Everything About AI
NextFin News - The economists debating artificial intelligence are arguing about productivity, inequality and growth, but the most immediate test is narrower: who gets the first rung of the economics career ladder? New evidence points in two directions at once. Stanford researchers say employment for workers ages 22 to 25 in highly AI-exposed occupations was about 19% below the path set by less-exposed peers as of June 2026, driven mainly by weaker hiring. A separate Danish study found no measurable employment or hours effect at workplaces that adopted chatbots. The contradiction is real, but it is not symmetrical: AI has not produced economy-wide displacement, yet it may already be changing who enters analytical professions.
Data cutoff: Nov. 11, 2026, using research and official labor projections published through that date.
The First Rung Is the Real Labor-Market Test
The headline question is whether AI destroys jobs. The more useful question is whether it destroys the process by which workers become experienced enough to do the jobs AI cannot yet perform reliably. Entry-level research roles are unusually exposed because they bundle repeatable tasks: cleaning data, summarizing papers, writing standard code, producing first drafts of charts and checking routine calculations. Those tasks are not the whole job, but they are a large share of the work through which a novice learns judgment.
That distinction explains why the latest evidence can look alarming without showing mass layoffs. Stanford’s Digital Economy Lab reported in August that employment for workers ages 22 to 25 in AI-exposed occupations stood about 19% below what it would have been if it had kept pace with similarly aged workers in less-exposed occupations. The researchers said the gap arose mainly through reduced hiring rather than increased separations, and that experienced workers showed no comparable gap. The result is a labor-market signal at the point of entry, not proof that firms are replacing their entire workforces.
The same Stanford work also found no evidence of widespread, economy-wide displacement. That qualifier matters. A firm can leave its senior economists in place, ask them to supervise more AI-generated analysis, and still recruit fewer junior analysts. The aggregate payroll can remain stable while the pipeline narrows.
Other evidence points the other way. Anders Humlum and Emilie Vestergaard’s National Bureau of Economic Research study of Denmark linked employer adoption surveys to administrative records and found precise null effects on earnings and recorded hours at the worker and workplace levels, ruling out effects larger than 2% two years after ChatGPT’s launch. At workplaces that encouraged chatbot use, their estimates also ruled out a differential change in the employment share of early-career workers larger than one-third of a percentage point. They found that the decline in early-career employment in exposed occupations was not driven by firms adopting chatbots.
This is not a clean refutation of the entry-level threat. It is evidence that exposure is not the same as adoption, and adoption is not the same as substitution. The first-generation impact may be showing up in hiring plans, occupational composition and the content of training rather than in layoffs or wages.
Official projections reinforce the mixed picture. The Bureau of Labor Statistics projects data-scientist employment to rise 33.5% from 2024 to 2034, adding 82,500 jobs, while computer and information research scientists are projected to grow 19.7%, or 7,900 jobs. Software-developer employment is projected to rise 15.8%, an increase of more than 267,000 jobs. At the same time, executive administrative assistants and general secretaries are each projected to decline 1.6%, and medical transcriptionists 4.9%. BLS says AI and information technology are expected to increase demand in some occupations while productivity gains limit demand in others. AI is therefore not a single labor-demand shock. It reallocates demand toward tasks that complement systems and away from tasks systems can perform more cheaply.
The key issue is whether junior research work belongs in the first category, the second, or both.
Why Productivity Gains Can Still Close the Career Ladder
The mechanism is not simply “AI makes analysts faster.” It is that AI changes the marginal value of a novice’s time before it changes the value of a senior economist’s judgment. When a senior researcher can produce a literature scan, a preliminary regression or a draft memo with a model, the firm may expand output without expanding junior headcount. The first-order effect is productivity. The second-order effect is a different staffing mix.
An NBER study of a generative-AI assistant used by 5,179 customer-support agents illustrates the tension. Access to the tool increased productivity by 14% on average, with the largest gains among novice and lower-skilled workers, while experienced and highly skilled workers saw minimal effects. The study also found higher retention and evidence that newer workers moved down the experience curve faster. But the authors explicitly said the study could not identify aggregate employment, wages or the skill composition of future hires. Firms could respond by hiring more novices because they become productive faster, redesigning jobs so fewer skills are required, or developing more capable systems that replace lower-skill work altogether.
“Our overall findings show that access to generative AI can increase the productivity and retention of individual workers,” the NBER study’s authors wrote. They cautioned that the data were not designed to reveal “changes in wages, overall labor demand, or the skill composition of workers hired for the job.”
That unresolved margin is the career-ladder problem. A junior analyst is not valuable only because of today’s output. The analyst is also an option on tomorrow’s senior capacity. Firms that optimize only for current output may underinvest in the workers who would have become tomorrow’s supervisors, domain experts and independent researchers.
Economics has a particular vulnerability because a large portion of the apprenticeship curriculum consists of tasks that can be imitated before they can be deeply understood. A model can generate plausible code without understanding identification. It can summarize a paper without knowing whether the research design survives a specification change. It can produce a chart without noticing that the denominator shifted. Senior economists remain accountable for these failures, but the training path that teaches juniors to detect them can become thinner.
This is why the distinction between complementarity and substitution matters more than the broad label “AI-exposed.” Stanford’s updated findings said employment declines were concentrated in occupations where observed AI use was more automating than complementary, while employment was flat or rising where use complemented workers, especially experienced workers. The same occupation can contain both channels. Routine coding may be substitutable; formulating a credible research question may be complementary. Data assembly may be substitutable; deciding whether a result is economically meaningful may not be.
The structural judgment follows from that mechanism. The short-run hiring cycle is cyclical: firms facing high rates or weak demand can pause entry-level recruiting, and that pause can reverse. But the task-level reallocation is structural. Once a firm has redesigned a workflow around automated first drafts, it has less reason to rebuild the old junior-heavy process when conditions improve. The technology does not need to eliminate every junior job to permanently reduce the number of junior jobs required per unit of research output.
That is the difference between a weak hiring market and a thinner ladder. The former mean-reverts with demand. The latter requires institutions to redesign training.
The Evidence Is Conflicting Because the Questions Are Different
The strongest case against an AI-driven entry-level shock is that the labor-market data do not yet show a causal displacement effect. Humlum and Vestergaard replicated declines in early-career employment in exposed occupations, but their difference-in-differences analysis found that the declines were not driven by firms adopting AI chatbots. Their results rule out a simple story in which chatbot adoption immediately removes junior workers from adopting workplaces.
Macro conditions provide a credible alternative explanation. Stanford researchers have addressed the role of interest rates and timing in updates to their work. If hiring in exposed occupations began weakening before generative AI became widely deployed, then the 19% gap cannot be attributed entirely to the technology. Sector-specific investment cycles, the post-pandemic normalization of remote work and tighter financial conditions may have hit early-career workers first because they are easiest to defer.
The counter-thesis is therefore serious: economists may be mistaking a macro hiring slowdown for automation. It attacks the central claim, not a side detail. If firms reduced junior hiring because demand weakened, the solution is not a new apprenticeship model; it is a recovery in labor demand. The Danish evidence strengthens this view because adoption did not produce a measurable workplace employment break.
But the counter-thesis does not explain away the full pattern. The Stanford result is concentrated by age and exposure, with experienced workers showing no comparable gap and the adjustment occurring mainly through hiring. That is exactly where a technology that substitutes for novice tasks should appear first. The NBER evidence also shows why neutral payroll data can coexist with changing work: wages and hours may remain stable while AI-related tasks spread, work is reorganized and adopters move into occupations where chatbots are more relevant.
A second-order implication is that the labor market may reveal the transition first through the price of experience. If firms use AI to make inexperienced workers productive faster, entry-level wages could be supported while hiring falls. If they use AI to avoid hiring novices, senior workers could capture more output and bargaining power while the supply of future senior workers contracts. Either way, a stable wage series would not settle the question.
The conventional focus on eventual productivity growth is too broad to answer the career question. The relevant expectation gap is between output per researcher and researchers per project. Even without assuming a particular productivity percentage, the mechanism is clear: if one senior worker with AI can complete the routine work previously assigned to a senior-junior pair, economic output can rise while the number of junior workers needed for a given research pipeline falls. Aggregate growth and individual opportunity can diverge.
That divergence is also why official occupational projections should not be read as a verdict for junior economists. BLS’s AI- and IT-linked occupational projections show strong growth in data science, operations research and computer-information research through 2034. Those forecasts indicate rising demand for some technical and analytical capabilities. They do not guarantee that the traditional route from assistant to independent researcher survives unchanged. New jobs can coexist with fewer apprenticeship jobs if the new positions require experience that the old system no longer supplies.
The falsifying signal for this structural thesis is specific. If, over the next two annual hiring cycles, AI-exposed occupations show a sustained recovery in 22-to-25-year-old hiring relative to less-exposed occupations, while AI-adopting firms expand junior employment rather than merely increasing output per senior worker, the case for a permanently thinner ladder weakens materially. A general rise in employment would not be enough; the test must be age-specific, exposure-specific and tied to adopting firms.
Until that signal appears, the evidence supports a narrower conclusion than “AI is taking all the jobs” and a stronger conclusion than “nothing is happening.” The first rung is under pressure even if the rest of the ladder remains standing.
What Economists and Employers Are Really Fighting About
The disagreement is partly empirical, but it is also about the unit of analysis. Erik Brynjolfsson, who led the Stanford work, has warned that continued declines in hiring among vulnerable workers could produce much larger labor-market effects if the pattern spreads. Chad Jones has emphasized the distributional stakes of an AI-led growth surge. Their concern is not that every occupation disappears; it is that the gains arrive unevenly, leaving workers without access to the experience needed to share in them.
“If unemployment rates for these workers continue to slide and if the falloff in hiring spreads to more jobs, we’re going to see much more striking effects,” Stanford economist Erik Brynjolfsson said at the 2026 SIEPR Economic Summit.
That warning is consistent with a growth model in which technology raises the return to complementary skills. It is not a forecast that mass unemployment is imminent. It is an argument that a small early signal can matter if it affects the accumulation of human capital.
The opposing emphasis is adaptation. Firms can redesign entry-level roles around verification, client context, data provenance and problem selection rather than routine production. A junior researcher who knows how to interrogate a model, reproduce a result and identify a hidden assumption may be more valuable in an AI-enabled team than a junior researcher who only performs the old manual workflow. But that redesign costs money, and the firms best placed to make it may be the largest employers with the strongest internal training systems.
The distributional risk is therefore not only between workers and capital. It is between firms that can create supervised learning environments and firms that simply cut junior tasks. It is between universities that teach statistical judgment and those that train students for a workflow whose low-level steps are increasingly automated. It is between workers who gain access to AI as a complement and workers whose first job disappears before they can learn to use it.
There is a feedback loop. If junior hiring falls, fewer workers acquire domain expertise. If domain expertise becomes scarcer, senior researchers become more valuable, but firms may respond by building more automation around them. That can further reduce the need for novices. The loop is not inevitable, but it is why the early-career margin deserves attention before aggregate employment statistics show a crisis.
Employers also face a hidden operational risk. AI-generated research can lower the cost of producing plausible analysis while increasing the cost of detecting subtle errors. Senior staff may spend less time creating outputs and more time auditing them. If the organization cuts juniors whose role included independent checking, it may save on production and lose on quality control. The short-term productivity gain can then be offset by model risk, compliance risk or reputational damage.
This is the strongest reason not to treat the junior researcher as a disposable input. The value of the role may migrate from production to verification, but the verification function still needs people who understand the underlying discipline.
Three Horizons for the Economics Profession
In the short term, sentiment and budgets will dominate. Firms under pressure can defer entry-level hiring because AI makes existing teams appear more scalable. This effect is likely to be most visible in job postings and offers, not layoffs. It is cyclical where it reflects weak demand and structural where the firm permanently changes its workflow.
In the medium term, the winners will be workers and institutions that combine domain knowledge with AI supervision. BLS projects 33.5% growth for data scientists and 21.5% for operations research analysts from 2024 to 2034, compared with 3.1% for total employment. Those figures show a rising demand for analytical work, but they also raise the skill threshold for the work that remains. The question is whether employers convert that demand into broad-based training or concentrate it among experienced hires.
In the long term, the profession faces a pipeline choice. The base case is a thinner but redesigned ladder: fewer routine assistants, more junior roles centered on validation, data architecture and research design, with senior economists supervising larger AI-enabled teams. The upside case is complementary growth: productivity lowers research costs, expands the number of projects and creates enough new work to offset task substitution. Its trigger would be evidence that AI-adopting firms are increasing both output and early-career hiring.
The downside case is a bottleneck. Firms reduce junior recruitment, universities continue supplying graduates for the old workflow, and the profession discovers several years later that it has fewer experienced researchers than it needs. Its trigger would be a widening age-specific employment gap in exposed occupations alongside stable or rising senior employment.
The practical beneficiaries of the base case are firms with strong data systems, clear audit procedures and the capacity to train workers in judgment rather than repetition. The exposed are workers whose first assignments consist mostly of tasks that can be automated and institutions that treat AI literacy as a software tutorial instead of a research discipline. No single employment number captures that asymmetry.
Policymakers and employers will need to watch three indicators together: age-specific hiring in AI-exposed occupations, employment at firms that actively adopt chatbots, and the share of research work devoted to verification rather than first-pass production. A recovery in the first measure would challenge the structural squeeze. Stable total hours with a widening junior gap would confirm that the labor market is absorbing the change through composition rather than layoffs.
AI has not yet delivered the economy-wide job destruction feared by its critics, and the Danish evidence argues against claiming that it already has. But the absence of mass displacement is not evidence that the career ladder is intact. The decisive contest is taking place before the layoff: at the moment a firm decides whether its next unit of research capacity should be a beginner, a senior worker with a model, or the model itself.
The first effect of AI on economics may not be fewer economists. It may be fewer chances to become one.
Explore more exclusive insights at nextfin.ai.
