aggregate employment holds up, first jobs don't

· originally published on LinkedIn

this is a translation of the spanish original · read the original

in april 2026, openai published a paper authored by economist alex martin richmond, with a foreword by the company's chief economist, ronnie chatterji. the document is called the ai jobs transition framework: mapping ai's near-term impact on jobs. its main thesis is that traditional measures of "ai exposure" are too blunt to predict which jobs will disappear in the short term. it proposes a framework with three additional dimensions (human necessity, demand elasticity, real chatgpt usage) and concludes that only 18% of jobs in the united states are at high risk of automation, against 46% that will see "little immediate change" [1].

table 1 of the paper is the strongest part. it compares unemployment rates by occupational archetype between the first quarter of 2024 and the first quarter of 2026. jobs classified as "high automation risk" rose 0.3 percentage points. jobs "with less immediate change" rose 0.6 points. the jobs least exposed to ai took double the impact of the most exposed. the paper's own conclusion:

"this is why, while ai may be linked to some changes in employment, it remains difficult to clearly link ai and the aggregate labor market."

that is the data point openai chose to put in a table and publish.

eight months earlier, in august 2025, another group of researchers published data that tell a different story. erik brynjolfsson, bharat chandar and ruyu chen, of the stanford digital economy lab, released canaries in the coal mine? six facts about the recent employment effects of artificial intelligence [2]. the paper uses administrative records from adp, the largest payroll company in the united states, with data through july 2025. what they found:

"early-career workers (ages 22-25) in the most ai-exposed occupations have experienced a 13 percent relative decline in employment even after controlling for firm-level shocks."

for software developers aged 22 to 25, the drop from the late-2022 peak was almost 20%. for workers aged 35 to 49 in the same occupations, there was no drop. in some cases, there was growth.

the contrast is direct. the openai paper says ai still doesn't show up in the aggregate labor market. the stanford paper says it shows up very clearly, but only in a very specific age cohort. both can be true at the same time. and that is where uruguay has a problem that no local media outlet is looking at through that lens.

why the two papers don't contradict each other (even though they seem to)

the openai paper measures averages. it groups more than 900 occupations into four archetypes, crosses those archetypes with chatgpt usage data, and looks at how total unemployment evolved by occupational cohort. the unit of analysis is the occupation, not the person. when openai says high-automation-risk jobs rose 0.3 pp in unemployment, it is averaging seniors and juniors within the same occupation.

the stanford paper does exactly the opposite. it holds occupations constant and disaggregates by age. within the same exposed occupation, seniors grow and juniors fall. when you average them, the effect gets diluted.

the generational effect is invisible in openai's data by methodological design. not because it doesn't exist, but because the chosen method doesn't look for it. and openai chose that method just when stanford's micro data had already been circulating for eight months.

the pattern is not stanford's alone. in september 2025, seyed hosseini and guy lichtinger, of harvard, published generative ai as seniority-biased technological change [3]. they used résumé and job posting data from 62 million workers in 285,000 american companies between 2015 and 2025. the central finding:

"following adoption, junior employment declines sharply in adopting firms relative to non-adopters, while senior employment remains largely unchanged. the junior decline is concentrated in occupations most exposed to genai and is driven by slower hiring rather than increased separations or promotions."

the specific number: -7.7% in junior employment at ai-adopting firms, compared with non-adopters, over six quarters from the first quarter of 2023. seniors don't move. and the drop is not from layoffs, it is from non-hiring.

bouke klein teeselink replicated the pattern in the united kingdom. anders humlum and emilie vestergaard, in denmark, found no aggregate effects on wages or hours, but they did find intense restructuring of tasks within the same companies [4]. what humlum and vestergaard call "still waters, rapid currents": on the surface nothing shows, underneath the water is moving hard.

four independent studies, four different methodologies, four countries, one converging conclusion: ai is affecting the age composition of work, not the aggregate volume. that is what the openai paper chooses not to see.

the "apprentice margin": what is breaking

brynjolfsson, chandar and chen gave the phenomenon a name: apprentice margin. the space where young workers historically accumulated practical experience and tacit knowledge inside companies. what happens when ai automates specifically the tasks that were given to juniors so they could learn.

the argument is old in physical trades. the apprentice electrician adds no net economic value in his first year, but the company carries him because it knows that in five years he will be a fully formed electrician. if the first-year tasks can be done with a robot, the company doesn't carry the apprentice. but the robot doesn't learn to be an electrician either. the result, a decade later, is that there are no new electricians.

the same reasoning applies to lawyers, accountants, programmers, journalists, consultants. all the professions where tacit knowledge is transmitted by doing, not by reading. all the professions where the uruguayan it sector concentrates employment.

in the uruguayan public debate this point hasn't entered yet. the stanford paper names it. the harvard one confirms it with compositional data. the conversation in uruguay remains split between "ai didn't break aggregate employment" (openai, advice, cuti at certain moments) and "ai is going to destroy all jobs" (press headlines, generic predictions). neither narrative captures what is actually happening.

the uruguayan case

the uruguayan it sector closed 2024 with 20,464 direct jobs, revenue of us$3.681 billion, and exports that in 2025 reached us$1.503 billion, completing five consecutive years of growth [5][6]. the aggregate numbers are good. cuti's annual survey, presented in december 2025, says so explicitly:

"in terms of employment, the sector maintained a stable trend, with close to 20,500 people employed, a slight increase from the previous year."

that is exactly the kind of average the openai paper looks at and the stanford paper doesn't. cuti's aggregate figure doesn't disaggregate by hiring age, doesn't separate juniors from seniors, doesn't measure the flow of first jobs. it measures the total stock. and the stock is stable.

the problem is what is falling apart beneath that average. in march 2026, the monitor of the consulting firm advice reported that uruguay has "the lowest competitiveness values since the year 2000" in the global services sector. the median salary in dollars for sector positions rose 34% since 2022 [7]. cuti is negotiating with the executive branch a joint working space on competitiveness, exports and qualified employment. cuti's president, amilcar perea, explained it in february 2026:

"until 2022 the rise in the exchange rate accompanied, to some extent, the increase in costs in pesos, mainly salaries. since then both variables began to decouple: costs in local currency kept growing while the value of the dollar lost ground." [8]

in may 2026, búsqueda published a piece by iara zinno on how ai is redefining junior work and questioning the "man-hours" model [9]. aníbal gonda, cuti's vice president of talent, told the magazine:

"before, you sold a senior consultant with 10 juniors or 10 more consultants. today you do it with two or three senior consultants with artificial intelligence tools to complement them."

and:

"companies no longer hire man-hours, what they hire is value, solutions."

guzmán sarachaga, of advice, added:

"many tasks that used to justify entry-level roles can today be assisted or, in some cases, not replaced, but very much accompanied by an artificial intelligence tool."

this is the same observation stanford and harvard make with administrative data, but told from the side of the uruguayan sector's sellers. the pyramid model of a senior + many juniors doing operational tasks is collapsing. not because there is less aggregate demand, but because ai does what juniors did and clients no longer want to pay for hours that produce less unit value.

no uruguayan data, to this day, measures the age composition of new hiring in the it sector. the ine measures aggregate unemployment. cuti measures total employment. cuti's it observatory has data on academic training from the mec [10] but not on first-job flows. nobody is looking at the data that matters.

three pillars collapsing in parallel

in november 2025, ceibal announced the end of jóvenes a programar (jap), its flagship technology training program for people over 18. it had sustained it for nine years with support from cuti, bid-lab and sector companies. the official statement:

"after almost a decade of work and more than 6,000 young people trained across the country, ceibal is ending its pioneering program jóvenes a programar (jap). building on the experience accumulated in this program, the technological university (utec) will develop a new technology training proposal, called pixel." [11]

the transition was announced. the details were not.

uruguay has an it sector facing three simultaneous structural blows. the first is the loss of exchange-rate competitiveness, already acknowledged by cuti and the government. the second is the silent neutralization of the tax incentives of free trade zones and similar regimes by the oecd's pillar two, which entered into force gradually and takes away from uruguay a good part of the regulatory arbitrage that sustained foreign services investment. the third is exactly the apprentice margin one: if companies stop hiring juniors because ai does junior work, the pipeline that produces seniors breaks. and uruguay has a pipeline that depends almost entirely on the first formal job in the private sector.

none of the three pillars is falling because of ai. the currency appreciation is domestic macro. pillar two is multilateral fiscal geopolitics. ai is the catalyst of the third pillar, but that pillar had been weakening for years due to public underinvestment in training beyond the formal education cycle.

the problem is not that uruguay is going to have a massive ai-driven unemployment crisis in 2026 or 2027. the openai paper is right that this doesn't show in the aggregates. the problem is that in 2030, 2032, 2035, when today's seniors start leaving the system (through retirement, migration, moves to other sectors), there won't be enough new seniors to replace them. because the juniors who were going to become those seniors never entered the system. that cost doesn't appear in any quarterly. it appears in ten years, when it can no longer be reversed.

the apprenticeship severance, as a recent substack essay circulating among economists called it [12], is hard to defend politically because its costs are invisible within the horizon of any organization governed by quarters. the company that cuts juniors in 2026 shows better margins in 2027. the cost, in missing expertise, arrives in 2040.

what is not being measured

the uruguayan public debate on ai and employment is being held with borrowed frameworks. when cuti cites studies, it cites the big american papers. when local media report, they cite openai, stanford, mckinsey. no uruguayan data disaggregates:

  • age composition of new hires in the it sector
  • duration of the first job in the it sector
  • conversion rate from jap/utec/udelar to a first formal job
  • evolution of the junior/senior ratio in software-exporting companies
  • difference in employment growth between genai-adopting and non-adopting companies

all that data exists in some system. the bps has contribution data by age and industry. the dgi has revenue data by company. cuti has data on member companies. none of it is cross-referenced. none of it is published through that lens.

meanwhile, public discourse polarizes between two extremes that don't describe reality. on the optimistic side, cuti says that "when there are technological advances, in the long run what ends up being generated is more work. different work, but more work." it is a defensible historical observation but it doesn't answer the specific question of what happens to uruguayan juniors in the next five years. on the pessimistic side, headlines talk about "ai destroys jobs" as if it were a homogeneous event. that doesn't describe anything useful either.

the empirical reality, according to four independent studies in four different markets, is that ai is redistributing work within occupations: toward seniors, away from juniors. that pattern is exactly the one the structure of the uruguayan it sector cannot absorb without structural damage in the medium term.

conclusions

the openai paper is a good technical document. it is also a document published by a company with a direct interest in keeping the public conversation about ai and employment from becoming more restrictive. its table 1 is built with a methodology that dilutes exactly the effect that stanford, harvard, klein teeselink and cuti's qualitative observation converge in identifying. that doesn't mean the openai paper is badly done. it means the methodological cut it chose has political consequences, and it is worth reading it with that lens on.

the stanford paper isn't the final diagnosis either. the adp data has its own biases, the 22-25 cohort includes dynamics not attributable solely to ai (post-pandemic labor market, interest rate adjustment, etc.). but the pattern repeats across four studies and four different methodologies. that is convergent robustness, not anecdote.

uruguay has an it sector that historically worked as the gateway to qualified employment for a very specific cohort: young people between 22 and 30, trained partially or fully in the public system (udelar, utec, jap, technical schools), placed in services-exporting companies. that gateway is narrowing for three simultaneous reasons, one of which is exactly what the american papers are measuring and the uruguayan data is not disaggregating.

the local public debate should be discussing how to measure this, not whether ai "destroys" jobs. the operational question is simpler: how many first jobs in the it sector were generated in uruguay in each of the last eight quarters, how did the median age of new hires evolve, and what percentage of software-exporting companies hired juniors in 2025?

none of those three questions has a publicly available answer. that is the real problem. not ai. the absence of data.

sources

[1] richmond, alex martin. the ai jobs transition framework: mapping ai's near-term impact on jobs. openai economic research, april 2026. available at https://cdn.openai.com/pdf/ai-jobs-transition-framework.pdf

[2] brynjolfsson, erik; chandar, bharat; chen, ruyu. canaries in the coal mine? six facts about the recent employment effects of artificial intelligence. stanford digital economy lab, august 2025 (revised version november 2025). https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/

[3] hosseini, seyed mahdi; lichtinger, guy. generative ai as seniority-biased technological change: evidence from u.s. résumé and job posting data. harvard university working paper, ssrn id 5425555, august 2025. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555

[4] humlum, anders; vestergaard, emilie. still waters, rapid currents: early labor market transformation under generative ai. nber working paper w33777, may 2025 (previously circulated as "large language models, small labor market effects"). https://www.nber.org/papers/w33777

[5] la industria tecnológica factura más de u$s 3.600 millones y reafirma su rol estratégico a nivel país. cuti annual survey 2024, presented december 2025. https://www.improfit.com.uy/post/uruguay-digital-la-industria-tecnológica-factura-más-de-u-s-3-600-millones-y-reafirma-su-rol-estrat

[6] conly. uruguay exportó usd 1.503 millones en servicios de software en 2025. april 2026. https://conlyapp.com/blog/exportaciones-servicios-uruguay-2025-pymes-tech

[7] servicios globales en jaque: uruguay registra su menor nivel de competitividad en 26 años. ámbito, march 2026. https://www.ambito.com/uruguay/servicios-globales-jaque-registra-su-menor-nivel-competitividad-26-anos-n6259148

[8] competitividad, exportaciones y empleo, el gobierno pone el foco en un sufrido sector tecnológico. ámbito, february 2026. https://www.ambito.com/uruguay/competitividad-exportaciones-y-empleo-el-gobierno-pone-el-foco-un-sufrido-sector-tecnologico-n6249521

[9] zinno, iara. la inteligencia artificial redefine el trabajo "junior" y cuestiona el modelo de "horas hombre". búsqueda, may 14, 2026. https://www.busqueda.com.uy/economia/la-inteligencia-artificial-redefine-el-trabajo-junior-y-cuestiona-el-modelo-horas-hombre-n5412960

[10] cuti. monitor laboral ti de uruguay. it observatory based on the mec's statistical yearbook of education. https://cuti.org.uy/wp-content/uploads/2021/02/monitor-laboral-ti-de-uruguay.pdf

[11] ceibal. jóvenes a programar de ceibal finaliza e inspira la creación del programa pixel de utec. november 2025. https://ceibal.edu.uy/institucional/articulos/jovenes-a-programar-de-ceibal-finaliza/

[12] the apprenticeship severance: how ai is breaking the expertise pipeline. smarter articles, may 2026. https://smarterarticles.co.uk/the-apprenticeship-severance-how-ai-is-breaking-the-expertise-pipeline

additional studies referenced in the body of the article:

← all writing