Chapter 5 — The Politics of AI Tools (plain-language version)
The tools are not neutral
The tools are not neutral. 💡 the shortest paragraph in this thesis and one I wrote last, when I had stopped trying to hedge it
Chapter 4 described how expertise in European infrastructure advisory is kept alive at three scales: the practitioner, the firm, and the European public-interest tradition. This chapter asks what happens to that system when the work passes through AI tools made under today’s commercial and political conditions.
Two arguments run side by side. The first is about sovereignty, by which I mean identity at risk: generic tools can make the work less European, and less any one firm’s own. The second is about extraction: a firm’s knowledge can flow out to AI vendors and return to the market as something anyone can buy. One mechanism sits behind both. Bernard Stiegler, a French philosopher, described how generic technical systems wear away the things that make a person, a firm or a tradition distinctively itself.
The evidence in this chapter is uneven, so I keep claims about how the system is built apart from claims about what any company has been shown to do. The mechanism applies to any advisory practice rooted outside the Anglo-American world. European infrastructure is where it shows most clearly, because duties to the public are most formally binding there. And the question is larger than productivity: in twenty years, will the conditions still exist for European infrastructure advisory to be a distinctive practice?
Sovereignty at the national scale: whose defaults are in the tool?
Most writing on AI sovereignty is about procedure: where data is stored, whose cloud is used, which rules apply. The argument here is different. Generic language models are built within particular cultures, and the ways of writing and reasoning that dominate their training data are the ones they produce most fluently and their users absorb most easily. Those patterns are mostly Anglo-American. 💡 a chapter on Anglo-American training distributions, written in English, by a french person, with a model trained on those distributions helping me type it up. I’m aware. It’s being managed.
Europe has its own public-interest tradition in infrastructure: engineering for public purposes, regulated utilities, public consultation, climate commitments, and duties to what the thesis calls the third box, the publics and future generations who live with infrastructure decisions. That tradition is in the training data only as a minority pattern. So European advisory work that passes through a generic tool tilts toward the dominant patterns. The tilt is small on any one project. It builds up across projects and people until it becomes the house style.
Nobody has to intend this. Sheer weight of numbers is enough. A study by Tao and colleagues of five successive GPT versions across 107 countries found that the models’ default answers consistently reflect the cultural values of English-speaking and Protestant European countries 1 Tao et al. (2024) Cultural bias and cultural alignment of large language models. PNAS Nexus. 3(9). .
Sovereignty at the firm scale: losing a distinctive voice
The same pull operates between firms, and research on AI in consulting, which looks at tasks and workflows, does not see it. Each firm has its own way of seeing, built over decades, and it separates that firm from others selling similar services.
Madrid, 2025. A Spanish firm is known for its reading of how the national energy regulator deals with network operators. A mid-career consultant drafts a section on regulatory risk with the help of an AI model. It is correct and well written, and it reads as if a global consultancy had written it for a generic European regulator. The managing partner tells him the client is paying for the firm’s reading of that regulator, and that he needs to see the difference in the document. 💡 every managing partner has said this, in some form, at some point in the third year of AI-assisted deliverable production
Tools used by every firm drift toward the same structures, framings and vocabulary. Juniors draft with them, mid-career consultants review with them, and seniors sign off work shaped by defaults all firms share. The projects through which a firm passed on its way of seeing are the ones now going through the tools. The critic Walter Benjamin wrote of the quality a work has because it belongs to a particular tradition and a particular here and now 2 Benjamin (1969) The Work of Art in the Age of Mechanical Reproduction. Illuminations. . A firm’s way of seeing is of that kind, and a system trained on everyone’s output averages it away. AI makes competent output widely available, a real gain. It cannot supply counsel whose authority comes from being formed within a tradition.
Firms have not visibly lost their voices yet. But the convergence is already measurable in the market. By mid-2026 the large firms had settled on the same two or three AI providers. One industry analysis observes that the announcements “mostly show a handful of large organizations arriving at the same shortlist of vendors within about seven months of each other” 3 The State of AI (2026) Accenture and Deloitte Are Selling the Same Brain to Every Company in Your Category. . Each firm presented its choice as a way to stand out. 💡 the same firms that sell differentiation to their clients have bought the same two or three tools as everyone else and announced it as a differentiator. You couldn’t brief it.
Extraction: the dark mirror
Chapter 4 described the cycle by which a firm keeps its knowledge alive: seniors put what they know into methods, playbooks and case studies, and juniors absorb them. The cycle assumed these documents stay inside the firm. AI workflows can break that assumption. When drafts and internal analyses are sent through a vendor’s system, whether they remain the firm’s depends on the contract and on whether customer content is used to train later models. 💡 Twenty years of knowledge-management programmes trying to get partners to write things down. The vendor managed it in one enterprise rollout. Different beneficiary, though.
I call this the dark mirror of the cycle. The knowledge flows outward, is absorbed into the vendor’s models, and returns as general capability sold to the firm, its competitors and its clients. This is a claim about how the system is built, not a proven fact about any vendor. Practice varies: some contracts exclude customer data from training, and others expose it, often without the customer fully understanding.
London, 2024. A firm is about to use a vendor’s summarising tool on regulatory filings. A senior partner asks for the contract to be checked, and the standard version contains a training-data clause. Three weeks of legal review produce an agreement that keeps client content out of training. Three months later a competitor is asked by a client to confirm the same thing and has no clear answer.
A 2026 study by Ngwenyama, Klein and Rowe shows the same pattern in research: academic publishers use AI-equipped platforms to capture researchers’ working knowledge, and one of them, Elsevier, keeps “perpetual irrevocable rights” over it 4 Ngwenyama et al. (2026) Platform capture of scientific knowledge production: publishers’ dominance, generative AI and subsumption of academic labor. European Journal of Information Systems. 35(5), 942–965. p. 956. . Their conclusion carries over: the remedy has to be collective, not firm by firm.
The defence reference case
In defence, European policymakers have treated sovereignty as a strategic matter for decades. The reasoning is that defence capability must not depend on the decisions of entities outside the European political community. Defence is a reference point, not a model to copy: it shows that this chapter’s thinking is already at home in European policy.
The same reasoning applies, in adjusted form, to energy, water, transport and urban planning, whose content depends on European commitments such as climate transition and affordable regulated services. The claim is that a political community has a say in decisions that affect its autonomy.
The integrated picture
Sovereignty and extraction are one mechanism at three levels. The European tradition tilts toward Anglo-American patterns. The firm’s output drifts toward common defaults. The firm’s written knowledge can flow out to vendors and back to its competitors and clients.
Mügge, studying EU strategy documents from 2018 to 2024, shows that policy has concentrated on legal authority over AI (regulation, data location, compliance) and neglected two other questions: are European practitioners truly empowered by the tools, and do European practices stay distinctive 5 Mügge (2024) EU AI sovereignty: for whom, to what end, and to whose benefit?. Journal of European Public Policy. 31(8), 2200–2225. ? The argument here sits in that gap.
Generic AI is produced by a few very large firms, mostly American, with commercial reasons to sell the same capability everywhere. These are political and economic conditions, and they could be different. Chapter 6 turns to software and law, where the mechanism has run long enough to leave evidence.