Chapter 3 - Theoretical Framework

3.1 What the framework is for

Let’s now try and assemble the analytical apparatus the productivity framing doesn’t 💡 supply. Two principal anchors carry the normative aspects: Stiegler on the proletarianisation of savoir-faire under technical capture, and Jonas on the imperative of responsibility toward futures. Supplementary anchors (MacIntyre on internal goods, Polanyi and Collins on tacit and contributory expertise, Dewey on publics formed through shared consequences) provide bounded support at specific points in the argument.

The chapter provides the philosophical apparatus through which the five next chapters (4–8), more analytical, frame their arguments, and the grounding from which Chapter 9 draws its recommendations. It does not itself develop the arguments about contemporary advisory transformation.

The chapter also introduces the distinction between tacit-as-constraint and tacit-as-formation, developed in §3.5, which concerns the unit of analysis rather than empirical evidence and marks where this work parts company with Tuczek et al.’s Task-GenAI Fit framework and adjacent information-systems research on AI in consulting.

The philosophical framework assembled in this chapter is not domain-specific. The Stieglerian account of proletarianisation and individuation, the Jonasian imperative of responsibility, and the Collins/Polanyi account of tacit and contributory expertise could be applied to advisory work in other professional contexts 💡 . As discussed in the introduction, this work’s specific contribution is to apply and develop the framework in European infrastructure advisory, its primary research site. Where this chapter makes general philosophical claims, they are claimed as general; where the analytical chapters apply those claims to European infrastructure advisory, that application is the specific-level work.

3.2 Stiegler: proletarianisation and individuation

Bernard Stiegler contributes two things here. He reworks the Marxist concept of proletarianisation into an analytical category that operates beyond the industrial-economic context Marx originally engaged. And he uses Simondon’s concept of individuation to characterise what proletarianisation operates on.

Proletarianisation, in Stiegler’s reworking, is the loss of savoir-faire (practical know-how, but also know-how-to-live (savoir-vivre) and ultimately know-how-to-think (savoir-penser)) as it is progressively captured into technical systems the worker does not own 1 2 . The Marxist account of proletarianisation focused on the loss of control over the means of production; Stiegler’s focuses on the loss of the content of expertise into technical systems whose ownership is institutionally elsewhere. The mechanism is not specifically industrial. It operates wherever practical know-how is being captured in this way, and Stiegler argues it runs across contemporary capitalism in ways the original Marxist framework did not anticipate.

The loss is cognitive and epistemic, and possibly ontological as well. 💡

Critically for the framework, Stiegler treats proletarianisation as selective. It displaces the expertise that constituted the worker’s professional identity without abolishing the worker. The carpenter whose knowledge of timber and joinery is captured into computer-aided manufacturing systems remains employed but does different work, its content altered by the capture. The same mechanism operates on advisory under contemporary AI mediation. The advisory profession remains; what advisory work is, what it produces, and what is being formed in the practitioners who do it are being altered. 💡

Individuation, in Stiegler’s reading drawing on Simondon 4 , is the process through which a person, a collective, or a culture comes to have distinctive form. It is never completed: across time, the individuating entity keeps integrating new content while preserving its distinctiveness. The proletarianisation Stiegler diagnoses is, in these terms, dis-individuation: the loss of distinctive form under technical capture, smoothed out into the generic forms that the captured expertise produces.

Individuation can be explored on three scales, and at each an identity is at stake. At the individual scale it is the consultant’s, as a particular kind of practitioner. At the organisational scale it is the firm’s, as a knowledge-managing collective with distinctive ways of seeing. At the national-cultural scale it is that of the European public-interest tradition under contemporary generic AI mediation. One mechanism connects the three: dis-individuation under generic technical mediation operates on all of them, with different visible effects.

Identity, from here on, names the result of individuation so far at each scale: the practitioner, the firm or the tradition as it currently stands. It is provisional, since the process that made it can still change it, and the argument does not defend it for its own sake. At the individual scale, Ibarra’s study of junior consultants and investment bankers moving into client advisory roles, to which Chapters 4 and 7 return, starts from the assumption, taken from Schein, that professional identity “forms over time with varied experiences and meaningful feedback” 5 .

The three-scale deployment is an analogical extension of Simondon’s framework rather than its strict application: Simondon constrains collective individuation to the transindividual domain through which psychic individuation proceeds, and his framework does not address national-cultural individuation directly. The extension is this work’s own, warranted by the parallel; the subsequent chapters justify it by what it enables.

Consulting presents a difficulty for the Stieglerian account, and it has to be met head-on. Stiegler’s proletarianisation thesis was built around the paradigm of embodied craft knowledge: the cabinetmaker’s tacit feel for grain and joint, the weaver’s knowledge of thread tension carried in the hands rather than in any articulable formula. This knowledge is in the body, acquired through years of sustained engagement, and the machine extracts it by encoding the movements, the sequences, the timing, reducing the craftsperson from productive contributor to operative.

Management consulting does not fit this paradigm straightforwardly. Consulting firms had substantially externalised their savoir-faire 💡 well before AI mediation arrived, and by their own deliberate institutional choice. The McKinsey 7S, the BCG matrix, the analytical playbooks of every major advisory firm are prior instances of consulting savoir-faire being inscribed into technical objects. The consulting firm has always been, in an important sense, a machine for externalising savoir-faire: that is the product the firm sells and the mechanism through which it scales 💡 . If the Stieglerian mechanism is the capture of embodied knowledge into technical systems, and if advisory knowledge was already substantially externalised before AI arrived, then we can say that AI mediation introduces something new only by naming what it captures that prior externalisation did not.

Contemporary practitioner stories sharpen the question. Renaud (2026), writing from practice, identifies the most consequential knowledge-sharing gesture of the current AI moment as the transmission of method in natural language (skills rather than implementations, posture rather than code) because such methods can be taken by another agent, bent to its context, and executed autonomously 6 . Prior externalisation inscribed consulting savoir-faire into legible artefacts for human practitioners to apply with developing judgment; skills-as-text inscribe it into executable methods that agents run without that judgment. The crossing is from legible-to-practitioners to executable-by-agents. Externalisation itself is old; that threshold is new.

The answer, which the tacit-as-formation distinction of §3.5 develops, is that the Stieglerian mechanism in consulting operates on temporal structure rather than on content. Prior externalisation captured the content of consulting savoir-faire: frameworks encoded analytical approaches; playbooks encoded what works in recurring engagement types; case databases encoded institutional memory 💡 . It did not, and could not, capture the formative engagement through which a junior consultant becomes capable of using those frameworks with genuine judgment. The duration of struggling with a tariff model from a blank document, the experience of ten failed memos 💡 before one lands, the absorption of a senior practitioner’s hesitation in a regulator meeting: all of this is anything but content. It is the temporal structure of a formation process whose medium is the difficulty and duration of engagement with consequential work under real uncertainty 💡 .

AI mediation collapses this temporal structure. When a junior consultant produces a competent first draft through AI assistance in hours instead of weeks, the content is present but the formative engagement is eliminated. Prior externalisation left the formation process intact while capturing its outputs; AI mediation substitutes for the process itself. This is a different kind of capture, and it is the proletarianisation we can diagnose in consulting: not the capture of what practitioners know, but the elimination of the conditions under which future practitioners will come to know it. The practitioner’s individuation (the process of becoming a particular kind of consultant with a particular quality of situated judgment) depends on the temporal engagement that AI mediation removes. It is the formation pathway that is proletarianised, rather than the practitioner already formed.

This application of Stiegler to consulting sits within a developing literature on AI-induced proletarianisation in professional and cognitive domains. Nony (2024) applies the Stieglerian apparatus to argue that AI is producing generalised cultural and cognitive proletarianisation across professional and creative domains 7 ; Alombert (2024) extends the pharmacological reading to argue that generative AI concentrates the risk of symbolic misery: the impoverishment of shared cognitive and cultural life under generic technical mediation 8 . This work adds the mechanism, the collapse of the formative engagement through which practical wisdom develops 💡 , together with the Jonasian intergenerational dimension: the obligations to future generations that make the consulting domain distinctive. 💡

3.3 Jonas: the imperative of responsibility

Hans Jonas contributes the imperative of responsibility toward futures. In The Imperative of Responsibility he argues that contemporary technological capacity has created a new ethical condition: human action now operates across scales that bind future generations, whose conditions of life present decisions are shaping, and the ethical commitments of the present must be suited to this condition.

The mechanism Jonas identifies is the long-horizon 💡 , irreversible character of contemporary technological action. Decisions made now about energy, infrastructure, climate, and ecological systems shape the conditions of life for generations whose interests cannot be represented through classical, contemporary political processes. The intergenerational asymmetry is constitutive: future generations cannot participate in decisions that bind them, yet they will bear the consequences across time horizons that exceed any single human life or political cycle. For Jonas, the ethical content of contemporary action must engage this asymmetry instead of abstracting away from it.

The asymmetry Jonas identifies is new, though it is sometimes misread as a version of the general duty to consider future interests that earlier moral frameworks could accommodate. Jonas’s claim is more radical. Traditional ethics assumed a fixed human condition: that what it meant to do well, to act rightly, to be a capable practitioner was sufficiently stable across generations to provide a reference point for present decision. Contemporary technological capacity breaks this assumption by altering the conditions under which future generations will form and exercise their own capacities. The ethical problem is that we are altering the scale, rather than failing to maximise future welfare on a stable one. The appropriate response is therefore not merely to “consider future interests” (which existing moral frameworks could absorb) but a categorical imperative of a different kind: do not foreclose the conditions under which future people can form and exercise the capacities that make them agents. For European infrastructure advisory, this imperative means preserving the formation pathways through which future practitioners will develop the situated judgment on which both their agency and the answerability of the work depend 9 .

Jonas also grounds the third-box constituency Chapter 2 introduced. The publics who bear the consequences of infrastructure decisions include the future generations whose conditions of life are being shaped now, and the answerability of European infrastructure advisory operates across this intergenerational dimension. In European infrastructure advisory practice the asymmetry takes a concrete form: a fifty-year asset commissioned in a five-year political cycle 💡 . This asymmetry is constitutive of the legitimate practice of European infrastructure advisory; advisory work that abstracts away from it has lost its grounding in the European public-interest tradition that defines the domain.

Anchoring in Jonas sets this work apart from the dominant tradition in contemporary AI ethics, which has crystallised around principlist approaches descending from medical ethics through bioethics into AI policy. Floridi’s canonical synthesis 10 articulates this tradition’s five principles 💡 and provides a reference framework for much contemporary AI governance, including the EU AI Act and adjacent European policy.

3.4 Supplementary anchors

Three supplementary anchors each carry a limited part of the argument.

MacIntyre on internal goods. In After Virtue 11 , a practice is a cooperative human activity through which goods internal to the practice are realised, where the goods are recognisable only through engagement with the practice itself. A practice’s value is internal to it, and its integrity depends on whether those internal goods go on being reproduced through its institutions.

We can draw on this account at the points in the analytical chapters where the firm’s distinctive way of seeing, or the senior practitioner’s professional commitments, need articulating in terms of internal goods rather than external markers. We need not take on MacIntyre’s wider Aristotelian-Thomist commitments; the practice-and-internal-goods apparatus does its work without them. MacIntyre’s own argument implies that consulting practice was substantially affected by external goods (client fees, growth pressures, competitive positioning) well before AI mediation arrived. The claim here is comparative rather than restorative: AI mediation is part of, and a distinct form of, internal-goods displacement, working through formation at a pace and with a comprehensiveness that market pressures could not previously achieve. The practice AI disrupts was never untouched. The concern is that AI eliminates the formation conditions that historically allowed the internal goods to be reproduced at all, imperfectly but genuinely, alongside and sometimes against the external goods.

Polanyi and Collins on tacit and contributory expertise. From Polanyi 12 comes tacit knowledge: the professional knowledge that is institutionally constitutive of expertise but cannot be fully articulated in a propositional form. From Collins 13 comes the distinction between contributory expertise (the ability to contribute to a practice through embodied engagement) and interactional expertise (the ability to engage with a practice through linguistic competence without the embodied contributory grasp). Together they supply the framework’s account of what professional expertise is and how it differs from competent linguistic performance.

These two anchors are also where the argument diverges from Tuczek et al.’s framework and the research adjacent to it. That research engages tacit knowledge instrumentally, as a constraint on what AI can substitute for. Here it is engaged as constitutive of professional formation, so that the question becomes what AI substitution does to the pathway through which contributory expertise is reproduced. §3.5 develops the distinction.

The distinction has a precursor in the Austrian economics tradition that predates Polanyi’s Personal Knowledge by thirteen years. Hayek (1945) distinguished scientific knowledge (explicit, generalisable, in principle centralisable) from ‘the knowledge of the particular circumstances of time and place’: local, contextual, practical knowledge that ‘cannot enter into statistics and therefore cannot be conveyed to any central authority in statistical form’ 14 . In Hayek’s terms, the generic model is exactly the ‘central authority’ whose ‘statistical information by its nature cannot take direct account of these circumstances of time and place’: the same incapacity the Collins/Polanyi account names from the sociology of expertise, now in the vocabulary of classical political economy. The epistemic insight is independent of Hayek’s political conclusions about market mechanisms. We deploy one claim only: dispersed local knowledge cannot be aggregated statistically without losing the particularity that makes it valuable.

The Social Doctrine tradition reaches the same point. Magnifica Humanitas, the first papal encyclical to engage AI directly (Leo XIV, 2026), distinguishes between the AI system’s ‘statistical adaptation based on data and feedback’ and the formation of a human person who ‘allows themselves to be shaped by life and grows over time through choices, mistakes, forgiveness and fidelity’, and so arrives from within the natural-law tradition at the distinction the Collins apparatus produces from the sociology of expertise. Genuine formation is constituted by temporal engagement with lived consequence, not by output equivalence 15 . 💡

Dewey on publics formed through shared consequences. In The Public and Its Problems 16 , a public is constituted through shared exposure to the consequences of decisions made by others: not a pre-existing political constituency but a political reality that comes into being when institutions recognise shared consequences as conferring political standing. Chapter 2’s introduction of the third box and Chapter 5’s sovereignty argument both draw on this account, since European public-interest practice operates partly through the institutions that recognise consequence-bearing publics as present in decisions that affect them.

3.5 Tacit-as-constraint versus tacit-as-formation

Let’s start with a concrete image, before the philosophy arrives. Reaching the sixth floor by lift delivers the destination: faster, without effort, adequately. What the lift does not deliver is the muscle that stair-climbing would have built. 💡 The lift provides the function without the formation, and that is the structure the tacit-as-formation distinction is designed to name: AI mediation of formative tasks delivers the output of formative engagement without the engagement itself, and the engagement is what this analysis concerns.

The distinction between tacit-as-constraint and tacit-as-formation is where this framework departs from the existing literature on AI in consulting. Tacit-as-constraint is the dominant reading in that literature. Tuczek et al. 17 engage Polanyi explicitly in their Task-GenAI Fit framework, citing him to explain why some consulting tasks resist codification and therefore resist AI substitution. Tacit knowledge figures as a constraint on what AI can substitute for. The constraint reading supports systematic mapping of tasks against substitutability, with tacit knowledge as the boundary condition that defines where substitution operates and where it does not. It is empirically productive, and Tuczek et al. deploy it carefully.

The other, tacit-as-formation, is the reading adopted here. Tacit knowledge is engaged as constitutive of professional formation: of how a consultant becomes a consultant, how expertise is reproduced across generations of practitioners, and how the integrated three-scale system Chapter 4 describes operates. It is the ground on which contributory expertise is built.

The two readings ask different questions about the same situation. Tacit-as-constraint asks: “which tasks resist AI substitution because of their tacit-knowledge character”? Tacit-as-formation asks: “what does AI substitution for formative tasks do to the formation pathway through which contributory expertise has historically been reproduced”? The questions are not in opposition and neither outranks the other; each brings its own analytical apparatus and leads to its own recommendations.

Our claim is that the formation reading is necessary for asking what contemporary AI mediation is doing to European infrastructure advisory 💡 . The constraint reading can say which tasks are being substituted; it cannot say what the substitution is doing to formation, because formation is not visible at the level of task substitution.

Deleuze and Guattari’s concept of the apparatus of capture 💡 describes as mechanism what the tacit-as-formation distinction identifies from the epistemic side. The training process of a large language model is an apparatus of capture in their precise sense. It takes in the textual traces of advisory practice: the deliverables, frameworks, methodology documents, industry analyses, and professional discourse that practitioners have produced and that have been accumulated as training data. These traces are themselves already representations: they are what advisory work looks like after it has been externalised as text. The apparatus captures the representation, codes it as tokens and their statistical relationships, and distils from the corpus a parametric model of how advisory discourse is organised. The product is the extract of the capture: the coded distillation of the patterns in the corpus, recirculated as advisory output.

The point is what the apparatus cannot reach. The junior who built the tariff model from scratch produced an artefact (the spreadsheet, the analytical output) and also underwent a formation: the temporal engagement with difficulty, error, and correction through which their practical judgment developed. The apparatus absorbs the artefact. It does not absorb the formation, which lies in the process that produced the artefact: a process that leaves no textual trace in the training corpus and that the apparatus therefore cannot code. Better capture techniques would not overcome this: any apparatus that operates on representations leaves the remainder unreached in principle.

Contemporary practitioner discourse reaches the same boundary from a different direction. Renaud (2026) observes that sharing expertise as natural-language method transmits posture rather than implementation and makes it executable by agents who can take it, adapt it, and run it without the practitioner who produced it 6 . The qualification introduced in the same exchange is the relevant one: natural language improves transmissibility but not reliability. Transmissibility cannot transfer the judgment required to apply the method faithfully when the context in play is one it was not designed for. That remainder (situated judgment resistant to capture) is what the apparatus cannot reach and what the formation reading names.

The epistemic account and the capture account describe one structure. On the first, the knowledge constitutive of professional formation is not knowledge-as-content (capturable as representation) but knowledge-as-process (the temporal engagement through which capacity develops); on the second, the training process is configured to absorb content and leave process uncaptured. The stake in AI-mediated consulting is therefore the process-remainder. Deleuze and Guattari’s apparatus of capture constitutes what it appropriates 18 ; my extension of their account is that formation, as a process, is the part it does not reach, and the part the recommendations must therefore deliberately maintain.

Guy Debord’s account of spectacular pseudo-cyclical time names this substitution in terms of time. He distinguishes irreversible historical time (the time of productive engagement, in which development is real and irreversible and through which a practitioner becomes) from spectacular pseudo-cyclical time, the temporal form produced when accumulated representation is recirculated as a substitute for lived activity. Spectacular time has the surface structure of historical time: tasks are completed, outputs appear, sequences unfold. It lacks the formative quality through which historical time constitutes the subject who passes through it. The AI-mediated advisory process generates spectacular time in Debord’s sense. The same junior, now producing that tariff model in hours by working with a generative AI system, has consumed spectacular time: they have experienced the output of formative engagement without having lived its duration. Debord’s critical move (the one this argument relies on) is to insist that the substitution cannot be detected at the level of output quality. The spectacular commodity often satisfies; the AI-generated model may be more technically adequate than the unassisted junior would have produced. 💡 .

Spectacular time does not constitute the subject who consumes it. The junior who consumed spectacular time producing three hundred deliverables over two years has the models. They do not have the formation that would have come from building them under conditions of formative difficulty. Temporal structure collapse, in Debord’s terms, is the elimination of formative time, whose irreversibility is its constitutive property and which no better instruction could recover 19 .

Vaneigem complements Debord at the level of being. Where Debord asks what kind of time the junior consultant is consuming, Vaneigem asks what kind of being is doing the consuming. The practitioner who moves through spectacular time accumulating deliverables but not formation is missing more than what the time would have produced: they are, in Vaneigem’s vocabulary, surviving the professional role rather than living it, inhabiting the functional form of advisory practice without being constituted by its genuine demands 3 .

A third line of theory makes the argument from the side of the receiver. Benjamin’s essay “The Storyteller” (1936) distinguishes between counsel (Rat) 💡 and information, which arrives complete, self-explanatory, with its explanation already attached. The distinction rests on one Benjamin develops in “On Some Motifs in Baudelaire” (1939) 20 : the separation of Erfahrung 💡 from Erlebnis 💡 . The formation process the tacit-as-formation reading describes is the one through which Erlebnis accumulates into Erfahrung: isolated encounters with consequential work sediment over time into the transmissible practical wisdom that counsel requires. The AI-mediated formation process substitutes Erlebnis without accumulation: the junior engages with AI-generated outputs rather than with the direct challenge of the work, and the Erlebnis of those engagements does not accumulate into Erfahrung, or does not accumulate at all. The apparatus of capture reduces Erfahrung to training data; what it returns (in the model’s output and in the junior’s AI-mediated engagement with it) is Erlebnis 21 20 .

A profession has written down the opposite ideal for itself. As AI reached Japanese audit, the national association’s journal held up the auditors who “keep training themselves, adjusting to the needs of the times, never being content with past experiences” 22 ; Goto reads it as one theme of a new collective identity, the auditor as “Continuous Self-disruptor” 22 , and Chapter 6 returns to the study. In Benjamin’s terms the ideal makes a virtue of Erlebnis: meet each new thing and move on. Counsel needs the reverse, experience settled enough to be drawn on, and the ideal does not say how an auditor would come to have it.

Morozov (2025) comes at the problem from political philosophy. Examining what adequate frameworks for governing AI systems require, he identifies a circularity: “the values we would use to govern these systems are themselves being formed through our encounters with those (ever fluid) systems” 23 . This is the governance version of temporal structure collapse: one cannot use pre-formed professional values to evaluate what AI mediation is doing to professional formation, because those values are themselves being reconfigured through practitioners’ ongoing encounters with AI-mediated practice. The same circularity explains the productivity framing’s characteristic blindness (its inability to detect the formation deficit behind adequate-looking outputs): the evaluative apparatus it uses is not independent of the AI mediation it purports to assess.

Temporal structure collapse (Stiegler), spectacular pseudo-cyclical time and ontological survival-mode (Debord and Vaneigem), the erosion of counsel by information (Benjamin), and the circularity of values-under-formation (Morozov) are five routes to one claim. Five routes are not five independent witnesses: Stiegler’s diagnosis shares much with the Situationist critique from which Debord and Vaneigem write, Benjamin stands within the same tradition of reflection on experience and technical mediation, and Morozov restates the problem in the language of contemporary political philosophy. Their convergence shows a common inheritance, and I claim no more for it than that 💡 .

The tacit-as-formation distinction answers the difficulty §3.2 raised. Relocating the site of capture from content, already externalised, to the temporal structure of formative engagement is what makes the Stieglerian mechanism grip a domain where content externalisation is a feature, not a vulnerability.

The tacit-as-formation reading has begun to acquire empirical support from cognitive science. Kosmyna et al.’s EEG study of essay-writing under LLM assistance 24 finds measurably weaker brain connectivity, lower memory recall, and reduced sense of authorship accumulating across repeated sessions with AI assistance: one study, preliminary, but pointing one way 💡 . Oakley et al.’s review of cognitive psychology and neuroscience 25 develops the memory paradox: as external aids become more capable, the internal memory systems on which expertise, critical thinking, and long-term retention depend risk atrophy, and effective human-AI interaction itself depends on the absorbed schemata that the AI is putatively replacing. Both papers operate at the cognitive level rather than the philosophical one, but they ground the claim this framework makes: displacing the engagements through which contributory expertise is built is not equivalent to displacing the tasks themselves, because the engagements are constitutive of the formation that makes future practitioner judgment possible.

Two further fields point the same way, and later chapters take them up. In medical education, Natali et al. (2025) find the formation-through-engagement capacities most exposed to AI-induced upskilling inhibition, the foreclosure of skill acquisition rather than the loss of skills already held (Chapter 6) 26 . In economic theory, Ide’s overlapping-generations model shows that entry-level automation can weaken the intergenerational transmission of embodied knowledge even where entry-level employment holds (Chapter 7) 27 .

At the epistemic-architecture level, Quattrociocchi, Capraro, and Perc (2025) map seven systematic fault lines between human judgment and LLM processing. The tool, on their account, operates as a stochastic pattern-completion system whose outputs are trajectories of linguistic plausibility rather than convergences toward judgment; their term Epistemia (the user’s experience of possessing an answer without having traversed the cognitive labour of judgment) intensifies rather than resolves as model scale increases 28 .

The profession itself has put a number on how much time formation now has to work with. The HEC Alumni white paper (Chapter 2) reports, citing LinkedIn Learning, that the half-life of technical skills has fallen from seven years in 2015 to two and a half in 2025, and reads it as a case for continuous retraining 29 . Read through the distinction this section draws, the figure says more. A half-life shorter than the Foundation stage means that formation, if it is formation of skills, can no longer complete within the life of what it forms. The two readings split on what follows. On the constraint reading the response is perpetual re-tooling: skills are replaced faster, and the practitioner is a stock of current skills. On the formation reading the speed of decay is the reason formation was never of skills in the first place: the Foundation years form judgment, the capacity to acquire, weigh and discard skills, and that capacity has no half-life of the kind the metric measures. Because it comes from inside the profession, the figure gives an empirical handle on the difference between a skill and a formation. 💡

The distinction also settles the framework’s relationship to Tuczek et al. and adjacent work. They are not wrong: the mapping they undertake is productive at its own level, the level of the task, which is not where formation shows. 💡

3.6 The working ontology of the LLM

The working ontology of the LLM is a specification rather than a complete philosophical account: a generic LLM is a socio-technical inscription of collective discourse, recirculated under specific conditions of ownership. 💡 . Its three components, in turn:

Socio-technical inscription. The model is an inscription of human discursive material into a technical system, not a technical artefact standing alone, and its operation depends on what has been inscribed. It is socio-technical in that it is technically structured but constituted by the social material it encodes. It follows that the model’s outputs are that material, recirculated through technical operation, and not the product of the technical system alone.

Collective discourse. The discursive material the model has inscribed is collective in two senses. It draws on human discourse at scales that no individual or institution has produced on their own: the wide corpus of internet-available text, scholarly publication, professional discourse, and adjacent material that constitutes the training distributions of contemporary models 💡 . And its content has been produced by communities of practice whose contributions are not individually attributable in the inscribed form. The outputs therefore recirculate collective work the model has not produced but has inscribed and reorganised.

Specific conditions of ownership. This inscription operates under political-economic conditions: the institutional form(s) of the firms that produce the models, the contracts that govern how the inscribed material was acquired 💡 , the commercial terms on which the models are deployed and monetised, and the wider political-economic context of model production. These conditions shape what is inscribed, what is recirculated, and what is foreclosed. Chapter 5 develops the point; here the ontology only records it.

The working ontology supports the later analytical chapters in three ways. First, it grounds the tool-agnostic stance: the analytical work operates on the class of socio-technical inscriptions that contemporary generic LLMs constitute, not on any particular instance. Second, the sovereignty argument: the cultural-discursive patterns the model inscribes affect the texture of the work the model mediates, and the political-economic conditions of inscription affect what is at stake. Third, the extraction argument: the dark mirror of the SECI model operates through the inscription mechanism, where firm-distinctive material is at risk of being inscribed into vendor systems whose conditions of ownership are institutionally outside the firm.

Quattrociocchi, Capraro, and Perc’s fault-line analysis (§3.5) gives the same object an epistemic description, the counterpart of what the working ontology names from the political-economic side: the model’s output is discourse-as-plausibility recirculated, not counsel-as-judgment 28 .

3.7 Our framework now assembled

The proletarianisation of savoir-faire through selective displacement into technical systems (Stiegler’s analytical apparatus) provides the core claim. Individuation, the process through which distinctive form is reproduced across the three scales (individual, organisational, national-cultural), provides the vocabulary for reading AI mediation as operating on formation rather than merely on output. Jonas grounds the obligation: answerability to futures whose conditions present decisions are shaping is the third box’s philosophical anchor.

The distinction between tacit-as-constraint and tacit-as-formation (§3.5) reframes the question. Tuczek et al.’s framework asks what AI cannot substitute for; this work asks what AI mediation does to the formation pathway through which expertise is reproduced.