sovereignty thinking developed here is available within European policymaking, and that the conceptual move from the procedural level to sovereignty has already been made in at least one infrastructure subdomain.

Warsaw, 2023. An advisory team is conducting a rail network resilience assessment for the national rail infrastructure manager using standard infrastructure methodology: redundancy analysis, demand scenarios, maintenance investment prioritisation. In the third client meeting, a figure whose position is listed on the briefing note as “infrastructure coordination” from a ministry asks a set of questions the team has not prepared for: not about passenger demand or maintenance cycles, but about freight corridor capacity under disruption scenarios, alternative routings for bulk materials, and recovery timelines for specific segments of the east-west corridor. The team lead recognises, without anyone saying so, that these are not transport planning questions. The engagement subsequently re-scopes. The resilience methodology the team has developed, designed for a commercial transport planning context, is not the methodology that applies when the infrastructure is also a strategic military logistics asset. The team leader has never been told this explicitly, but she understands it from the questions asked in the third meeting; no standard methodology, and no model trained on commercial transport planning, would have flagged the shift, because it lies in who is asking. 💡

The sovereignty argument in defence has historically been about strategic autonomy: the conditions under which European defence capability depends on systems controlled by extra-European entities, and what that dependency means for the autonomy of European defence decisions. It has been built into institutional architectures: domestic procurement preferences, technology transfer requirements, restrictions on foreign ownership of strategic defence firms, and, increasingly, specific provisions on AI capability in defence contexts. The logic is that defence capability cannot be subject to the operational decisions of entities outside the European political community without compromising the sovereignty of European defence policy.

The same logic, we can argue, applies to other infrastructure subdomains where the European public-interest tradition is constitutive of the domain’s content. Energy infrastructure decisions are not strategically equivalent to defence decisions, but their content nonetheless depends on European public-interest commitments (to climate transition, to social tariffs, to regulated affordability, to consumer protection) that the sovereignty logic engages. Water infrastructure, transport, urban planning, and the other subdomains in the family Chapter 4 described all share, to varying degrees, this characteristic.

Institutions recognise sovereignty unevenly in these other subdomains. Some EU policy has begun to engage AI sovereignty in non-defence contexts (the European AI Act, certain provisions of the Green Deal, some procurement frameworks), but the recognition of sovereignty as constitutive of European public-interest practice in infrastructure is less developed than in defence. This chapter articulates the logic that would extend defence-style sovereignty thinking to those subdomains; Chapter 9 develops the recommendations to European policymakers that follow from it.

The analogy has limits. I am not calling for defence-equivalent sovereignty architectures across all infrastructure subdomains. Defence has features (secrecy requirements, kinetic capability, military application) that other subdomains do not share, and the institutional architecture of defence sovereignty cannot be transferred wholesale. My claim is that the logic of sovereignty (the recognition that the political community has standing in decisions that affect its autonomy) is institutionally available in defence and applicable, in modulated form, to other infrastructure subdomains. Chapter 9 works out the modulation in the same recommendations.

5.6 The integrated political economy

The two arguments the chapter has developed (sovereignty and extraction) are connected through a single underlying mechanism: dis-individuation under generic technical mediation, operating at multiple scales of the integrated three-scale system.

At the national-cultural scale, generic AI mediation tilts the texture of European infrastructure advisory toward dominant Anglo-American patterns, eroding the distinctiveness of the European public-interest tradition. At the organisational scale, it tilts firm-level outputs toward common defaults, eroding the firm’s distinctive way of seeing as reproduced through the SECI cycle. At the level of the firm’s externalised knowledge, it opens pathways through which firm-distinctive content flows outward into vendor systems whose generic capabilities are then recirculated to competitors and clients. It is one mechanism, operating at two scales and on the firm’s externalised knowledge, and in each case it bears on what the three-scale system Chapter 4 described exists to reproduce.

Political economy is not the only way to make this argument; formal economic theory converges on it. Daron Acemoglu, with Dingwen Kong and Asuman Ozdaglar 10 , develops a formal dynamic model of human learning under agentic AI whose conclusions align with the framework deployed here. Their distinction between general knowledge (community-level, accumulated, complementary to human effort) and context-specific knowledge (individual, idiosyncratic, substitutable by agentic AI) maps onto this framework’s own: expertise reproduced through the integrated three-scale system versus interactional fluency produced by AI mediation.

Their learning externality gives the argument formal backing. Human effort jointly produces private decision-quality and a public good, the community’s general knowledge stock, and that public good is what makes future human effort productive. Transposed to advisory work: the system through which European infrastructure advisory has reproduced its expertise depends on the formative engagements that transmit it, and AI mediation substituted for those engagements depletes the ground future advisory work would rest on. Their central result, that under sufficiently accurate agentic AI the system tips into a knowledge-collapse steady state where general knowledge vanishes despite high-quality individual decisions, is the formal-theoretic analogue of the argument developed above.

The convergence is informative, but the two approaches differ in method. Acemoglu et al. work at the level of information aggregation: their general knowledge is community-aggregated information, statistical in character, and the collapse is the depletion of an information stock. This framework works at the level of formation and individuation: the expertise at risk is not statistical information but the contributory expertise reproduced through the SECI cycle, embodied in firm-distinctive ways of seeing, sustained across generations through the formative engagements that make junior practitioners into senior ones. The two levels complement each other. Acemoglu et al. show formally that an information-aggregation system tips toward collapse under specific conditions; the philosophical argument here shows that a system of professional formation faces analogous conditions, and that what is at stake includes the European public-interest tradition rather than merely the aggregation of information about it. The formal economic argument supports the philosophical and political-economic one without replacing it.

Ide and Talamàs’s two-region model of global knowledge work provides a trade-theoretic complement to the learning-externality argument 11 . Their central distinction is between autonomous AI (which substitutes for problem-solving judgment at the top of the knowledge hierarchy) and non-autonomous AI, which augments human workers without replacing their judgment. For the sovereignty argument the implication is distributional: under autonomous AI deployment, their model shows that sophisticated AI reinforces the knowledge-advanced region’s comparative advantage in high-value problem-solving services. The mechanism that puts European infrastructure advisory’s public-interest distinctiveness at risk is dis-individuating in Stiegler’s sense and also distributional in the formal economic sense that Ide and Talamàs make explicit: the regions that produce generic AI are positioned to consolidate their comparative advantage in advisory services by deploying the systems they produce.

Mügge’s analysis of EU AI sovereignty discourse clarifies the institutional dimension of this argument 12 . His examination of EU Commission strategy documents across 2018–2024 shows that EU policy has consistently prioritised jurisdictional sovereignty (regulatory authority over AI systems, data residency requirements, compliance mechanisms) over what he identifies as epistemic sovereignty: whether European practitioners are genuinely empowered by the AI systems they use, and whether European professional practices retain their distinctiveness under generic mediation. The sovereignty argument developed here operates in the institutional gap Mügge identifies. The procurement architecture the chapter’s political-economic analysis calls for is an instrument of epistemic and not merely jurisdictional sovereignty, and the existing EU policy architecture has not yet made that distinction.

The conditions that produce the mechanism are also the same at every scale. Generic AI is produced by a small number of very large firms, mostly headquartered in the United States, in Anglo-American cultural contexts, with training distributions dominated by the same patterns and commercial incentives that favour broad deployment of generic capabilities across markets. These conditions are not given. They are political-economic and could be different, and part of this chapter’s work is to make visible what would have to change. The structural claims (that distributional dominance dis-individuates minority traditions, that vendor relationships open pathways through which firm knowledge can flow outward) are well-supported; how far they are realised varies by jurisdiction, vendor, firm, and deployment configuration, and I have flagged that variation where it arises.

Chapters 7 to 9 develop the implications for each stage of European infrastructure advisory and for the possible futures the profession faces.

First, Chapter 6 asks whether the mechanism is visible in two adjacent professions (software and law) where it has been operating long enough to leave empirical traces the argument can carry forward.