A junior consultant equipped with an AI assistant produces today work that only a senior would have delivered five years ago. It’s documented, measured, and presented as a victory. What the same studies note in a footnote: that senior never existed. They didn’t have time to develop.
The debate on AI and employment has long been organized around a simple question: which professions will disappear? The answer, until now, has been reassuring for white-collar workers: machines automate repetitive tasks, humans keep the rest. This interpretation is now incomplete. A series of recent works, including an article available on arXiv since August 27, 2025 (“Training for Obsolescence?”, arXiv:2508.19625), shifts the problem. The question is no longer “which professions” but “at what point in a career.” And the answer is uncomfortable: AI improves beginners, leaves seniors indifferent, and precisely destroys the intermediate positions through which the former became the latter.
The Essential Points
- Across the ten largest British and German cities between 1991 and 2021, intermediate-level positions declined markedly, while the extremes of the labor market – managerial and highly skilled jobs on one side, low-skilled jobs on the other – progressed or held steady.
- The “barbell effect” documented by Acemoglu and Loebbing (2022) and confirmed by Brynjolfsson et al. (2023) shows that this polarization is not a narrative but a measurement: European urban labor markets are emptying from the middle.
- Generative AI improves junior performance on analytical and writing tasks but offers little to experienced profiles – whose gains come less from tools than from accumulated judgment.
- The real risk is not short-term mass unemployment, but the drying up of the flow through which an organization renews its expertise: without intermediate learning, companies end up with a summit without a body.
- The lever for response is organizational: make tacit knowledge explicit, design skill development pathways that no longer rely on passive observation of tasks that AI absorbs.
The Middle That Has Been Hollowing Out for Thirty Years
To understand what AI is doing to the European labor market, we must first look at what previous technologies have already done. Data from the ten largest British and German cities over the period 1991-2021 – analyzed by Oesch et al. for the Institute for Employment Research (IAB) – tells a coherent story. Managerial and professional jobs have grown substantially. Low-skilled jobs have held steady or grown modestly. Intermediate positions – secretaries, accountants, technicians, administrative officers – have declined markedly. This is not an anecdote: it is the structural geography of the European urban labor market over three decades.
Acemoglu and Loebbing (2022) gave a name to this phenomenon: the “barbell effect.” The middle disappears, the extremes endure. Brynjolfsson et al. (2023) clarified the dynamics by showing that productivity gains from generative AI concentrate primarily among less experienced workers – with minimal impact on senior profiles. These intermediate positions were not merely boxes on an organizational chart. They were the rungs of a progression: the place where a junior became someone useful at the next level.
Automation in the 2000s and 2010s had already attacked these positions. Generative AI accelerates the movement, but by introducing a new variable: it now also targets the cognitive tasks of these levels, not just procedural tasks. Writing a first draft, synthesizing a document, producing a preliminary analysis: these are precisely the exercises through which a young graduate learned to think within a professional framework.
What AI Actually Does to Junior Learning
The article “Training for Obsolescence?” (arXiv:2508.19625, submitted August 27, 2025) poses the question directly: if AI improves the performance of beginners, what remains for them to learn? The answer is more nuanced than the title suggests, but it is disturbing.
Experimental studies – conducted notably by NBER teams with junior lawyers, consultants, and developers – show that generative AI compresses the time needed to produce an acceptable deliverable. A junior equipped rivals a non-equipped intermediate profile on the final result. This is real. But what these studies measure is the quality of the deliverable, not understanding of what underlies it.
The difference is massive. When a junior spent three years writing memos under supervision, reformulating analyses, having their output corrected and annotated by a senior, they were not just producing documents: they were internalizing a way of structuring a problem, prioritizing what matters, recognizing a reasoning error before it becomes a judgment error. This tacit learning – “tacit knowledge” in economic literature – is difficult to describe but easy to spot when it is absent. A profile with five years of real experience is not a profile with two years of good tools. They have gone through situations where tools were insufficient.
AI does not eliminate this need for learning. It eliminates the exercises that made it possible. The junior produces faster, but trains less. This is not a hypothesis: it is the mechanical consequence of the fact that low-level tasks, once assigned to train employees, are now delegated to a model.
The Broken Trajectory: From Junior to Senior Without the Middle
The central problem is not that juniors work differently. It is that the progression toward expertise follows a path that passes through stages, and some of these stages are being erased.
We can formalize this simply. A cognitive career follows three phases. First, the exposure phase: one executes simple tasks, observes how they fit into a larger system, makes mistakes and learns to recognize them. Next, the integration phase: one takes charge of entire projects, manages uncertainty, arbitrates between conflicting constraints. Finally, the judgment phase: one decides where data does not. This is the value of the senior.
Generative AI is excellent at compressing the exposure phase. It is marginally useful in the integration phase. It is little or not useful in the judgment phase – which NBER studies confirm by showing that performance gains decrease with experience. A senior using AI does not become noticeably more productive than a senior without AI on tasks that define their value. However, a junior using AI to skip the exposure phase never enters the integration phase with the necessary cognitive resources.
This is the paradox of the broken trajectory: the tool that accelerates the start prevents arrival. And it is not an individual problem. It is an organizational problem. Companies that today train their juniors via AI will find themselves in five years with profiles brilliant on short deliverables and helpless when facing complex decisions. The summit without a body.
This phenomenon resonates with what we know about remote work and informal transmission dynamics: work environments that eliminate friction often eliminate with it the channels of unplanned learning.
What Companies Are Doing – and What They Could Do
Organizations are not passive. Some have identified the problem and are beginning to address it, with varying degrees of ambition.
The most common response is enhanced mentoring: explicitly designate seniors as responsible for junior progression, independent of deliverables. This is better than nothing. It is insufficient if juniors no longer have concrete tasks to work through together with their mentor.
A more structured response consists in making explicit what was tacit. Consulting firms like McKinsey and BCG have begun decomposing judgment skills into transmissible protocols: how does one structure an ambiguous problem? What signals indicate that an analysis is incomplete? How do you recognize when a client is not asking the right question? These competencies, once learned by osmosis and by error, become formal training modules. It is not perfect – osmosis had a richness that the module does not fully restore – but it is a real adaptation.
An increasing number of European companies have introduced mandatory rotations on high-uncertainty projects for junior profiles: assignments where AI is deliberately underutilized, not because it would be forbidden, but because the problem is too poorly defined for it to help. This is a way to maintain the conditions for intermediate learning within an environment that makes them mechanically disappear.
The fundamental question is that of organizational design. Who is responsible for developing the skills of a junior profile in an environment where training tasks are outsourced to a model? This responsibility existed before by default: the natural flow of work produced it. It must now be designed, explicitly assigned, and measured. Organizations that have not formalized it within three to five years will be in difficulty at the precise moment when they most need profiles capable of autonomous judgment.
Polarization Is Not Inevitable
The barbell effect has been documented since the 1990s. It has survived several technological cycles. This does not mean it is inevitable – it means it is structural if nothing is done to change it.
Public policies can slow or modify the dynamic. Continuing education is the most often cited lever, but it is too vague to be actionable as such. What works is training targeted at judgment skills and uncertainty management – precisely what AI cannot transmit. Germany, through its dual system of vocational training, has infrastructure that allows this type of adaptation: apprenticeship structurally integrates learning through practice in real and varied contexts. The question is whether this model can be extended to intermediate-level cognitive professions, not just technical trades.
In the United Kingdom, debate on AI-assisted learning in universities touches on a connected question: if students delegate writing and analysis exercises to models starting from undergraduate level, they arrive on the labor market with diplomas but without the cognitive muscles these exercises were supposed to develop. Several British universities – Oxford and Imperial College among others – are experimenting with assessment formats that constrain in-person exercise precisely to recreate the conditions for cognitive learning. This is a partial response, but it is a response.
The debate between Acemoglu, who worries about technology gains being captured by capital at the expense of labor, and Brynjolfsson, more optimistic about long-term human-machine complementarity, remains open. The same data can lead to opposite conclusions depending on the analytical framework used – which is precisely the issue here: neither the catastrophism that predicts massive disappearance of cognitive jobs nor the blissful optimism that sees in AI a mere productivity multiplier accounts for what the data shows about individual trajectories.
What Individuals Can Anticipate
Organizations will adapt – or will not adapt – their training practices. Individuals early in their careers do not have to wait. A few guidelines emerge clearly from the literature on learning and cognition.
The first: use AI as a verifier, not a producer. Write first, compare afterward. The exercise of production, even imperfect, is what builds the muscle. The gap between one’s own version and the model’s is a source of learning; the model alone is not.
The second: deliberately seek situations where the tool is insufficient. Ambiguous projects, difficult clients, contradictory constraints are uncomfortable and formative. Avoiding them because AI handles simple cases well amounts to training only on easy exercises.
The third: invest in mentoring relationships. Seniors who have gone through the integration and judgment phases carry knowledge that models do not reproduce. Access to this knowledge, in an environment that makes it less visible by default, becomes a scarce resource.
The question that will remain open in the coming years is that of the market itself: if judgment skills become rarer because fewer people have been able to acquire them, will organizations that have invested in their internal transmission gain a decisive advantage over those that have not? The data on polarization suggests yes. But the advantage is visible only when the need is felt – that is, too late for organizations that waited.
Sources
- “Training for Obsolescence? The AI-Driven Education Trap” (arXiv:2508.19625, submitted August 27, 2025): https://arxiv.org/abs/2508.19625
- Acemoglu & Loebbing, “Automation and New Tasks: How Technology Displaces and Reinstates Labor” (2022), NBER: https://www.nber.org/papers/w25684
- Brynjolfsson et al., research on human-AI complementarity and productivity gains distribution (2023), NBER
- Oesch et al., studies on polarization of European urban labor markets, Institute for Employment Research (IAB)
- Oesch, Morris & Westenberger (2025), “Polarised upgrading in UK & German cities”: https://pmc.ncbi.nlm.nih.gov/articles/PMC12718246/
- Acemoglu & Loebbing (2022), “Automation and Polarization”, NBER WP 30528: https://www.nber.org/system/files/working_papers/w30528/w30528.pdf
- Brynjolfsson, Li & Raymond (2023), “Generative AI at Work”, NBER WP 31161: https://www.nber.org/papers/w31161
- BCG (2026), “When Everyone Uses AI, Companies Risk Losing Critical Skills”: https://www.bcg.com/publications/2026/when-everyone-uses-ai-companies-risk-critical-skills