Essay · Artificial Intelligence
The Layer That Teaches: compression, abstraction, and the formation of judgement in software engineering
Keywords

Abstract
Earlier abstractions in computing removed capabilities from the professional that were not missed, because they bore on the execution of the task, and this paper examines what changes when abstraction comes to bear on the formulation of the solution. The thesis is that compression is abstraction and that abstraction applied to the foundation degrades learning. The argument first addresses formation and employability, disputes the premise that AI literacy, as currently taught, solves the entry of junior professionals into the market, and distinguishes its mechanism from the deskilling described by the labour-process literature. The evidence shows performance gains alongside losses in comprehension, though the hypothesis that the harm concentrates among those who know least lacks the measurement that would settle it. The conclusion about juniors does not depend on that hypothesis, because the asymmetry lies in the stock of formation, where those already formed lose an advantage and those not yet formed never arrive. To avoid repeating predictions the history of computing has already seen fail, the paper offers an operational criterion of foundation, the minimum layer whose absence prevents diagnosing failure at the level immediately above. It argues, finally, that today’s senior engineer is valuable partly for having been formed without AI, that this stock does not replenish itself, and that from this follows the rule of treating AI as a tool and not as a methodology, with the caveat that delegating method is a bet on a shrinking stock, not an announced ruin.
Keywords: engineering education, learning and automation, deskilling, tacit knowledge, AI assistance, formation of judgement.
1. Introduction
The average adult today cannot make fire, although anyone can obtain it in under a second with a lighter that costs what a coffee costs. Making fire by friction, which for tens of thousands of years was the boundary between surviving and not surviving, became a survivalist’s pastime without anyone mourning the loss.
The trade of friction for the lighter was a good one for two reasons that usually go unnoticed. The abstracted capability stopped being necessary in order to judge the result, since you need not know how to make fire to know that the food burned. And the compression was local, because other spheres of life kept demanding first-hand learning, and the species kept learning, only other things.
The question that occupies me is what happens when the two conditions fail together, when the abstracted layer is the one needed to judge what the abstraction produced and compression reaches nearly every sphere at once. The first condition already fails, and the argument depends on it alone; the second is conjecture, because none of the conclusions about software engineering requires compression to be simultaneous.
Software engineering is where that failure becomes visible first, being the domain of fastest adoption and densest measurement, with more than half of enterprise use of AI tools while other professions register in single digits [1]. My thesis is that the senior engineer of 2026 is distinguished less by domain knowledge than by understanding the layer the agent abstracts, having been formed before assistance, and that this stock does not renew itself; hence the rule of using AI as a tool, never as a methodology. In formation, industry’s answer to the collapse in junior hiring has been AI literacy as operational fluency, which, on the reading I propose, does not replace the foundation, because the available evidence associates early, unrestricted exposure to assistance with performance without comprehension.
I run an engineering company and teach computing at a school of higher education, positions that the thesis of human formation as a scarce resource favours. Since declaring that interest does not fix the problem, I looked, for each central claim, for the evidence that would cost me something, and several claims came out of that test with a narrower reach than I would have preferred to give them.
2. Method and scope
This is a work of synthesis and argument, without a primary study, and its contribution is conceptual, resting on three elements: the mechanism of the abstraction of the foundation, the operational criterion of what counts as foundation, and the distinction between using AI as a tool and using it as a methodology. The data gathered here, from the psychology of learning and human factors to labour economics and software engineering, serve to show that the thesis is plausible and where it can be tested, not to prove it, because the subject is three years old and the available evidence is scarce, short-horizon and partly commercial.
Three rules governed the use of sources, the first being to trace every numerical claim to its primary source and report the real scope of the finding. The second was to identify preprints and vendor reports as such, because an argument about epistemic rigour cannot rest on material treated as firmer than it is, and the third was to discuss within the body of the argument the evidence that contradicts the thesis, of which there are two cases in the learning literature and one in the strategy literature.
Tracing the sources showed that several of the most cited statistics in executive material on AI and engineering do not survive checking [2], less through bad faith than through chain citation, in which each link cites the previous one. Appendix A lists the eight claims I traced and what the primary source actually supports, and the references are complete so that the reader can do the same with mine.
Since the thesis speaks of a horizon of years and the available empirical base measures sessions and months, it has the standing of a grounded hypothesis, with the scope limits declared alongside each claim they restrict.
3. Compression is abstraction
3.1. What compression removes
I call compression any technology that reduces the time between intention and result, and abstraction what it does to the path, whose steps go on existing, only hidden. The lighter encapsulated friction in a mechanism the user need not understand, without making it unnecessary.
The claim that there is no compression algorithm for experience [3] is usually read as praise for lived experience, but it says something stronger. Information and procedure compress; experience does not, because it is the residue of a lived process, and when technology compresses the process the very residue that mattered fails to form.
The psychology of learning treats that residue as a variable and reaches it along three paths. Through the generation effect, an item the subject produces is retained better than the same item merely read [4]. Through cognitive load theory, building the schemas that distinguish expert from novice consumes the capacity that solving the problem demands, and solving and learning compete. Through desirable difficulties, conditions that worsen performance during practice improve retention and transfer afterwards [5], [6]. In all three, friction is part of the mechanism of learning, which is uncomfortable for any productivity tool applied to a learner.
Judgement, without qualification, here designates the cognitive sense of the word, the capacity to assess whether an output is right, as distinct from the economic sense, which classifies tasks by how codifiable they are, and from the organisational sense, which names the decision about method and architecture within a team.
Education research calls that sense evaluative judgement and treats it as something a curriculum can develop deliberately, instead of leaving it as a residue of experience [7]. Faced with systems whose internals cannot be inspected, the same line of work proposes teaching students to work with the black box, which requires the standards of quality that this judgement supplies [8]. I differ on where the standard comes from, because that pedagogy assumes a learner who acquires standards while working with the box, and in software the standards for formulation were acquired by having formulated, which is the acquisition the assistant now performs. Whichever is right, the construct is measurable, and that removes one excuse for not measuring it.
3.2. Why this compression differs from earlier ones
The obvious objection is the lighter, according to which humanity has always abstracted and whoever lamented it was almost always wrong. The calculator did not destroy mathematics, nor did the compiler destroy programming or the IDE the engineer, and it remains to explain why now would be different.
The difference lies in the layer each technology abstracted and in the one it left to the human. The calculator removed arithmetic execution and left the formulation of the problem intact; the compiler removed translation to instructions without touching algorithm design; the IDE removed syntax recall and left the programmer the decision of what to write. The generative assistant removes decomposition, the choice of approach and the entire first version, and returns to the human only acceptance or rejection of the result (Figure 1).

Figure 1. How far each compression reaches. Earlier technologies stopped below formulation. The generative assistant reaches it, and returns to the human only the layer that depends on it.
I read this as a difference of kind, and not only of degree, because my hypothesis is that it was formulation that produced the capacity to judge, and it is formulation that now leaves, handing the human only judgement. If the hypothesis is right, the loop closes on itself, since judging the output well would require having walked the path the output dispenses with.
Bainbridge described this structure four decades ago as the irony of automation: the more advanced the system, the more critical the operator’s intervention in anomalies becomes, and it is automation itself that removes the practice by which the operator would maintain that competence [9]. Human factors research added the attentional mechanism of complacency, which underlies uncritical acceptance of the machine’s recommendation [10]. The difference is one of temporal order, because that case concerned operators who already knew how to operate and were losing the practice, and this one concerns apprentices who will never have it.
3.3. What the labour-process literature already established
The claim that technology removes the path by which skill is formed is fifty years old and gave rise to the vocabulary of deskilling. Braverman argued that separating conception from execution was the purpose of mechanisation, through which management captures the worker’s knowledge, encodes it into procedure and machine, and returns to the worker a task stripped of discretion [11]. There, the separation is imposed from above and resisted; here, it is adopted from below, tool by tool, by the very person whose formation it interrupts, and it arrives as a favour. There, the worker’s output does not improve as discretion falls; here, it improves immediately, which makes the loss hard to see and hard to argue against.
Attewell’s review corrected the deskilling thesis by showing that it over-generalised, because the same period upgraded skill in other occupations, and that the level of analysis, among task, job and occupation, decided the answer [12]. The correction applies to my own claim, because when junior technology work was moved out of rich countries between 2000 and 2015, the prediction that seniority would dry up did not appear in aggregate employment [13], and the thesis must be tested at that level too.
Zuboff’s field studies of computerised factories and offices are the closest historical analogue [14]. Once work was mediated by symbols on a screen, operators lost action-centred skill and had to build another, intellective skill, by interpreting the representation, which happened only when they were given access to the underlying data and time to interpret it. The outcome was a design decision, a conclusion also reached by a field experiment in which pedagogical safeguards largely mitigated the learning harm caused by unrestricted access to AI [15]. And the operators who had run the plant by hand read the screen in a way their successors could not, which is the same stock argument that holds today for the senior engineer.
Collins supplies the epistemic half by showing that the collective part of tacit knowledge is acquired by socialisation into a practice, which instruction cannot do [16]. With Evans he separates interactional expertise, that of someone who speaks the language of a domain without being able to contribute to it, from contributory expertise, which requires having practised [17]. Taught as operational fluency, AI literacy produces interactional expertise in a practice the student has never performed, convincing to the student and to the interviewer alike.
Against that lineage the contribution still available is narrow, because none of the three mechanisms, organisational in Braverman, representational in Zuboff and epistemic in Collins, describes a tool that performs the formulation step itself and returns a complete and plausible artefact for acceptance or rejection. Nor does any of them pose the question of stock and flow, which asks whether the next worker will be produced at all, since the tasks by which newcomers entered are the first the tool absorbs.
3.4. The precedent this argument must face
Arguments like this one have been made before, by people with greater technical authority than mine, and they failed. In 1975, in manuscript EWD498, Dijkstra wrote that “it is practically impossible to teach good programming to students that have had a prior exposure to BASIC: as potential programmers they are mentally mutilated beyond hope of regeneration” [18]. Backus records that in 1954 the community did not believe automatic programming could produce efficient code, because efficiency was human judgement and human judgement could not be automated [19].
Both predictions had the structure of my argument, in which abstraction removes the professional from the foundation and destroys the formation of judgement, and they were falsifiable and failed. My argument lacks, as Dijkstra’s did, a criterion of foundation independent of the layer in which the writer was trained; without it, the thesis says only that the next generation knows less of what mine learned, which every generation says. I offer the following criterion, verifiable:
Foundation is the minimum layer whose absence prevents diagnosing failure at the level immediately above.
The criterion leaves whole layers behind without nostalgia, and that is what separates it from generational complaint. Microcode and manual register allocation left the foundation because almost no current failure requires descending to them, whereas concurrency semantics and computational cost stayed, because a large class of production failures is diagnosable only by those who understand them.
Applied to the present case, the criterion locates the risk in formulation, the layer to which the characteristic failure of a system built with assistance belongs. It shows up as a plausible solution to the wrong problem, as a structure that works in the demonstrated case and breaks in the neighbouring one, or as an architectural decision nobody explicitly made.
An analysis by Apiiro, an application security vendor, of roughly 62,000 enterprise repositories offers the first measurement of that prediction. In code written with assistance, syntax errors fall by 76% and logic bugs by 60%, while architectural design flaws rise by 153% and privilege-escalation paths by 322% [20]. If the result is confirmed, the defect will have changed layers without diminishing, in the direction the criterion predicts; for now, with the methodology only partly published, it is evidence consistent with the thesis rather than an established finding. BASIC abstracted a layer whose absence did not prevent diagnosing the level above, and my reading is that the assistant abstracts the very layer on which diagnosis depends.
The criterion has two limits, the first being that it depends on the current distribution of failure modes, itself partly produced by the tools in use, which makes it useful for discarding nostalgia and weak for prediction. The second is that its empirical support amounts, for now, to that single analysis, and an adequate test would compare defect classes across matched projects, with assistance assigned and defects classified blind to it. To refute the argument, it is enough to exhibit a relevant class of failures in assisted systems that is diagnosable without the formulation layer.
3.5. Is the capacity the individual’s or the system’s?
When I say there will be nobody left to judge, I treat the capacity to judge as a property of persons, and a consolidated tradition denies that premise. In transactive memory, a group’s knowledge is each member’s memory plus a directory of who knows what [21]. In distributed cognition, a ship’s position is computed by a sociotechnical system of which no single navigator comprehends the whole, and the memory of an aircraft’s reference speeds belongs to the cockpit, not to the pilots [22].
The lighter example works against me here, because nobody knows how to make fire and even so society’s capacity to control fire rose, institutionalised in fire services, building codes and fire engineering. The capacity migrated from the individual to the system, and the system is better than any individual was.
The question becomes whether the capacity can be institutionalised, and here the generative assistant differs from the artefacts that distributed cognition studied. The sextant, the takeoff checklist and the compiler are deterministic and auditable, and whoever operates them can inspect and correct the capacity they carry. The generative model has none of those properties, because its output varies between runs, its justification is produced after the decision and the error is not localisable in the artefact.
There is an attempt under way to return the agent to the category of institutionalisable artefacts, by making deterministic and auditable the layer that decides what it may do. That layer settles permission but does not say whether the output is right, so the half of the objection concerning safety has a built answer, and the half concerning judgement depends on continuous quality evaluation in production, the least mature piece of the stack. Until that piece has the same rigour, the sextant analogy holds only for the half of the objection that does not worry me.
3.6. The objection from scale
The second condition that made the lighter trade benign was locality, by which the social sum of learning shifts rather than falls, a premise implicit in nearly every optimistic argument about AI and work that evaporates when compression is simultaneous. If the same technology abstracts formulation in programming, in writing and in diagnosis at once, and across every level of training, no compensating sphere remains where first-hand learning would go on. What the generality of the model puts in question is the existence of somewhere to displace to, more than human adaptability.
I did not locate empirical evidence on the aggregate effect of simultaneous compression across domains, for the simple reason that the phenomenon is three years old, and so the hypothesis stands as conjecture; what the evidence supports is the in-domain effect.
4. What the learning evidence supports
4.1. The pattern that repeats
Five independent studies, with different populations, tasks and methods, find the same pattern: assistance improves the product and worsens formation (Table 1).
Table 1. Experimental evidence on AI assistance and learning.
| Study | Population | Finding |
|---|---|---|
| Field experiment, PNAS [15] | About 1,000 high-school students | With access during exercises, performance rises 48% with the open interface and 127% with the tutor with safeguards. With access withdrawn, the open-interface group performs 17% worse than those who never had access. The safeguards eliminate almost all the harm |
| Randomised experiment, BJET [23] | 117 university students | The assisted group improves its essay grade and does not improve knowledge or transfer. What changes is self-regulation, outsourced to the model, which the authors call metacognitive laziness |
| Laboratory study with eye tracking, ICER [24] | Novice programmers | AI widens the distance between strong and weak: those who already knew what they wanted to write gained, and those who did not had their metacognitive difficulties aggravated and ended with inflated self-assessment, which the authors describe as an illusion of competence |
| Controlled experiment [25] | Software engineers learning a new library | Conceptual comprehension, code reading and debugging ability worsen, with no significant average efficiency gain. Those who delegated entirely gained productivity at the cost of not learning the library. In a related experiment, 50% against 67% correct on a comprehension test |
| Survey of real uses, CHI [26] | Knowledge workers, 936 real uses | Higher confidence in the tool is associated with lower self-reported critical effort, and the effort that remains shifts from producing to verifying |
The unequal effect among novice programmers comes from the laboratory and is contradicted by the field evidence on productivity, so it does not prove that juniors suffer more. In the experiment with students, the strongest causal evidence available, the harm appears only when the crutch is withdrawn, and while the assistant is present the performance indicator registers the problem with its sign inverted, which is worse than hiding it.
A study using electroencephalography found weaker brain connectivity and lower ownership of one’s own text among those who wrote with an assistant, which suggests a mechanism without establishing one, since it is a preprint with eighteen participants in the critical session and has been challenged on statistical power, reproducibility and reporting transparency [27], [28].
4.2. The gradient the literature measures is not the gradient the thesis needs
It would be tempting to say that the harm of assistance concentrates among those who know less, linking the learning effect to the collapse in junior hiring, but all the evidence on heterogeneity by skill level points the other way (Table 2).
Table 2. Studies estimating effects by skill level or seniority.
| Study | N | Design | Direction of the effect |
|---|---|---|---|
| Developers at three companies, Management Science [29] | 4,867 devs | Three field RCTs | Adoption and gains higher among recent hires and junior positions; no effect among those with longer tenure |
| Customer support agents, QJE [30] | 5,172 | Staggered difference-in-differences | +34% for novice and low-skill workers; effect near zero for the experienced |
| Consultants on knowledge tasks [31] | 758 | Pre-registered field RCT | +43% below the performance average and +17% above; and a 19 percentage point drop on a task outside the capability frontier |
| Professionals on writing tasks, Science [32] | 453 | Pre-registered RCT | Inequality between workers decreases |
| Programmers on a programming task [33] | 70 completers | RCT, single task | Less experienced developers gain more, with marginal significance |
The more than eleven thousand subjects taken together, all with the same sign, do not refute the concern about formation, because these studies measure productivity while the question is about learning. Nothing prevents assistance from raising juniors’ output more while reducing their learning more, and the mechanism of compression predicts exactly that: the greater the distance between what the person knows and what the tool delivers, the greater the immediate gain and the greater the portion of the path not walked.
On learning, the interaction term with prior knowledge has never been estimated, which I checked in the two available meta-analyses. One, with 23 studies and 27 effect sizes, finds a significant effect on productivity, g = 0.33 with an interval of 0.09 to 0.58, and an effect indistinguishable from zero on learning, g = 0.14 with an interval of −0.18 to 0.47 [34]. The other, with 35 controlled studies, reaches the same shape, with performance at 0.86 and comprehension at 0.16, p = 0.41, and records the lack of analysis by prior knowledge as a declared limitation [35].
The picture, one of a null average effect with high heterogeneity rather than of a negative effect, is compatible both with harm concentrated in one subgroup, as I propose, and with no effect at all, and greater harm for those who know less remains a hypothesis awaiting an estimate nobody has produced.
The conclusion about juniors does without that hypothesis and requires only that assistance dissociates performance from comprehension, which is measured and holds at any level, and that the junior is the one who still has to form. The asymmetry comes from stock position, since seniors who degrade lose part of an advantage they already have, whereas juniors who do not form never reach the position from which they would degrade. If the effect is uniform, the warning extends to the erosion of those already formed; if the effect is progressive, as the productivity evidence suggests, juniors deliver more without forming, and their delivery becomes an even worse indicator of the capacity they are building.
4.3. Two tensions the thesis must face
The first tension is the expertise reversal effect, whereby instructional scaffolds that help the novice lose effectiveness and come to harm the advanced learner [36]. Read hastily, the finding would absolve assistance, but what it establishes is that optimal support depends on prior knowledge, and the generative assistant does not modulate, offering maximum, undifferentiated support to someone with no prior knowledge at all. Moreover, it offers a finished solution, which dispenses with the path, rather than a worked example, which exposes it.
The second tension comes from the experiment with high-school students, in which the tutor that refuses to hand over the answer and guides the student produces a larger performance gain than the open interface and also preserves learning [15]. No deterministic version of the thesis survives this, because what degrades formation is AI designed to deliver a result rather than to guide a path, and the prescription becomes one of design.
5. The market and the literacy fallacy
5.1. Two curves in opposite directions
Rather than shrinking or growing as a whole, the technology labour market is splitting by seniority (Table 3).
Table 3. Labour market indicators: the two simultaneous directions. The sources differ in rigour, and include vendor reports, a news item and an opinion piece; the scope and caveat of each row are stated in the middle column.
| Indicator | Scope and caveat | Source |
|---|---|---|
| −73% in entry-level hiring | European tech, year on year. P1-level hiring rate falling from 35% to 8% | [37] |
| −50% or more in new-graduate hiring | Large technology firms, against the pre-pandemic level | [38] |
| −7.7% junior headcount at AI-adopting firms | 62 million résumés. The revised working paper reports about −9% | [39] |
| −20% employment among developers aged 22 to 25 | Against the 2022 peak. The 35 to 49 brackets keep growing | [40] |
| +1.3 million AI-related jobs | Global, two years, LinkedIn data. An opinion piece published by the WEF, not a report | [41] |
| 67,000 open engineering roles against 52,000 layoffs | Simultaneous, Q1 2026, base of 9,000 companies | [42] |
| AI Engineer as fastest-growing role | List of the 25 fastest-growing US roles in 2026 | [43] |
The first four rows show the entrance closing, and the last three show demand for those who already know growing, over a period that coincides with the post-pandemic correction and with high interest rates. No study fully isolates AI from the cycle, and the one that comes closest, comparing adopting with non-adopting firms, still measures adoption by proxy [39]. Attributing the whole decline to AI goes beyond the data, but the shape of the split is supported, and what separates the two trajectories is seniority.
5.2. The closest control case we have
The split by seniority has a precedent in a natural experiment, when, between 2000 and 2015, the West moved the junior layer of technology work abroad under an argument almost identical to today’s, holding that without entry work nobody learns and seniority dries up.
A study of about half a million workers showed that firms with their own offshore centres came to employ 8% less of their domestic workforce in tradable occupations because they stopped hiring, not because they laid people off [44]. Five years later, the review of the ACM task force found no measurable impact on aggregate technology employment in developed countries [13].
Bank tellers followed the same shape, with fewer tellers per branch, more branches and a remaining task of different content and greater value [45]. I found no study of the effect of offshoring on the formation of seniors at the firms that did it, and that absence is weak evidence that no collapse occurred, but fifteen years without a visible seniority crisis is a fact that my thesis has to explain.
My explanation is that, under offshoring, entry work went on being performed by people who formed while performing it, and the pipeline moved country without stopping, with an effect of geographic reallocation on the world production of seniors. The difference is falsifiable: if the entry work absorbed by agents reappears elsewhere as formative human work, we will have another reallocation rather than an interruption.
The comparison also shifts the question of whose formation is at stake, because offshoring took entry work and its rungs to India, Brazil, the Philippines and elsewhere, whose seniority is now part of the global supply. If agents absorb that work, the loss falls first there and on markets with less capacity to subsidise formation, which changes who pays first without changing the mechanism. The series I use are from the United States and Europe as a limitation of evidence, since that is where the measurement exists.
The episode also weighs against what I propose, because if the pipeline is aggregate, whoever trains pays alone for a benefit that leaks to the whole market, and the firm that does not train hires ready-made from the one that did. Under offshoring, firms did that arithmetic correctly, one by one, until the aggregate arithmetic changed, and since my propositions are addressed to the individual firm, I must say why it would pay.
While seniority was abundant, buying was cheaper than forming and the free-rider’s arithmetic held; with senior vacancies open for months and the wage premium rising, forming internally begins to compete with buying externally, and the reason to pay becomes economic before it is moral. That reason answers only in part, because it depends on scarcity persisting and because formation returns more slowly than almost any business plan. The rest is a coordination problem, historically solved outside the firm by accreditation and sector rules, and without that what I propose depends on firms willing to pay for a good they do not capture whole.
5.3. Why literacy is not the answer
The answer now circulating for the entry problem has two parts that sound sensible. The first is to keep hiring juniors, as Matt Garman, CEO of AWS, argued when he called their substitution by AI “the dumbest thing I’ve ever heard”, since without them nobody will have learned anything in ten years [46]. I agree with that conclusion and disagree with the second part, which makes AI literacy what renders juniors employable and, for three reasons, confuses two competences.
The first reason is that literacy is competence in operation, and formation is competence in judgement. Orchestrating an agent, writing a good prompt and chaining tools is a real and teachable skill, but it acts by construction on the abstract layer and does not produce the schema that allows one to assess whether the output is right. That is why performance rises while comprehension falls, and it is comprehension that the organisation will need in five years.
The second reason is that literacy is the shortest-lived skill in the set, since the interfaces, models and orchestration patterns of 2026 will not be those of 2029, whereas the foundation, from data structures and computational cost to concurrency semantics and domain modelling, lasts for decades. Training the young primarily in what expires first is hard to defend even in economic terms.
The third reason is that literacy does not solve the problem it means to solve. Literate juniors would become employable again because they produce like mid-level engineers, and the productivity studies by seniority confirm that they do, but that present capacity comes from the tool, and the firm could equally obtain it by giving the same tool to someone who already has judgement. What would remain for the junior is the price difference between the two options, which narrows as the tool improves.
Since the design of assistance changes the outcome, rejecting all literacy would be incoherent, and the objection applies to shallow literacy, understood as operational fluency in prompts, tool chaining and this year’s interfaces. From what I have examined, that is what most short courses, corporate upskilling tracks and platform certifications teach, but this is observation rather than measurement, since only a content analysis of those curricula, which does not exist, would say whether the target is as common as I claim. The uncertainty does not reach the short half-life of that fluency, nor the fact that it does not produce judgement.
The literacy that would solve the problem is something else, and the literature already indicates its shape. In the study of engineers learning a new library, three of the six interaction patterns identified preserved learning even with the tool available, because they kept the learner cognitively engaged in the path [25]. Teaching them is teaching metacognition about the use of the tool rather than its operation, which calls for its own curriculum and assessment.
5.4. The legitimate periphery is being dismantled
Lave and Wenger called legitimate peripheral participation the mechanism by which a craft reproduces itself, in which the novice enters through real, low-risk tasks and migrates toward the centre as the community trusts them with more [47]. It is by this route that communities of practice transmit what they know, including what never comes to be articulated in a manual [48].
The Dreyfus brothers’ model of skill acquisition arrives at the same place by showing that experts stop operating by explicit rules precisely because they have accumulated concrete cases [49]. The periphery is that mechanism, and it is also the first layer that assistance absorbs: the simple CRUD, the missing test, the small bug, the obvious fix.
Continuing to hire juniors is therefore necessary and insufficient, because setting juniors to operate the agent preserves the post and destroys the trajectory, and reaching this conclusion requires only the dissociation between performance and comprehension, without assuming greater harm for those who know less. What has to be preserved are real, consequential, low-abstraction tasks through which the novice enters. If humans learn from other humans by watching, asking questions and repeating, as was written about another transformation [3], the organisation has to decide what is left for the novice to watch, ask about and repeat.
5.5. The periphery is also being bought from outside
A second movement appears in the investment literature rather than in the literature on formation. Bek argues that the next generation of AI companies will sell outcomes to the end client rather than tools to the professional, capturing services budgets roughly six times larger than software budgets [1]. The entry wedge is already-outsourced work, high in volume and low in judgement, whose budget already exists and whose replacement requires no internal reorganisation.
Seen from the perspective of formation, that list of targets coincides, in almost every profession, with the tasks through which the junior used to enter: the standard NDA for the lawyer starting out, procedure coding in hospital billing, top-of-funnel screening in recruitment. Without addressing this, Bek’s thesis describes from the capital side the phenomenon I describe from the formation side, with the periphery removed from inside, through the adoption of assistants, and from outside, through the replacement of the supplier who performed it.
Hence the same fact carries opposite valences, and what the investment case reads as structural scarcity of professionals the profession reads as depletion of the stock of those who can judge.
6. Propositions for formation
These propositions presuppose the formation hypothesis, according to which the difficulty of the path formed the judgement of today’s seniors rather than merely selecting those who could bear it. P1, P4 and P5 depend directly on that hypothesis and lose their footing if selection is the dominant mechanism; P2 rests on its own experimental evidence and P3 concerns measurement, and both hold in either case. Each proposition is formulated to be testable in the practice of whoever adopts it.
P1. Deliberate low abstraction at the start of the trajectory. The curriculum and the first year of a career should contain an unassisted core, defined by a formation objective: building the data structure, debugging without suggestions, reading someone else’s code without a summary. The prohibition applies only to that core, where friction is itself the objective [6]. Whoever adopts P1 should measure attrition alongside the performance of those who remain, because that measure is what distinguishes friction that forms from friction that merely filters, and filters with bias.
P2. Assistance with safeguards, not free assistance. A tutor that guides the student without giving away the answer outperformed the open interface and preserved learning, and for that reason whoever gives tools to beginners should configure tutor mode as the default [15].
P3. Assess by reconstruction and transfer, not by delivery. A delivery indicator that improves while formation worsens has stopped measuring formation, and reconstruction recovers what delivery hides by asking the learner to explain, to modify under a new constraint, or to debug a corrupted version of their own artefact.
P4. Preserve the legitimate periphery as engineering policy. The organisation decides which real tasks stay human for formative reasons and sustains them as an investment with a declared cost, a name that survives a bad quarter better than generosity. As with the transfer of junior work offshore, part of the return leaks to the market and the firm that adopts the policy alone subsidises those that do not, so the proposition holds while buying seniority is expensive and stops holding if the market becomes slack again.
P5. Treat formation as the firm’s responsibility, not only the school’s. The idea that the team you have is the team you need was born in the move to the cloud, with capacity built inside rather than bought ready-made [50], [51]. Training on the new tool is no longer enough, however, because the old foundation also has to be protected.
7. What AI compresses in the economics of software
7.1. Hard to do is no longer hard to copy
Competitive advantages in software used to split into those that were hard to do and those that were hard to get (Figure 2). Compression attacks the first and spares the second, because it reduces the time it takes to do things without touching the time things take to happen: years of operation still take years, and a regulatory process does not accelerate because the applicant writes code faster. If a competitor reproduces in weeks what took two years, what remains defensible lies outside the code.

Figure 2. Compression reduces the time to do, not the time to happen. What remains defensible is in the lower band.
7.2. The prize is not in software
With services budgets roughly six times larger than software budgets, whoever sells a tool competes for the smaller of the two, and Bek predicts that the next trillion-dollar company will be a software company masquerading as a services firm [1].
Bek locates where this happens first by distinguishing two kinds of professional task: in intelligence tasks, the answer follows from known rules and patterns; in judgement tasks, it depends on deciding what should be done. Writing code is mostly intelligence, and knowing what to build next is judgement, so the more intelligence a profession contains, the sooner selling outcomes beats selling tools.
Two consequences follow, and the first weakens the defence of the hard to get, because if the moat is proprietary data and that data accumulates by performing service work that is already outsourced, the position of whoever generates the data can be bought, even though the time the data takes to exist does not compress. The second reaches the premise that judgement is the durable human residue, since Bek asserts that “Today’s judgement will become tomorrow’s intelligence” [1].
7.3. The bottleneck moved
Adoption is almost universal, with 90% of developers using AI [52], while confidence in the accuracy of outputs fell from 40% to 29% in one year [53]. When almost everyone uses what few trust, the most plausible reading is that the work the tool was meant to eliminate is migrating into verification.
On the business-outcome side, an MIT report attributes measurable impact to only 5% of generative AI pilots [54]. It is a preprint whose methodology has been criticised, but even with a generous discount the distance between adoption and impact is too large to be attributed to model capability.
At the level of the individual, a controlled trial found experienced open-source developers, on codebases they knew, taking 19% longer with AI while believing they had been 20% faster [55]. With 16 developers and 246 tasks the study does not generalise, and what matters is the sign of the perception error, because restructuring decisions are being made on the perceived gain.
In the artefact, review became more expensive because AI-assisted code reaches it with more defects: 10.83 issues per pull request against 6.45 for human ones, across 470 real cases [56], and, across 62,000 enterprise repositories, three to four times more code with ten times more security issues [20]. The methodologies differ and do not fuse into a single number, although that fusion already circulates (Appendix A).
In the working day, 82% of the engineers followed over six months reported writing less code, with a significant shift from creation to verification [57], a self-reported result that converges with the finding that assistance shifts critical effort from producing to verifying [26]. Each of these indications is weak on its own, but they come from different sources and methods and point the same way, and if generation accelerated without verification keeping pace, hiring more generation capacity does not help, because the constraint will have moved to the other side.
8. Today’s senior is a historical artefact
The current explanation for the senior engineer’s value in the age of AI is mastery of the client, the regulation and the product, and it is incomplete, because professionals with deep domain knowledge and no engineering training already produce, with assistance, artefacts that compete with those of developers: a lawyer and a cardiologist placed among the winners of a hackathon with thirteen thousand applicants [58]. If domain sufficed, that result would not be possible.
The professional literature gives a better answer with the convergence of profiles, which I read as two symmetrical movements, that of the specialist who broadens because AI absorbs their narrow execution and that of the generalist who deepens with specialist-grade tools [59]. Both arrive at a position in which value lies in the ideas of the software more than in the code, since, with building now fast, more time goes into deciding what to build [60].

Figure 3. The convergence of profiles, as the professional literature describes it, and what it leaves out.
The convergence correctly describes what the valuable professional does (Figure 3), but it does not explain how that capability came to exist, and whoever has to produce the next one depends on that answer.
My hypothesis is that what distinguishes the senior engineer of 2026 is historical, because they were formed before assistance, writing the hard implementation, debugging without suggestions, reading other people’s code without a summary, being beaten by concurrency and watching the system fall without understanding why until they understood. That trajectory would have produced schemas about the layer the agent now abstracts, and with them they would judge a non-deterministic output, recognising failure patterns one recognises only having stood beneath the abstraction, rather than comparing the code with what they would have written.
Three consequences follow from that origin, and the first is that the stock does not replenish itself, because of the position the junior no longer occupies and not because assistance does greater harm to those who know less, which the evidence by skill level does not support (Table 2). The productivity gains of senior teams with agents are real, and the problem is temporal, with a quantitative form even where the numbers are missing. The stock of those who can judge formulation is drained by retirement, by departure from engineering and by context obsolescence, factors I do not dispute. Replenishment is the product of the rate of hiring into consequential work and the fraction of those hired who form judgement while doing it. The first rate is falling (Table 3), and nobody measures the second, although I hold that assistance lowers it while raising output, and any of these terms may be wrong. A firm that hires and forms juniors as before is not exposed; one that cuts entry hiring and hands the remaining juniors an assistant is exposed twice, and its delivery metrics will not show it.
The second consequence is that the senior’s merit is partly an accident of era, mine included, because I was formed on fundamentals for lack of an alternative. This moves formation policy from individual virtue to system design, since, if the competence came from the absence of a shortcut and the shortcut is now free, reproducing it requires deliberately reintroducing the constraint that used to come for free.
The third consequence is an uncertainty I cannot resolve today and that changes the prescription. I said seniors were formed by low abstraction, but the data are equally compatible with their having been selected by it, in which case those who did not cross the barrier left the profession and what we observe is the residue of a filter rather than the product of a formation.
The two stories produce the same observations about current seniors and recommend opposite interventions. If the mechanism is formation, friction forms people; if it is selection, it merely filters, and filters with bias, because tolerance for friction is unevenly distributed and correlates with background, available time and support network. In that case, the deliberate low abstraction of P1 ends up producing fewer people, from where few already came, instead of more judgement.
This boundary condition can only be resolved by a design that follows an early-career cohort for twenty-four months or more, with arms by exposure regime and a measure of transfer instead of task completion, measuring attrition alongside the performance of those who stayed.
9. What if the frontier moves?
Judgement would be the durable human residue, what remains when execution is absorbed, and Bek denies that premise by asserting that “Today’s judgement will become tomorrow’s intelligence” [1]. If he is right, the prescription I defend has an expiry date and protects the last position before it falls too.
His argument does not rest on a promise about future model capability, because systems that perform intelligence work inside a profession accumulate proprietary data about how that profession decides, and it is that accumulation which converts judgement into intelligence, through a data-acquisition mechanism with a track record of working. I accept the mechanism and disagree with the conclusion on three points, all checkable against his own argument.
The first point is that the mechanism is parasitic on what it consumes, because the frontier moves only while the system observes human judgement being exercised, and if the profession stops forming those who exercise judgement, the source dries up before the conversion completes. The thesis assumes a continuous stock of observable judgement, which is precisely what is at risk.
The second point is that the mechanism works where there is a pattern to extract, which differs from there being a decision to make. The categories Bek picks as priorities, such as hospital billing coding and standardised contract issuance, were already rule-governed, with human judgement entering as a transaction cost and not as a choice, which is why their fall was predictable before generative AI.
The third point is temporal, because the conversion, if it happens, happens profession by profession, over years, and depends on data that must be generated before it can be learned. The window of declining human stock and incomplete conversion coincides with the five-to-ten-year horizon that concerns me, and within it the thesis, even if correct, postpones the problem and transfers it to whoever holds the data.
He does convince me, however, on the distinction between intelligence and judgement, which is more operational than mine, and it is the one I adopt as a delegation criterion.
10. AI as tool, not as methodology
10.1. The distinction
The prescription depends on separating tool from methodology, a distinction usually treated as rhetorical that has operational consequences. Methodology decides how work is organised, how the problem is decomposed and how decisions are made and revised, whereas a tool executes within an already decided method. The same technology can occupy both places, and the difference shows up only on the horizon in which the organisation has to form whoever decides (Table 4).
Table 4. The same technology in two distinct places in the system of work.
| Dimension | AI as tool | AI as methodology |
|---|---|---|
| Problem decomposition | Human, explicit, reviewable | Emergent from the session with the agent |
| Acceptance criterion | Defined beforehand, by whoever answers for the system | Defined by the plausibility of the output |
| Architecture | Documented human decision, with the agent implementing within it | By-product of what the agent generated and what passed the tests |
| Role of review | Verify against a prior criterion | Discover the criterion while reviewing |
| Effect on formation | The decision path stays human and observable | The path disappears and the result remains |
| Failure mode | Localised, attributable, correctable error | Diffuse erosion, in which nobody knows why the system is the way it is |
The second column works only while the stock of seniors formed by exposure to the foundation lasts. Contrary to what public discussion assumes, the risk does not lie in our not having trained juniors in AI, but in their not having passed through the low abstraction of the foundation, which leaves them unable to occupy the first column when the current occupants leave.
Risky, in this case, does not mean unviable, because two exits remain open to whoever operates this way. The capacity to judge may migrate from people to the system, institutionalised in continuous evaluation and executable policy, or the frontier between intelligence and judgement may move far enough for the first column to shrink, as Bek would have it. Neither is demonstrated, and the second column consumes a stock whose replacement is falling, so adopting it is a bet, and whoever makes it should know what they are betting on.
The distinction between intelligence and judgement [1] gives the cut-off criterion, applied to the task and not to one’s confidence in the model. A task whose answer follows from a known rule can be delegated without transferring method, whereas a task whose answer depends on deciding what should be done cannot, even when the output is plausible, because delegating it moves methodology inside the model.
10.2. What the distinction implies in practice
Specification and architecture stay human and written because they preserve the decision path, which forms whoever comes next and allows what was done to be audited. The agent operates inside a declared envelope, with scope, permissions and acceptance criteria defined before execution and verified outside the inference loop (Figure 5).
Two allocations follow from this, and the first is to give review an owner and a budget, because if the bottleneck migrated from production to verification, treating review as a residual expense optimises the wrong side of the constraint. The second is to require, before assisted output is accepted in production, sign-off from someone who has the foundation, since the competence to judge is now the scarce resource.
11. Structure and governance
11.1. Who builds, who operates, who sustains
The separation between an engineering group that builds and an operations group that runs was already an anti-pattern before AI, because a team that makes critical technical decisions and wakes another team to fix things when something goes wrong has no ownership of what it made [61].
Agents make the separation unsustainable, because an agent’s behaviour is non-deterministic and debugging it requires understanding the system prompt, the context window, retrieval and the reasoning chain, while an operations group organised to restart services and file tickets sees only the symptom. The diagnostic is whether the operations team can read prompts, trace reasoning chains and assess retrieval.
For delivery, the pod that works is a small team of experienced engineers with agents and end-to-end ownership of a workflow, Conway’s law applied to the stream-aligned team, and the trend goes far enough that the two-pizza team is already said to have grown too large for building software with agents [62]. The format requires a platform layer, because end-to-end ownership with agents is impossible without infrastructure that treats them as first-class citizens [63].

Figure 4. The pod and the organisation do not share an optimal shape. The inverted pyramid describes a delivery pod well; the hourglass is the only one that also reproduces the input the pod depends on.
A senior pod with agents delivers well and forms nobody, and an organisation made only of such pods delivers for a few years and then has nowhere to draw whoever forms the next one (Figure 4). The hourglass sustains both because it keeps the base, which produces the model’s scarce input, even though it costs more in the short run and is the first thing to be cut in a bad quarter.
The observation that speed is an executive choice [61] holds for the shape of the organisation, except that the effects of that shape appear only after the term of whoever chose it.
When part of the intelligence work migrates to suppliers selling outcomes, the buying organisation needs fewer people to execute and more people to specify and verify, so the pressure falls on the same scarce competence.
11.2. Governance as infrastructure
If the agent operates inside an envelope, the envelope has to be executable, and the convergence between regulators and product stacks suggests the framing is structural. Singapore’s Infocomm Media Development Authority organises agentic AI governance into bounding risks upfront, making humans meaningfully accountable, implementing technical controls across the lifecycle and enabling end-user transparency [64], and the logging and traceability obligations of Regulation (EU) 2024/1689 ask for substantially the same [65]. The four questions the envelope answers and the mechanism each one demands are set out in Appendix B.

Figure 5. The agent’s envelope. The decision about what may be done is evaluated outside the inference loop, at the authorisation point, and is therefore not reachable by content entering the context.
The layer most implementations lack is the one that decides what the agent may do. Asking the system to behave differs from preventing it from acting in the position at which the control operates, and not in its content, which is why authorisation languages with analysable semantics matter more here than in conventional applications [66], [67].
Technically, this materialises the separation between tool and methodology, with the method outside the model, written by humans and auditable, and the model executing inside it.
12. Limitations
The central thesis is a grounded hypothesis rather than a finding, because no available study follows a cohort of developers long enough to measure the formation of judgement, which is the dependent variable of the argument.
The evidence base is young and partly commercial, and that fragility holds for both sides of the dispute. Vendor reports with only partly published methodology support the findings on code quality and on the market bifurcation [20], [37], [38], [56], and preprints support others, among them the main finding on formation [25], [27], [54], [57], so a reader who discounts the evidence for the thesis should discount the evidence against it in the same proportion.
The evidence that qualifies the thesis receives an answer in the design of tool and task, but the evidence that confronts its premise, that today’s judgement becomes tomorrow’s intelligence, receives an answer that is only argumentative, and that answer therefore remains open.
Since the thesis predicts consequences outside the measurement horizon, it has the shape of a self-protecting proposition, and three observations, should they occur, count against it:
- An interaction term between prior knowledge and exposure to assistance, estimated with adequate power on a transfer measure, with a non-negative sign.
- The rate of promotion to senior, at intensively adopting organisations, remaining within the historical range over five years, corrected for the economic cycle.
- No difference in judgement quality, under blind assessment of architectural decisions, between engineers formed before and after 2023, controlling for years of experience.
All the employment series cited come from the United States or Europe, and the structure of the Brazilian market, with a different proportion of outsourcing and service exports, may produce different dynamics, so treating these curves as predictive for Brazil would be extrapolation.
No aggregate series supports the argument that simultaneous compression eliminates the sphere where first-hand learning would be compensated, and no other part of the argument is as speculative as that one.
13. Conclusion and agenda
The lighter was a good trade because you did not need to know how to make fire in order to judge fire, and because the compression was local. The trade we are making now satisfies neither condition, since the abstracted layer is the one that produced the competence to judge and the compression is simultaneous across nearly every sphere, and without those conditions the analogy that it was always like this and always turned out fine does not bear the weight placed on it.
In the person and in the organisation the same mechanism operates, with the indicator improving while the capacity worsens, and the prescription is also the same at both scales, because at both what is protected is the capacity to judge, the first to be lost and the last to return. In formation, it consists in reintroducing the constraint that used to come for free, with assistance that guides and assessment by reconstruction. The constraint forms if the mechanism is formation and merely filters if it is selection, and the safeguarded tutor that, in a field experiment, improved performance without sacrificing learning weighs in favour of the first hypothesis [15]. In engineering, the method decision stays on the human side, and the model executes inside an envelope verified outside the inference loop.
Bek’s objection still stands, because, if today’s judgement becomes tomorrow’s intelligence, this prescription protects a temporary position. Protecting a temporary position differs from protecting nothing by the time it buys, and the decision about what to do with that time has to be made within it.
The central question of the agenda is whether assistance degrades the formation of judgement over a long horizon, and under what design of task and tool that degradation disappears. The minimum design follows an early-career cohort for twenty-four months or more, with arms differentiated by exposure regime and measures of conceptual transfer and debugging rather than task completion. Two further questions remain open: is there, under simultaneous compression, a compensating sphere where first-hand learning goes on happening? And, if the frontier between intelligence and judgement does move, what happens to the mechanism that moves it when the source of observable human judgement begins to dry up?
Until those studies exist, every organisation cutting its base is running the experiment without a control group and without recording the result, and I wrote this because mine is one of them.
Statements and Declarations
Funding. No funding was received for this work.
Competing interests. The author directs an engineering company and teaches at a school of computing. The thesis that human formation remains the scarce resource is consistent with both positions, which is declared in the introduction and treated there as a reason for adversarial sourcing rather than as a disclosure that settles the matter.
Ethics approval. Not applicable, because the paper is a work of synthesis and argument and involves no human participants.
Data availability. No new data were generated. Every empirical claim is traced to a published primary source listed in the references, and Table 5 records the claims in circulation that could not be traced.
Use of generative AI. A large language model was used for translation between Portuguese and English, copy-editing, reference formatting and structural revision, and it produced the first draft of the passage on the labour-process literature, which the author then verified against the primary texts and rewrote. The argument, the selection of sources and the claims are the author’s, and no text was accepted without verification against the cited source.
Appendix A. Evidence hygiene
Tracing the sources of this paper, I found that several of the most cited statistics in executive material on AI and engineering do not survive checking [2]. The pattern repeats and comes not from bad faith but from chain citation, in which each link cites the previous one and nobody returns to the source. Table 5 lists the claims verified directly. The verdict “not located” does not mean that the claim is false, but that it should not serve as a basis for decision until whoever presents it supplies the reference.
Table 5. Claims in circulation and what the primary source supports.
| Claim in circulation | What the primary source supports | Verdict |
|---|---|---|
| 2.74× more vulnerabilities across 62,000 repositories | The 62,000-repository scope belongs to a study reporting about ten times more security issues. The 2.74× factor is from another, over 470 PRs. Two studies are fused into one citation [20], [56] | Partial |
| 73% of tasks with five or more steps hallucinate, attributed to Stanford HAI | I located no publication from the institute with that claim. Its material on hallucination concerns legal models, at a rate of one in every six queries | Not located |
| 17% less comprehension, n ≈ 1,500, attributed to “Bedard et al.” via HBR | The finding belongs to another study, with n = 52: 50% against 67% correct, that is, 17 percentage points. The speed difference was not significant [25] | Partial |
| Nine months compressed into 76 days with five times fewer people | The project cited exists, and the located datum is roughly ten times in code throughput, with attribution and date different from those in circulation [68] | Not located |
| 13% of agents have strong visibility | It is 13% of companies that report strong visibility into how AI touches their data. Subject and object swapped [69] | Partial |
| 1,587% growth in agent skills | The reference report lists AI Engineer as the fastest-growing role, and I did not locate that percentage in it [43] | Not located |
| 5.8% unemployment among recent graduates, “worst since 2013” | The value corresponds to the first quarter of 2025 and is described as the highest level since 2021 [70] | Partial |
| 1.3 million new AI posts, attributed to a WEF report | It is an article signed by a LinkedIn executive and published by the WEF, based on LinkedIn data [41] | Partial |
Appendix B. Four governance questions and their mechanisms
Table 6 expands the framing of agent governance discussed in the body of the paper, with the four questions an executable envelope has to answer and the mechanism each one demands.
Table 6. Four governance questions and their mechanisms.
| Question | Mechanism |
|---|---|
| Who is this agent? | Workload identity bound to a human supervisor. The agent never exceeds the permissions of the human who answers for it |
| What is it allowed to do? | Deterministic policy evaluated before tool access, outside the inference loop |
| Is it working correctly? | Continuous evaluation in production and complete reasoning-chain traces |
| Can we prove it? | A record of every enforcement decision, with cost, latency and error dashboards |
References
[1] J. Bek, ‘Services: The new software’. Sequoia Capital, Mar. 2026. Available: https://sequoiacap.com/article/services-the-new-software
[2] Industry executive material, ‘Executive presentations and reports on artificial intelligence and software engineering’. 2026.
[3] J. Allen, ‘A 12 step program to get from zero to hundreds of AWS-certified engineers’. AWS Cloud Enterprise Strategy Blog, Jul. 2017. Available: https://aws.amazon.com/blogs/enterprise-strategy/a-12-step-program-to-get-from-zero-to/
[4] N. J. Slamecka and P. Graf, ‘The generation effect: Delineation of a phenomenon’, Journal of Experimental Psychology: Human Learning and Memory, vol. 4, no. 6, pp. 592–604, 1978, doi: 10.1037/0278-7393.4.6.592.
[5] R. A. Bjork, ‘Memory and metamemory considerations in the training of human beings’, in Metacognition: Knowing about knowing, J. Metcalfe and A. P. Shimamura, Eds., Cambridge, MA: The MIT Press, 1994, pp. 185–206. doi: 10.7551/mitpress/4561.003.0011.
[6] R. A. Bjork and E. L. Bjork, ‘Desirable difficulties in theory and practice’, Journal of Applied Research in Memory and Cognition, vol. 9, no. 4, pp. 475–479, 2020, doi: 10.1016/j.jarmac.2020.09.003.
[7] J. Tai, R. Ajjawi, D. Boud, P. Dawson, and E. Panadero, ‘Developing evaluative judgement: Enabling students to make decisions about the quality of work’, Higher Education, vol. 76, no. 3, pp. 467–481, 2018, doi: 10.1007/s10734-017-0220-3.
[8] M. Bearman and R. Ajjawi, ‘Learning to work with the black box: Pedagogy for a world with artificial intelligence’, British Journal of Educational Technology, vol. 54, no. 5, pp. 1160–1173, 2023, doi: 10.1111/bjet.13337.
[9] L. Bainbridge, ‘Ironies of automation’, Automatica, vol. 19, no. 6, pp. 775–779, 1983, doi: 10.1016/0005-1098(83)90046-890046-8).
[10] R. Parasuraman and D. H. Manzey, ‘Complacency and bias in human use of automation: An attentional integration’, Human Factors, vol. 52, no. 3, pp. 381–410, 2010, doi: 10.1177/0018720810376055.
[11] H. Braverman, Labor and monopoly capital: The degradation of work in the twentieth century. New York: Monthly Review Press, 1974.
[12] P. Attewell, ‘The deskilling controversy’, Work and Occupations, vol. 14, no. 3, pp. 323–346, 1987.
[13] M. Y. Vardi, ‘Globalization and offshoring of software revisited’, Communications of the ACM, vol. 53, no. 5, p. 5, 2010, doi: 10.1145/1735223.1735224.
[14] S. Zuboff, In the age of the smart machine: The future of work and power. New York: Basic Books, 1988.
[15] H. Bastani, O. Bastani, A. Sungu, H. Ge, Ö. Kabakçı, and R. Mariman, ‘Generative AI without guardrails can harm learning: Evidence from high school mathematics’, Proceedings of the National Academy of Sciences, vol. 122, no. 26, p. e2422633122, 2025, doi: 10.1073/pnas.2422633122.
[16] H. Collins, Tacit and explicit knowledge. Chicago: University of Chicago Press, 2010.
[17] H. Collins and R. Evans, Rethinking expertise. Chicago: University of Chicago Press, 2007.
[18] E. W. Dijkstra, ‘How do we tell truths that might hurt?’ E. W. Dijkstra Archive, University of Texas at Austin, 1975. Available: https://www.cs.utexas.edu/~EWD/transcriptions/EWD04xx/EWD498.html
[19] J. Backus, ‘The history of FORTRAN I, II, and III’, ACM SIGPLAN Notices, vol. 13, no. 8, pp. 165–180, 1978, doi: 10.1145/960118.808380.
[20] I. Nussbaum, ‘4x velocity, 10x vulnerabilities: AI coding assistants are shipping more risks’. Apiiro Blog, Sep. 2025. Available: https://apiiro.com/blog/4x-velocity-10x-vulnerabilities-ai-coding-assistants-are-shipping-more-risks/
[21] D. M. Wegner, ‘Transactive memory: A contemporary analysis of the group mind’, in Theories of group behavior, New York: Springer, 1987, pp. 185–208. doi: 10.1007/978-1-4612-4634-3_9.
[22] E. Hutchins, Cognition in the wild. Cambridge, MA: The MIT Press, 1995. doi: 10.7551/mitpress/1881.001.0001.
[23] Y. Fan et al., ‘Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance’, British Journal of Educational Technology, vol. 56, no. 2, pp. 489–530, 2025, doi: 10.1111/bjet.13544.
[24] J. Prather et al., ‘The widening gap: The benefits and harms of generative AI for novice programmers’, in Proc. 2024 ACM conf. On international computing education research (ICER), 2024, pp. 469–486. doi: 10.1145/3632620.3671116.
[25] J. H. Shen and A. Tamkin, ‘How AI assistance impacts the formation of coding skills’, Anthropic, Jan. 2026. Available: https://www.anthropic.com/research/AI-assistance-coding-skills
[26] H.-P. Lee et al., ‘The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers’, in Proc. 2025 CHI conf. On human factors in computing systems, 2025. doi: 10.1145/3706598.3713778.
[27] N. Kosmyna et al., ‘Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task’. arXiv:2506.08872, 2025. doi: 10.48550/arXiv.2506.08872.
[28] M. Stankovic, E. Hirche, S. Kollatzsch, and J. N. Doetsch, ‘Comment on: Your brain on ChatGPT’. arXiv:2601.00856, 2025. doi: 10.48550/arXiv.2601.00856.
[29] K. Z. Cui, M. Demirer, S. Jaffe, L. Musolff, S. Peng, and T. Salz, ‘The effects of generative AI on high-skilled work: Evidence from three field experiments with software developers’, Management Science, 2026, doi: 10.1287/mnsc.2025.00535.
[30] E. Brynjolfsson, D. Li, and L. Raymond, ‘Generative AI at work’, The Quarterly Journal of Economics, vol. 140, no. 2, pp. 889–942, 2025, doi: 10.1093/qje/qjae044.
[31] F. Dell’Acqua et al., ‘Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality’, Harvard Business School, Working Paper 24-013, 2023. doi: 10.2139/ssrn.4573321.
[32] S. Noy and W. Zhang, ‘Experimental evidence on the productivity effects of generative artificial intelligence’, Science, vol. 381, no. 6654, pp. 187–192, 2023, doi: 10.1126/science.adh2586.
[33] S. Peng, E. Kalliamvakou, P. Cihon, and M. Demirer, ‘The impact of AI on developer productivity: Evidence from GitHub Copilot’. arXiv:2302.06590, 2023. doi: 10.48550/arXiv.2302.06590.
[34] S. Maier, M. Gunzenhäuser, J. Schweisthal, M. Schneider, and S. Feuerriegel, ‘A meta-analysis of the effect of generative AI on productivity and learning in programming’. arXiv:2605.04779, 2026.
[35] M. Alanazi, B. Soh, H. Samra, and A. Li, ‘The influence of artificial intelligence tools on learning outcomes in computer programming: A systematic review and meta-analysis’, Computers, vol. 14, no. 5, p. 185, 2025, doi: 10.3390/computers14050185.
[36] S. Kalyuga, P. Ayres, P. Chandler, and J. Sweller, ‘The expertise reversal effect’, Educational Psychologist, vol. 38, no. 1, pp. 23–31, 2003, doi: 10.1207/s15326985ep3801_4.
[37] Ravio, ‘Early career hiring is down 73%: Why it is time to redesign entry-level roles for the AI era’. Ravio Blog, Jul. 2025. Available: https://ravio.com/blog/early-career-hiring
[38] A. Bantock, ‘The SignalFire state of tech talent report 2025’, SignalFire, May 2025. Available: https://www.signalfire.com/blog/signalfire-state-of-talent-report-2025
[39] S. M. Hosseini Maasoum and G. Lichtinger, ‘Generative AI as seniority-biased technological change: Evidence from U.S. Résumé and job posting data’, Harvard University, SSRN Working Paper 5425555, 2025. Available: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555
[40] E. Brynjolfsson, B. Chandar, and R. Chen, ‘Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence’, Stanford Digital Economy Lab, Aug. 2025. Available: https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine
[41] D. Shapero, ‘AI has already added 1.3 million jobs, LinkedIn data says’. World Economic Forum, Jan. 2026. Available: https://www.weforum.org/stories/jobs-and-the-future-of-work/ai-has-already-added-1-3-million-new-jobs-according-to-linkedin-data/
[42] Implicator.ai, ‘Software engineering openings jump 30% even as 52,000 tech workers lose jobs’. Reportagem sobre dados do TrueUp (9.000 empresas), 1º trimestre de 2026, Apr. 2026. Available: https://www.implicator.ai/software-engineering-openings-jump-30-even-as-52-000-tech-workers-lose-jobs/
[43] LinkedIn News, ‘LinkedIn jobs on the rise 2026: The 25 fastest-growing roles in the U.S.’ Jan. 2026. Available: https://www.linkedin.com/pulse/linkedin-jobs-rise-2026-25-fastest-growing-roles-us-linkedin-news-dlb1c
[44] P. Tambe and L. M. Hitt, ‘Now IT’s personal: Offshoring and the shifting skill composition of the U.S. Information technology workforce’, Management Science, vol. 58, no. 4, pp. 678–695, 2012, doi: 10.1287/mnsc.1110.1445.
[45] J. Bessen, Learning by doing: The real connection between innovation, wages, and wealth. New Haven: Yale University Press, 2015. doi: 10.12987/9780300213645.
[46] S. Sharwood, ‘AWS CEO says using AI to replace junior staff is the “dumbest thing i’ve ever heard”’. The Register, Aug. 2025. Available: https://www.theregister.com/2025/08/21/aws_ceo_entry_level_jobs_opinion/
[47] J. Lave and E. Wenger, Situated learning: Legitimate peripheral participation. Cambridge: Cambridge University Press, 1991. doi: 10.1017/CBO9780511815355.
[48] M. Polanyi, The tacit dimension. Chicago: University of Chicago Press, 2009.
[49] S. E. Dreyfus, ‘The five-stage model of adult skill acquisition’, Bulletin of Science, Technology and Society, vol. 24, no. 3, pp. 177–181, 2004, doi: 10.1177/0270467604264992.
[50] J. Allen, ‘12 steps to get started with the cloud’. AWS Cloud Enterprise Strategy Blog, Jan. 2018. Available: https://aws.amazon.com/blogs/enterprise-strategy/12-steps-to-get-started-with-the-cloud/
[51] J. Allen, ‘The future waits for nobody: My Capital One journey to the AWS cloud’. AWS Cloud Enterprise Strategy Blog, Jun. 2017. Available: https://aws.amazon.com/blogs/enterprise-strategy/the-future-waits-for-nobody-my-capital-one-journey-to-the-aws-cloud/
[52] DORA and Google Cloud, ‘2025 DORA report: State of AI-assisted software development’, Google Cloud, Sep. 2025. Available: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report
[53] E. May, ‘Mind the gap: Closing the AI trust gap for developers’. The Stack Overflow Blog, Feb. 2026. Available: https://stackoverflow.blog/2026/02/18/closing-the-developer-ai-trust-gap/
[54] A. Challapally, C. Pease, R. Raskar, and P. Chari, ‘The GenAI divide: State of AI in business 2025’, MIT NANDA, Jul. 2025.
[55] J. Becker, N. Rush, B. Barnes, and D. Rein, ‘Measuring the impact of early-2025 AI on experienced open-source developer productivity’, METR — Model Evaluation; Threat Research, Jul. 2025. Available: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
[56] CodeRabbit, ‘State of the AI vs. Human code generation report’, CodeRabbit, Dec. 2025. Available: https://www.coderabbit.ai/whitepapers/state-of-AI-vs-human-code-generation-report
[57] A. Vella and K. Blincoe, ‘The impact of AI coding assistants on software engineering: A longitudinal study’. arXiv:2605.23135, 2026.
[58] Anthropic, ‘Meet the winners of our built with Opus 4.6 Claude Code hackathon’. Anthropic (blog), Feb. 2026. Available: https://claude.com/blog/meet-the-winners-of-our-built-with-opus-4-6-claude-code-hackathon
[59] U. Joshi, G. Venkatraman, and M. Fowler, ‘Expert generalist’. martinfowler.com, Jul. 2025. Available: https://martinfowler.com/articles/expert-generalist.html
[60] A. Ng, ‘The product management bottleneck’. The Batch, DeepLearning.AI, issue 349, Apr. 2026. Available: https://www.deeplearning.ai/the-batch/issue-349
[61] J. Allen, ‘Untangling your organisational hairball: Well-controlled’. AWS Cloud Enterprise Strategy Blog, May 2024. Available: https://aws.amazon.com/blogs/enterprise-strategy/untangling-your-organisational-hairball-well-controlled/
[62] D. Shipper, ‘The two-slice team’. Every, Chain of Thought, Feb. 2026. Available: https://every.to/chain-of-thought/the-two-slice-team
[63] L. Galante, ‘Ten platform engineering predictions for 2026’. PlatformEngineering.org, Dec. 2025. Available: https://platformengineering.org/blog/10-platform-engineering-predictions-for-2026
[64] Infocomm Media Development Authority, ‘Model AI governance framework for agentic AI, version 1.0’, IMDA, Singapore, Jan. 2026. Available: https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/press-releases/2026/new-model-ai-governance-framework-for-agentic-ai
[65] European Parliament and Council of the European Union, ‘Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (artificial intelligence act)’. Official Journal of the European Union, Jun. 2024. Available: https://eur-lex.europa.eu/eli/reg/2024/1689/oj
[66] J. W. Cutler et al., ‘Cedar: A new language for expressive, fast, safe, and analyzable authorization’, Proceedings of the ACM on Programming Languages, vol. 8, no. OOPSLA1, 2024, doi: 10.1145/3649835.
[67] Amazon Web Services, ‘Amazon Bedrock AgentCore now includes policy (preview), evaluations (preview) and more’. AWS What’s New, Dec. 2025. Available: https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-bedrock-agentcore-policy-evaluations-preview
[68] J. Magerramov, ‘Amazon Bedrock Mantle and developing at the speed of AI’. AWS Podcast, episode 753, Jan. 2026.
[69] Cyera Research Labs, ‘2025 state of AI data security report’, Cyera, Sep. 2025. Available: https://www.cyera.com/research-labs/2025-state-of-ai-data-security-report
[70] Federal Reserve Bank of New York, ‘The labor market for recent college graduates’. Quarterly statistical series, 2025. Available: https://www.newyorkfed.org/research/college-labor-market
Comments
Every comment is moderated before it appears here. Nothing is published automatically.
Loading…