Who Will Become
the Next Senior? Generative AI appears to be absorbing not just specific categories of tasks, but parts of the pathway through which the next generation of seniors is formed.

The decline of junior hiring is commonly read as job displacement.

Tools like GitHub Copilot, ChatGPT, and Claude now perform substantial portions of the code generation, debugging, and documentation that were frequently assigned to entry-level engineers. Concurrently, junior developer hiring has contracted sharply. These trends are commonly read as job displacement — the efficient substitution of human labor with automated tools, accompanied by corresponding adjustments in headcount.

We do not dispute that this displacement is occurring. We would argue, however, that the tasks currently being absorbed by GenAI — minor bug fixes, routine implementation, debugging — are not simply entry-level tasks that organizations can eliminate without further issues and concerns.

What GenAI is absorbing is the developmental scaffolding through which juniors grow into seniors.

What GenAI is absorbing is the developmental scaffolding through which juniors grow into seniors: the early stages of an apprenticeship-based development pathway, where juniors learn through real-world experience alongside more experienced developers, and expertise is passed down from one generation to the next.

When this scaffolding is absorbed, what may disappear is not only entry-level employment but also the very mechanism through which the next generation of seniors — skilled professionals — is formed.

Building on the macro-level hiring decline documented in numerous quantitative analyses, the contribution of our study lies not in confirming this phenomenon but in identifying its underlying mechanism and the factors contributing to its persistence.

Semi-structured interviews with 14 software engineers

We recruited 14 participants through purposive sampling: eight juniors (J1–J8), defined as individuals within two years of completing an undergraduate ECE/CS degree, at the threshold of entering the workforce; and six seniors (S1–S6), professionals with at least six years of industry experience, drawn from diverse organizational contexts — large enterprises, mid-sized companies, and startups. Interviews were conducted in Korean and analyzed using Reflexive Thematic Analysis.

Rather than serving as a representative national case, South Korea functions as an important critical case in which these dynamics emerge earlier and more clearly: work-related AI usage reaches 51.8%, almost double the U.S. rate; open recruitment has collapsed; and Korea's labor market structures displace adjustment costs onto labor market entrants rather than incumbent workers.

A foundational pattern of Absorption — and three consequences

Empirically, we identify a foundational pattern of Absorption: GenAI redirects entry-level work into senior–AI workflows. From this pattern, three consequences follow, each corresponding to one of our three research questions, as the diagram below summarizes.

Absorption of junior opportunities into senior workflows

A foundational pattern observed across senior participants was that GenAI tools had dramatically expanded their individual capacity, enabling them to handle tasks that would previously have been delegated to others, particularly entry-level engineers. This had a direct structural consequence: the entry-level work that had historically provided learning opportunities for entry-level engineers was no longer reaching them.

Having significantly reduced his junior headcount over the last year, S4 observed that their departure had had no measurable impact on productivity. This perspective was not unique to him: among senior participants (S1–S6), it was consistently observed that AI tools had effectively absorbed the more routine, lower-complexity tasks — e.g., code generation, initial debugging, and documentation. As S5 put it: “These days, things are being replaced so quickly that juniors are increasingly not being given the opportunity to do those things.”

A boundary condition. S1, at a large enterprise, was the one exception. He too had become able to handle most of the tasks previously done by juniors — yet he noted that “from the junior's perspective, the work they do and what they learn hasn't really changed much. The only difference is that AI now exists.” In contrast to the absorption pattern the other seniors described, juniors in his organization continued to do similar work, with AI added as a supplement rather than as a replacement. Rather than reading this as a simple outlier, we consider it as indicating a boundary condition for the absorption mechanism: absorption operates most freely where institutional structures for junior development are weak or implicit.

Erosion of competence through lost failure experiences

A central finding for both junior and senior participants was that, as they began using GenAI — whether in school or at work — failure no longer served as a source of learning. As GenAI absorbed the lower-complexity tasks that once characterized early-career workflows, the productive struggle embedded in those tasks also disappeared. All junior participants (J1–J8) noted that they could now achieve good results easily, without the struggle they once experienced; what emerged in its place was output without understanding.

J1 put the same problem more starkly: “I don't know what I don't know.” And J8 provided the most illustrative example, recalling a project presentation at her internship where she used unverified, AI-generated content:

Taken together, these accounts reflect what Kapur's theory of Productive Failure would suggest: when learners are shielded from the imperfect, effortful attempts they must endure to achieve mastery, they not only lose knowledge but also the ability to recognize the boundaries of their own competence. Bjork's concept of desirable difficulties makes the parallel point: the conditions that appear to slow progress — struggling, failing, and revising — produce durable and transferable learning. By eliminating these conditions, GenAI yields what J4 described as a distinct difference between two university courses covering the same subject:

The students, the academic field, and the evaluation criteria are all the same; the only difference is whether the assignment can be completed using GenAI. The grade-equivalent outcomes are what make the case analytically significant: they show that the institutional signal of learning — the grade — can remain stable while what is actually learned diverges sharply. In one course, the work produced durable domain knowledge; in the other, only fluency in operating a tool. Here, the current assessment structure in education fails to identify that difference.

Notably, junior and senior participants — standing at opposite ends of the experience spectrum — used nearly identical phrasing, “knowing what should not be done,” to describe the missing competency:

Moreover, what people can do in workplaces is fundamentally related to the skills formed by this failure process. S5, a senior engineer at a mid-sized company, described this point as follows:

Such verification — knowing what is wrong, what does not work — is the human skill that failure-based learning produces. Yet juniors, by being shielded from failures, are no longer developing it. GenAI accelerates output production while eroding the very experiences through which juniors would learn to evaluate that output.

Collective pressure as a barrier to productive struggle

Why is productive struggle being eliminated in the first place? This is not due to individual choice, but rather stems from a structural dynamic that operates within the educational environment itself. A consistent pattern across junior participants was that the use of GenAI in university coursework is no longer a matter of personal choice. It had become the norm for most students — meaning individual students found it difficult to resist using GenAI, regardless of their own intentions.

The pressure J4 describes is structural: the grading curve itself, as well as the fact that peers are using GenAI, has become the mechanism of coercion. J5 expressed the same norm — “Not using it is just stupid.” J7 framed it as anxiety — “If I don't use it, I feel like I'm falling behind on my own.” Crucially, this pressure operated not only among students but also at the institutional level in their universities. Faculty had effectively overlooked the practice, as J3 noted: “The professors know students are using ChatGPT anyway… I don't think there was a single professor who told us not to use it.”

J6 explained the institutional logic behind this silence as follows:

What makes this consequential for junior skill development is precisely identified by J5 — who recognized what was being lost, and used AI anyway, because the educational structure left no viable alternative:

Albeit without using the theoretical term, J5's account describes the desirable difficulty proposed by Bjork. Not all junior participants felt this pressure to the same extent — J2 described his ongoing resistance, “I keep trying not to take those shortcuts.” Yet J2 also acknowledged its limit: “I'm trying not to, but I'm so pressed for time that I end up doing it anyway.” Individual intention, in other words, was undermined by structural constraint, suggesting that collective pressure exerts a stronger influence than personal preference.

Misaligned perspectives across experience levels

Senior participants were not indifferent to the difficulties of juniors, and many expressed concern. But a consistent pattern across senior participants was that the current situation appeared manageable. Underlying this perception was an implicit assumption: that juniors who had accumulated direct, hands-on experience would develop into competent professionals, combining that experience with AI fluency, much as the seniors themselves had.

What this framing does not account for is whether the conditions it assumes are actually in place where juniors are educated — the very classrooms where GenAI has become the default mode of completing coursework. S5 himself acknowledged that juniors no longer encounter the very experiences he treats as a prerequisite, observing that “the problem is that we can't wait for that time anymore.” Yet this recognition coexists with the earlier assumption; S5 does not appear to recognize that these two statements are in tension.

Juniors experienced this same reality from a position that made its consequences immediate. While seniors viewed the situation as difficult but manageable, juniors experienced it as a process in which the opportunity to make knowledge their own had been denied. J2 described this directly: “The less time you invest, the less you get back — it feels like the core is hollowed out.” J4 articulated the same experience in terms of ownership:

From the senior side, the same reality appeared in fundamentally different terms:

This divergence is not simply a matter of differing opinions or experience levels. Rather, it reflects the situated cognition tradition — the idea that one's position within a practice shapes what is perceptible and what remains invisible. Seniors, who have already crossed the developmental threshold, perceive the current situation as one they can handle with ease; juniors, on the verge of crossing it, encounter the same situation as one in which crossing has become more difficult. Each group sees what its position makes visible; what the other experiences remains, to a significant degree, outside its field of perception. And the seniors' optimism was directed not at the juniors currently navigating the transition, but at a future generation who will grow up with AI from the beginning. S2 drew the distinction explicitly: “The current juniors are a bit in-between, honestly.”

Yet in the same breath, S2 expressed confidence:

Junior participants recognized that they were not the intended beneficiaries of this optimism. As J5 observed: “We've experienced both worlds. But the ones who are doing well these days have been using it since middle and high school — I don't think they'd even notice there's a problem.” And S4, who recognized the problem with unusual clarity, captured the impasse:

“Who is going to become the next senior? This is a serious problem, and nobody has a solution.” — S4

Those in a position to act — seniors and organizations — do not fully perceive the structural conditions that juniors are navigating. Those who experience those conditions most directly — juniors — are not in a position to change them. When the perception of the problem and the experience of the problem do not coincide, the conditions for self-correction are structurally absent.

Absorption, and an already-unprotected pathway

We go one step further to argue that these mechanisms did not arise solely due to the emergence of GenAI, but rather that they have been added onto the already unprotected development pathway of juniors. Senior participants described their own developmental trajectories not as the product of formalized training, but as the incremental accumulation of hands-on work that, in retrospect, was not systematized. The conditions for development depended on individual circumstance, not institutional design.

Yet this is the same S4 who described that very work — the kind that had constituted his own developmental trajectory — as substitutable by a “100,000-won Claude subscription.” The contradiction does not present itself as one: both statements make sense within their own contexts.

The absorption proceeds through general organizational rationales — efficiency calculations, headcount planning, productivity assessments. As S4's remarks suggest, no one needs to realize that this process is effectively blocking the pathways through which they themselves grew. This lack of awareness is made structurally possible because the costs of adjustment arise at a different point in time, and for a different group, than the decision-makers. In this way, the absorption continues, with no one in the right position to recognize or protect the pathway; what was already unprotected is further eroded.

Because the erosion operates at organizational and institutional scales, individual-level interventions — such as mentorship, advice, and encouragement — cannot, on their own, overcome this asymmetry. Sustaining the conditions for junior development requires structural intervention rather than individual remediation.

Implications for institutional design

If self-correction does not occur on its own, the conditions for junior development must be re-established through institutional intervention rather than relying on individual initiative. S6's account illustrates what such intervention can look like in practice — an onboarding design her team calls “first step, one step, big step.” The first step is small and concrete: finding a missing period in a production project, putting the period in, committing it, opening a pull request, getting reviewed, and pushing it all the way to production — so the newcomer sees their own contribution become part of the final product. Subsequent steps expand into larger tasks juniors must approach with increasing self-direction, until, in a final stage, the team simply asks them to “find a problem and try to solve it.”

The key lesson from this case — that the environment for training juniors should not be left to develop naturally but must be systematically designed — has significant implications across three areas: university education, evaluation criteria for juniors, and workplace onboarding.

This isn't unprecedented. Other high-stakes fields already design against automation-driven skill decay:

In both cases, institutional design preserves developmental conditions that automation would otherwise erode. Without a comparable design, the development pathway through which the next generation of seniors might be formed will continue to attenuate.

Three principal risks of misreading

BibTeX

@misc{yu2026seniorgenerativeaierodes,
      title={Who Will Become the Next Senior? How Generative AI Erodes the Development Pathway in Software Engineering},
      author={Sumin Yu and Taesup Moon},
      year={2026},
      eprint={2607.17067},
      archivePrefix={arXiv},
      primaryClass={cs.CY},
      url={https://arxiv.org/abs/2607.17067},
}