Introduction
The first sign of AI-driven displacement will not look like a factory closing. It will look like a graduate programme cut in half. Nobody is fired in the dramatic version of the story. The traditional hiring pipeline just narrows.
That is the transition problem. AI does not need to be good enough to replace a senior worker outright. It only needs to be good enough to replace the tasks that made junior workers worth hiring while they were learning. I think it has crossed that threshold in some cognitive markets, especially software and professional services. Not everywhere. Not cleanly. But enough that the next few years are likely to be messy.
I propose the market is likely to move through the following sequence. First, frontier AI tools become just good enough to handle codified junior work: routine coding, first-draft writing, summarisation, research, testing, document review, data cleaning, financial modelling, and other tasks whose outputs a senior can check. Second, firms keep experienced workers because judgement, taste, client context, and accountability remain scarce. Third, junior hiring weakens because the short-run business case for beginners deteriorates. Fourth, the training pipeline thins. Fifth, the damage compounds across cohorts: this year's graduates compete with last year's unplaced graduates, and both respond by acquiring more credentials for a shrinking pool of first-rung jobs. The economy may still grow. The labour market may still look healthy in aggregate. But a new graduate cohort finds that the first rung has moved upward.
The usual economic warning is fair: do not commit the lump-of-labour fallacy, the idea that there is a fixed amount of work to be divided among workers. David Autor's classic answer to automation panic is still the right starting point: machines substitute for some tasks, complement others, raise output, and often increase demand for labour elsewhere. The blocked-ladder argument does not deny that. It asks a different question: what happens in the years before the new work arrives, and what happens if the new work requires experience that the old apprenticeship system has stopped producing?
This is not a prediction of permanent technological unemployment. It is a claim about sequence. AI is likely to change the skill ladder before it changes the unemployment rate.
The Threshold Has Moved
The important fact about current AI tools is not that they are perfect. They still hallucinate, lose context, make brittle architectural choices, over-edit, under-review, and mistake quantity of work for system quality. They do not have senior taste. Anyone who has used them on a large codebase knows the feeling: the model can generate a lot of correct-looking work quickly, but the human still has to decide whether the work should exist.
The important fact is that they are barely good enough. Barely is enough.
Claude Code can read files, run commands, make changes, use test suites, and handle the workflow from issue to pull request. Anthropic describes Claude Sonnet 4.5 as capable of 30-plus hours of autonomous coding in enterprise settings, with improved testing and long-running agentic work. Karpathy's autoresearch follows the same shift: an agent, under broad guidance from an expert, edits a real training file, runs a five-minute experiment, keeps the change if validation improves, reverts it if not, and loops overnight. The system is narrow. It is not an independent scientist. But it performs the apprentice loop: try a (given) plausible idea, test it, keep it, discard it, repeat.
That is why software is the warning case. Current agents are still unreliable for open-ended engineering judgement, but they are already strong at many of the tasks juniors used to do while becoming engineers. They can write the test, sketch the component, find the API call, refactor the obvious duplication, draft the migration, and keep going while a human works on something else. A senior engineer can reject bad work faster than a junior can produce good work from scratch. Once that becomes true often enough, the junior role has to be justified as training, not just production.
The same pattern appears outside software. Junior law, consulting, finance, accounting, research support, and many administrative roles contain a large amount of codified, reviewable work. The model does not need to be a partner, manager, or senior analyst. It needs to be a cheap, tireless first-pass worker whose output can be checked by someone with judgement.
This is the uncomfortable middle. AI is not good enough to remove the senior. It is increasingly good enough to remove the firm's immediate need for a junior.
The Evidence
AI has not yet produced broad labour-market collapse. The best frontier tools have only recently become capable enough to sit inside real workflows for hours at a time, and aggregate data are too blunt to catch a missing graduate vacancy. A 2026 NBER survey of almost 6,000 executives in the United States, United Kingdom, Germany, and Australia found that around 70% of firms actively use AI, but over 80% reported no effect on employment or productivity over the previous three years. A Danish administrative study found precise null effects on earnings and recorded hours two years after ChatGPT adoption, even for intensive users and early-career jobs. Those findings should kill the lazy version of the argument. They do not kill the blocked-ladder version.
The blocked-ladder thesis predicts a specific pattern: young workers in exposed occupations should be hit before older workers in the same occupations. On that margin, the evidence is sharper.
The useful question is not “are firms firing everyone?” It is “where is the first margin moving?” Brynjolfsson, Chandar, and Chen answer that with ADP payroll data. Early-career workers aged 22 to 25 in highly AI-exposed occupations suffered a relative employment decline, while more experienced workers in the same occupations were stable or grew. The losses are concentrated where AI appears to automate tasks rather than augment workers. That is not a generic AI panic result. It is the blocked ladder in data form: the exposed young are hit before the exposed old.
The same pressure is visible in hiring expectations. Entry-level workers are not simply being asked to do old junior jobs. They are being asked to arrive already fluent with the tools that are eating parts of those jobs. In the United States, NACE's 2026 spring update found that more than one-third of entry-level jobs now require AI skills, nearly triple the share from fall 2025. Nearly 60% of employers assign interns projects involving AI tools. Only 11% report discussing AI as a way to replace some positions, which is a real limit on the replacement claim. But more than a quarter say AI has already reduced the need for tasks performed by entry-level workers. That is task compression. The bottom of the job is being redesigned before the top is.
Industry reports describe reduced entry-level recruiting in the sectors one would expect. Law firms remain profitable but have reduced summer-associate and junior hiring. Professional services and finance show the same barbell pressure: fewer routine junior roles, more demand for experienced workers who can own the client, the risk, or the final judgement. These reports do not prove AI causality by themselves. Demand cycles, interest rates, offshoring, and the post-Covid correction matter. But they line up with the mechanism.
The strongest version of the claim is therefore this: AI is already changing the economics of entry-level cognitive work, and it is likely to continue doing so as models improve. The effect will be uneven. It will be hidden by macro noise. But it will show up first in the places where junior work is codified, digital, and reviewable.
Why Firms Will Undertrain
The obvious objection is also the best one. If AI blocks junior hiring now, won't firms eventually run out of senior workers? And if seniors become scarce, won't wages rise until firms start training again?
Some of that will happen. The market is not stupid. But it may happen too slowly, and the reason is that apprenticeship is not a normal input.
A junior role has always been half job, half training mechanism. The junior produces work, but the firm also pays for mistakes, feedback, supervision, and ramp time. That bargain works when the junior's output is useful enough to offset part of the training cost. AI weakens that bargain. If a model can generate the first draft, the boilerplate code, the diligence summary, or the reconciliation faster and cheaper, the junior loses the production half of the role before losing the developmental need for it.
The social value of apprenticeship remains. The private value falls.
That wedge is the whole problem. Firms do not capture all the value of training because workers can leave and competitors can poach. This is the standard human-capital problem made worse by AI: the training cost is paid by one firm, while the experienced worker may later be bought by another. When the junior also contributes less immediate production, undertraining becomes the rational private choice.
Senior attention is the binding resource. A senior can review a junior's work, or review five AI-generated drafts, or use the model directly to produce something near-final. The opportunity cost of teaching rises. This is why “firms need future seniors” is not enough. They may need them in the abstract and still choose not to train enough of them this quarter.
The pain will not be evenly distributed. When firms hire fewer juniors, they do not usually spread the reduction evenly across every university. They retreat to the safest signals. Cambridge, Oxford, Harvard, MIT, Stanford, and similar institutions will still place students because employers trust the brand, alumni networks are dense, and many students arrive with family capital behind them. The squeeze falls hardest on the next tiers down. In a normal market, a firm might recruit widely: elite universities, strong regional universities, conversion courses, less obvious candidates. In a narrow market, it buys certainty. The ladder remains visible from the top and disappears from the middle.
That matters because the students most dependent on the ladder are often the least protected from its loss. A wealthy graduate can wait, but they won't need to. They'll be hired. A graduate without family support, at a less recognised university, cannot turn a two-year hiring freeze into a lifestyle choice. They need the broad entry-level market to work.
This is not just unfair. It is inefficient. The economy will still need people with judgement. It will still need people who understand clients, systems, institutions, edge cases, and taste. But judgement is not downloaded. It is trained through exposure. If the first exposure disappears, we should expect a shortage later.
The Messy Middle
The clean stories are wrong in both directions. The optimistic story says productivity rises, new work appears, and workers move. The catastrophic story says AI takes the jobs and the game ends. The more likely path is uglier: both are true for a while. That is what I mean by the messy middle: the period between today's labour market and any AI-enabled abundance people might reasonably hope for.
AI will continue to improve. The frontier models of 2026 are not the frontier models of 2024 with better marketing. They run longer, use tools better, write more code, test more of their own work, and operate inside real workflows. Their judgement is still uneven, but the error profile is improving. The bad design choices are less cartoonish. The hallucinated API calls are less constant. The need for human review remains, but the amount a reviewer can supervise is rising.
That creates a period in which output expands before opportunity does. A startup can do more with four people. A law firm can produce more drafts per partner. A research team can run more experiments. A senior developer can ship more code. GDP can rise. Profit margins can rise. The stock market can celebrate. At the same time, the person trying to become a senior developer cannot get the first job.
The labour-market damage then stacks by cohort. If the class of 2026 cannot find graduate roles, they do not vanish before the class of 2027 arrives. They remain in the queue, often with a year of underemployment, applications, unpaid projects, or a master's degree added to the CV. The next cohort competes with them, and the cohort after that competes with both. A weak entry-level market therefore creates its own backlog. It also creates credential inflation. More graduates apply to postgraduate programmes because study is the respectable waiting room in a bad labour market. The degree that once opened the door becomes the degree required merely to stand near it.
The instability risk is not mainly absolute poverty. It is frustrated expectation. Gurr's relative-deprivation theory is useful here: grievance grows when there is a gap between what people believe they should be able to get and what they can actually get. Bartusevicius and van Leeuwen sharpen the point: poor individual prospects are a stronger predictor of support for political violence than inequality by itself. The blocked ladder creates exactly that psychology. It tells a cohort: the economy is richer, the technology is miraculous, but your path is closed.
That is a bad bargain to offer people.
What Policy Should Do
Policy should not try to freeze old jobs in place. It should preserve the developmental function of work and spread the cost of transition. I would build the response around three instruments.
First, measure the ladder and tell the public what is happening. Governments should create an AI transition observatory that tracks early-career employment by age, occupation, region, firm size, and exposure. The key indicators are graduate vacancies, internship conversion, backfills, wage progression, underemployment, postgraduate deferral, and movement into unrelated work. Waiting for unemployment is too slow. The signal is cohort intake. The body should also publish a public ledger of AI's effects: the benefits, the harms, the distributional consequences, and the remaining uncertainty. It should not be an industry propaganda arm. But if public discourse becomes only resentment, societies may throw away real gains in medicine, science, accessibility, education, and public services. The model is closer to an independent public broadcaster or a serious policy institute than to a ministry press office: trusted enough to say both “this helped” and “this hurt.”
Second, subsidise apprenticeship directly. The best policy in this paper is an entry-level training credit. Firms should receive a time-limited subsidy for hiring new entrants into roles with real supervision, training content, wage floors, and progression standards. The subsidy should require net additional headcount in the relevant job family, retention gates, and clawbacks for churn. This is not corporate charity. It pays for the social value of training that firms will underprovide when AI lowers the immediate production value of beginners. It also gives large employers permission to experiment. Nobody knows exactly what the AI-era junior role should be. The credit reduces the cost of finding out.
The new roles may not look like the old ones. I predict them to be more human and organisational: running events, coordinating clients, building communities, handling messy in-person operations, maintaining relationships, gathering requirements, testing products with users, and translating between technical systems and real people. Models will get better at coordination, but the physical and social world is still harder to automate than the production of text, code, and analysis. The policy should not guess the right job titles. It should pay firms to discover roles that genuinely train people.
Third, make layoffs and payroll compression less fiscally free. This should not be a general tax on companies for using AI, and it should not apply merely because a firm operates in an AI-heavy industry. The trigger should be behavioural. If a large firm reduces covered labour cost in a job family while output, revenue, or margins hold up, it pays a payroll-gap levy into a transition fund unless it can show genuine contraction or another exempt cause. The point is broad and simple: if a firm converts labour cost into automation surplus, it should help fund the workers and cohorts pushed into transition.
The levy should be assessed on corporate groups, not shell entities; include contractors and off-payroll labour where the firm still controls the function; and be capped in duration so it does not become a permanent tax on productivity. It should be piloted in high-exposure sectors before expansion. The design is intentionally imperfect but administrable: it does not require the state to prove that a specific model replaced a specific person. It observes payroll compression where the business is still producing, then asks the firm either to contribute or explain.
The fund should pay for three things. First, transition allowance before reemployment, administered through existing benefit infrastructure where possible. Second, wage insurance after reemployment where the new job pays less. Third, entry-level hiring credits of the kind described above. This borrows from trade-adjustment assistance and pro-worker AI ideas, but the trigger is domestic labour displacement in high-exposure functions rather than import competition. It will not be perfectly self-funding. No serious insurance scheme is. Firms fail, restructure, and avoid rules. The right standard is not mathematical perfection; it is whether the scheme makes displacement less fiscally free than it is now while making junior hiring cheaper than it would otherwise be.
The transition body should also provide public AI access. If AI becomes a basic tool for job search, learning, and small-business formation, access cannot depend entirely on private subscriptions.
Conclusion
The blocked ladder is a better warning sign than mass unemployment. It is earlier, quieter, and more specific. It also fits the technology we actually have: AI that is not yet a senior professional, but is already strong enough to consume much of the work by which juniors used to become one.
I expect the next phase to be messy. Some new roles will appear. Some firms will over-cut and rehire. Some sectors will discover that AI raises demand enough to expand junior work again. But in the exposed cognitive markets, the default path is likely to be thinner intake, higher expectations for beginners, and more value captured by experienced workers and capital.
The answer is not to stop the technology. The answer is to keep the ladder. Measure it, subsidise it, and make the firms that profit from pulling labour out of production help pay for the transition. If the economy eventually creates better work, good. But people cannot wait ten years for a first rung.