Data current through June 2026. The 15 load-bearing claims in this essay (key employment figures, RCT effect sizes, reversal findings) were each adversarially checked by 3 independent verifiers (downloading primary sources, checking numbers word-for-word, searching for revisions and counter-evidence): 13 passed as stated, 2 were corrected per the verifiers (BLS "never declined" softened to "dipped slightly, then hit a record"; NY Fed figures aligned to a single data vintage), 0 were refuted. Citations that did not go through verification carry source-grade annotations inline; numbers from vendors and interested parties are flagged as such.
The 2025-2026 evidence on AI and junior engineers contains a surface contradiction, and it is more informative than any one-sided claim.
On one side, the lab. The largest randomized controlled experiment on AI coding to date (Cui, Demirer, Jaffe, Musolff, Peng & Salz — three field experiments at Microsoft, Accenture, and a Fortune 100 firm, 4,867 developers, published in Management Science in February 2026) found that developers given the AI tool completed 26.08% more tasks (SE 10.3%) — and that less-tenured developers gained the most: output rose 27-39% for the short-tenure group versus 8-13% for long-tenure developers (the authors flag these splits as noisy and not statistically significant at conventional levels). The peer-reviewed sibling result comes from customer support: Brynjolfsson, Li & Raymond's study of 5,172 agents, published in the Quarterly Journal of Economics, found AI assistance raised average productivity 15%, with gains around 30% for the least experienced workers and essentially nothing for agents past their first year. Two bodies of causal evidence, one direction: AI is an amplifier for novices, not veterans.
On the other side, the labor market. The Stanford Digital Economy Lab's "Canaries in the Coal Mine" study (Brynjolfsson, Chandar & Chen, ADP payroll microdata, November 2025 version, verified) measured: employment of software developers aged 22-25 fell nearly 20% from its late-2022 peak by September 2025, while employment of older developers kept growing. Venture firm SignalFire's data (vendor-graded; direction corroborated elsewhere, magnitudes soft) is starker: its 2026 report puts new-grad/entry-level hiring at its twelve "Tech Majors" roughly 65% below 2019. The Pragmatic Engineer's practitioner data: new grads fell from roughly three in ten engineering hires at larger companies in 2023 to about one in ten in 2025.
Put the two sides together and you have the question this essay answers: if AI helps juniors most, why are junior jobs contracting first? Which half of "junior engineers are disappearing" is true, which half is panic, and which half is something else wearing the story as a costume?
Half the noise in this debate comes from metric confusion: every side quotes a real number, but the four numbers measure different things. Line them up first.
The stock (how many people hold the job) — at a record high. BLS Occupational Employment and Wage Statistics (official, verified and corrected): total US software developer employment was about 1.657 million in May 2023, dipped 0.15% to about 1.654 million in May 2024 (within sampling error), then rose to about 1.688 million in May 2025 — the latest reading is the highest ever recorded. As a statement about the stock, "software engineering is disappearing" has no evidentiary support at all.
Job postings (how many people employers advertise for) — deep trough, now recovering. Indeed's software-development postings index (Feb 2020 = 100, verified): peaked at 234 in February 2022, bottomed at 61 in May 2025, and stood at 72.5 in late June 2026 — still 27.5% below the pre-pandemic baseline, but up roughly 10% year over year. Note where the trough sits: 2025, not immediately after ChatGPT. The confounders section returns to this timeline.
The entry port (how many offers new grads get) — the hardest-hit gauge. SignalFire's 2025 report: new graduates fell to 7% of Big Tech hires, down more than half from pre-pandemic; the 2026 report deepens it to roughly 65% below 2019 (a VC's own dataset built on scraped professional profiles, no disclosed error bounds). The job-seeker mirror: New York Fed data on recent college graduates (official, aligned to one data vintage per verification), computer science majors show 7.0% unemployment and computer engineering 7.8% — second-to-fourth highest of all majors (versus about 5.6% for recent graduates overall). Yet in the same table, CS *under*employment — working jobs that don't require the degree — is just 19%, among the lowest of any major. The door narrowed; the room behind it did not get worse.
Age-resolved payroll employment (who is actually on the books) — only the youngest bracket is falling. This is the decisive gauge, because BLS publishes no occupation-by-age employment series; ADP payroll data is currently the only source that can slice software developer employment by age (the Canaries study — working paper, not yet peer-reviewed, verified): the 22-25 bracket fell nearly 20% from its late-2022 peak while every older bracket kept growing. The dashboard update through April 2026 shows the divergence still deepening — employment of 22-25-year-olds in highly AI-exposed occupations contracting at 3.8% annually, while the same age group in the least-exposed occupations grows about 2% (figures relayed via Fortune; media-grade).
The four gauges compress into one sentence, the first foundation of this essay: the "disappearance" is a flow phenomenon, not a stock phenomenon. The occupation is not shrinking; the entrance to it is — and the squeeze is concentrated on the youngest, most AI-exposed cell. Any argument that ignores the stock/flow distinction — whether "engineers are done" or "the data shows everything is fine" — is using one gauge as camouflage for another.
Labor economics' standard equipment for automation is the Autor, Levy & Murnane (2003) task model: computer capital "substitutes for a limited and well-defined set of human activities, those involving routine (repetitive) cognitive and manual tasks; and complements activities involving non-routine problem solving and interactive tasks" (NBER abstract, checked verbatim). Because the unit is the task, the model predicts recomposition of jobs, not elimination. The 2003 taxonomy filed coding under "nonroutine cognitive" — LLMs have moved a sizable portion of it into the automatable column, a migration the original model did not anticipate; the framework survives, the table entries change.
Acemoglu & Restrepo formalized the direction question as a race between two forces: the displacement effect (automation takes tasks away from labor) versus the reinstatement effect (new tasks are created and handed back). Which wins is an empirical question, not a theoretical one — the theory explicitly permits both outcomes (Journal of Economic Perspectives, 2019). So both "the task model predicts juniors vanish" and "the task model says juniors are fine" are misquotes: the model predicts a race, not a winner.
For the junior question specifically, the sharpest theoretical prediction comes from Autor & Thompson's 2025 "Expertise" paper (NBER working paper), estimated on four decades of occupational data: when automation eliminates an occupation's inexpert tasks, wages in that occupation rise, employment falls, and the expertise bar for the remaining work rises; eliminating expert tasks flips every sign. What AI coding tools remove from software engineering is precisely the former — boilerplate, simple CRUD, first-pass bug fixes: the exact task bundle junior engineers have traditionally used to enter the field. Applying the framework to AI coding is an inference (the paper frames AI in these terms but its estimates are historical): it predicts not the death of software engineering but a smaller entrance, a higher bar, and better pay — which is exactly the shape of the four gauges in Section 1.
If the narrowing entrance were purely cyclical, the market would self-correct; the trouble is a mechanism that can make it self-locking. Arrow's 1962 learning-by-doing is the foundation (checked verbatim against the original): "Learning is the product of experience. Learning can only take place through the attempt to solve a problem and therefore only takes place during activity." Doeringer & Piore's internal labor markets (1971) supply the organizational half: firms hire through limited ports of entry and fill senior rungs by internal promotion — junior hiring is an investment with a deferred payoff.
Connect both to AI: if agents absorb the doing, novices lose the medium of learning-by-doing; if firms close the entry ports, the ladder breaks at its first rung. Ide's 2026 apprenticeship model (arXiv working paper, pure theory) formalizes the intuition: entry-level automation is privately profitable for each firm but can be socially costly, because it severs the transmission of tacit knowledge from experts to novices — a loss that appears on no single firm's income statement. That is the shape of a market failure: every firm rationally hires fewer juniors, and the industry collectively runs short of seniors five years later. Matt Beane's field research (HBR 2019; The Skill Code, 2024) documented the pre-AI version of the same mechanism across 30+ professions: intelligent machines separate novices from experts and block on-the-job skill formation.
The counter-theory comes from Autor himself (Noema, 2024, an essay): AI as an expertise leveler that could "extend the relevance, reach and value of human expertise to a larger set of workers" — on this reading AI should qualify more people, not fewer, for high-value work. The two Autors are not in contradiction: the leveler thesis is about who can do the work, the expertise framework is about how many the firm needs — wages and employment can move in opposite directions. That is precisely what makes the econometric battle in the next section worth watching.
The 2025-2026 empirical literature splits into two camps — and the dividing line is not ideology but data type.
Canaries (Stanford, ADP payroll microdata, verified): workers aged 22-25 in the most AI-exposed occupations show a 16% relative employment decline versus the least exposed (November 2025 version; the August original said 13% — the headline moved between versions). It survives firm-time fixed effects, meaning firm-wide shocks — layoffs, rates, tax changes hitting everyone at the company — cannot account for it. The decline concentrates in occupations where AI usage is automation-dominant rather than augmentation-dominant; adjustment runs through employment, not wages; results are robust to dropping tech firms entirely and to excluding remotable occupations.
Hosseini & Lichtinger (Harvard, 62 million résumés plus postings, SSRN working paper, verified) gave the phenomenon its name: seniority-biased technological change. Beginning in 2023Q1, junior employment at generative-AI-adopting firms declined sharply relative to non-adopters while senior employment kept rising — driven primarily by slower hiring, not separations. Two independent data sources, two methods, one shape: the entrance moved, not the stock.
Yale's Budget Lab (CPS household data, verified): the broader labor market "has not experienced a discernible disruption since ChatGPT's release 33 months ago"; its May 2026 synthetic difference-in-differences re-test through Q1 2026 "generally find[s] no statistically or economically significant effects as of yet" on employment or wages in AI-exposed occupations. EIG: from 2022 to early 2025, unemployment rose less for the most AI-exposed workers (+0.30pp) than the least exposed (+0.94pp). The New York Fed (Lightcast postings plus an Anthropic-usage exposure metric, May 2026, verified): within highly exposed occupations, "labor demand for junior and senior roles ... is moving broadly in parallel," and the high-vs-low exposure gap predates 2022 with no post-2022 break. LinkedIn's official report (January 2026, verified): "Entry-level roles have not been disproportionately impacted relative to experienced roles" and "we do not see AI impacting entry-level roles—yet" — a report worth savoring, since it flatly contradicts the "bottom rung of the career ladder is breaking" op-ed that LinkedIn's own executive Aneesh Raman had published in the New York Times eight months earlier. Same company, two voices.
This is not a muddle of half-for, half-against. The two camps' instruments differ systematically:
In other words: if the true phenomenon is precisely "a hiring-flow contraction for the youngest cohort," payroll data should see it and postings and household data should not — and that is exactly the observed distribution of evidence. The null results are not a contradiction; they are what limited instrument resolution looks like. Economist Jed Kolko's review of this literature supplies the right timestamp: "None of this research is—nor could be—the last word." The research, he writes, is still in the first inning.
The payroll camp has also narrowed its own claim: in their February 2026 response note (verified), the Canaries authors conceded that under the broadest controls, the employment decline is statistically significant only after 2024 and that "part of the timing of this decline is due to factors other than AI" — the authors themselves narrowed the original late-2022 origin story. The same note strikes back at the interest-rate critique: AI exposure is negatively correlated with occupational rate sensitivity (construction, the most rate-sensitive sector, has the lowest AI exposure) — "the interest rate hypothesis points the wrong direction."
The "it's macro, not AI" camp holds four genuine cards. Each deserves face value — and then an audit of what it cannot explain.
Timing. Indeed's own analysis (verified): nearly half of the decline in US tech postings from their early-2022 peak occurred before ChatGPT was publicly released. Two Google economists (interested party, noted) add: postings in the highest AI-exposure quintile peaked in March-April 2022, tracking Fed tightening rather than any model release; Census data show fewer than 10% of large firms even planned to use AI in production as of late 2023. Apollo chief economist Torsten Slok's version: the recent-grad unemployment gap opened in early 2022 and is "far more likely a product of the general low-hire, low-fire labor market than of a technology that companies had barely begun to deploy when the gap emerged."
Tax law. Section 174 forced five-year amortization of US software R&D salaries from 2022 (fifteen years for offshore); a Tax Foundation analyst modeled roughly 20,000 software jobs lost (a policy estimate, not an ex-post measurement), and Microsoft paid $4.8 billion in additional 2023 tax. The OBBBA restored immediate expensing on July 4, 2025 — creating a natural experiment that cuts both ways: the aggregate rebounded (software engineer postings up ~11% year over year by March 2026 per Citadel's analysis of Indeed data; IT/CS postings +14.2% in April 2026 per ZipRecruiter), while the junior share kept falling (entry-level share of IT postings 8.1%→7.4%, senior share 38.8%→43.1%; relayed via media coverage). An age-neutral tax change cannot explain an age gradient that kept deepening after the tax was fixed.
The overhiring correction. Per layoffs.fyi: 262,735 tech layoffs in 2023 (the peak), 152,922 in 2024, 122,549 in 2025 — declining every year, with the bulk of the purge predating capable AI coding agents.
Offshoring. India's Global Capability Centers are projected to hire 510,000 people in 2026, up 11% year over year in H1 — and early-career (0-3 years) hiring is growing 18% annually (job-platform data). Junior software work did not vanish; part of it moved to Bengaluru. Of the four confounders this is the least-rebutted: like AI, it can explain a US-specific junior squeeze — and even OBBBA's retained 15-year foreign amortization penalty has not slowed it.
The composite verdict on the confounders — the second foundation of this essay: the macro cocktail explains the water level, not the composition. Rates, Section 174, and the overhiring payback are enough to explain the 2022-2025 collapse and recovery in aggregate hiring; but all three are age-neutral, sector-wide forces — and the most striking shape in the data is compositional: within the same firm, over the same months, only the youngest-by-most-exposed cell falls (inside firm-time fixed effects), and the same exposure taxonomy predicts nothing about young workers' employment in earlier periods, including the COVID unemployment spike (the Canaries placebo test, verified). The AI-attributable share remains an inference — but the "pure macro, zero AI" account owes an explanation for that gradient, and nobody has produced one.
Return to the opening contradiction: RCTs say AI helps novices most; the market says novice jobs contract first. How are both true?
Step one: draw the evidentiary boundary of the RCTs precisely. The Copilot experiments measure task output of already-employed developers; the QJE study measures resolutions per hour of already-hired agents. No experiment measures whether firms respond by hiring more or fewer novices — not a flaw, just a question experiments of this design cannot answer. The QJE authors put both branches on the table themselves (checked verbatim): "In the longer run, firms may respond to increasing productivity among novice workers by hiring more of them or by seeking to develop more powerful AI systems that replace labor altogether." The same experimental evidence supports two opposite hiring predictions — the fork lies in the firm's production-function response, which the experiment does not identify.
Step two: see which branch shows up in the data. The payroll evidence of Section 3 says that in US software, 2023-2026, firms picked the second branch. Novices became more capable — but whatever "novice + AI" can do, "senior + AI" does faster and with less review overhead, and the seniors are already on payroll. When AI raises the floor of junior output, it simultaneously lowers the cost of not hiring juniors. The technology that helps novices most is the same technology that makes novices most dispensable — on the firm's ledger these are one sentence, not two. No paradox: complementarity at the individual level, substitution at the organizational level, because the unit that gets hired is not a task but a bundle.
Step three: the long-term invoice this resolution generates — the empirical interface of Section 2.3's apprenticeship economics. If novices aren't hired, learning-by-doing cannot happen; if novices are hired but agents do all the formative work, it equally cannot happen. Learning science has produced its first causal datapoint (PNAS 2025, ~1,000 high-school math students, verified): unguardrailed GPT-4 raised practice scores 48%, but once access was removed those students scored 17% worse than peers who never had AI — while a guardrailed tutor variant largely eliminated the harm. Extrapolating to engineers is analogy, not evidence — and it coexists with the QJE counterexample, where novice gains reflected "durable worker learning" (skills persisted through AI outages): whether AI is poison or nutrient for skill formation depends on the usage architecture, not on AI itself. The most telling number in the industry is the one nobody measures: as of July 2026 there is no credible data anywhere on whether junior-to-senior promotion time has changed. An entire industry is debating whether the ladder is broken without instrumenting the ladder.
One headstone to set in passing: the most-cited paper claiming AI helps top performers most (Toner-Rodgers, MIT, materials scientists) was disavowed by MIT in May 2025 for fabricated data — MIT stated it has "no confidence in the provenance, reliability, or validity of the data." As of July 2026, no published RCT shows seniors gaining more than juniors from agentic coding tools; "seniors orchestrate agents better" remains a vendor-telemetry hypothesis, not an experimental result.
What companies said and what they did diverged sharply across 2025-2026 — and the shape of the gap is itself evidence.
The proclamation phase (H1 2025). Anthropic CEO Dario Amodei warned via Axios that AI could eliminate half of all entry-level white-collar jobs and push unemployment to 10-20% within one to five years (note the source grade: the numbers are Axios's rendering of the interview, not a verbatim quote — and an AI lab CEO has a commercial stake in AI-capability narratives). Meta's Zuckerberg predicted AI functioning as a mid-level engineer during 2025; Salesforce's Benioff announced no new software engineer hiring for 2025, citing a 30% productivity gain from Agentforce (vendor-self-reported, no methodology); Amazon's Jassy wrote in an official memo that generative AI "will reduce our total corporate workforce" over the next few years (primary source, but not specific to engineering or entry level).
The boomerang phase (H2 2025-2026). Klarna, which had marketed its AI as doing the work of 700 agents, began rehiring humans in May 2025, its CEO admitting the company went too far. Benioff disowned the "white-collar apocalypse" framing by July. IBM announced in February 2026 it would triple US entry-level hiring, its CHRO stating the pipeline logic plainly: "The companies three to five years from now that are going to be the most successful are those companies that doubled down on entry-level hiring in this environment" — firms that skip juniors will be poaching everyone else's mid-levels. Robert Half: 32% of hiring managers who eliminated a role primarily because of AI later rehired the same or similar role; Orgvue: 55% of leaders who did AI-driven layoffs concluded it was a mistake (all three surveys from staffing-industry vendors who benefit from rehiring narratives — noted). AWS CEO Matt Garman was bluntest: replacing junior employees with AI is "one of the dumbest things I've ever heard" — juniors are the least expensive employees and the most fluent AI users — and Amazon is hiring 11,000 interns and new grads in 2026.
What remains after netting proclamations against boomerangs is three structural facts. First, seniorisation (PwC's 2026 Global AI Jobs Barometer, consulting-vendor grade): entry-level roles are not vanishing but being redefined — entry postings in highly AI-exposed occupations are 7x more likely to demand traditionally senior skills, "seniorised" entry roles have grown 35% since 2019 while traditional entry openings fell only 10%. The job exists; it is no longer designed for a 22-year-old blank slate. Second, the training-ground tasks are being absorbed by agents (vendor telemetry, unverified): Anthropic's Economic Index classifies 79% of Claude Code conversations as automation, concentrated in simple UI work, CRUD apps, and bug fixing — precisely the traditional junior leveling zone. Third, the training pipeline's business model died first: the coding bootcamp industry collapsed through 2024-2025 (Epicodus and Code Fellows closed; 2U exited after a 40% enrollment drop; high-reputation Rithm stopped taking applications, calling 2024 the worst entry-level market in its decade-plus) — business models that depend on abundant junior openings were this ecosystem's canaries, and they died before any official statistic moved.
Historical base rates supply the final calibration. In October 2016 Geoffrey Hinton said: "People should stop training radiologists now. It's just completely obvious that within five years deep learning is going to do better than radiologists… It might be ten years." A decade later, the 2026 residency Match offered a record 1,478 radiology positions with a 97.6% fill rate, average radiologist pay reached $571,000, and Hinton conceded to the press that he had spoken too broadly and gotten the timing wrong. But the precedent bills both ways: the prediction itself wounded the funnel — radiology residency applications fell 14% from their 2023 peak even amid record demand. A frightening forecast does not need to come true to scare away a cohort. ATMs coexisted with rising teller employment for fifteen years (tellers per branch fell from ~21 to ~13, branches got cheaper, banks opened more) until mobile banking finally delivered the cut; spreadsheets erased ~400,000 bookkeeping-clerk jobs while ~600,000 higher-judgment accounting jobs were added. The occupation surviving is the historical norm — and the routine tier absorbing all the losses is the historical norm too. The question was never whether the profession disappears; it is who owns the bottom rung.
The first six sections convert into three action lists. Each recommendation is tagged with the evidence it stands on — and stops where the evidence stops.
The essay compresses to seven claims, ordered by evidence strength:
Criteria worth watching through 2027: whether the Canaries gradient spreads to the 26-30 bracket (the natural corollary of seniorisation); whether postings data develop the junior/senior fork the New York Fed currently does not see; whether IBM-style reversals become the norm or stay anecdotes; and whether any organization starts publishing junior-to-senior promotion-time data. The first company to put the career ladder on its balance sheet will be more informative than every proclamation combined.
Employment and hiring data: BLS OEWS (SOC 15-1252, May 2023/2024/2025) · Indeed Hiring Lab software-development postings index (FRED: IHLIDXUSTPSOFTDEVE) and Bernard, "The US Tech Hiring Freeze Continues" (July 2025) · NY Fed, The Labor Market for Recent College Graduates (Feb 2026 update) · LinkedIn Economic Graph, Labor Market Report (Jan 2026) · SignalFire State of Tech Talent 2025/2026 (VC grade) · Orosz, The Pragmatic Engineer (June 2026) · CRA Taulbee Survey (June 2026) · layoffs.fyi
Econometric studies: Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine?" (Stanford Digital Economy Lab, Nov 2025 version + Feb 2026 response + Canaries Dashboard) · Hosseini & Lichtinger, "Generative AI as Seniority-Biased Technological Change" (SSRN 5425555) · The Budget Lab at Yale (Oct 2025; May 2026 SDID) · Eckhardt & Goldschlag (EIG, Aug 2025) · Audoly, Guerin & Topa (NY Fed Liberty Street Economics, May 2026) · Iscenko & Curto Millet (Google, "Looking for the Ladder," Jan 2026) · Kolko (PIIE, Mar 2026) · Massenkoff & McCrory (Anthropic, Mar 2026, vendor grade)
Experimental evidence: Cui, Demirer, Jaffe, Musolff, Peng & Salz, "The Effects of Generative AI on High-Skilled Work" (Management Science, 2026) · Brynjolfsson, Li & Raymond, "Generative AI at Work" (QJE 140(2), 2025) · METR (arXiv:2507.09089; Feb 2026 update) · Paradis et al. (Google, arXiv:2410.12944) · Bastani et al., "Generative AI without guardrails can harm learning" (PNAS 122(26), 2025) · MIT's disavowal of the Toner-Rodgers paper (May 2025)
Theory: Autor, Levy & Murnane (QJE 2003) · Acemoglu & Restrepo (AER 2018; JEP 2019) · Acemoglu, "The Simple Macroeconomics of AI" (2024) · Autor & Thompson, "Expertise" (NBER WP 33941, 2025) · Autor, Noema (2024) · Ide, "Automation, AI, and the Intergenerational Transmission of Knowledge" (arXiv:2507.16078) · Arrow (1962) · Doeringer & Piore (1971) · Beane, "Learning to Work with Intelligent Machines" (HBR 2019) / The Skill Code (2024) · Deming, Ong & Summers (NBER WP 33323) · Bessen (ATMs and tellers) · NPR Planet Money #606 (spreadsheets)
Industry behavior: Amodei via Axios (May 28, 2025, paraphrase grade) · Zuckerberg via JRE (Jan 2025) · Benioff (Feb 2025; walk-back July 2025) · Jassy memo (June 17, 2025) · IBM entry-level expansion (Feb 2026) · Garman via Platformer (June 2026) · Robert Half / Orgvue / Careerminds boomerang surveys (2025-2026, staffing-vendor grade) · PwC Global AI Jobs Barometer 2026 (consulting-vendor grade) · Strada (May 2026) · NACE Job Outlook 2026 · LeadDev AI Impact Report 2025 · Anthropic Economic Index (Apr 2025 software report; Mar 2026; Jun 2026, vendor grade) · Stack Overflow Developer Survey 2025 · DORA 2025 · Course Report bootcamp reviews · Klarna/IBM/Duolingo/Shopify walk-back timelines · NRMP Match 2026 radiology data · Hinton's 2016 remarks / Fortune May 2026 interview