A synthesis of Science and Technology Studies theory, philosophy of technology, and the political economy of software engineering in the Large Language Model era
→ Original framing: The following reorientation of the deskilling debate is proposed here. The crisis narratives surveyed are documented in the literature; the claim that they ask the wrong question is original.
Every few years, the software engineering field generates a new crisis narrative. In the late 1960s and 1970s it was the software crisis: complexity outrunning method. In the 1990s, it was Year 2000. In the 2010s, it was the mythology of the “10x engineer.” Now, it is deskilling: the worry that Artificial Intelligence coding tools are eroding the craft, homogenizing output, and producing a generation of developers who can prompt but cannot think.
This is the wrong question. Not because deskilling is imaginary some version of it is real but because the framing assumes a baseline of skill that was never universally distributed in the first place. It assumes a field organized around craft, now threatened by automation. The field was never organized around craft. It was organized around a small, substrate-producing elite and a large majority doing work that nobody has adequately theorized, compensated, or formed.
The right question is older and harder: Why does software engineering have no formation ladder? Why, after seven decades as a field, is there still no road from junior developer to architect that does not depend on network access, institutional luck, or the willingness of a specific senior engineer to transmit what they know? And what does the Large Language Model era reveal about that absence that was not visible before?
The answer is structural, political, and geographically distributed in ways the deskilling literature almost entirely ignores. This essay maps that structure.
The deskilling literature academic papers in AI & Society, practitioner essays on Substack, and Science and Technology Studies analyses applying Stiegler’s proletarianization framework to Copilot is largely produced by people whose professional identity is bound to craft: Senior engineers, Academics, Technical architects, et al. People whose standing in the field derives from the legitimacy of deep knowledge as the universal currency of software work.
This is not a conspiracy. It is a structural condition of knowledge production. The people positioned to theorize a field are rarely the people most injured by its political economy.
→ Synthesis: The observation above draws on the sociology of knowledge (Mannheim, Bourdieu) and is well established in Science and Technology Studies. What follows departs from the existing literature.
What the craft literature cannot quite bring itself to say because saying it would undermine the legitimacy of the knowledge that got its authors where they are, is as simple as: most of the current software field runs fine without craft.
→ Original concept: The “software nurse” category does not appear in the Science and Technology Studies or software engineering literature surveyed. It is proposed here as an analytic category to name the structurally absent subject of the deskilling debate.
The majority of software developers are what we might call software nurses. They maintain legacy Create, Read, Update, Delete applications. They configure other people’s frameworks. They implement specifications written by others, inside constraints set by others, on infrastructure owned by others. They are not failing to be craftsmen. They have a different relationship to the work professional, competent, bounded and that relationship is not a deficiency. It is simply what most of the field actually is.
→ Synthesis: The following two paragraphs apply Stiegler’s proletarianization framework and Simondon’s alienation theory, both drawn from existing literature, to observe that the software nurse is absent from their implicit subject.
The software nurse is structurally absent from the deskilling literature. When Stiegler’s proletarianization framework is applied to Artificial Intelligence coding tools, the implicit subject is always the craftsman: the developer who had savoir-faire to lose. When Simondon’s alienation theory is applied to abstraction layers, the implicit subject is always the technician who understood the machine from within and is now being pushed outside. The nurse, who was never “inside” in that sense, simply does not appear.
→ Original argument: Applying the Gramscian concept of the “class document” to the Science and Technology Studies deskilling literature itself rather than to a technology or institution is a move not found in the surveyed literature.
This is the literature’s central blind spot. It theorizes a minority’s identity crisis while performing concern for a universal condition. It is, in the precise Gramscian sense, a class document not a lie, but a partial truth that happens to serve the interests of those producing it.
Law and Medicine are an imperfect comparison, yet worth making carefully. The medecine ladder credentializes as much as it forms a licensed nurse is not automatically a competent nurse, and the curriculum does not guarantee the knowledge it claims to transmit. But even a ladder that partly credentializes is structurally different from no ladder at all. The credential creates a floor, a shared vocabulary, and a defined relationship between levels that exists independently of who you know. Software engineering has neither the floor, nor the vocabulary, nor the relationship. The comparison is not between medicine’s ideal and software’s reality; it is between an imperfect formation structure and the near-complete absence of one.
The care assistant knows something. The nurse knows more, and more precisely. The General Practitioner knows more broadly. The specialist knows more deeply. The researcher pushes the boundary of what any of them know. Each level has formal recognition, defined compensation, a known relationship to the levels above and below, and a curriculum that makes the transition between levels possible without requiring that you know the right people.
Software engineering has almost none of this. The typical career ladder runs: junior, mid, senior and then… a cliff. The transition from senior to staff or principal is almost nowhere governed by a verifiable curriculum. It is governed by visibility, sponsorship, the accident of having worked on something that mattered to enough people, or the historical luck of being at the right company when it scaled. This trajectory mirrors the art world far more than medicine. Without objective credentials to settle disputes over quality, the field defaults to subjective trends; this is why framework competitions often resemble pop-culture fandoms rather than engineering decisions.
Certifications exist, but they certify platform loyalty rather than craft. An Amazon Web Services Solutions Architect certification is not portable across infrastructure providers; it is portable only within Amazon Web Services’ gravitational field. A nurse’s license is portable across hospitals, health systems, and countries. The platform certification encodes dependence into the credential itself.
The open-source world was supposed to provide the meritocratic alternative contribution histories as a visible curriculum. But the platformization of software development through GitHub converted that into a reputation economy: stars, followers, commit graphs. Visibility metrics replaced knowledge metrics. You could be the most deeply knowledgeable person in a domain and remain invisible, while someone with a popular but shallow repository accumulated the social capital that functions as a proxy for competence.
→ Original argument: The shadow tier analysis that follows extends Ensmenger’s historical stratification framework and Barley and Orr’s sociology of invisible technical work to name specific contemporary categories not found as a unified account in the surveyed literature.
But the ladder does not only end at senior; it also conceals what exists below junior. The software nurse population is itself stratified and that internal stratification is just as illegible as the cliff at the top.
Software Quality Engineers accumulate deep systemic knowledge of how software fails in practice. They understand failure modes, edge cases, and integration pathologies that developers who only write code rarely encounter. This knowledge is organizationally classified as verification work rather than engineering: structurally subordinate, poorly compensated relative to its consequence, and without a curriculum that would let its practitioners name what they know or climb toward what they understand. Support engineers develop diagnostic expertise that often exceeds developers’ understanding of their own systems. They inhabit the gap between documentation and reality, between what the software is supposed to do and what it actually does under unanticipated conditions. That knowledge is tacit, organizationally invisible, and non-transferable upward. Technical sales engineers translate between customer need and technical possibility a position requiring both domain knowledge and systemic comprehension but are classified outside the engineering organization and outside the formation conversation entirely.
Bootcamp graduates occupy a specific slot in this shadow stratification. Empirical research consistently shows them clustering in front-end and full-stack web development roles, with a salary gap relative to Computer Science degree holders that persists over time. The ceiling they hit is not epistemic; it is organizational, credential-dependent, and network-gated. Harvard Business School research found that even after companies dropped degree requirements from job postings, fewer than one in seven hundred new hires without degrees actually benefited: companies changed the posting, not the hiring practice. The paper ceiling is real and operates independently of ability.
What the field has, then, is not an absence of hierarchy but an absence of legible hierarchy. The tiers exist. They are operative daily. They determine compensation, access, and career ceilings with the precision of any formal structure. What is absent is an official acknowledgment that these positions form a continuum, any road between them, and any name for what it would mean to move.
→ Original argument: The following claim that the “genius exception” was used to naturalize the absence of a formation ladder, conflating two distinct problems is not found in the surveyed literature on software professionalization, including Ensmenger’s historical account.
None of this was inevitable. It reflects choices or more precisely, the absence of choices, the failure to build a road that would have required effort from people who had no incentive to build it. The people who could have formalized the ladder senior engineers, technical leads, academic computer scientists benefited quietly from its absence. Informal selection looks like meritocracy while encoding network access. The “genius exception” was used to naturalize the absence: you cannot curriculum your way to genius, therefore curriculum is beside the point. This conflation of two entirely different problems: 1) how to form competent practitioners and 2) how extraordinary talent emerges served, both justify building neither track.
The proof is in the biographies. Dijkstra stumbled into computing through a newspaper advertisement for a three-week course in 1951; at his wedding in 1957, the registrar refused to accept “programmer” as a profession. Kay dropped out of college, joined the Air Force, encountered computers through an aptitude test, and synthesized molecular biology, child psychology, and Engelbart’s 1968 demo into a vision of personal computing that no curriculum could have prescribed. Knuth nearly became a musician. These are not stories of formation. They are stories of collision between an individual temperament and an institutional moment that was never designed to be reproduced. Medicine does not need to explain how someone became a cardiologist through a newspaper advertisement. The fact that software’s highest achievements are irreducible to any curriculum is not a charming feature of the field’s youth. It is evidence that no one with the power to build the road has ever needed it built.
Extraordinary people had to be extraordinary precisely because the road didn’t exist. Kay, Dijkstra, and Knuth are not counterexamples to the political gate; they are its most legible symptom. In a system with a formation path, their abilities would be remarkable but explicable. In a system without one, they can only be mythologized. And mythologization serves a function: it makes a structural vacancy look like a natural landscape. The road doesn’t exist because only exceptional people could walk it. The inversion is the truth: the exceptional were necessary because the road was never built.
→ Synthesis: The end-user programming literature (Scaffidi, Nardi, Burnett) and the computational thinking debate (Wing, Guzdial, Tedre) are well established. The following argument that the spread of programming into other fields strengthens the case for a formation ladder rather than weakening it is an original extension applied to the Large Language Model context.
The absence of a formation ladder would be a contained problem if programming remained inside the software profession. It does not. It never did.
In 2005, Christopher Scaffidi, Mary Shaw, and Brad Myers estimated that by 2012, fifty-five million Americans would be end-user programmers that is around eighteen times the number of professional programmers. The vast majority were spreadsheet users: accountants, analysts, project managers, scientists, doing real computational work inside tools that were never classified as programming environments. Bonnie Nardi’s ethnographic work in the 1990s showed these users were not accidentally programming; they were deliberately solving problems, building models, and constructing logic. The spreadsheet was their medium. VisiCalc in 1979 and Lotus 1-2-3 in the 1980s had already done what current Artificial Intelligence tools are claimed to be doing: putting computational power in the hands of people who were not trained as programmers.
This is the baseline the deskilling debate ignores. Programming has always extended beyond engineers. What changes with each generation of tools is the scale, the domains, and the legibility of the gap between using and understanding.
Today, that gap is crossing into domains where the consequences are harder to contain. Bioinformaticians write Python scripts that process genomic data without software engineering formation. Digital humanities scholars build text corpora and network graphs without code review practices. Parametric architects generate building geometries through visual programming environments whose outputs enter construction workflows. Computational social scientists analyze political behavior with statistical code they write but did not formally study. In each case, programming has entered the domain as a tool before any formation structure arrived to govern its use.
Generative Artificial Intelligence accelerates this pattern and raises its stakes simultaneously. A 2026 study published in Science, tracking thirty million GitHub commits, found that approximately twenty-nine percent of Python functions in their United States sample were Artificial Intelligence-generated by the end of 2024 with productivity gains accruing only to senior developers, while early-career practitioners saw no measurable benefit despite being the most enthusiastic adopters. Research on non-programmers assessing Artificial Intelligence-generated data analyses found participants could not reliably detect safety-critical flaws even when warned. The tool lowers the barrier to production while raising the barrier to evaluation.
The Cardiopulmonary Resuscitation analogy is precise here. Most people know the recovery position enough to recognize that someone has collapsed and to do something rather than nothing. Fewer know Cardiopulmonary Resuscitation enough to intervene in a cardiac arrest with some probability of effect. Fewer still are paramedics able to diagnose, triage, and act under uncertainty with professional accountability. The recovery position is valuable. It is not Cardiopulmonary Resuscitation. Cardiopulmonary Resuscitation is valuable. It is not paramedicine. Knowing which of these you are doing matters enormously when the stakes are high.
Programming has spread into other fields at the “recovery position” level. Enough to produce output. Not enough to evaluate it, govern it, or take responsibility for it in any formal sense. The bioinformatician writing a genomic pipeline, the digital humanist building a text classifier, the financial analyst generating Artificial Intelligence-assisted models each is doing real computational work with real consequences, in a formation vacuum that the software profession never resolved for its own practitioners.
This is the argument for the ladder that the deskilling debate never makes: it is not about protecting craft. It is about the fact that the problem has escaped the building. When programming was contained within a professional community however poorly organized the consequences of the formation gap were internal. Now that programming has diffused into biology, architecture, law, social science, and financial analysis, the consequences of having no shared vocabulary for what different levels of competence mean are external, diffuse, and largely invisible until something fails.
→ Synthesis: The following argument applies Actor-Network Theory’s concept of inscription (Latour, Akrich) and the framework homogenization critique, both established in the literature, to the Artificial Intelligence coding tool context. The legibility argument that follows is an original extension.
The homogenization concern that Artificial Intelligence tools are encoding dominant patterns into every suggestion, crowding out alternatives, and inscribing a particular worldview into the act of programming itself is real. But it predates Artificial Intelligence by at least two decades.
Rails did not just make Model-View-Controller popular; it made Model-View-Controller feel like how web applications work. Spring did not just implement dependency injection; it made Java developers think in Spring idioms before they thought in object design. React did not just popularize component trees; it restructured how frontend developers reason about interfaces. The framework inscribes a worldview, and sufficient adoption makes the worldview invisible.
The mechanism Actor-Network Theory calls inscription human decisions encoded into technical artifacts that then act on future users has been operating in software since the first popular framework. The Large Language Model acceleration is real, but it is a continuation, not a rupture. What changed is legibility. You could read the Rails source. You could understand its conventions, disagree with them, fork them, replace them. The inscription was transparent. With a large language model trained on GitHub’s public repositories, the homogenizing force is statistical and opaque. You cannot read what Copilot thinks a web application looks like. You can only observe what it suggests.
→ Original observation: The structural irony of Berkeley Software Distribution-under-macOS that Apple’s platform lock was built on a public commons and its implication for the Large Language Model era’s foreclosure of that commons is an original argument not found in the surveyed literature.
More importantly, you could build underneath frameworks. The framework sat on top of a language, which sat on top of a runtime, which sat on top of an operating system that was, at least in significant part, public and inspectable. The Berkeley Software Distribution layer underneath macOS is the structural irony that the platform consolidation story usually omits. Apple’s lock the decision to make macOS a luxury product with a sealed interface was built on a commons it did not create and does not own. Darwin is Berkeley Software Distribution-derived. Berkeley Software Distribution is Unix. Unix is public in the sense that matters: its logic is legible, its ideas are forkable, its descendants run everywhere.
This is why the Berkeley Software Distribution enthusiasts ended up at Apple rather than nowhere. The substrate was real, public, and comprehensible. The serious people found their way to it. HyperCard’s killing was a political move dressed in marketing language: a decision about who gets to be a producer rather than a consumer of software.
The Large Language Model era attempts something different. The substrate is not being built on a commons. The weights are proprietary. The training data is proprietary or disputed. The inference infrastructure is proprietary. When the “vibe coding” factory kills the legible layer when the interface becomes a natural language prompt and the thing underneath is a black box in the precise Latourian sense: there is no Berkeley Software Distribution moment available. No point at which the serious person can reach underneath the product and find something real to build on. The commons that made climbing theoretically possible, even when practically difficult, is what is being foreclosed.
→ Synthesis: The cost gravity argument draws on Hintjens’ observation about infrastructure cost curves and the historical pattern of production democratization through price reduction.
Pieter Hintjens observed that competition drives cost down, and falling cost means the barrier to owning the means of production eventually drops low enough that individuals or small groups can afford what was previously institutional. That is the historical pattern: the printing press, the personal computer, the cheap cloud instance, the consumer Graphics Processing Unit. Each cost reduction cycle opened a window in which new actors could enter production rather than merely consumption.
The Field-Programmable Gate Array, the homemade Tensor Processing Unit, the NanoGPT-scale model these are the current candidates for that window. If you can run a meaningful model on hardware you built or fully understand, you recover something like the Berkeley Software Distribution layer: an inspectable, ownable foundation. Karpathy’s original NanoGPT point was Simondonian before it knew it you can understand the whole thing, which is a move against black-boxing, an insistence on the technician’s relationship to the machine from within.
→ Original thesis: “Cost gravity without formation is not democratization; it is cheaper consumption” is proposed here as a compound claim. The distinction between cost reduction and democratization of production, applied specifically to the absence of a formation ladder, is not found in the Hintjens literature or the Science and Technology Studies cost-reduction literature.
But cost gravity without formation is not democratization. It is cheaper consumption.
Cheaper consumption means the falling cost of the factory reaches you as a product you use, not a tool you control. Cheaper Graphics Processing Unit inference means you pay less per Application Programming Interface call. You do not own the model. You cannot inspect what changed between versions. You consume the output. The ease is real. The dependency is also real, and it deepens as the tool becomes more capable and more central.
For the software nurse, cheaper consumption means their job gets easier and their leverage gets smaller simultaneously. They produce more with less effort which looks like empowerment but the thing they are producing runs on infrastructure they understand less than before, controlled by fewer actors than before. The price signal says this is yours. The knowledge structure says this is not yours.
The medicine comparison clarifies the mechanism. Cost gravity in medicine cheap diagnostic equipment, portable ultrasound, Artificial Intelligence-assisted imaging reading genuinely extends capability down the ladder because the ladder exists. The nurse can use the cheaper tool meaningfully because the nurse was trained to a known standard and occupies a defined position in a formation structure. The software nurse has no such formation. The cheaper hardware reaches them as a consumer product, not a production tool. The falling cost lowers the entry price for people who already have the knowledge to walk through the door. For everyone else, it lowers the price of standing outside and consuming what those inside produce.
→ Original framework: The consumption-before-agency-before-autonomy sequence, applied to the geographic distribution of technical formation, is proposed here. The Global South technology access literature documents the distribution asymmetry; the specific framing as an epistemic ladder problem tied to software formation is original.
The asymmetry described so far is not evenly distributed. It follows existing lines of geography, language, and economic position with the precision that structural conditions always exhibit when left unaddressed.
The smartphone reached rural Kabylie before the laptop did. That is cheaper consumption: people got connected, got access to services, got a voice in networks. They did not get the ability to build the network. The cost of the device fell to near zero while the cost of understanding the infrastructure stayed where it was or increased. Connectivity was democratized. Production was not.
The Large Language Model era runs the same sequence at higher speed and with higher stakes. The Application Programming Interface reaches the Global South before the formation that would allow someone to own what runs on it. The prompt interface arrives before the curriculum that would let a developer in Abidjan, Lahore, or Tizi Ouzou understand what they are prompting, evaluate its output critically, modify its behavior, or build an alternative. The consumption layer is accessible. The production layer is geographically, linguistically, and economically concentrated in a small number of places.
→ Original concept: “Knowledge serfdom” as a precise structural term locating the mechanism of exclusion in the absence of formation rather than in legal or geographic barriers is proposed here. The serf analogy as applied to epistemic conditions in software is not found in the surveyed literature.
This is knowledge serfdom in its precise structural form. The medieval serf was not excluded from the land by a locked door. They were excluded by the absence of the formation that would let them operate as anything other than a serf. The lord did not need to guard the knowledge. The knowledge simply did not transmit because there was no road for it to travel. The serf could see the field. They could work the field. They could not own the field, not because ownership was prohibited, but because the conditions that produce ownership formation, accumulated knowledge, positional leverage were systematically absent from their life.
The software equivalent: the developer in the Global South can consume the Application Programming Interface, can use the vibe coding interface, can build applications on top of the factory. They cannot govern the factory, cannot inspect its substrate, cannot influence what it inscribes, cannot build the alternative. And unlike the medieval situation, this is not primarily a legal or geographic barrier. It is an epistemic one. The formation road does not exist. The curriculum was never built. The ladder ends before they can reach it.
→ Synthesis: The Science and Technology Studies horizontal dream draws on Latour’s symmetric treatment of actors, Free/Libre and Open Source Software commons literature, and Bounegru’s platformization research. The following critique distinguishing knowledge circulation from knowledge formation is an original extension of that tradition.
The Science and Technology Studies horizontal dream flat networks, symmetric actors, open source as commons was a genuine intellectual and political project that produced real things. Linux. Wikipedia. The web standards process at its best. But horizontal network theory describes how knowledge circulates. It does not resolve how knowledge forms. You can flatten the governance structure of a contributor community without flattening the epistemic distance between someone who understands Transmission Control Protocol/Internet Protocol and someone who uses a browser. The network is horizontal. The knowledge ladder is not. And the assumption embedded in the horizontal hope that participants already have formation is precisely the assumption that excludes the Global South software nurse before the conversation begins.
→ Original distinction: “A curriculum is not content” separating formation structure from information availability is proposed here as a precise claim. The abundance of software learning content is well documented; the argument that content abundance does not substitute for a formation road is original.
A curriculum is not content. Every field has content books, tutorials, documentation, Stack Overflow answers, now Large Language Model explanations. Software engineering has more content than any practitioner can consume in a lifetime. What it does not have is a road: a sequence, a formation structure, a thing that takes you from software nurse to architect in a way that does not depend on being in the right network at the right time, working at the right company, or catching the attention of the right senior engineer.
The road would require several things that nobody with institutional power has had any interest in building. It would require agreeing on what architects actually know which surfaces the political question of who gets to define the knowledge that counts. It would require creating transitions between levels that are verifiable independently of network access which threatens the informal selection mechanisms that benefit incumbents. It would require investment in the formation of people who will not necessarily generate returns for the organizations doing the investing which conflicts with the short-term logic of most software employers. And it would require taking the software nurse seriously as a subject of professional development rather than a cost to be optimized which the craft literature has never quite managed to do.
→ Original framing: Positioning the Large Language Model era as a mirror rather than a cause revealing a pre-existing structure rather than creating a new one is a reframing not found in the surveyed Artificial Intelligence-and-labor literature, which predominantly treats the current moment as a rupture.
The Large Language Model era does not cause this absence. It reveals it, and it makes the stakes higher. The foreclosure of the inspectable commons, the acceleration of consumption before agency, the geopolitical concentration of production these are not new dynamics. They are existing dynamics running faster, with less friction, on infrastructure that is less legible than what came before.
The consolidation of software knowledge is not coming. It already happened. The current moment is when it becomes undeniable when the cost gravity argument, which promised democratization, visibly produces cheaper serfdom instead; when the horizontal network dream, which promised equal access, visibly encodes existing inequalities at scale; when the deskilling debate, which performed universal concern, is revealed as a conversation the field is having with itself about itself, ignoring the majority it has never theorized.
The road nobody built is not a technical problem. The knowledge of how to build it exists. The political will to build it has never been assembled, because the people who could assemble it have always had other roads to walk roads that existed precisely because someone, at some point, decided their formation mattered enough to build one.