Chapter 5 — The Politics of AI Tools (plain-language version)

The tools are not neutral

Chapter 4 described what the three-scale system looks like when it’s working — how individual formation, firm identity, and domain character constitute each other. This chapter asks what the political economy of contemporary AI mediation is doing to that system.

Two arguments run through the chapter. The first is about sovereignty: AI tools carry cultural and discursive patterns that don’t respect national or firm-level distinctiveness. The second is about extraction: the knowledge firms externalise through AI-mediated workflows can flow outward to vendors and back into the market as generic capability. Both arguments operate through the same underlying mechanism — generic AI tools, because they’re generic, tend to smooth away distinctiveness.

The chapter is the most politically direct in the thesis. Some of its claims are structural arguments based on documented patterns; some are more speculative and it says so.

Sovereignty: what gets lost when the tools carry someone else’s assumptions

Here is the sovereignty argument.

Large language models are trained primarily on large bodies of text. The text that forms the majority of any current general-purpose AI model’s training data is predominantly Anglo-American in origin — English-language, shaped by American and British professional and cultural conventions. This isn’t controversial; it’s documented in the academic literature on AI training data and cultural bias.

What follows is structural, not intentional. The model produces Anglo-American patterns most fluently. When European infrastructure advisers use generic AI tools to draft regulatory analysis, they are using tools whose defaults are shaped by a different professional tradition from the European public-interest practice described in Chapter 2. The tools don’t carry European regulatory instincts, don’t default to the European public-interest tradition’s way of weighing long-term systemic risk, don’t frame problems the way a practitioner formed in the Dutch or French or German advisory tradition would.

This doesn’t mean the outputs are wrong. Often they’re adequate. The concern is cumulative: across hundreds of engagements, across dozens of practitioners, across multiple firms, the texture of European advisory work gradually tilts toward the dominant patterns in the training data. Not dramatically on any single engagement. Steadily, structurally, over time.

Tao et al.’s research, which analysed successive versions of a major AI model across 107 countries, found that the model consistently defaults to values resembling English-speaking, Protestant European contexts. The European regulatory and public-interest registers are present in the training data — but they’re minority patterns, not the defaults.

The framework calls this national-cultural dis-individuation: the gradual erosion of distinctiveness under the weight of a tool whose defaults pull toward something different.

Sovereignty at the firm level: the distinctive voice being smoothed away

The same mechanism operates at the level of individual firms.

Each consulting firm has developed, through years of work, a characteristic way of seeing problems — specific framings, methodological reflexes, sectoral instincts, vocabulary in which the firm recognises a situation. Chapter 4 described how this is produced and reproduced. This distinctiveness is what makes one firm meaningfully different from another even when they’re selling similar services.

Generic AI tools produce outputs that converge toward common defaults. The model was trained on the whole corpus of consulting output, and its defaults represent the average of that corpus, not any particular firm’s distinctive approach.

Madrid, 2025. A Spanish infrastructure advisory firm has built its reputation on a specific reading of Spanish regulatory dynamics — formed through two decades of experience with CNMC, the specific relationships the firm’s senior consultants have developed, knowledge built through years of engagement with particular officials and processes. An Applied-stage consultant uses an AI tool to structure the regulatory landscape section of a client report. It’s structurally correct. It’s well-written. It reads as if written by a team advising a generic European regulatory authority. The firm’s managing partner reads it and asks whether it was AI-assisted. The consultant says partly. The partner doesn’t say this is wrong. He says: the client is paying for our reading of CNMC, not a reading of what a generic NRA would do. Those are different things. I need to see the difference in the document.

The consultant’s output was adequate. It didn’t demonstrate what the firm had built over twenty years.

The philosopher Walter Benjamin’s concept of aura is useful here: the quality of authentic presence that derives from something being embedded in a specific tradition. An expert senior adviser’s recommendation carries this quality — it comes from a particular person’s sustained engagement with a particular regulatory tradition, and the client can feel that embeddedness. AI-generated analysis lacks this quality; it carries the generic marks of the training corpus rather than the marks of specific formation. Benjamin himself noted the ambivalence: mechanical reproduction makes competent output widely available, which is a real gain. What it can’t reproduce is the auratic quality of genuine counsel. Both the gain and the loss are real simultaneously.

Extraction: knowledge flowing the wrong way

The second argument is about what happens to firm knowledge when it flows through AI-mediated workflows.

The traditional assumption in consulting knowledge management is that the frameworks, case studies, analytical documents, and playbooks that a firm develops stay within the firm. They’re the firm’s intellectual property, transmitted to the next generation of practitioners through training and apprenticeship.

AI-mediated workflows create a structural pathway through which this assumption can fail.

When firms send working materials — drafts, frameworks, internal analyses, client documents — through AI tools operated by external vendors, the firm’s externalised knowledge is no longer entirely within the firm’s control. Depending on the vendor’s terms of service and the deployment configuration, that material may be used to train subsequent models. What the firm has carefully built over decades can, structurally, become part of the generic capability that the vendor then sells to the firm, to competitors, and to clients.

The firm’s distinctive knowledge stops being distinctive.

London, 2024. An infrastructure advisory firm deploys an AI summarisation tool for regulatory filing review. A senior partner asks IT to check the vendor’s agreement before processing client materials. The standard enterprise tier includes a training-data clause. The partner escalates. Three weeks of legal review follow. The firm ends up on a bespoke agreement explicitly excluding client content from the vendor’s training distributions, with audit rights built in. This cost additional legal fees and delayed the deployment. Three months later, a peer firm receives a client request asking whether their AI tools use client content for training. The peer firm doesn’t have a clear answer.

The thesis is careful about this argument. It doesn’t claim that all firms’ knowledge is currently being extracted into vendor training data. It claims the structural pathway exists, that analogous patterns are well-documented in adjacent industries (surveillance capitalism, platform governance), and that the consequences if it happens are significant enough to treat as a strategic concern. The frameworks used in existing consulting research don’t engage extraction as a concern at all.

When the SECI model — the virtuous cycle through which firms convert tacit knowledge into explicit assets and back again — runs in reverse, firm-distinctive knowledge flows outward. The thesis calls this the dark mirror of the SECI model.

The defence comparison

European policymakers have already recognised something like the sovereignty argument in one infrastructure sector: defence.

Defence sovereignty is institutionally established — there are domestic procurement preferences, technology transfer requirements, restrictions on foreign ownership of strategic defence firms. The logic is that defence capability can’t be subject to decisions made by entities outside the European political community without compromising European autonomy.

The thesis doesn’t claim energy or water advisory should be treated exactly like defence. But the underlying logic — that some decisions are too consequential to be dependent on systems and institutions outside European control — applies, in modulated form, across other infrastructure sectors. Chapter 9 develops what policy instruments might follow from this.

Warsaw, 2023. An advisory team conducting a rail network resilience assessment begins receiving questions from a Ministry official that aren’t about passenger demand or maintenance cycles. They’re about freight corridor capacity under disruption scenarios, alternative routings for bulk materials, recovery timelines for specific segments. The team lead recognises, without anyone saying so directly, that these are not transport planning questions. The engagement subsequently re-scopes. The resilience methodology designed for commercial transport planning is not the methodology that applies when the infrastructure is also a strategic logistics asset. Nobody states this explicitly. The team lead understands it.

Economic theory converges on the same concern

Two economic research programmes arrive, from very different directions, at conclusions consistent with this chapter’s argument.

Acemoglu, Kong and Ozdaglar built a formal mathematical model of how human learning works when AI tools are available as a substitute for human effort. Their central result: under the right conditions, heavy AI substitution can tip a professional community toward a “knowledge-collapse steady state” — where collective expertise vanishes despite individual outputs continuing to look competent, because no one is doing the work of building and transmitting the shared knowledge the community depends on. The thesis’s concern about European advisory practice is the philosophical analogue of this formal economic result.

Mügge’s analysis of EU AI policy documents found that EU strategy has consistently focused on jurisdictional sovereignty — regulatory authority over AI, data residency requirements — while largely neglecting what he calls epistemic sovereignty: whether European practitioners are genuinely empowered by the tools they use, and whether European professional practices retain their distinctiveness. The thesis’s sovereignty argument occupies exactly the gap Mügge identifies.

The integrated picture

The two arguments — sovereignty and extraction — are different expressions of the same underlying dynamic. Generic AI tools don’t respect distinctiveness. They produce outputs that gravitate toward distributional averages, and cumulative use produces dis-individuation at multiple levels: the individual practitioner’s analytical voice, the firm’s distinctive character, the European professional tradition’s public-interest orientation.

The conditions that produce this are not fixed. They’re political-economic conditions that reflect choices: who builds AI tools, with what training data, under what ownership structures, deployed under what terms. Those conditions could be different. Making them visible is part of what the chapter’s argument is for.

Chapter 6 looks at whether the same pattern is visible in professions that have had AI tools operating at scale longer than consulting — software development and law.