Methodology note: Round 1 of this issue split the question into 8 research lines (official projections auditing themselves / measured outcomes for US recent graduates / three-way contest over AI exposure / Chinese official statistics / Chinese commercial surveys and employment quality / field-of-study effects vs. institution effects / the licensing moat, pro and con / oversupply mechanisms and historical precedents), producing 169 raw claims. Round 2 split the load-bearing claims into 60 groups, each independently checked by 3 votes, refute-by-default, for 180 votes in total: the first half (108 votes) returned 17 HOLDS / 86 CORRECTED / 5 REFUTED, the second half (72 votes) returned 4 HOLDS / 66 CORRECTED / 2 REFUTED; after merging under the rule "when the three seats disagree, take the strictest," this converged to 3 groups HOLDS, 53 groups corrected, 4 groups overturned. Round 3 assigned each of 12 single-source load-bearing empirical findings both a contradiction-search seat and a methods-audit seat with veto power, for 24 seats in total, returning 9 VETO / 9 RESTRICT_USE / 3 UPGRADE_MULTISOURCE / 2 CONTRADICTED — for 10 of the 12 items submitted, the core number was withdrawn or downgraded. There were four key retractions: ① the methodological spine originally planned for the whole essay was withdrawn outright by both seats (a MAPE figure from a 2005 external evaluation conflicts head-on with BLS's own self-evaluation over the same period, the same 338 occupations, and the same definitions, and back-computing from the full error distribution the latter publishes shows the figure is arithmetically impossible), replaced by the BLS official evaluation series; ② all of the Stanford "canary" study's headline figures are barred, and it may not serve as an empirical support for "AI is compressing entry-level jobs"; ③ Microsoft's AI applicability scores may not serve as a quantitative support, and none of its two-decimal figures may enter comparative sentences in the body text; ④ the core testable signal originally planned for the China side (the reversal in the first derivative of the number of schools newly adding AI programs across Ministry of Education batches) was overturned and vetoed by primary data, and substitute signals are listed separately. Evidence grades: [Multi-source] = ≥2 independent sources concur; [Single-source, verified] = primary document traceable; [Contested] = independent sources conflict; [Vendor claim] = an interested party's own statement, direction usable, magnitude not load-bearing; [Commercial survey] = a for-profit firm's proprietary sample, interests must be flagged; [Unverified] = sourcing failed. All evidence is as of July 2026.
Operational anomalies that must be disclosed: the second half of the Round 3 workflow hung at seat 11 of 12; the main process stopped it and salvaged the completed verdicts from the run log, and the missing seat was re-run separately (the re-run seat obtained a primary document not previously obtained; its conclusion does not conflict with the prior 11 seats). Three further classes of tooling artifact have been identified: an "access blocked" result on a Ministry of Education domain was in fact a redirect loop caused by the scraping tool's protocol upgrade — that site never actually blocked anything; the summarization model inside the web-scraping tool fabricated four nonexistent verbatim quotations from one arXiv paper, so every verbatim PDF quotation in this project has been re-checked against locally extracted source text; the same summarizer also misread a three-column table from 2010 across columns and produced a fake list, which has since been falsified against six sources. Two primary documents were confirmed unobtainable (the Ministry of Education's 2019 annual program-filing annex is a scanned image with no text layer; a certain commercial survey firm's original 2010/2011/2025/2026 volumes are behind a paywall).
This essay observes three self-imposed constraints it will not depart from: for China and the US, compare only direction and structure, never place the numbers side by side, and attach a construct-difference note to every cross-national comparison; the destination is a list of fields with evidence grades attached, not a ranking; make no ten-year forecasts, staying instead with structural changes that have already occurred and with test signals you can observe yourself. The readers of this essay are 17- and 18-year-olds choosing a field of study, and their parents — a precise number with a decimal point will be carried into a life decision as if it were a fact, so it is better for the essay to have one table fewer than one false decimal more.
This essay gives no ranking of majors, and does not tell you "what will be hot over the next decade." That is not modesty; the hardest single piece of evidence in this issue closes off that road directly: official ten-year occupational projections do contain real information at coarse levels of aggregation, but at the detailed-occupation level they are barely better than "assume nothing changes" — and choosing a major happens precisely at the detailed level.
So this essay does something else: it takes apart, one by one, the definitional bases of the figures most often imported from popular narratives into admissions advice; it strictly separates structural changes that have already happened from things that are merely projected; and it ends with a list organized by criterion, with an evidence grade attached to every cell, plus ten claims you can test yourself.
The US Bureau of Labor Statistics (BLS) publishes ten-year employment projections by occupation every two years, and it is the only producer of such projections anywhere in the world that regularly grades itself. These evaluation pages are published by BLS itself, the method is public, they include a naive benchmark, and they truthfully report the cases where BLS loses to the naive model — this is classic against-interest evidence, and its credibility is higher for that reason.
[Multi-source] Per BLS's own published evaluations: in the 2012–2022 round, at the detailed-occupation level BLS's total error was 17.1, against 17.8 for the naive model of "each occupation's employment share stays unchanged," which makes no prediction at all; out of 748 detailed occupations BLS beat that naive model in only 432 (58%). In the 2006–2016 round, BLS correctly called whether an occupation would grow or decline 78% of the time, but correctly called whether an occupation would grow faster than the economy as a whole only 57% of the time — on a two-way question, the baseline is 50%.
[Single-source, verified] An earlier round (1988→2000) points the same way: across 9 major groups BLS's mean absolute percentage error was 5.86% (5.29% weighted) against 13.8% and 11.8% for two naive models — BLS was clearly better; but across 338 detailed occupations, the rank correlation between projected and actual growth was only 0.43 (n=338; the 95% interval implied by a Fisher z transformation is roughly 0.34–0.51, and because errors are correlated across occupations the true interval is wider).
The quintile transition table is the best translation of this evidence: of the 67 occupations BLS placed in its "fastest-growing fifth," 48% (about 32) did land in the actual fastest fifth — the random hit rate is 20%; but 19% (about 13) fell into the actual slowest two-fifths.
So the correct reading is not "the BLS list is guesswork" — it is clearly better than guessing; it is that "what it gives you is a coarse signal roughly 2.4 times better than random, not a list you can pick a major from." An occupation listed among the fastest growers has roughly a one-in-five chance of ending up at the bottom ten years later.
First, in these evaluations the projected total growth rate is itself frequently wrong (2006–16: 10.4% projected vs. 3.6% actual; 2012–22: 10.8% projected vs. 13.2% actual), because BLS projections are built on a full-employment assumption and do not predict recessions. A substantial share of the percentage error at the occupation level comes from getting the total wrong, not from "being unable to tell which occupations will grow."
Second, the naive model is an extremely low bar — "each occupation's employment share stays unchanged." "BLS beat the naive model" supports only "the projections contain real information and are not waste paper"; it does not support "accurate enough to choose a major by." At the same time this must be spelled out: BLS itself writes that "there are no comparable projections which are not in some way derived from BLS projections" — no independent occupational projection comparable to it exists anywhere in the world, so it has never been tested against a genuine rival.
Third, the method changed along the way (the 2016–26 edition onward replaced the replacement-needs algorithm; the SOC classification was reclassified from 2010 to 2018; from 2019 modeled estimates were adopted), so the error magnitudes from 1988–2000 can serve only as a prior expectation for the latest edition, not as a verified fact about it.
Fourth, two evaluation gaps must be written down together: the 2010–20 round was never evaluated (BLS's official reason is COVID's impact on 2020 data, not the classification revision); and for the 2014–24 round BLS did not even evaluate its labor force projections or its macroeconomic projections — meaning that the link "labor force projections anchor employment" has not been officially tested at all in the most recent round.
[Multi-source] The three figures for directional accuracy have different denominators and may not be listed side by side: 86% is the 2014–24 round, 22 major occupational groups; 77% is the 2012–22 round, likewise 22 major occupational groups; 69% is the 2014–24 round, 16 major industry sectors — the first two are comparable, the third is not. Understating growth is consistent across both rounds: for 2014–24 the projected average occupational growth was 6.5% against 12.9% actual; for 2012–22, 10.8% projected against 13.2% actual.
[Multi-source] Three self-stated qualifiers, verbatim and load-bearing:
On uncertainty, the accurate statement is: the current ten-year projections publish a single-scenario point estimate only, with no confidence intervals in the statistical sense; but BLS has published scenario ranges before (in 2021 it published two separate pandemic scenarios for 2019–29; in the 1980s–90s it published high/moderate/low variants), and since the late 1990s the routine ten-year projections have no longer published alternatives.
[Single-source, verified] The cell most worth writing out from that same BLS internal study is the digital camera: BLS had judgmentally lowered its projection for photographic-processing occupations (2004–14 projected −23.6%), while the actual figure fell from a 2004 peak of 86,300 to 28,800 in 2014 (−66.6%), and to just 9,200 by 2023. Even in the most favorable case — a clear technological path already visible in the data — BLS still understated the decline by roughly a factor of three.
[Not obtained from multiple sources; exhaustive-search negative result] Canada's COPS, the EU's Cedefop, and the UK's Working Futures all publish ten-year detailed occupational projections, and for all three, no public quantification of ex-post error can be found — Employment and Social Development Canada even states on its own site that COPS "is not a forecasting system but rather a method providing signals." Globally, the long-term occupational projection trade holds itself almost entirely unaccountable for its own accuracy in public; BLS is one of the few willing to do so. (The contradiction-search seat logged 16 search angles item by item, including all the misses; the only non-US ex-post evaluation of this kind found was a Dutch self-assessment, cut by type of education and thus not comparable to detailed occupations, judged "a lead, not evidence.")
[Single-source, verified] A completely independent 2025 study (a US Naval Postgraduate School working paper, using the text of the Occupational Outlook Handbook from 1946–1996) found that the third of occupations viewed most favorably grew about 50 percentage points more over the following three decades than the least favored third. So this essay cannot be written as "official occupational judgments contain no signal"; only as "its increment over 'look at the last ten years' trend' is nearly zero (R² 0.148 vs. 0.139), and the precise ranking is not credible."
This section is the methodological spine of the whole essay. What it gives is not "don't trust the officials" but an aggregation-level gradient: the coarser, the more credible; the finer, the less credible — and the decision you have to make sits at the finest end.
This is the hardest-evidence section in the essay. Three seats each independently recomputed all 832 detailed occupation lines, with all residuals ≤0.1 and zero exceptions; the identity below is not a correlation, it is a definition. This is also the one piece of evidence in this essay you can verify yourself in a single afternoon.
[Multi-source] Per Table 1.10 of the BLS Employment Projections, 2024–34 edition, the column header reads verbatim "Occupational openings, 2024–34 annual average" — this is an annual-average flow over a ten-year period, not a stock at any point in time, and not a ten-year cumulative total either.
Cell by cell for all occupations: 2024 employment 169,956.1 thousand; 2034 employment 175,167.9 thousand; ten-year net change +5,211.8 thousand (+3.1%); annual average labor force exits 8,155.0 thousand; annual average occupational transfers 10,187.1 thousand; annual average total separations 18,342.1 thousand; annual average openings 18,863.3 thousand.
The compositional identity: openings = annual average separations + ten-year net change ÷ 10. Recomputed: 18,342.1 + 521.18 = 18,863.28. So separations contribute 97.2% and net growth contributes only 2.8%.
One further definition must be known: "This estimate of openings does not count workers who change jobs but remain in the same occupation." — changing employers within the same occupation is not counted at all. And BLS itself states that this is a different construct from the "job openings" on hiring sites: "The JOLTS job openings metric is a stock measure… In contrast, occupational openings are projections of the number of positions that are created and filled on an annual basis… There is therefore no way to create an annual average metric of job openings from JOLTS that could be compared…" Not interchangeable, not divisible into one another, not to be plotted on the same chart.
[Multi-source] Within the 18,342.1 thousand separations, occupational transfers are 10,187.1 thousand (55.5% of separations) and labor force exits only 8,155.0 thousand (44.5%). In other words, most of the so-called "openings" are not retirements clearing a seat; they are incumbents moving to a different occupation. In the same BLS Q&A: "For declining occupations, not all workers who separate need to be replaced," and BLS states that the measure's usefulness extends only to giving a sense of the magnitude of expected openings, especially relative to other occupations.
[Multi-source, our own computation] Ranking the 832 detailed occupation lines (excluding summary rows) by annual average openings in descending order, 13 of the top 15 have median annual wages below the all-occupations median ($49,500, May 2024 basis, covering only nonfarm wage and salary employees and excluding the self-employed — this basis is slightly mismatched with the all-jobs basis of openings); the two above are general and operations managers ($102,950) and heavy truck drivers ($57,440). Six of the top 15 have shrinking ten-year net employment (retail salespersons, cashiers, waiters and waitresses, customer service representatives, general office clerks, and secretaries and administrative assistants).
There is an extremely useful criterion here: a ratio of separations to openings above 100% is strictly equivalent to that occupation's ten-year net employment shrinking (verified by three seats across 832 rows). But note that it holds only under the current method — the older method used for 2014–24 and earlier truncated the growth term to zero for declining occupations, so applying this criterion across editions is wrong.
BLS has acknowledged something similar, but the range of the claim must be stated precisely: its actual wording is "Each of the occupations in chart 1, like most at this education level, had wages below the median for all occupations." — chart 1 is the top 10 occupations by openings within the "no formal educational credential" tier, not the economy-wide top 15; and the same article reaches the opposite conclusion for higher education tiers.
[Multi-source] The 2014–24 edition's column header reads verbatim "Job openings due to growth and replacements, 2014-24" — no annual average; this is a ten-year total, of 46,506.9 thousand. The 2016–26 edition's header reads "Occupational openings, 2016-26 annual average," with a value of 18,981.5 thousand. Annualized, 46,506.9 ÷ 10 = 4,650.69 thousand/year, and 18,981.5 ÷ 4,650.69 = 4.08×. If you simply set 46,506.9 next to 18,981.5, you would read it as "cut by 59%" — which is exactly the illusion this item exists to expose.
BLS has passed judgment on its own old method: "BLS used a cohort-component method… This method is no longer in use because BLS identified statistical and conceptual issues with the implementation of this method that compromised the accuracy and validity of the resulting estimates." And a harder sentence: "Published estimates from the 2014–24 projections and prior years should not be compared with the 2016–26 projections…"
Attribution must be tightened: on a share-of-employment basis, the old method gives 3.09%/year and the new one 12.16%/year, so the jump attributable to the method change is about 3.9×, and charging the whole 4.08× to the method overstates it by about 4% (the remainder comes from the difference in base-year employment between the two editions). [Not obtained from multiple sources] Each of the three seats searched independently and found no published academic challenge to, or independent quantitative evaluation of, the new method — written here as "not found," not as "does not exist."
Interests at stake: BLS is both the party that executed the method change and the only party that has passed judgment on it; its self-statement that "the old method had statistical and conceptual flaws" is simultaneously a self-justification for discontinuing it.
[Multi-source] Grouping the 832 detailed lines by "typical entry-level education" and summing (three seats recomputed independently, and all eight tiers agree figure by figure): the shares of annual average openings are high school 35.81% / no formal educational credential 33.34% / bachelor's 17.67% / postsecondary nondegree award 6.16% / some college, no degree 2.33% / associate's 1.76% / master's 1.68% / doctoral or professional degree 1.25%.
The three degree tiers together account for only 20.6% of annual average openings, yet 29.9% of 2024 employment and 58.0% of the ten-year net increase. "Some college, no degree" is the only one of the eight tiers with a net decline (−157.0 thousand, −3.0%).
The mechanism is the separation rate, not the number of jobs: annual average separation rates form a monotonic gradient — doctoral 4.67% < master's 7.25% < bachelor's 7.29% < associate's 9.10% < postsecondary nondegree 10.52% < some college 10.83% < high school 10.96% < no credential 15.47% — with the exit-rate and transfer-rate components pointing the same way. That is: the relatively small number of openings at degree levels comes from low separation rates, not from a shortage of jobs.
This cuts both ways, and both must be said: on one side, degree-level jobs are more stable; on the other, the annual entry aperture relative to the stock is narrower — good news for those already inside, not necessarily for the new graduate trying to get in.
One field qualifier: "typical entry-level education" is an attribute assigned by BLS analysts to each detailed occupation, not the actual educational distribution of incumbents, nor the hiring requirement for the job.
[Multi-source] The combined group "software developers, quality assurance analysts, and testers": 2024 median annual wage $131,450; 2024 employment 1,895,500; 2024–34 +15%, +287,900; annual average openings about 129,200. Software developers reported separately show "15.8 percent" and "267,700 new jobs," with a separate median annual wage of $133,080 (QA analysts and testers separately: $102,610). The combined group's growth rate is slightly lower than that of developers alone precisely because the testing portion, the part most easily automated, grows more slowly — a substantive argument for the topic of the next section.
For contrast: computer programmers, 2024 employment 121,200, 2024–34 −6%, annual average openings about 5,500, with BLS writing verbatim, "All of those openings are expected to result from the need to replace workers who transfer to other occupations or exit the labor force."
But "programmer jobs down 60% = programming work is disappearing" is a false inference — to a large extent this is label migration, and the software developer row in the same table refutes it directly. Likewise, "data scientists, 245,900 in 2024" cannot be read as growth: the 2018 classification revision was the first to give it a separate code; before that these people were classed under "mathematical science occupations, all other." The accurate statement is: the official classification system may lag nearly a decade in recognizing an emerging job title — not "BLS failed to foresee data science."
BLS's two sentences about AI coexist in the same document, and neither may be quoted alone: the favorable one — "The growing adoption of AI technologies… is another factor that will fuel strong job growth among computer and mathematical occupations."; the unfavorable one — "…and resulting productivity gains are expected to dampen labor demand in a variety of fields, such as sales, design, and administrative support." And their nature must be labeled: this is a ten-year modeling baseline (base year 2024), not measured employment, with no error bands; at the time that round was released, adoption data for most agentic coding tools did not yet exist.
"Computer science is already saturated" is one of the most widely circulated narratives of the past three years. Reconciled cell by cell, the result is: the numbers are broadly true, but every one of them is weaker, shakier, and more dependent on definitions than the retellings suggest.
[Multi-source] The New York Fed's The Labor Market for Recent College Graduates, "Outcomes by Major" (released 2026-02-04, data year 2024 American Community Survey, ACS), covers 73 majors. The unemployment rate and underemployment rate are defined for ages 22–27 holding a bachelor's degree or higher (including master's and doctoral holders); the median wage is defined for full-time workers holding a bachelor's degree only — two different populations, which may not be hung on the same hook. All figures exclude those currently enrolled.
A few key cells (unemployment rate / underemployment rate / early-career median wage / mid-career median wage): computer engineering 7.783 / 15.835 / $90,000 / $131,000; computer science 6.992 / 19.127 / $87,000 / $120,000; nursing 2.147 / 12.781 / $70,000 / $87,000; elementary education 1.18 / 16.213 / $45,000 / $55,000; biology 4.292 / 51.13 / $45,000 / $83,000; psychology 4.985 / 48.291 / $45,000 / $72,000; fine arts 7.655 / 58.87 / $45,000 / $72,000; economics 3.524 / 33.094 / $72,000 / $115,000; finance 2.758 / 27.758 / $70,000 / $112,000; all majors 4.211 / 39.35 / $58,000 / $87,000.
On rankings: computer science's unemployment rate is 4th highest out of 73 majors, its underemployment rate 9th lowest, its early-career median wage 2nd highest. Nursing's underemployment rate of 12.781 is the lowest in the table (second is aerospace engineering at 14.7%), but its unemployment rate of 2.1% ranks only 10th and is not distinctive — "nursing is a sure thing" can be hung on the underemployment rate, but hanging it on the unemployment rate is contradicted by the data. Education fields take all three lowest unemployment slots (special education 0.7%, other education 1.1%, elementary education 1.2%). Nursing's mid-career median wage ranks 32nd out of 73; the early-to-mid increase is +24% for nursing, +50% for all majors, +38% for computer science, +22% for elementary education.
Two disciplines for reading the table: the denominator of the underemployment rate is employed graduates, and the unemployed are not in it; wages are in current nominal dollars and may not be differenced directly across editions. Also to be flagged: on 2025-10-31 this page carried the notice "Due to the suspension of necessary data, The Labor Market for Recent College Graduates has not been updated as scheduled," and the current unemployment-rate note states that "October 2025 results are estimated due to missing data."
[Multi-source; three seats each retrieved the raw snapshots of all four CSV editions from a web archive, and all 20 cells agree] By release date / data year: 2023-02-10 / 2021 data, CS 4.8 / 19.1 / $73,000; 2024-02-22 / 2022 data, 4.267 / 16.654 / $78,000; 2025-02-20 / 2023 data, 6.056 / 16.456 / $80,000; 2026-02-04 / 2024 data, 6.992 / 19.127 / $87,000. The population definitions are verbatim identical across the four editions, so comparability over time holds.
Three things must be said together:
The load-bearing combined statement: using a uniform 2022→2024 window, CS unemployment and underemployment deteriorated in the same direction, and the only figure moving the other way is the nominal starting wage — whose growth is still below that of all graduates.
[Single-source, verified, and the source has an explicit policy-advocacy interest] The only party to put interval estimates on this table is the Economic Innovation Group (EIG) in Washington. Its researcher writes verbatim: "In most cases, the confidence intervals around these estimates of recent grad unemployment are so large that it makes no sense to use them to make firm conclusions about the returns to particular majors, let alone use them to make policy decisions." — deleting the opening "In most cases" would promote the author's "in most cases" into a universal claim, committing exactly the fault the author is criticizing. Three intervals verbatim: computer engineering "somewhere between four percent and 11 percent"; physics "somewhere between two percent and 12 percent"; public policy "from zero percent to 13 percent."
The interest must be disclosed alongside the numbers: EIG is a policy-advocacy organization, not a neutral statistical agency; it was founded in 2015 with funding from tech investors and has long advocated expanding high-skilled immigration; the piece opens by framing the dispute as bearing on "the future of STEM education, AI's impact on the labor market, and even high-skilled immigration." The conclusion "the CS job market has not collapsed" points the same direction as its funders' and its own policy position. This does not invalidate its statistical work (the interval recomputation method is legitimate and there is a public code repository), but it must be disclosed when cited.
[Multi-source] The across-edition volatility can itself be recomputed: from the 2023 data year to the 2024 data year, all 73 majors can be matched by name, and the median absolute change in the unemployment rate is 1.2520pp, with exactly 20 majors changing by more than 2pp. Extreme cases: early childhood education 1.298→6.593, nutrition sciences 0.441→4.536, art history 3.047→6.688, foreign languages 4.040→1.579 (in the opposite direction). By comparison CS moved only +0.936pp and computer engineering +0.245pp.
So "consistently high CS/CE unemployment reproduces reliably" must be downgraded: the current CS/CE elevation is supported by only two consecutive readings; and computer engineering's elevation itself originated in a single-year jump (2022 data 2.311% → 2023 data 7.538%); EIG's 95% interval for it is 4%–11%, and 7.538 and 7.783 sit inside the same interval, so "reproduction" adds no statistical information. The strongest permissible statement is: CS/CE readings have been elevated for two consecutive years, but at the magnitude of sampling noise this does not yet amount to a stable trend.
A double standard must be guarded against here: if you invoke noise to explain art history's one-year blowup, you must apply the same yardstick to the CS/CE two-year elevation.
[Bounded enumeration, not an unbounded negative] The New York Fed does not provide, in any of the public release channels for this table, sample sizes, standard errors, or confidence intervals — the four data CSVs have no such columns, full-text searches of the interactive page's notes and Q&A return zero hits, and the downloadable historical tables contain no mention of sample size; nor do the two methodological papers named in the Q&A publish major-level sample sizes for this web table. A harder piece of positive evidence: the New York Fed's own methodological source, when doing analysis by major, uses 13 broad major categories with three years of pooled ACS, whereas the web version expands this to 73 detailed majors with a single year of ACS — an expansion never accompanied by any published sample size or methodological note.
[Multi-source] The 2025 business-media story that set off the discussion (first published 2025-05-16, updated 2025-06-05, under two different headlines) carries a correction notice at the end, verbatim: "Correction: This story has been updated to correct the data in a chart." — that is, the chart data in the first version was wrong and was corrected afterward; and what went viral was chiefly that very chart.
The definitional mishap must also be described precisely: the story did state the source and year for the by-major data; what was unlabeled was the 5.8% figure on the same page — which comes from a different, monthly survey, seasonally adjusted with a three-month moving average, with a single-month reference period, covering all majors combined. The mishap lies in "two surveys placed side by side with only one of them labeled," not in "the year was not stated." The two bases really are non-comparable: for the same 2023 calendar year, the monthly survey's twelve-month average recent-graduate unemployment rate is 4.377%, while the same year's ACS figure for all is only 3.640% — a gap of 0.74pp.
[Multi-source] Moreover, neither of that sentence's two numbers still holds: at the time of the story they were 5.802 / 4.600; the 2025-08 archive shows 5.710 / 4.542; the current version (2026-07-24) shows 5.672 / 4.521 — this series is revised retroactively. The accurate statement is that the story is "one release cycle / one data year behind," not "two years out of date."
[Single-source, verified] Georgetown University's Center on Education and the Workforce, in an October 2025 report, writes that "unemployment among recent graduates with a bachelor's degree in computer science is 7.2 percent, up from a low of 5.3 percent in 2014" — the two windows are each three-year pooled data and must not be treated as single years; its "recent graduates" are ages 22–26, not interchangeable with the New York Fed's 22–27. Its other figure, "6.8% unemployment for recent graduates in computers, statistics, and mathematics," is one of 16 broad fields (computer engineering is included within it); "highest within STEM" holds but must carry the "within STEM" qualifier — across all 16 fields the highest is arts at 8.9% and the lowest education at 2.8%.
The key characterization: this report and the New York Fed use the same underlying microdata, differing only in age window and pooling strategy, and do not constitute an independent measurement. It can only be written as "different cuts of the same data source point the same way, which lowers the probability of single-year sampling noise," never as "two independent institutions corroborate each other." Furthermore, two publications from the same institution give mutually contradictory starting values for the same statistic (5.3% in the report text, 4.3% on the blog), so the safer approach is to use only the 7.2% level and not the change.
Interests at stake: the Center's position is explicitly pro-degree, this report's funders all have raising degree attainment as their mission, and the report's headline conclusion points the same way as the institution's and the funders' interests.
Its supply-side numbers, on the other hand, are clean: bachelor's degrees conferred in computers, statistics, and mathematics rose from 53,499 in 2009 to 138,755 in 2023 (+159%), the fastest of the 16 fields. But this series uses the Center's own field taxonomy, which is broader than the Department of Education's standard classification, and is neither interchangeable with nor splice-able onto the long series in Section 5.
This is the section that most requires restraint, and the weakest empirical territory in the essay. The conclusion first: over the past three years, hiring for some white-collar entry-level jobs in the US has indeed cooled markedly, and this is visible in multiple data sources; but on the three questions "was it AI," "how far did it fall," and "will it last until you graduate," no study can currently give a reliable answer, and any statement offering a specific percentage exceeds what the evidence can support.
[Multi-source (at the construct level)] Nearly all discussion of "will AI take your job" conflates three entirely different things:
One: theoretical exposure. The representative work is the OpenAI team's GPTs are GPTs, which measures "tasks whose time could be halved at equal quality by direct use of a large model, plus, at half weight, tasks that could be halved only after an additional software layer is built." The authors write in their own abstract: "We do not make predictions about the development or adoption timeline of such LLMs." The body adds that "technical feasibility does not guarantee labor productivity or automation outcomes." The figures are fixed at the journal version's 1.8% / 46% (occupation basis); the working paper's body gives the corresponding 3% / 49% (worker basis), revised downward in the journal version. The journal version is a three-page policy-forum piece with no methodological detail whatsoever — anyone citing it for methodological detail is in fact citing the preprint. The annotation quality is itself a load-bearing defect: human–model agreement is 65.6%–82.1%; the model labeled 86 occupations fully exposed against only 15 by humans (5.7×); and the human and model top-five lists have zero overlap on both bases. But the authors' rebuttal must be shown alongside: "Still, we observe a high degree of agreement between human ratings and GPT-4 ratings at the occupation level…" The annotators were company insiders and its outsourced alignment annotators; not one was a practitioner in the corresponding occupation. Three of the four authors list OpenAI affiliations; the fourth is at the Wharton School with no OpenAI affiliation.
Two: vendor usage logs. The unit of observation is the conversation, not the person; one vendor's first release covered only consumer subscriptions and excluded enterprise customers. The publisher's own statements: "we don't argue that the uses in our dataset are a representative sample of AI use in general"; "We can't know for certain whether someone using Claude for a task was completing a task for work."
Three: measured employment. Population surveys, payroll records, administrative data, job postings — each with its own blind spots, discussed below.
[Multi-source] There is also a citation-chain problem that must be named: many studies that appear to "corroborate each other" share the same independent variable (the same machine-scored exposure measure), not the data and not the conclusion. All they can corroborate is that "after grouping by this exposure measure, employment diverges"; they cannot corroborate that "this exposure measure actually measures AI's impact." The hardest link in the chain: two downstream studies adopt precisely the ratings that the original paper explicitly states are not its primary results — "In this analysis, we present results from human annotators as our primary results." — they use the machine annotations the original authors themselves do not treat as their main result.
The direction of the interests here is counterintuitive, and worth writing out on its own: at the high-exposure end stand three OpenAI authors; on the "automation usage is rising" side stands a vendor using its own logs and its own classifier; while the most systematic rebuttal on the "AI is not killing entry-level jobs" side was written by Google's chief economist and his team's AI-and-the-economy lead, with an acknowledgments list including Alphabet's president and chief investment officer, Alphabet's general counsel, and a Google senior vice president. → The heuristic "vendors = exaggerating AI's impact" does not hold; the deepest vendor fingerprints are precisely on the "AI is not killing entry-level jobs" side.
It is not that the experts are fighting; it is that two microscopes at different magnifications are pointed at different things.
[Single-source, verified; descriptive monitoring, no inferential statistics] The low-power lens (the Yale Budget Lab looking at economy-wide occupational composition via the monthly population survey) asks: has the overall distribution of the occupations Americans work in been rearranged? The answer so far is no, and the pace of change remains within the range of historical technology shocks. They compute an "occupational dissimilarity index" monthly, comparing structural change since November 2022 against the computer in January 1984 and the internet in January 1996 on the same yardstick. Two standing conclusions hold verbatim: "While the occupational mix is changing more quickly than it has in the past, it is not a large difference and predates the widespread introduction of AI in the workforce." and "Currently, measures of exposure, automation, and augmentation show no sign of being related to changes in employment or unemployment." The text itself limits its claim: "our analysis is not predictive of the future." The New York Fed, using entirely different data (job postings), independently arrived at the same sentence — the decline in postings for high-exposure occupations began before ChatGPT.
But this lens's resolution means it cannot see a change within an occupation where "the incumbents are still there but newcomers can't get in" — as long as the occupational share is unchanged, it reads flat. This is not an outsider's nitpick; it is their own admission: a 2026-05-07 analysis states that the survey is "somewhat underpowered for subgroup analysis, like the 22–27 year old recent college graduates," and that when the 22–27 sample was actually tested, "too few occupations met the threshold of 50 observations per quarter." (Any citation of this monitoring must lock the version number and date; it is a rolling publication.)
[Single-source, verified; not peer reviewed] The high-power lens (a Danish firm-level difference-in-differences) asks: at the same time, in the same country, did earnings and hours diverge between firms that adopted AI and firms that did not? The answer is no, and this "no" is measured precisely — the pooled specification rules out effects larger than 2%, and the dynamic specification larger than 3%. But by construction this lens can only see between-firm differences — if an entire industry reduced entry-level hiring in step, it would register as zero. The authors themselves concede that "difference-in-differences cannot capture sector-wide general equilibrium effects."
That neither lens saw anything does not mean there is nothing. Each has a blind spot, and they share it: the low-power lens cannot see entry flows within an occupation, and the high-power lens cannot see changes that are uniform across an industry. And the shrinking of entry-level opportunity falls exactly inside that shared blind spot.
[Single-source, verified; official agency] Falling inside that blind spot is a measurement by US Census Bureau researchers using unemployment insurance administrative records: cutting by age, in high-exposure industry × state cells after ChatGPT's release, new hiring of 22–24-year-olds dropped immediately by about 9%, cumulating over ten quarters into an employment decline of about 12%, and almost entirely through "not hiring" rather than "firing."
[Single-source, verified; not peer reviewed] And the Danish evidence indicates that what is happening in that crack is not necessarily AI's doing: the researchers have two survey waves covering 25,000 workers and 7,000 workplaces, linked to national administrative records, letting them split trends by whether a firm actually adopted AI. The result: entry-level positions in high-exposure occupations in Denmark are indeed declining, but not at the firms that adopted AI. They computed an upper bound: AI adoption can explain at most 0.14 percentage points of the total 0.6 percentage point decline, about one quarter. The remaining three quarters they attribute, speculatively, to the macro environment or to precautionary "freeze hiring first" contraction.
So the correct statement is "AI explains at most a small part of it," not "AI has zero effect on entry-level jobs." Anyone retelling this study as the latter is misciting it.
The Danish study's other qualifiers must all be written down: the treatment variable is workers' self-report of whether their employer encourages use, not adoption intensity; the 2024 wave invited 115,000 people and recovered about 25,000 (about 22%), so self-selection exists and the authors have reweighted; the roughly 25,000 in each wave are independent cross-sections, not the same people tracked for two years; and "saves 3% of hours" is users' self-report, not measured productivity.
Nordic homogeneity warning: the three countries that corroborate the Danish result (Denmark, Finland, Norway) are all Nordic, sharing strong unions, high dismissal costs, and centralized wage bargaining — the Finnish researchers themselves attribute the difference to "the Nordic labor market model and strong employment protection." Among the countries reporting declining entry-level positions — the UK, Sweden, the US — US dismissal costs are close to zero. The institutional differences and the data differences overlap completely, and nobody can separate them. So Denmark's "zero" cannot be transplanted to the US, and still less to China. What it establishes is "AI adoption does not necessarily cut entry-level jobs," not "entry-level jobs will not be cut." All three of the studies above are unrefereed.
[Vendor/proprietary data + single source + not peer reviewed; Round 3 double-seat VETO] The most widely circulated quantitative study of "AI is compressing youth employment" was vetoed by both seats in this issue, and none of its headline figures may enter the body text. The weak conclusion that can be retained goes only this far, and must carry all the qualifiers:
In a private panel covering millions of US payroll records (the client-firm composition is not representative of the US economy), since late 2022, in occupations placed in the highest tier by an LLM-scored "AI exposure" measure, the number of employed 22–25-year-olds declined relative to low-exposure occupations at the same firms, while no comparable decline appeared in any age group above 30. This is a descriptive fact, not a causal conclusion. The same authors have publicly acknowledged that once the fullest macro controls are added, this decline becomes significant only after 2024; and that when the sample is extended back to 2018, the occupations flagged as "most exposed" by this exposure measure were already growing more slowly as of 2020 — before generative AI existed. The current statistical design cannot distinguish "AI substituted for entry-level jobs" from "firms froze hiring in these occupations, and a four-year-wide age band of 22–25-year-olds naturally drains under a hiring freeze." As of July 2026 the study remains an unrefereed working paper, the data are proprietary, and it cannot be replicated externally.
Interests at stake: the data provider is a paying corporate partner of the lab in question and co-brands a public dashboard with it; the paper has no conflict-of-interest statement and no data-availability statement; from the first version in 2025-08 through 2026-07 there is still no record of journal acceptance; and the authors' own dashboard states that what it measures is correlation, not causation.
Independent comparisons give a wide spread of magnitudes: [Single-source, verified] the Dallas Fed, using the monthly population survey, sees only the employment share of 20–24-year-olds in the highest-exposure occupations falling from 16.4% to 15.5% (about one third the magnitude of the former), and states explicitly that the data "do not show AI causing widespread labor market disruption"; the Yale Budget Lab and the Economic Policy Institute, using different exposure measures, both find no relationship between exposure and employment; two institutions each used commercial job-posting data to test directly whether, within the same high-exposure occupation, junior positions fell more than senior ones — and both independently found that they did not; one study using state unemployment insurance administrative records finds that deterioration in high-exposure occupations began as early as early 2022, months before ChatGPT; and another New York Fed study attributes 64% of the rise in young college graduates' unemployment to remote work.
[Single-source, verified; not peer reviewed; causality unresolved] Another study, using state GDP and employment data, computes that a one-standard-deviation increase in AI exposure corresponds to about 7% higher output (the event-study reading at the end of the sample is about 10%) and about 3.9% higher employment. But this evidence cannot be used to say "AI has not affected employment," for three reasons. First, its effects begin in 2021, which the authors attribute to enterprise AI tools, a year and a half before ChatGPT — it is not measuring the same thing at all. Second, decomposing exposure, all of the employment gain comes from collaborative exposure; substitutive exposure has no effect on employment at all (and the authors concede a significant pre-trend in that column). Third, and most importantly, on the same industry × state cells and the same administrative records, the Census Bureau study gets the opposite sign as soon as it cuts by age. +3.9% is an aggregate reading; it masks the age recomposition inside, rather than refuting it.
Three definitional corrections are required: "productivity +10%" is wrong — that is output, and the authors' abstract says 7%; "wages +4.8%" is wrong — 4.8% is the total wage bill (employment × compensation per worker), and the real wage effect the authors estimate separately is only 0.9–1.1% — writing it as "wages +4.8%" overstates the wage effect roughly fourfold; and the distributional conclusion must be carried along: only about 29% of the output increment flows to labor, and the labor share falls about 5 percentage points per standard deviation of exposure. The safe weak conclusion goes only this far: at aggregate resolution, US state × industry cells with high AI exposure show both output and employment rising after 2021, with no aggregate contraction.
These three widely cited figures are exactly how the author himself combined them in a popular commentary piece, while the paper's headline number is 7% — this set of figures entered the evidence chain after a second round of packaging.
[Commercial survey / platform basis] A hiring platform's research arm reports verbatim on 2026-07-08: "Even after the recent rise, software development job postings remain about 27.5% below their pre-pandemic level, while overall job postings are essentially the same as in February 2020." It also contains the sentence this essay most needs to quote: "This suggests demand is growing for experienced professionals who can work with AI, not necessarily a broad-based recovery across all software roles." The increment decomposes as: "senior roles account for 71% of the net increase… and AI-related titles for 37%, with the two categories overlapping" — this is the composition of the net increase from 2025-05 to 2026-05, not the structure of the stock, and the two categories overlap and cannot be added.
But this evidence's framing problem must be written out at the same time: the base date for the report's rebound magnitude is, per the chart note verbatim, "the release date of a certain AI coding product" — and the index on that day is very nearly the lowest point of the entire window, so starting a step away from the historical low mechanically maximizes the measured rebound; the report's own headline is a directional judgment that the underlying level data do not support. Interests at stake: the platform's commercial interest lies in "job-posting counts are a meaningful labor market signal"; the index covers only its own postings; platform share drift contaminates the index; and it continuously revises seasonal factors and restates history, so the same sentence yields different numbers depending on when it was scraped. Its industry classification uses proprietary normalized job titles and cannot be aligned with official occupational classifications.
[Vendor claim + single source + not peer reviewed; Round 3 VETO on quantitative use] One unrefereed preprint uses roughly 100,000 consumer chat records to measure "what share of an occupation's work activities appear in consumer chat logs." On this basis, education occupations overlap markedly more than healthcare-support occupations, with the gap a multiple rather than a few percentage points.
But this direction is to a large extent determined by the instrument: explaining, teaching, and drafting materials are the default output form of a chatbot to begin with (the paper's own over-representation multiples give 6.6× for "teaching subject content" alone), while turning a patient, transferring, drawing blood, and monitoring vital signs cannot in principle appear in a text conversation. The paper states explicitly that it does not know the user's own occupation; its classification pipeline even has a dedicated flag for "is this user a student doing homework," yet it never reports that proportion and runs no sensitivity analysis excluding homework conversations — and "postsecondary teacher" is precisely the cell most easily inflated by students doing homework. The score's discriminating power comes almost entirely from the "coverage" component, so what it measures is topic overlap, not capability, and certainly not substitution risk. The authors warn verbatim that reading high applicability as "will be automated, will lose jobs, or will lose wages" — "This would be a mistake." Interests at stake: all five authors are from that company's research institute, the data are its own product logs, the classifier is its own model, ethics review was internal, the three human annotators used for validation were the paper's own authors, the raw conversations cannot be released, and not a single cell can be reproduced externally. After publication the paper was widely rewritten by media into "Company X lists the N occupations AI will replace," forcing the authors to publish a separate clarifying blog post.
So this essay retains from it only one mechanistic judgment and no numbers: the part of teaching work that can be rendered into text (lesson preparation, explanation, item writing, grading, drafting materials, answering questions) overlaps heavily with current chatbot capabilities; the part of nursing work that can be rendered into text is far smaller. So the blanket claim that "nursing and education are both in the automation-resistant camp" does not stand — but this does not mean the employment outlook for education majors will worsen, which is a separate question answerable only with measured employment and wage data.
[Three-source comparison; the strongest weak conclusion that is load-bearing] At the nursing and care end, low AI exposure is robust — two vendor measurements and one non-vendor measurement agree. The education end is highly dependent on which education you mean: postsecondary and graduate teaching has markedly higher exposure than K–12 teaching, and this within-field difference is consistent across all three datasets as well. The real dividing line is not "education vs. nursing" but "how much of the work is information production that can be rendered into text."
[Single-source, verified; the only non-vendor measurement] The counter-evidence that must be written down alongside: the International Labour Organization's May 2025 working paper (based on a sample of nearly 30,000 tasks from the Polish occupational classification + a survey of 1,640 incumbents + an expert panel, using no AI product logs at all) places primary school teachers, secondary school teachers, nursing professionals, and general practitioners in the same "unexposed" category; its knowledge-work comparison group (financial analysts, data entry clerks) is nearly twice as high as the highest teaching category, university teachers. "Education is close to knowledge work" does not hold on this basis either.
[Single-source, verified; third-party methodological audit] A cross-platform audit demonstrates that the rank correlation of occupational orderings across platforms is only 0.43–0.79 (and that at the coarse level of 22 major groups), while within one vendor's own data, 9 of 22 major occupational groups crossed a quintile boundary over 14 months. Switch vendors' logs, or even switch quarters, and the ordering changes.
Once more, because this section is the easiest to quote out of context: over the past three years, hiring for some white-collar entry-level jobs in the US has indeed cooled markedly, and multiple data sources show it; but on the three questions "was it AI," "how far did it fall," and "will it last until you graduate," no study can currently give a reliable answer. This essay does not use this section to argue anyone into or out of any particular major.
The cobweb model — students chase the current employment signal, and four years later graduate together into an oversupply — is not a theoretical deduction; it is history that can be counted year by year. Three cases each illustrate a different shape.
[Multi-source, Round 3 UPGRADE_MULTISOURCE]
Per the mandatory-disclosure aggregates of the American Bar Association's (ABA) Section of Legal Education, of the 46,776 graduates in the class of 2013, 26,653 (57%) held full-time, long-term, bar-passage-required positions at about nine months after graduation; excluding the 775 positions funded by the law schools themselves, the figure is 25,878 (55%). In the same year, the National Association for Law Placement (NALP), independent of the ABA, measured with its own survey system: overall employment for that class was 84.5%, and bar-passage-required positions accounted for 64.4% of graduates whose employment status was known — tied with the class of 2012 as the lowest in NALP's records; of these, the share that were full-time and lasting at least a year was 59%. Two questionnaires, two institutions, two denominators, pointing to the same trough and the same turning-point year.
By the class of 2025, the same measure had returned to 83% (29,928/36,206).
But two-thirds of that recovery arc comes from the denominator, not the numerator: over the same period the number of positions rose only from 26,653 to 29,928 (+12%), while the number of graduates fell from 46,776 to 36,206 (−23%). Had the class of 2025 been as large as the class of 2013, the ratio would be 64%, not 83%. In other words, the repair came mainly not from the industry taking more people but from the schools sending fewer — which is exactly the supply-side clearing the cobweb model describes, but it is not the same as "the legal job market has recovered."
Six qualifiers, none dispensable:
[Multi-source] The cooling at large firms has two further independent corroborations: NALP notes that multiple firms shrank their 2024 and 2025 summer associate recruiting; and in March 2026 the Law School Admission Council published a piece by a former NALP executive director stating that the latest data show hiring at the largest firms is slowing "measurably," and that the median number of offers for 2026 summer 2L programs is the lowest on record. The same piece supplies a colder coordinate: the total number of entry-level legal jobs is still about 5% below 2007.
The strength of the statement must be lowered one notch: law school is the only case in this essay that completed all four steps — overheating, collapse, supply contraction, and ratio repair; the fifth step, a new round of admissions overheating, currently has only early signs on the input side and does not yet constitute evidence. In particular, the 2026 application surge must not be written up as the start of a new cobweb cycle: the widely circulated "applicants +33%" is a reading from roughly a 15% sample early in the season, and the publisher's own qualifiers are "extremely early data" and "broadly directional at best"; by mid-June 2026 the year-over-year applicant increase was only about +8.5%. Citing 33% without citing 8.5% is selective quotation.
Moreover, the publisher's own list of drivers, in order, is: more schools adopting early decision and rolling admissions; over-competition in the previous cycle prompting earlier submissions; the political environment; and employment uncertainty from the economy and AI. In its survey of test takers, the top three motivations are "to help others," "to advocate for social justice," and "to gain valuable skills" — "rising demand for legal jobs" does not appear at all.
The correct landing point: the new expansion on the application side is not driven by a demand signal; it is driven by risk aversion and identity. What this weakens is not the cobweb model itself but the version of it in which "students rationally chase the current employment signal." The statement that is genuinely load-bearing, and stronger, is: the supply side is re-expanding (42,817 first-year students in fall 2025, the largest class since 2012) while the demand side is softening (large firms slowing, total positions falling, federal hiring frozen) — and that divergence is exactly the precondition of the mismatch around 2010. Close this passage with "early warning," not "confirmation."
For prospective students, the landing point is this: in uncertain years, law school becomes a safe harbor, and a safe-harbor influx is not automatically corrected by wage signals — which makes it more worth watching than the cobweb model's prediction, not less.
Interests at stake: the data are produced by the parties that directly benefit from rankings and admissions (the law schools themselves); the test administrator's application growth is directly tied to its business. One more note: the people filling in the ABA and NALP forms are the same law school career offices, reporting on the same graduates — "independent instruments, shared upstream," not two fully independent observations; and for the class of 2025 only the ABA is currently measuring.
[Multi-source; three seats each fetched the raw HTML and parsed all 58 academic-year rows] Per the National Center for Education Statistics' Digest of Education Statistics long series on degrees conferred in computer and information sciences (bachelor's degrees):
The correct causal reading of the lag (this strengthens rather than weakens the cobweb argument): the dot-com bubble burst in March 2000, yet the degree peak came in 2003-04, about four years later — but that is the four-year pipeline delay of the degree. The 2003-04 peak cohort enrolled around 1999–2000, i.e. chose their field at or before the top of the bubble, and was not a reaction to the crash; the actual behavioral response shows up in degrees conferred only from 2004-05, and the −36.1% decline starts exactly there.
Two uncertainties must be flagged: 2021-22 is preliminary; and the entire long series has been retroactively recoded using the 2020 classification of instructional programs, a classification break that must be disclosed when comparing across forty years. Note also the breadth of the basis: this classification is far broader than computer science itself (it includes artificial intelligence, information technology, information science, networking, information systems security, and so on), while data science and data analytics are not in it.
[Commercial / association survey] On leading indicators, an industry association's annual survey gives a split signal: the turn to negative in newly declared undergraduate computing majors holds only for the pure-CS basis within an 83-unit five-year cohort; widened to 95 units across all computing fields, 2025 new enrollment is flat to slightly up (+1.2%) — both directions must be shown side by side. Moreover the survey's own sampling frame is shrinking (units invited fell from 314 in 2024 to 226 in 2025), so absolute numbers are not comparable across years; the association itself attributes the graduate-level decline to visa and immigration barriers, so it cannot be read as a signal about US domestic demand; and its numbers are on a wholly different basis from the national statistical series above, and must not be drawn on the same chart or in the same sentence. Interests at stake: the association's data are self-reported by member units and the sampling frame is an invitation list the association itself decides — a self-reported, self-selected sample.
[Industry-survey basis] Petroleum engineering's peak of 2,615 comes from a voluntary industry questionnaire administered by one university, not from a census; the official series peaks at about 2,151, so the only permissible wording is "about 2,150 around the 2017–18 academic year." The circulating claim that "the number of institutions offering it fell from 35 to 20" is contradicted by official data: the number of institutions conferring the degree rose from 17 in 2008 to 25 in 2022. The contrary fact must be written down as well: by 2023 lower-division enrollment was already +13%, and 7 of the 10 largest programs saw enrollment rebound. Interests at stake: the surveying party and the surveyed parties belong to the same admissions-interest community within the discipline, and the results are circulated through the industry association's magazine.
[Single-source, verified (first-hand count); the media-compiled series = unverified] In China, program instances of "data science and big data technology" (数据科学与大数据技术) fell from a peak of 250 schools in the 2017 approval year to 13–15 schools in the 2024 approval year, a decay of over 90% in eight years; the peak came in the second year after the program was created. The year-by-year figures may not be quoted verbatim — the media-compiled series uses inconsistent bases from year to year, with an error of roughly ±10–15% per year, and the official 2019 annex is a scanned image with no text layer and cannot be counted. The broad conclusion — peak, final year, and magnitude of decay — is solid; the specific figure for any one year is unusable.
⚠️ For this essay's inference rules, this curve is a counterexample, not supporting evidence. Big-data program instances peaked around 2017, yet demand for data jobs grew strongly for at least five more years afterward. A leading indicator issuing a false alarm five years early is the norm for administrative counts of this kind, not the exception. And after the decay it stabilized at a floor of about 30 a year rather than going to zero, which is the steady state after a finite pool is exhausted and has nothing to do with market signals — the cumulative total by 2023 is about 790 institutions, roughly 62% of all undergraduate institutions nationwide, a fact the rhetoric of "a 90% collapse" conceals.
Here is something this essay must state honestly: the cobweb's shape has historical evidence; its parameters do not.
[Multi-source] The engineering-academy report appendix widely cited as the basis for a "four-year adjustment cycle in the engineer market" is only two pages long, contains a single figure, and has no data table, no regression, no coefficient, no elasticity, and no statistical estimate of any kind; its wording is that "the market is likely to stay at point b," and its figure's own note says it is adapted from a popular-economics blog. The "four years" is an assumed degree-production period and is stated conditionally; the "one year" is the shortest lag over which the supply quantity can be changed (the visa application cycle), not the whole adjustment period, and the original refers to the temporary skilled-immigration visas expanded by the Immigration Act of 1990, not to any one visa category specifically. The final half-sentence must be kept: "the cycles are both faster and associated with smaller wage changes" — talking about "faster cycles" while omitting "smaller wage swings" overstates the cobweb's threat to students.
[Single-source, verified] The load-bearing qualitative conclusion of one US study is that degree completions are "most strongly related to wages observed three years earlier, when students were college freshmen"; its elasticity is about 0.67, but the paper reports no standard errors and its R² is only 0.02 — wages three years earlier explain about 2% of the variation in degree output, which the author calls "modest," and the heterogeneity figures come from 8 public universities in a single state. [Multi-source (working paper version)] A French study simulating a 10% rise in expected earnings obtains elasticities of 0.09 for science, 0.14 for humanities and social sciences, and 0.12 for law/economics/management, with the authors' characterization: "These elasticities appear to be very low, which means that the choice of a major is mainly driven by non-pecuniary factors" — statistically significant but small in magnitude, not a zero response.
⚠️ These two sets of figures may not be set side by side as a cross-national contrast of "elastic in the US / inelastic in France" — the US study's own authors reject that contrast: the French study's percentage-point effects are in fact larger, and the elasticities run the other way purely because the denominators differ (one uses 3 broad groups with large base shares, the other 6-digit detailed codes with an average share of only 2.07%). The original text reads, "we largely share this conclusion."
Conclusion: students do respond to wage signals, but weakly, slowly, and mainly under the influence of non-pecuniary factors. This weakens both the panic that "everyone will pile in" and the optimism that "the market will correct itself."
This is the most original criterion correction in this issue, and the basis for the column structure of the list in Section 11.
The popular criterion is "practice requires a license → supply is limited → overshoot is restrained." This criterion does not hold, and it directly contradicts this essay's own cases:
What determines overshoot is not "whether practice requires a license" but "whether training capacity is physically or human-resource constrained" and "the marginal cost of expansion." A license that binds after graduation (law, pharmacy) imposes no constraint on admissions at all; only a constraint at the entrance restrains overshoot (nursing: clinical placement slots, preceptors, and faculty are all limited simultaneously). And accrediting bodies themselves loosen procyclically during a boom.
At the same time, "overshoot is almost certain to be restrained" is downgraded to "the amplitude and speed of overshoot are markedly smaller and slower": constrained supply can only lengthen the lag and flatten the amplitude; it cannot guarantee that overshoot does not occur.
This criterion is itself a product of this essay's reasoning, not any institution's conclusion — labeled as such. Each of the three counterexamples has primary support.
[Multi-source] The American Association of Colleges of Nursing's (AACN) turned-away series (the same metric on the same fact sheet): 65,766 in 2023 → 80,162 in 2024 → 92,672 in 2025. The full 2025 wording, verbatim: "U.S. nursing schools turned away 92,672 qualified applications (not applicants) from baccalaureate and graduate nursing programs in 2025 due to insufficient number of faculty, clinical sites, classroom space, and clinical preceptors, as well as budget constraints." Of these, 6,496 are master's and 10,359 doctoral, so the entry-level baccalaureate level is about 75,817 — using 92,672 to describe the bachelor's-level training bottleneck overstates it by about 22%.
Even AACN itself appends the parenthetical "(not applicants)" after the number, but media retellings almost universally drop that parenthesis. One applicant may apply to multiple schools, so the true number of people turned away is markedly lower than that value, and AACN has never published a count of applicants turned away.
⚠️ One point that runs against this essay's own claim and must be written out too: AACN's October 2025 faculty survey gives a national faculty vacancy rate of 7.2%, while the same page states that "the 10-year national average full-time faculty vacancy rate was 7.64%." → 2025's 7.2% is below the ten-year average, so it cannot be used as evidence that "the bottleneck is worsening"; it supports only "the bottleneck is long-standing." These two numbers must always appear together.
Interests at stake: AACN is the trade association of US baccalaureate-and-higher nursing schools, and its fact sheet openly states that it is seeking federal appropriations for faculty development programs and has long lobbied on related legislation; the larger the number of applications turned away (the numerator), the better for its appropriations case, and the fact that it chooses to publish "applications" rather than "applicants" must itself be disclosed alongside the number. The faculty survey is a self-reported survey the association distributes to its members (80.3% response rate).
This is the section with the strongest academic evidence in the essay — three Quarterly Journal of Economics papers, one Journal of Human Resources paper, one Economic Inquiry paper, all published and peer reviewed. But the limits on cross-national extrapolation are severe, discussed below.
Splitting "does the major matter?" into three separately answerable questions, the answers are: differences between fields are real and not small in magnitude, but they are largest in the tails and small near the median; institution effects fall essentially to zero once self-selection is controlled (the upper tail excepted); and the dispersion within a single field is no smaller than the dispersion between fields.
[Multi-source, peer reviewed] The early version (1999 working paper, published 2002) found that students who attended more selective institutions earned no more than students admitted to comparable institutions who chose to attend a less selective one; but an institution's average tuition was significantly positively correlated with earnings; and the return to an elite school "appears to be greater for students from more disadvantaged family backgrounds."
The updated version (2011 working paper, published 2014), re-estimated using Social Security administrative earnings records: "the returns to other college characteristics (the Barron's Index and net tuition) are substantial in the basic model…but small and never statistically distinguishable from zero in the self-revelation model." The net-tuition coefficient goes from .058 (significant) in the 2002 version → .041 (not significant) with self-reported earnings → .033 with administrative earnings, and across the whole 1983–2007 period runs "generally between 0 and .02."
⚠️ The direction must be stated correctly: many retellings get this backwards. The tuition premium is precisely the one the new version eliminates using administrative data. The correct statement: the tuition premium exists only in the 2002 version (self-reported earnings); switching to long-horizon administrative earnings data, it goes to zero along with the selectivity premium. (A further note: "net tuition" is sticker tuition minus average financial aid — it is a price, not institutional spending.)
Heterogeneity that still holds: excluding historically Black colleges, for minority students in the 1989 cohort a 100-point higher entrance exam score corresponds to a 12% earnings return and moving up a selectivity tier to 14%; for those whose parents averaged 12 years of education, a 200-point higher score corresponds to 5.2% higher earnings, whereas for those whose parents averaged 16 years "there was virtually no return."
The sample limitation must be quoted together with the authors' own defense: the sample comes from 27 institutions, most of them highly selective; but the same passage immediately continues, "estimates…were similar to — indeed, slightly higher than — those based on a nationally representative dataset." Quoting only the first sentence and deleting the defense is truncation in the opposite direction.
[Multi-source, peer reviewed] Per a study published in the Quarterly Journal of Economics in February 2026: attending the 12 Ivy-Plus institutions (the Ivy League plus Stanford, MIT, Duke, and Chicago) rather than an "average flagship public university" raises a student's probability of reaching the top 1% of the income distribution by 50% (the published version's figure; the original working paper had 60%, revised down), "nearly doubles" the probability of attending an elite graduate school, and "almost triples" the probability of working at a prestigious firm.
The null results matter just as much: "The impact of Ivy-Plus admission on reaching the top quartile of the distribution is small and statistically insignificant"; the effect on log earnings is modest, "consistent with the findings of Dale and Krueger (2002)." The mean effect is $101,000 higher average earnings at age 33, but the original text introduces it with "As a result of these upper-tail impacts" — the mean is pulled up by the upper tail.
Two qualifiers: the comparison group is 9 flagship publics in the authors' own sample, not the average of all US flagship publics; and this must not be written as "elite schools barely matter for the median" — the paper does not directly report a median effect, and the accurate statement is "no significant effect on the probability of reaching the top 25%, and a modest effect on log earnings."
[Multi-source, peer reviewed] A Norwegian study uses nearly the complete records of the centralized higher education admissions system for 1998–2004 plus education and tax registers, instrumenting with admission-cutoff discontinuities while holding the next-best alternative fixed; the outcome variable is annual earnings in the eighth year after application. The abstract, verbatim: "different fields have widely different payoffs… For many fields the payoffs rival the college wage premiums, suggesting the choice of field is potentially as important as the decision to enroll in college."
The frequently omitted second half must be written with it: "by choosing Science instead of Humanities, individuals almost triple their earnings early in their working career. By comparison, choosing Science instead of Engineering or Business has little payoff." — i.e. the conclusion that "field matters" is carried mainly by the "science vs. humanities" end, and among science, engineering, and business there is almost no difference.
The roles of institution and peer quality are minimal: the correlation between return estimates before and after controlling for predicted institution is 0.84, and 0.98 before and after controlling for predicted peer quality.
⚠️ Denominator-swap warning (this is where the value of this item lies): the abstract uses "college wage premiums," while the corresponding sentence in the body reads "rival the usual estimates of college earnings premiums" — the comparison is to the standard estimates in the literature, not to the paper's own raw descriptive mean difference. And carrying "choosing a field matters as much as whether to attend college" into the Chinese or American context swaps not only the denominator but the entire shape of the wage distribution: Norway has centralized admissions, free public tuition, and marked wage compression, so between-field income gaps are compressed to begin with; and the estimate itself is a local average treatment effect for compliers near the cutoff. (The US college premium is fixed here at the official Education Pays 2024 basis of +66%: median weekly earnings of $1,543 for a bachelor's against $930 for high school.)
[Multi-source] A think tank project's study (sample: 80 majors, workers holding exactly a bachelor's degree and no graduate study, a synthetic cohort constructed from cross-sections, not restricted to full-time workers and including part-timers and those with spells of unemployment during the year; all lifetime cumulative earnings are discounted to present value at a 3% annual rate — without the word "discounted," readers will take the million-dollar figures for total wages actually received):
The framing must be got right: A and B are logically entirely compatible (different cross-sections of the same distribution) and are not "two sentences in the same report that contradict each other", but rather "two sentences serving opposite rhetorical purposes while remaining logically consistent." The judgment that "quoting only A yields 'the major doesn't matter' while quoting only B yields 'the major determines everything'" does hold.
The report's self-stated limitations must be quoted: "earnings differences across majors are driven by many factors and do not necessarily reflect a wage premium for that particular major. The estimates cannot distinguish why graduates in certain majors earn more than others." Its own first conclusion is precisely that "a college degree—in any major—is important." Interests at stake: this project is a policy-advocacy project, not a statistical agency.
[Must be relabeled] Another US study of earnings by field must not be called "the causal version," nor placed alongside the Norwegian discontinuity design — the authors state that they use a selection-on-observables model, explicitly note that it requires a "strong assumption," and note that in the US there are "few opportunities to use a regression discontinuity approach." Its results: relative to liberal arts, quarterly returns range from $983 for communications to $7,901 for engineering and architecture (16–20 years after high school graduation); and quantile treatment effects show "substantial (and differential) ex-ante risk."
[Multi-source, peer reviewed] Per a 2020 Quarterly Journal of Economics study (sample: four-year college graduates aged 23–50 and employed full-time during 2009–2017; the omitted group is "all other majors, including education"): "computer science and engineering majors earn about 45% more early in their career, but only 33% more by age 50. The earnings advantage for business majors declines from around 38% initially to 20% by age 50. In contrast, the earnings premium grows over time for life and physical sciences and social sciences majors."
Two sentences that must be quoted as a pair: "In levels, earnings growth is rapid for all college graduates, regardless of major. However, while computer science, engineering and business majors are earning substantially more in their mid-twenties…, this advantage is greatly diminished by age 40." — quoting only the first and cutting the next is exactly the error this essay is guarding against.
The paper's own mechanism conclusion is the single most useful sentence in this section: "Declining relative returns is a feature of STEM jobs, not majors." The share of computer science and engineering majors employed in STEM occupations "declines from 59% at age 26 to 41% by age 50," and those 18 percentage points are almost entirely offset by movement into non-STEM managerial occupations. The authors' own footnote must also be carried: the rapid lifetime earnings growth of life and physical sciences majors is partly due to their very high graduate-school rates, and restricting to bachelor's-only holders slows that growth.
Locked conclusion (downgraded version): the three studies above use different baselines, different measures, different samples and age windows, and different dimensions, and cannot directly corroborate or refute one another. All that can be said is: there is currently no evidence on a comparable basis supporting "field-of-study effects decay generally" — not "the three point in inconsistent directions, therefore no decay exists."
[Multi-source] Three different definitions of "field match," with different numerators and denominators; any arithmetic or comparison across them is wrong:
| Basis | Definition | Value |
|---|---|---|
| A · Respondent self-assessment | How related the job is to the highest degree held (three categories), National Science Foundation National Survey of College Graduates | All: closely 53.7% / somewhat 27.1% / not related 19.1%; for those whose highest degree is a bachelor's, 43.6% / 31.1% / not related 25.3% (only this row approximates a bachelor's-level match rate) |
| B · Analyst cross-coding | Whether the job directly matches the undergraduate field (binary), New York Fed staff report, restricted to metropolitan areas, ages 16–64, excluding all graduate degree holders | About 27% directly matched; computer science 33% directly matched / 73% in jobs requiring a bachelor's degree |
| C · Whether employed in a STEM occupation | Census Bureau basis, ages 25–64 | Engineering 52%, computer/math/statistics 51%, biological/environmental/agricultural 16%, psychology 10%, social sciences 9% |
⚠️ The nesting of the two wage coefficients must be stated clearly, and must not be described as "five times as important": in that 2010 cross-sectional urban wage regression, "the job requires a bachelor's degree" corresponds to about +25% in wages, and on top of that "the job matches the field of study" corresponds to a further +5% — the original reads verbatim, "which in principle is on top of the wage premium for a College Degree Match." They stack rather than being mutually exclusive, and the ratio of coefficients is not a ratio of importance. And these two coefficients are not causal estimates (the paper's topic is agglomeration economies, and the matching variables are merely controls). On the self-reported basis, the mismatch penalty alone is 11–17%, and the two bases are not interchangeable.
[Single-source, verified] A study published in 2024 shows that from 1993 to 2019 the educational mismatch rate fell only slightly, from 19% to 17%, while the wage penalty for mismatch rose 51%. But it uses the same self-reported indicator as basis A, is not an independent measurement, and is an observational regression, not causal. (Its title is "occupation–education" mismatch, not "occupation–field of study" mismatch.)
[Multi-source; the 18% figure is single-source, verified] The share of the workforce covered by state licensing laws rose from under 5% in the early 1950s to 25% in 2008; adding local and federal licensing brings it to 29% (that figure rests on a telephone survey with a sample of only about 2,500). For an official large-sample comparison: the Census Bureau's Survey of Income and Program Participation (n≈58,000) shows that about 20% of civilian employed persons aged 18–64 hold a license, and 28% hold a license or certificate.
The wage premium must be presented as a ladder and never as a single value: 18% (the 2013 published version, a cross-sectional correlation not controlling for education and training) → 10–15% (the same author's 2015 think tank report, restricted to hourly wages and to "universally licensed occupations") → 5–8% (occupations licensed in only some states, the cleanest identification of the group).
The interstate mobility penalty must be taken from the published version: the preferred specification in the 2020 American Economic Journal: Economic Policy paper (restricted to long-distance movers, with a comparison group of occupations subject to a national examination) is "7 percent less likely to move between states." The working paper's 36% used "all other occupations" as the comparison group; the naive estimate without the long-distance restriction is −58% — citing 36% without noting that it is the crudest estimate is exactly the kind of citation this essay criticizes.
⭐ The single most useful cell in this section for choosing a field: that paper itself classifies registered nurses / nurse practitioners into the national-examination group and K–12 teachers into the state-specific-examination group; among its occupation-level results, "Pharmacists and teachers have the lowest relative rates, at −47 and −39 percent." → The mobility penalty exempts nurses and hits teachers squarely. Put differently: if you are unsure which state (or province) you will work in later, the geographic lock-in cost of a teaching license is far higher than that of a nursing license.
On licensing's wage effects for nurses and teachers, the original reads verbatim: "the impact of licensing on earnings is murky, with some studies finding small effects and others finding none" — that is "the evidence is unclear," not "the effect does not exist"; and in quoting this sentence the second half must not be dropped: "by limiting entry or making it more difficult for an individual to be hired for a job in another state."
Interests at stake: the field's leading critic is simultaneously the author of policy-advocacy texts and the principal source cited in the White House report, and his two vote counts — "9 of 11 studies find price increases" and "only 2 of 12 find quality improvements" — both come from a literature set selected by his own academic circle, and must be labeled alongside the numbers (moreover those 11 estimates come from only 5 papers, 4 of them dental studies from 1978–2000; the quality count, on the other hand, checks out exactly). Independent evidence from the opposite position: another study using a quasi-experiment based on staggered state legislation finds a small positive wage effect of registered nurse licensing.
On nurse licensure compacts: the accurate wording is "43 jurisdictions have enacted legislation and about 40 have implemented it"; it must not be written as "nurses move freely" — a study using data on 1.8 million nurses finds that adopting a compact has no measurable effect on nurse labor supply or geographic mobility.
[Multi-source] On the US official major-occupational-group basis, the fastest growing is healthcare support at +12.4% (2024 median annual wage $37,180), with computer and mathematical at +10.1% called "second fastest" by the officials themselves; healthcare practitioners and technical is only +7.2%. The economy-wide average is +3.1%. In absolute terms the two healthcare groups together add +1,715,200 jobs, against only +545,600 for computer and mathematical. (Do not write this as "the official numbers do not support healthcare being the highest-growth track" — the same press release states verbatim that "Healthcare and social assistance is projected to have the largest job growth and be the fastest growing industry sector (+8.4 percent)." On the industry basis, healthcare is both the largest and the fastest.)
The four K–12 teaching occupations are all in decline for 2024–2034 (median annual wage / 2024 employment / rate of change): kindergarten and elementary school teachers $62,310 / 1,539,800 / −2%; middle school teachers $62,970 / 633,700 / −2%; high school teachers $64,580 / 1,094,500 / −2%; special education teachers $64,270 / 559,500 / −1%. But preschool teachers are the exception: $37,120 / 555,100 / +4% — quoting only their low pay without saying they are growing is an omission with directional consequences for prospective students.
The primary sentence for "openings ≠ growth" (verbatim from the kindergarten and elementary teachers page): "Despite declining employment, about 103,800 openings…are projected each year… All of those openings are expected to result from the need to replace workers who transfer to other occupations or exit the labor force" (middle school 40,500, high school 66,200, special education 37,800). The educational instruction and library major group as a whole is +0.6%, a slight increase, not a contraction; and "all from replacement" holds only for those 103,800 kindergarten and elementary teacher openings.
The load-bearing mechanism statement: a stock of 1,539,800 generates 103,800 replacement openings a year (about 6.7% annual turnover), and the replacement flow alone is enough to absorb new graduates.
⚠️ One thing that must not be written as a causal sentence: "low unemployment coexists with negative growth because the supply side is shrinking just as fast" — none of the three seats could trace this to a primary source, and the two numbers are on different bases (the unemployment rate of elementary education majors is by-major data, while −2% is an employment projection for the occupation). An equally plausible explanation is that teacher licensure blocks the entrance, so those who take an education degree but never enter a classroom move into other work and are counted as underemployed (elementary education 16.2%) rather than unemployed. This can only stand as a hypothesis awaiting proof.
[Multi-source] US official 2024–2034 projections: nurse practitioners 320,400 → 448,800 (+40.1%), 2024 median annual wage $129,210, entry-level education a master's degree, 29,500 annual openings for that occupation alone, ranked 3rd on the "fastest growing occupations" table. For comparison: software developers +15.8%, median annual wage $133,080 — higher than nurse practitioners; registered nurses +4.9%, 189,100 annual openings, median $93,600.
Meanwhile another federal agency, the Health Resources and Services Administration (December 2025 edition), projects nurse practitioner supply adequacy at 126% in 2028 / 152% in 2033 / 175% in 2038, verbatim: "the supply of nurse practitioners (NPs) is projected to exceed demand over the projection period; however, distribution remains the most important issue."
The two are not necessarily contradictory: the former measures realized employment (approximating the demand side), while the latter models supply and demand separately. But it must be labeled as a model projection rather than an observation, and this model is highly sensitive to the base year — the prior edition (March 2024, with a pandemic-affected base year) estimated 192% for 2036, and while the new edition revises downward it still shows substantial surplus; in the same revision, the projected 2038 registered nurse shortage was revised down roughly threefold, from 337,970 full-time equivalents in the old edition to 108,960, while the licensed practical nurse (LPN) shortfall widened in the opposite direction. The widely circulated "300,000 nurse shortage" does indeed come from the superseded old edition.
[Multi-source (at the level of the values)] Nurse turnover offers a beautiful methodological case: the intention indicator and the behavior indicator moved in opposite directions from 2021 to 2024. On the intention side (the national nursing workforce survey), the share of registered nurses planning to retire or leave nursing within five years went from 22.1% in 2020 → 28.7% in 2022 → 39.9% in 2024 (the wording of that item changed in 2024, so the series cannot be read as continuous; the burnout indicator, by contrast, fell from 45.2% to 35.4%). On the behavior side, hospital registered nurse turnover actually ran 27.1% at the 2021 peak → 22.5% in 2022 → 18.4% in 2023 → 16.4% in 2024 → 17.6% in 2025 — 2025 is a rebound of 1.2 percentage points, not a continued decline, and the driver of that rebound (rising retirements) points the same way as the "planning to retire" half of the intention data.
The two indicators do not describe the same people: one is a sample survey of license holders nationwide (intentions — and with a worse response rate than is usually cited: 16.9% combined for the mail arm, 9.7% for the email arm, as low as 2.2% in individual states, so 39.9% should be treated as an upper bound); the other is a benchmarking survey of 527 voluntarily enrolled acute care hospital employers (behavior).
Interests at stake: [Vendor claim] the publisher of the behavior-side data is a nurse recruitment and retention services vendor, and the report's executive summary prints sales copy and a phone number outright ("Every RN hired saves $66,081. … Contact …"); its methods section states that hospitals were "invited to participate," enrolled voluntarily, and are a non-probability sample. Both its per-turnover cost figure and its national nurse shortage estimate are self-estimates based on its own questionnaire — a vendor claim: direction usable, magnitude not load-bearing.
The biggest China–US difference is not the level of the numbers; it is reproducibility. This is the correct landing point for the cross-national comparison in this essay, and the only honest one.
[Multi-source] The accurate statement is: China has no publicly released, nationwide, comparable, continuous numerical employment indicator by field of study — no unemployment rate by field, no starting wage by field, no education-mismatch rate by field; exhaustive searching found no official release of any of the three. But "there is no official employment data by field of study at all" is too strong a statement: employment data by field do exist inside the system, are collected, and are used for administrative action (red and yellow cards, orders to suspend admissions, program abolition), and the authorities have publicly released ranked outputs — what those outputs make public is rankings and name lists, not denominators or values.
Three instances, all traceable to primary documents:
At the provincial and institutional levels there are indeed published destination rates by broad discipline or by program: Hunan publishes undergraduate employment rates by broad disciplinary category; Anhui publishes the 21 programs with relatively low destination rates; Jiangxi issues a yellow card when the destination rate falls below 50% and a red card, with an order to suspend admissions, when it falls below 50% for two consecutive years, and Sichuan sets yellow/red cards at two/three consecutive years below 50%; institutional annual employment quality reports generally include destination rates by program.
⭐ The hardest piece of positive evidence: page 29 of the Statistical Indicator System for Monitoring and Evaluating Chinese Education (2025 Edition) (《中国教育监测与评价统计指标体系(2025 年版)》), issued by the Ministry of Education's Department of Development Planning, breaks the indicator "graduate destination rate" (毕业生毕业去向落实率) down verbatim as "by level; by form of employment; by type of provider" — there is no "by field of study." Across all 108 pages, the finest disciplinary dimension for any indicator is "by broad discipline area" (category level). On the same page the officials state a limitation of the destination rate: it "cannot fully reflect the actual person–post match of graduates," or their long-term development — this is the best primary basis for the fact that the authorities acknowledge having no underemployment indicator.
⭐ The correct methodological landing point: China has data by field of study but does not publish the values; the US by-field microdata come from the Census Bureau's American Community Survey (which added the relevant question in 2009), are public and recomputable, and the New York Fed's product is derived from them (note: the Bureau of Labor Statistics itself does not publish unemployment rates by field of study). → The asymmetry is in reproducibility, not existence.
Construct-difference note (required at every cross-national comparison): China's "graduate destination rate" has total graduates as its denominator, and its numerator includes those going on to further study, those in flexible employment, freelancers, and the self-employed — that is, a graduate is counted as "settled" if they enrolled in a master's program, took gig work, or opened an online shop. The US unemployment rate has the labor force as its denominator, and the underemployment rate has employed graduates as its denominator. These are three different constructs. This essay therefore never puts the destination rate and the US unemployment or underemployment rates in the same table, on the same axis, or in the same sentence — that is the usage in this whole essay most likely to cause real harm, because a prospective student will read it directly as "this Chinese major has better job prospects than that American one."
[Multi-source, Round 3 UPGRADE_MULTISOURCE] The General Office of the Ministry of Education's Notice on Further Improving the Statistics and Verification of Employment of Graduates of Ordinary Higher Education Institutions (Jiao Xue Ting Han [2021] No. 19, 教学厅函〔2021〕19 号) was drafted on 2021-05-10 and issued on 2021-05-12. Its final page is marked "(this document is disclosed upon request)," so it is not in the Ministry's proactively disclosed catalogue on its government portal; the full text quoted here is taken from the forwarded copy issued by the Hunan Provincial Department of Education and from the full text posted on the employment website of Shandong University (a ministry-affiliated university and one of the document's addressees), two mutually independent sources, verbatim identical after whitespace-level comparison. (That the ministry's own site does not proactively publish the full text is consistent with the document's "disclosed upon request" marking and does not cast doubt on the text's authenticity.)
Per the 2021 edition of the Definitions and Standards for Graduate Destinations:
Note 1 of the document limits the minimum wage standard to those three categories. The original text says only "pay must meet the local minimum wage standard," without specifying monthly or hourly, and without setting any minimum hours or duration.
⭐ Here is the load-bearing point: in the broadest tier of the official destination rate, the supporting materials are provided by the graduate themselves and adjudicated by their own university's employment office — the auditor and the audited belong to the same organization. Opening an online shop (code 75) counts toward the destination rate and need not meet the minimum wage; doing e-sports or all-media operations (code 76) counts on the basis of materials the graduate signs themselves, once approved at the university and school levels.
The data-release control clause can be carried verbatim as load-bearing: "Before publicly releasing their province's graduate destination rate, provincial employment authorities must verify the data with the Ministry of Education's Department of College Student Affairs, and may not publish without such verification. Universities may not provide their own employment data to other departments or organizations without the consent of the provincial employment authority."
Time limitation: the 2021 formula should be described as "the basis in force for the graduating classes of 2021 through 2023"; from 2024, per Jiao Jiu Ye Ting Han [2024] No. 11 (教就业厅函〔2024〕11 号), a new destination classification is used, the merged tier of "flexible employment" and the entire coding system are no longer in use, and the current formula is "employment-with-an-employer rate + self-employment rate + freelance rate + further-study rate." The thresholds have been tightened at the operational level and remain in force: the current standard requires that employment through other forms of hiring have a term of at least 6 months, and that e-commerce entrepreneurship supply transaction records reflecting normal operation of the online shop.
⭐ A new fact that can be carried in passing: both Jiao Xue Ting Han [2021] No. 19 and Jiao Jiu Ye Ting Han [2024] No. 11 are absent from the Ministry of Education's proactively disclosed catalogue — the two operational documents that determine what counts as "employed" are both disclosed-upon-request documents.
Interests at stake: the document was issued by the General Office of the Ministry of Education, and that same body is the indicator setter, the data aggregator, the verifier, the release gatekeeper, and the party being evaluated; the release-control clause was written by the party being measured.
One further point that must be set alongside: the document's "three prohibitions" (do not set unrealistic targets, do not pile on quotas level by level, do not tie the destination rate alone to counselors' performance pay) are themselves official evidence that quota-pushing exists, and are worth more to this essay's argument than the better-known "four prohibitions."
At the same time this must be made clear: "loose recognition" is an implication of the text, not an official admission — the document's tone throughout is one of strengthening statistical discipline. The two must be stated separately.
[Round 3 VETO on the original wording] The most widely circulated calculation divides the official flexible-employment rate by a figure from a commercial survey, yields a "three- to fourfold discrepancy," and treats it as evidence of the scale of falsification. This calculation is wrong, and this essay does not use it. The reason is that the official "flexible employment" and the commercial survey's "flexible employment" are two sets with an extremely high degree of non-overlap:
| Component | Official flexible employment | The commercial survey's flexible employment |
|---|---|---|
| Employment through other forms of hiring (has wage records, no contract, mostly full-time) | Included | Classified as "employed full-time" |
| Freelance work | Included | Included |
| Self-employment / entrepreneurship | Not included (counted in the entrepreneurship rate) | Included |
| Part-time employment | Not itemized | Included |
The only intersection is "freelance work," and the authorities never publish a standalone figure for that code, so a same-basis comparison cannot be completed with public data. It is entirely possible that the residual equals "employment through other forms of hiring" — the share of young workers nationwide without a written labor contract is far higher than that residual. With no upper bound available for this item, no lower bound can be set for the "falsified share" — and the lower bound is 0.
⭐ A load-bearing substitute conclusion with an independent measurement: [Single-source, verified; independent institution / independent sampling / academic peer review] For the same graduating class (2021) and the same all-levels basis, official administrative data report flexible employment at about 16%, while an independent national sample survey by the research group at Peking University's Institute of Economics of Education (34 institutions, 20,269 responses) measured 11.4% — the official figure is about 4.9 percentage points higher, roughly 40% higher in relative terms. The same survey also shows the class of 2021's destination rate at only 76.5%, with only 32.1% having secured a position with an employer. This direction is consistent with the Ministry of Education's own August 2023 statement that it would "strictly review each graduate's employment materials, with a focus on verifying flexible employment and related data": flexible employment is indeed the cell in the administrative basis most prone to inflation, but the magnitude of that inflation is around 40%, not three- to fourfold.
(⚠️ The 76.5% must not be subtracted directly from the official destination rate — the reference points differ (before leaving campus vs. six months after graduation vs. the two official statistical reference dates), the sampling frames differ, and the educational composition differs. One may say "the two presentations give readers different pictures"; one may not say "the two numbers falsify each other.")
[Single-source, verified] Article 4 of Document No. 19 reads verbatim: "Each year in early September, the Ministry of Education commissions the National Bureau of Statistics to conduct a sample survey of graduate employment status, the results of which will be circulated to each locality." And the National Statistical Survey System for the Employment Status of Graduates of Ordinary Higher Education Institutions (National Bureau of Statistics, 2020-12-16) hardwires "not designed to face the public" into the system: results are "fed back by official letter to the education administrations of the 31 provinces (autonomous regions, municipalities)."
→ The only official independent measurement capable of checking the data universities report is never publicly released. (This essay makes no conjecture about that survey's sample size, sampling frame, or the direction of its results.)
[Multi-source] On falsification, both the official statements and the official silences are checkable: the Ministry of Education's 2023-08-04 release states that "where verification confirms violations such as fake signed contracts or fake certificates, the relevant departments shall be instructed to deal with them strictly according to rules and discipline, and the responsibility of the universities and individuals concerned shall be pursued, so as to protect graduates' lawful employment rights and interests," and that it would "focus on verifying flexible employment and related data" and investigate with "zero tolerance." ⭐ But when three seats independently and exhaustively searched from multiple angles for "notices of disciplinary outcomes" and "numbers of confirmed cases," everything returned was a reprint of that same release, with no named notice, confirmed case count, or disciplinary outcome announcement of any kind. → "Falsification is acknowledged; no public report of how much was found" can be written directly (using "no public notice found," not "does not exist"). The authorities' own stated criteria for detecting falsification are themselves usable material: "flexible employment rate, freelance rate, and entrepreneurship rate too high; share of other forms of hiring too high; suspicious clustering of employment at small firms."
[Single-source, verified + multi-source (monthly values)] The old series ends with the urban surveyed unemployment rate for ages 16–24 of 21.3% in June 2023 (released 2023-07-17, the highest since the statistic began in 2018; publication suspended as announced on 2023-08-15); the new series begins with the December 2023 unemployment rate for the 16–24 labor force excluding students of 14.9% (introduced 2024-01-17). The difference between these two numbers must not be read as improvement — they are not comparable.
⭐ The key blank is a documented fact: the official Explanation Concerning the Improvement of Surveyed Unemployment Rates by Age Group contains not one unemployment rate figure anywhere in its full text, and offers no side-by-side old/new figures, no dual computation over an overlap period, no back-cast series, and no conversion factor. The only quantitative material it provides is the denominator stock: averaged across the months of 2023, students accounted for over 60% of the urban population aged 16–24, nearly 62 million people, and non-students for over 30%, about 34 million. → The statistics bureau explained only "who was removed and how many," and never gave the resulting unemployment rate for the same period. The official reason for the change reads verbatim: "Given China's national circumstances, the main task of students in school is study, not part-time work; including students in the age-group breakdown would mix young people looking for part-time work while enrolled with those looking for work after graduation, and would not accurately reflect the employment and unemployment situation of young people who have genuinely entered society and need work."
The year-over-year figure must be given at the same time: June 2026 is 14.9%, while June 2025 on the same basis was 14.5% — 0.4 percentage points higher than a year earlier. "Three consecutive monthly declines" is the seasonal shape of every March→June; the year-over-year direction is upward, and giving only month-over-month without year-over-year is materially misleading. (Only these two monthly figures are used here; for the same month, ages 25–29 were 7.1% and ages 30–59 were 4.0%.)
Limits on obtaining the evidence must be stated as they are: since January 2024 these figures no longer have a standalone release page, the primary source is the National Bureau of Statistics data release database, all three seats were denied access, and the monthly values were confirmed by cross-checking same-day citations from multiple independent media outlets. This decline in publishability is itself worth writing into the methodology.
Interests at stake: the National Bureau of Statistics is both the setter of the definition and its sole publisher; the definitional change occurred after the old series hit a record high and publication was suspended, and that sequence itself must be disclosed to readers.
One misunderstanding that must be cleared up: "the ILO's one-hour employment standard" is not a reason China and the US are non-comparable — China's surveyed unemployment rate shares a lineage with the US monthly population survey, and the statistics bureau itself states that the one-hour standard is "used to define whether someone is employed, not whether their employment is 'sufficient'." The non-comparability lies in the by-field dimension and in the numerator construct, not in the one-hour standard.
Given that the values cannot be obtained, three structural facts remain load-bearing and checkable.
[Official platform basis + primary commercial survey document; the strongest item in this section, since administrative records are not subject to sampling problems]
So what can be established? Volume. 955 institutions now offer computer science and technology and 661 offer software engineering (per the Ministry of Education's Sunshine Gaokao platform, as of November 2025; the figure is drifting — one commercial admissions platform listed 960 as of 2026-06), and the two programs together graduate more than 100,000 students a year; from 2018 to 2022, 1,047 new program instances were added in the computing category, the most of any program category. (A "program instance" — 专业点 — is one major offered at one institution; it is a count of school–program combinations, not of enrollment.) These come from the Ministry of Education's program filing and approval records, which are administrative statistics, not a sample survey. What computing majors face today is a market whose supply side was substantially amplified within five years — this is the one part that requires no discounting.
Before giving the second structural change, one extremely widely circulated claim — vetoed outright by the methods-audit seat in this issue — must be dealt with.
[Round 3 VETO; the passage below was drafted by the audit seat as finished text and is reproduced as written]
On the question of whether computing majors in China are cooling, first a thing the reader must know: almost every domestic ranking of employment rates by major uses the "graduate destination rate," and that indicator's numerator counts enrollment in graduate school, domestic or abroad, as employed. Which means programs with high further-study rates automatically look like they have "good employment." And in that firm's own data, computing is the program category within the engineering disciplines with the lowest graduate-school rate (13.6% for the class of 2022, against 32.4% for materials and 24.1% for electronics and information in the same year). So subtracting the destination rate of computing from that of history or foreign languages, whose further-study rates are far higher, compares not difficulty of employment but differences in further-study structure. The widely circulated line that "computing graduates now have a harder time finding work than humanities graduates" is essentially a product of this definitional mismatch, and this essay does not use it.
That data source also has several methodological gaps that cannot be ignored. MyCOS (麦可思) is a commercial survey company formerly listed on the NEEQ ("New Third Board"; delisted in 2020), with data-monitoring services accounting for 85% of its revenue, and its paying customers are precisely the universities it surveys and evaluates; its public materials contain no disclosure of interests. Its technical report gives only a sample size of 135,000 and has never published the total number of invitations issued or the response rate — a numerator with no denominator; on the most damaging bias of all, that "people who found jobs are more willing to respond," the report offers only the assertion that "testing found no self-selection sample bias," without specifying the comparison variables, the test method, or the p-value; its weighting uses resampling, and after weighting the regional distribution matches the National Bureau of Statistics data exactly — but weighting of this kind can in principle only correct observable structural differences such as region and institution type, and cannot correct bias correlated with the employment outcome itself. Not once in the entire volume does a confidence interval or standard error appear, so whether the few percentage points separating computing's destination rate from the national average are even distinguishable from the unknown sampling error cannot be judged. The corresponding rankings are likewise not reproducible.
Interestingly, in the same report, computing's monthly income six months after graduation (6,863 yuan, against an undergraduate average of 5,990), employment satisfaction (80%, average 77%), and job–field relatedness (77%, average 74%) are all above the undergraduate average, and six of its constituent programs — network engineering, information security, software engineering and others — have repeatedly made the firm's "green card" list over the past five years, while history has been a "red card" program for several consecutive years in the same system. One indicator says cold, three say hot — which itself shows that no single employment-rate ranking can support a ten-year decision about a field of study.
On that firm's most famous product — the red/yellow/green card lists — the methodological conclusion reached in this issue is more useful than any single year's list.
[Multi-source (for the list text), with one distinction to preserve] The firm's 2010 undergraduate red-card list contained ten programs: animation, law, biotechnology, biological science and engineering, mathematics and applied mathematics, physical education, bioengineering, computer science and technology, English, and international economics and trade; the 2011 list is identical. That year's green card list of nine is entirely resource-extraction and heavy industry: geological engineering, port/waterway and coastal engineering, naval architecture and ocean engineering, petroleum engineering, mining engineering, oil and gas storage and transportation engineering, mineral processing engineering, process equipment and control engineering, and hydrology and water resources engineering — and the firm explicitly recommended that year that institutions "cut admissions to red-card programs and increase green-card admissions correspondingly."
Three mandatory qualifiers:
① This must not be written as "the report was bearish on computing back then and has been proven wrong today." In the same release that assigned the red card, the firm wrote: "Computing programs have a large number of unemployed graduates, but demand for talent in the relevant industries is very strong. In software alone, the annual talent gap runs into the hundreds of thousands." The judgment at the time was "a mismatch between supply quality and demand," not "the industry is failing." ② Red card = top ten by absolute number of unemployed graduates. Contemporary reporting states explicitly that the list came from "the top 10 by number of unemployed," and English was 1st, computing 2nd, and law 3rd — precisely the three largest programs by graduate volume that year. This criterion is mechanically biased against large programs. ③ The load-bearing content should move to two harder places: (i) the green-card reversal — the nine resource-extraction programs recommended that year are today prime "avoid" territory, and on this side there is no disclaimer from the publisher; the reversal is clean and one-directional; and (ii) an extrapolation assumption the publisher wrote down and that was subsequently falsified — both the 2010 and 2011 editions state verbatim: "According to National Bureau of Statistics data, since 2000 the enrollment structure across broad undergraduate and junior-college program categories has been essentially fixed. On that premise, a program's employment and unemployment situation has a certain inertia, and absent intervention, fundamental change is unlikely." The entire early-warning power of the lists rests on that sentence, and the fifteen years since have directly falsified it.
⭐ And the hardest methodological conclusion, one that depends on no reversal narrative at all, is this: in the firm's own definitions, the wording for the red card and the yellow card differs by only five characters (the yellow card is the same sentence with "other than the red-card programs" prepended). That is, there is no substantive criterion distinguishing red from yellow at all; there is only a ranking cut line. And the first criterion in the ranking is the absolute number of unemployed graduates, which must equal "that program's non-employment rate × the national number of graduates in that program" — the size term has no error and is enormous, the rate term has sampling error and is tiny, so the ordering is driven mainly by how many people are in the program. The lists have never published a sample size, confidence interval, or weight threshold for any program (across the 235-page volume, the terms "weight," "threshold," and "evaluation standard" all return zero hits; "unemployment volume" appears only 4 times in the whole book and is never defined; the stated criteria are "unemployment volume, destination rate, salary, employment satisfaction and other employment indicators" — that "and other" leaves the criterion set itself open-ended). → Structurally, the red/green card lists cannot distinguish "this program is bad" from "this program is big," and no candidate should treat them as a basis for application. The firm itself writes that "red, yellow, and green card programs reflect the national overall situation; individual provinces and institutions may differ" — which is itself the publisher's own admission of non-extrapolability to the individual. The rules also change without notice: the 2023 edition states in-book that "some programs added in large numbers in recent years (such as artificial intelligence, data science and big data technology, and robotics engineering) are not yet included, as there is not yet employment data on a meaningful scale or trend for their graduates," yet robotics engineering already appears on the green-card list in the 2024 edition — a rule change one year later with no explanation whatsoever.
Interests at stake: the firm is a for-profit commercial survey company; its founder holds 61.61%; it delisted from the NEEQ in 2020 (and has had no mandatory disclosure obligation since); 85% of its revenue comes from data-monitoring services sold to universities, departments, and provincial education bureaus; universities are simultaneously its paying customers and its sample source; the red/yellow/green card list is its most viral annual marketing asset; and the 2010 red-card list was disclosed by media on May 10 while the book's official launch was June 2 — the list was pushed into the gaokao application season three weeks before the book was published; and the list's policy recommendations point directly at its own sellable services. All of its data are commercial-survey data and may not be treated as load-bearing alongside official statistics. One distinction must also be preserved: the multi-source confirmation of the red/green card lists confirms the text of the lists, not an independent multi-source measurement of that year's graduate employment situation — which does not exist in China and cannot be constructed.
[Multi-source (direction); Round 3 VETO on the magnitude]
The Hubei Province Public Recruitment Announcement for Primary and Secondary School Teachers, published annually by the Hubei Provincial Department of Education and the Provincial Department of Human Resources and Social Security, shows the provincially unified recruitment plan declining for four consecutive years: 11,653 in 2023, 9,257 in 2024, 5,799 in 2025, 2,740 in 2026.
These four numbers describe one province and one channel only and are not equivalent to the total number of teaching posts, for three reasons. First, the plan covers unified recruitment into establishment posts (编制 — budgeted, permanent public-sector headcount slots) for kindergarten and compulsory education (including rural compulsory-education teachers recruited independently by localities), but excludes the Special Post Teacher Plan, the placement of publicly funded normal-university graduates fulfilling their service agreements, off-establishment contract hires, private schools, and regular high school teachers. Second, the 2026 post table no longer includes kindergarten teacher posts or posts reserved for demobilized military personnel, whereas the 2025 table did; the two years are on different bases, neither announcement discloses the category composition, and the figures cannot be converted to a comparable basis — which is why this essay gives no percentage decline. Third, recruitment plan numbers are a product of establishment quotas and fiscal budgets; establishment freezes, fiscal pressure, and county-level staffing reallocation policies could all depress them even with school-age population unchanged.
The same direction is corroborated on other independent bases: the annual recruitment announcements of the Jiangxi Provincial Department of Education and Department of Human Resources and Social Security show the provincial plan falling year by year from 7,821 in 2023 to 3,957 in 2024, 2,146 in 2025, and 1,190 in 2026; and the national rural compulsory-education "Special Post Plan" allocation issued by the General Offices of the Ministry of Education and the Ministry of Finance fell from 37,000 in 2024 to 20,976 in 2025 and 8,594 in 2026.
This evidence therefore supports only a directional judgment: establishment-post channels for replenishing teachers are tightening in some Chinese provinces. It cannot show that the total number of primary school teaching posts in China has already contracted, nor can one province be extrapolated to the country.
In particular, recruitment cuts must not be juxtaposed with the decline in primary school enrollment to claim that together they prove "posts are already contracting" — the causal timing does not work: the 2024 decline in primary school enrollment affects the first-grade cohort entering that year, and that cohort will not finish primary school until around 2030; the 2026 recruitment cut cannot mechanically have been "already" caused by the 2024 enrollment decline — at most it is a forward-looking contraction of establishment quotas. Describing an anticipatory policy action as an already-realized change in the stock is a tense error.
⭐ Negative results reported as they are: a dedicated search was run for whether any province's 2026 provincial teacher recruitment plan rose, and no contrary official announcement was found — this may be written as "no counterexample found," and must not be written as "no exception nationwide."
Interests at stake: the publishing bodies are provincial education and human-resources departments, and the announcements serve talent-attraction programs — if the publisher is biased, the motive runs toward overstating rather than understating, so a declining signal is unlikely to be manufactured by the publisher. But the conflicts of interest along the transmission chain are severe: the whole narrative of "recruitment halved in many places" and "hundreds-to-one competition for some subjects" is carried by admissions-consulting social media accounts whose monetization directly benefits from manufacturing anxiety about choosing a major; the three seats between them could substantiate only a single province, and that magnitude has been vetoed. This item should be treated as "a single point surviving from a list of high-conflict-of-interest sources," not as "independent evidence."
[Multi-source; three seats each fetched the same release page, and all 19 industry categories × 2 bases agree cell by cell] Per the National Bureau of Statistics' Average Annual Wages of Employees in Urban Units in 2025 (2026-05-15):
Four definitional notes, none dispensable: (a) total wages are pre-tax and include individually paid taxes, social insurance, housing fund contributions, and housing/utility charges; (b) coverage is limited to legal-person units with 5 or more employees, so self-employed businesses and freelancers are not in the sample — and those two categories are precisely an important absorber of youth employment; (c) the release page contains not a single instance of the word "median" — these are weighted arithmetic means; (d) they are averages across all incumbent employees, not new graduates' starting salaries.
Construct differences for cross-national comparison: the statistics bureau does publish the median of household per capita disposable income, but does not publish median wages by industry; whereas the US by-field table gives medians. A mean and a median are not the same quantity and may not be placed side by side. This essay compares direction only: both sides show that IT's relative advantage persists while its growth rate has been overtaken by other industries.
[Single-source, verified + multi-source] A third straight decline in graduate-school applications coincides with the first reduction in the national civil service intake: 3.43 million registered for the 2026 national master's entrance examination, 450,000 fewer than 2025's 3.88 million (−11.6%), the third consecutive annual decline since 2024, and it should be described as "the lowest since 2020 (3.41 million)" (full series: 2.90 million in 2019 → 3.41 million in 2020 → 3.77 million in 2021 → 4.57 million in 2022 → a peak of 4.74 million in 2023 → 4.38 million in 2024 → 3.88 million in 2025 → 3.43 million in 2026). Over the same period, the 2026 national civil service examination planned to hire 38,100, 1,602 fewer than the previous year's 39,700, the first reduction since 2019; 3.718 million passed qualification review (an all-time high, +8.8%), giving a ratio of qualified applicants to planned hires of about 98:1 (about 86:1 the previous year) — this rise comes simultaneously from a rising numerator and a shrinking denominator, squeezed from both ends. A further note: the age ceiling for general positions was relaxed from 35 to 38, and for new master's and doctoral graduates from 40 to 43, with the official rationale being the gradual raising of the retirement age; relaxing the age limit mechanically inflates the applicant base, and the authorities have not broken out how much of the increase it accounts for.
A definitional reminder: 3.718 million is the number passing qualification review, not the number who registered, and certainly not an acceptance rate; the graduate exam's registration count and the civil service exam's qualification-review count are on different bases and may not be compared side by side. Only one inference is load-bearing: the fact that "graduate applications −11.6% and civil service qualifications +8.8% happened at the same time" refutes the narrative that "young people are withdrawing from competition across the board." (A news agency article titled "A Return to Rationality in Applications" carries the subtitle "expert analysis" — it is an interpretive frame supplied by invited experts, not an official interpretation; the Ministry of Education's release that day gave figures only and offered no interpretation. This in fact strengthens the position that neither explanation can be falsified.)
[Multi-source] Definitional facts on program abolition and admissions suspension: the Ministry of Education, 2025-04-22, verbatim: "nationwide, institutions added 1,839 program instances, adjusted the degree category or program length of 157 program instances, suspended admissions for 2,220 program instances, and abolished 1,428 program instances"; 2026-04-28, verbatim: "during the 14th Five-Year Plan period, institutions nationwide added 10,200 undergraduate program instances and abolished or suspended 12,200… a cumulative adjustment ratio of over 30%, and this year the national program adjustment ratio exceeded 10% for the first time." The Catalogue of Undergraduate Programs (《普通高等学校本科专业目录》) currently covers 13 disciplinary categories, 92 program categories, and 883 program types; there are 62,800 undergraduate program instances nationwide.
Three definitional reminders: the unit is program instances, not people; suspension ≠ abolition; and the national catalogue contains only 883 program types in total — useful for contrasting with media misreadings like "more than three thousand majors have been cut." The correct reading of the stock of program instances is: actively enrolling instances are contracting while the stock of instances is still growing on net (a net increase of 411 in the 2024 cycle), and at the same time the number of graduates is rising (9.09 million for the class of 2021 → 11.79 million for 2024 → 12.22 million for 2025 → a Ministry of Education projection of 12.70 million for 2026). In the catalogue's structure, the number of program types is a clean net increase of 38 with none deleted; what has been reorganized is the framework of disciplinary categories and program categories.
⭐ The locked policy-chain statement: an employment rate that is too low is an officially stated trigger for "suspension of admissions," and the stated condition for abolition is "no admissions for five consecutive years" — the two are linked by the same policy chain. Primary basis: Jiao Gao [2012] No. 9, Article 26, "…where the employment rate is too low, the higher-education authority in charge shall order the institution to rectify within a time limit and suspend admissions"; Jiao Gao [2023] No. 1, Article 13, "where conditions for provision are seriously inadequate, teaching quality is poor, or the employment rate is too low, admissions shall be ordered suspended and rectification required within a time limit," and Article 16, "programs at institutions that have not admitted students for five consecutive years shall be abolished"; Jiao Gao Si Han [2025] No. 3, "for programs and closely related programs that are numerous in the region and have low employment rates, additional offerings shall in principle no longer be supported"; Jiao Jiu Ye [2024] No. 5, "implement a red- and yellow-card alert system for programs with low employment quality" and "use graduate employment status as an important basis for allocating institutional resources, evaluating teaching quality, and setting admissions plans." → So "many abolitions/suspensions = bad employment" is not folk speculation; it is the intended outcome of the policy design. What candidates really need to be warned about is the lag — it can take five years or more from a too-low employment rate to eventual abolition.
⚠️ Where this item does readers the most harm is the abolition rankings: [Unverified; underlying table of unknown provenance] the Ministry of Education's annual notice annexes for 2023 and 2024 contain no abolition list at all (full-text searches of both annexes return zero hits for "abolish"), and the source table behind the various circulating "top-5 abolished programs" lists is of unknown provenance. Locked statement: whether a program's name contains the words "information / computing / network," whether it belongs to the computing category, and whether it is being phased out are three separate things. The programs leading abolition counts are management and mathematics programs carrying "information" in their names (information management and information systems is a management program; information and computing science is a science program), along with programs with no IT flavor at all such as public administration, marketing, and product design; no core computing program appears in any of these top-5 lists. Citing any abolition ranking requires stating three things at once: ① there is no denominator for an abolition rate (nowhere is an abolition rate given); ② suspensions do not enter the abolition ranking (in the same year 2,220 suspensions exceed 1,428 abolitions, and suspension is often the precursor to abolition); ③ the underlying table's provenance is unknown.
[Commercial survey] Hiring platforms' rankings are customer-acquisition and brand-marketing products; the denominator of the data is their own platform's users; there is no external audit and no disclosed methodology. Four construct warnings:
One further definitional fact can be nailed down: the denominator of one platform's "share who have received an offer" is verbatim "among new graduates with job-seeking plans," which excludes those going on to further study, preparing for civil service exams, or deliberately delaying employment, so it cannot be used to back out an employment rate. And its industry growth ranking is year-over-year growth in industry job postings, not majors and not headcount hired — treating industry growth as a basis for recommending a major is exactly the construct slippage this essay criticizes; moreover the report carries its own anti-hype qualifier: "a high growth rate does not mean the largest total number of positions," and citing the growth ranking without citing that sentence amounts to making an inference on the publisher's behalf that the publisher itself declined to make.
⚠️ First, the nature of this section, stated flatly: the list below is this essay's way of organizing its material, not any institution's conclusion. It is not ordered by "hot/cold," it assigns no score to any field, and it does not predict which field will do well in ten years. What it does is apply the conclusions of the previous ten sections to three transferable criteria — the criteria matter more than the list, because in four years the list will be stale and the criteria will not.
The three criteria are: the codability of the skills, capacity constraints on the training side, and supply elasticity and current penetration. The "load-bearing evidence" in each column attaches only to that column's criterion, and promises no employment outcome whatsoever.
Criterion source: Section 4. The dividing line is not "technical vs. non-technical," nor "education vs. nursing," but "how much of the work is information production that can be rendered into text."
| Field / direction | Position within this column | Load-bearing evidence | Evidence grade |
|---|---|---|---|
| Nursing, rehabilitation therapy, medical imaging technology, and other hands-on medical care | Low-codability end (robust) | Two vendor log measurements and one non-vendor measurement (the ILO, based on occupational task sampling and surveys of incumbents, using no AI product logs) all place nursing occupations in the low-exposure group; turning a patient, transferring, drawing blood, and monitoring vital signs cannot in principle appear in a text conversation | [Multi-source (direction)] + [Vendor claim (two sets)] + [Single-source, verified (non-vendor measurement)] |
| K–12 education | Low end, but the gap from the row above is far smaller than popular claims suggest | The only non-vendor measurement places primary school teachers, secondary school teachers, nursing professionals, and general practitioners in the same "unexposed" category | [Single-source, verified] |
| Postsecondary teaching, writing, translation, junior legal drafting, general copywriting | High-codability end | Teaching, explaining, and drafting materials are the default output form of a chatbot to begin with; three datasets consistently show postsecondary and graduate teaching markedly more exposed than K–12 teaching | [Vendor claim + single-source, verified] — no rankable number may be given in this column |
| Software engineering / computer science | Internally split; cannot be filed as a whole | In the official ten-year projections, the growth rate for software developers alone is higher than for the group combining QA analysts and testers — the most easily automated testing portion grows more slowly, direct evidence that the split occurs within the field; on the job-postings side, the net increase is concentrated in senior roles and AI-related titles | [Multi-source (official projection basis)] + [Commercial survey (postings side)] |
The strongest qualifier for this column: no per-occupation "AI exposure" score is load-bearing — the rank correlation among three measurements at the detailed-occupation level is only 0.427; a third-party audit shows cross-platform rank correlations of 0.43–0.79 (and that at the level of 22 major groups); and within one vendor's own ordering, 9 of 22 major groups crossed a quintile boundary within 14 months. Switch vendors' logs, or even switch quarters, and the ordering changes. [Single-source, verified; third-party methodological audit]
Criterion source: Section 6. The question is "what extra inputs does a school need to open a new program instance," not "does this line of work require a license."
| Field / direction | Position within this column | Load-bearing evidence | Evidence grade |
|---|---|---|---|
| Nursing, medicine, and engineering fields requiring labs or accredited clinical beds | Expansion constrained physically and by human resources | Three consecutive years of the trade association's turned-away-applications series; the stated reasons are faculty, clinical sites, classroom space, preceptors, and budget | [Multi-source (values)] — but the interests must appear with the numbers: the publisher is a trade association, and the larger the number turned away the better for its appropriations case; it publishes "applications," not "applicants" (the parenthetical is its own, and media retellings generally drop it); master's and doctoral levels are about a fifth of the total, so using the total to describe the bachelor's-level bottleneck overstates it by about 22% |
| Same as above | But "the bottleneck is worsening" does not hold | The same page states that the 2025 faculty vacancy rate of 7.2% is below the ten-year average of 7.64% | [Single-source, verified] — these two numbers must always appear together |
| Law, management, most computing fields, most humanities | Low marginal cost of expansion | Three counterexamples: US law requires an accredited school plus a license and is still the textbook overshoot case; pharmacy has all three thresholds and still nearly doubled and produced a glut; nurse practitioner programs grew several-fold within a decade. On the China side: computing added the most new program instances of any program category in 2018–2022 | [Multi-source (each counterexample has a primary source)] + [Primary commercial-survey document (the China-side new-instance count)] |
| The criterion itself | — | "Whether practice requires a license" is not a criterion; a license that binds after graduation (the bar exam, the pharmacist exam) imposes no constraint on admissions, and only a constraint at the entrance restrains overshoot; and accrediting bodies loosen procyclically during a boom | A product of this essay's reasoning, labeled as such; not any institution's conclusion |
The strongest qualifier for this column: constrained capacity can only lengthen the lag and flatten the amplitude; it cannot guarantee that overshoot does not occur. "Overshoot is almost certain to be restrained" is wrong; the correct statement is "smaller in amplitude, slower in speed."
Criterion source: Sections 5 and 10. The question is "over the four years between your enrolling and your graduating, how many people will end up with the same label as you."
| Field / direction | Position within this column | Load-bearing evidence | Evidence grade |
|---|---|---|---|
| China: computer science and technology, software engineering | Penetration already extremely high | 955 institutions offer computer science and technology and 661 offer software engineering (official platform basis, as of November 2025); the two together graduate more than 100,000 a year; computing added 1,047 new program instances in 2018–2022, the most of any category | [Official platform basis; administrative records, not subject to sampling problems] — the strongest cell in this section |
| China: artificial intelligence | The label is no longer scarce | As of the 2024 approval cycle, about 626 undergraduate institutions nationwide had filed this program, roughly half of the country's 1,257 undergraduate institutions | [Single-source, verified (first-hand count of the primary annex)] — the only robust implication is that the label itself is no longer scarce. It cannot tell you anything about supply and demand for jobs in 2030 |
| US: computer and information sciences | Two complete cycles already run; currently in the third at a high level | Forty-year series: two peak-to-trough declines of −42.1% and −36.1%; the third ascent is +186% from the trough; the degree pipeline has a four-year delay, so the behavioral response appears in degrees conferred about four years after the crash | [Multi-source; three seats checked all 58 academic-year rows cell by cell] |
| US law | The only case to complete all four steps; currently a divergence of "supply re-expanding + demand softening" | The class of 2013 trough was computed independently by two institutions, two questionnaires, and two denominators; two-thirds of the repair came from the denominator; the fall 2025 first-year class of 42,817 is the largest since 2012; the slowdown at large firms has three independent corroborations | [Multi-source (UPGRADE_MULTISOURCE)] |
| China: teacher-training fields (the establishment-post channel) | One channel is tightening, but the magnitude is not credible | Three independent bases (two provinces' recruitment announcements + the two central ministries' Special Post Plan) point the same way; a dedicated search found no contrary provincial announcement | [Multi-source (direction)] — the magnitude was vetoed by the methods-audit seat; this essay gives no percentage decline, and one province cannot be extrapolated to the country |
| Resource extraction and heavy industry (mining, petroleum, naval architecture, etc.) | A sample of historical reversal | The nine programs given green cards and recommended for expanded admissions by a commercial survey firm in 2010–2011 are today prime "avoid" territory; on this side there is no disclaimer from the publisher, and the reversal is clean and one-directional | [Multi-source (list text)] — but note: multi-source confirmation of the list confirms "the text," not an independent measurement of employment outcomes |
| Weaknesses of the criterion itself | — | The cobweb's shape has historical evidence; its parameters do not: the widely cited "four-year adjustment cycle" comes from a two-page appendix containing no statistical estimate of any kind; and students' elasticity with respect to wage signals is very low (one US study has an R² of only 0.02; a French study's authors call it "very low," with non-pecuniary factors dominant) | [Single-source, verified + multi-source (working paper version)] — the two sets of elasticities may not be set side by side across countries |
First, where the returns come from (Section 7): institution effects fall essentially to zero once self-selection is controlled (the upper tail excepted — elite institutions mainly affect the probability of reaching the top 1% of income, elite graduate schools, and prestigious firms, while the effect on the probability of reaching the top 25% is "small and statistically insignificant"); between-field gaps are real but carried mainly by the "science vs. humanities" end, with almost no difference among science, engineering, and business; and within a single field, moving from the 25th to the 75th percentile typically doubles or even triples lifetime earnings, and that span is wider for lower-earning fields. [Multi-source + peer reviewed] Limits on cross-national extrapolation: Norway has centralized admissions, free public tuition, and marked wage compression, so between-field income gaps are compressed to begin with, and the estimate itself is a local effect for compliers near the cutoff; the US percentile study includes part-timers and people with spells of unemployment during the year, while the decay study is restricted to "full-time employed" — when the two are juxtaposed, that difference is decisive.
Second, return decay is a property of occupations, not of majors: the share of computer science and engineering majors employed in STEM occupations falls from 59% at age 26 to 41% at 50, and those 18 percentage points are almost entirely offset by movement into non-STEM managerial occupations. The premise that "choosing this major means doing this for life" is not supported by the data itself. [Multi-source + peer reviewed]
One last emphasis: what hangs in every cell above is evidence about a criterion, not a promise about employment. The correct use of this list is to test the recommendations you read elsewhere — not to take it as a recommendation.
Ordered by strength of evidence.
This subsection is the most valuable methodological output of this issue, which is why it comes first.
(1) "How many schools newly added an artificial intelligence program in the Ministry of Education's annual batch."
Many people will tell you to watch this number and see whether it falls. The advice sounds testable but is unusable, for four reasons:
For contrast: program instances of "data science and big data technology" peaked in 2016–2017, yet demand for data jobs went on growing for more than five years afterward. A leading indicator issuing a false alarm five years early is the norm for administrative counts of this kind, not the exception.
So this essay does not list it as a test signal. If you see someone elsewhere using this number to tell you that AI programs "have already peaked," that is an argument you can now dismantle yourself.
An institutional fact you will run into (worth knowing, but it is not a signal): from the 2025 batch onward, the Ministry of Education changed its release granularity — in April 2026 it published only the Catalogue of Undergraduate Programs (2026), with no accompanying school-by-school list; searching the policy document library by the title "results of undergraduate program filing and approval" returns only the eight documents for 2017–2024. The lists were sent down to provincial departments and institutions, but the public must piece things together from provincial education departments' press releases and third-party organizations' incomplete compilations. The narrowing of release granularity is worth knowing in itself, but it cannot substitute for a signal.
(2) Any single year's employment-rate ranking of majors.
Reasons in Section 10.2: the main indicator in domestic rankings has further study in its numerator, so programs with high further-study rates automatically look like they have "good employment"; and the first criterion for the red and yellow cards is the absolute number of unemployed graduates, which is driven mainly by how many people are in the program. Structurally, rankings of this kind cannot distinguish "this program is bad" from "this program is big."
(3) "Occupation X will have a gap of XX0,000 over ten years."
Reasons in Section 2: this number is a ten-year annual-average flow, of which 97.2% comes from separations, and 55.5% of those separations are moves into a different occupation. A large "gap" almost never means abundant opportunity; it mainly means people leave fast.
Strength of evidence: ★★★★★ [Multi-source, an identity, fully reproducible by the reader]
A test you can run yourself: open the BLS employment projections' occupation.xlsx and find Table 1.10. Take any occupation you have seen described in admissions advice as having a "huge gap," and divide that row's annual openings by its total separations. If the ratio is ≥ 1, that occupation's total employment is shrinking over the next decade — the entire "gap" comes from people leaving. Then divide openings by that occupation's 2024 employment stock, and you get its annual turnover rate; the higher the turnover, the larger the "gap," and it has nothing to do with the industry expanding. The corresponding move on the China side: when reading a program's employment quality report, first ask "who is in the denominator" — all graduates, or "graduates with job-seeking plans," or "those already employed full-time"? The three give completely different pictures.
Strength of evidence: ★★★★★ [Multi-source, multiple official evaluation rounds, with a naive benchmark, against-interest]
How to test: go to bls.gov/emp/evaluations and open any round's "Occupational Projections Evaluation." Look at the major-group scorecard first, then the detailed-occupations scorecard, and compare the official model's win rate against the naive model's. You will see with your own eyes the gap the same evaluation gives at two levels of aggregation. A quicker move: take the list of "20 fastest-growing occupations" from the projection edition of ten years ago and look up each one's actual employment today. Mind the baseline: drawing 20 at random from 338 occupations, you would expect about 1.2 hits — the standard is not "is the hit rate high" but "is the hit rate clearly above 1.2 out of 20."
Strength of evidence: ★★★★☆ [The criterion is a product of this essay's reasoning, but the three counterexamples and the association data are all primary]
How to test: for each field you are considering, ask three questions. ① What extra inputs does a school need to open a new program instance? Only classrooms, slides, and a few reassigned faculty (law, management, most computing fields) → supply can expand extremely fast; clinical placement slots, preceptors, labs, accredited beds (nursing, medicine, some engineering) → expansion is physically constrained. ② Does the constraint bind at the entrance or at the exit? A license that binds after graduation imposes no constraint on admissions. ③ Will accrediting bodies loosen procyclically during a boom? (Historical answer: yes.) Public data you can watch: in the US, the nursing colleges association's annual "applications turned away" and "faculty vacancy rate" — remembering that applications ≠ applicants, and that these must be compared to the ten-year average rather than to last year; in China, the urgently-needed-program list and program-warning list published each year by your province's education department, plus the number of institutions offering a given program in your province — once the institution count approaches the province's total number of undergraduate institutions, that label's scarcity is gone.
Strength of evidence: ★★★★☆ [Multi-source + peer reviewed, but with severe limits on cross-national extrapolation]
How to test: wherever you are given a "major income ranking," ask whether it gives percentiles. Rankings that give only medians or means systematically conceal the information that matters most to you — within a single field, moving from the 25th to the 75th percentile typically doubles or even triples lifetime earnings, and that span is wider for lower-earning fields. A concrete move: in the New York Fed's "Outcomes by Major" table, look at the early-career and mid-career columns together, and then look at that field's underemployment rate. Fields with high underemployment but low unemployment (biology 51%, psychology 48%, fine arts 59%) tell you their graduates can find work but cannot find work that requires the degree — which predicts your actual situation better than the unemployment rate does.
Strength of evidence: ★★★★☆ [Multi-source + published in 2026; but the comparison group is 9 flagship publics from the authors' own sample]
How to test: ask yourself "which part of the distribution is the outcome I want in?" If the goal is the top 1% of income, an elite graduate school, or a prestigious firm, the choice of institution has a substantive effect; if the goal is the top 25%, the best study available measures an effect that is "small and statistically insignificant." An observable signal: when reading a target institution's graduate destinations report, do not look at average salary; look at the absolute number of upper-tail destinations (how many went to a handful of target employers or programs), then divide by the total number of graduates from that institution in that field.
Strength of evidence: ★★★★☆ [Multi-source, a forty-year series; the law school four-step arc is dual-sourced]
How to test: for any claim that "field X is cooling / recovering," demand a series of at least five consecutive years. A single year's direction is not a signal — the median absolute year-over-year change in unemployment rates by field is about 1.25pp, and 20 of 73 fields change by more than 2pp in a year. Checkable on the US side: the National Center for Education Statistics' Digest of Education Statistics long series on computer and information sciences degrees conferred (1964-65 to the present), where you will see two complete declines of −42% and −36%. Checkable on the China side: the annual admissions plan books by program published by each province's admissions examination authority (countable school by school, corresponding directly to headcount supply), and the year-by-year change in admission ranks and cutoff scores for your target program (each province's score-distribution table plus institution-program-group cutoff lines) — these two are primary data you can look up yourself during application season, and they have no "school count ≠ headcount" unit mismatch.
Strength of evidence: ★★★☆☆ [The policy-chain provisions are multi-source; the ranking's underlying table is unverified]
How to test: abolition is harder behavioral evidence than addition, because it carries a real cost. The Ministry of Education's annual batch includes abolition and suspension figures (in the 2024 cycle, 1,428 program instances abolished and 2,220 suspended nationwide). But before citing any abolition ranking you must ask three questions: ① is there a denominator for the abolition rate? (Nowhere has one ever been given.) ② Do suspensions enter the ranking? (No — and suspension is often the precursor to abolition.) ③ Where does the underlying table come from? (The Ministry of Education's annexes for 2023 and 2024 contain no abolition list at all.) A move closer to you: check your own province's program-warning list and urgently-needed-program list (already covering 473 program types), and whether your target institution's program has ever received a provincial yellow or red card (Jiangxi issues a yellow card below a 50% destination rate and a red card, with suspension of admissions, after two consecutive years below 50%). A provincial list is closer to the decision you have to make than the national count of newly added schools.
One more thing that can serve only as a supplement, and whose coverage must be stated: the Public Notice of Undergraduate Program Applications published on the Ministry of Education's government services platform each August–September is the earliest reading of "where institutions want to crowd in this year." ⚠️ But it covers only state-controlled programs and programs outside the catalogue (i.e. the approval track), not the filing track — artificial intelligence goes through filing and is not within the notice's scope at all. So it cannot be used to watch AI programs.
Strength of evidence: ★★★★☆ [Multi-source, verbatim provisions of normative documents, verified as character-for-character identical across two independent chains]
How to test: do not use the national flexible-employment rate as a baseline for subtraction (that mistakes definitional differences for falsification). The correct move: open your target institution's Graduate Employment Quality Report and make a within-institution cross-program comparison — look at the sum of "flexible employment + awaiting employment + not currently seeking employment" for a given program, and whether its flexible-employment share is markedly higher than that of other programs at the same institution. A within-institution comparison is unaffected by definitional differences, and that is the usable criterion. Institutional background: freelance work is verified "on the basis of supporting materials signed and confirmed by the graduate themselves, adjudicated by the responsible officers of the employment offices at the university and school levels" — the verification materials are produced by the person being counted, and the auditor is the unit that is itself being evaluated; e-commerce entrepreneurship counts toward the destination rate and need not meet the minimum wage.
Strength of evidence: ★★☆☆☆ [Contested, heavily vendor-sourced, with the quantitative statements triply vetoed in Round 3]
How to test: for any claim that "software jobs are recovering / collapsing," ask three things first: ① what is the base date? (That widely circulated rebound magnitude uses a base date that is exactly the trough of the series — and happens to be an AI vendor's product launch day.) ② where is the level? (Even after the rebound, software development job postings remain about 27% below February 2020.) ③ which seniority tier does the increment come from? (Of the net increase from 2025-05 to 2026-05, senior roles account for 71% and AI-related titles 37%, with the two overlapping — growth is concentrated in senior and AI-related roles, which is worst for the entry point.) What you can check yourself: that platform's research arm provides a public data repository with daily series by industry, so you can re-base and recompute yourself. Warning: this index is revised retroactively, and the same sentence "up Y% since date X" will give different numbers depending on when it was scraped — any citation must lock the retrieval date.
Strength of evidence: ★★★★☆ [Multi-source; all three bases have primary definitions]
How to test: every time you see a "field match rate" or "field relatedness," first identify which of the three it is: ① respondents' self-assessed relatedness to their highest degree (among those whose highest degree is a bachelor's, 25.3% report "not related"); ② analysts' cross-coded determination of whether the job directly matches the undergraduate field (about 27% directly matched; computer science 33%); ③ whether one is employed in a STEM occupation (engineering 52%, computer/math/statistics 51%, psychology 10%). The three have different numerators and denominators, and any arithmetic or comparison across them is wrong. On the China side, "job–field relatedness" has those already employed full-time as its denominator — not all graduates, and not the administrative "field-matched employment rate." Also note one very common misreading: "the job requires a bachelor's degree" corresponds to about +25% in wages, and "the job matches the field" corresponds to a further +5% on top of that — they are two nested thresholds, and the ratio of coefficients is not a ratio of importance.
After the originally planned "newly added AI schools" signal was vetoed, the substitute signals designated by the audit seat have constructs much closer to the quantity you actually care about:
| Substitute signal | Why it is better | Where to look |
|---|---|---|
| Enrollment plan headcounts by program | Corresponds directly to headcount supply, with no "institution count ≠ headcount supply" unit mismatch | Each province's admissions examination authority's annual admissions plan book, countable school by school |
| Year-by-year change in admission ranks and cutoff scores for your target program | Directly reflects candidates' collective judgment, and you can look it up yourself during application season | Each province's score-distribution table plus institution-program-group cutoff lines |
| Appearance of relevant program codes in the annual batch's abolition/suspension lists | Abolition is harder behavioral evidence than addition, because it carries a real cost | The Ministry of Education's annual Results of Program Filing and Approval |
| Each province's "urgently needed program list" and "program warning list" | Already covering 473 program types; which program your own province puts on the warning list is closer to the decision you have to make than the national count of newly added schools | Provincial education departments, published by July 31 each year per Jiao Gao Si Han [2025] No. 3 |
| The total planned headcount in your province's annual Public Recruitment Announcement for Primary and Secondary School Teachers (March–April) | Checkable year by year, and a primary administrative document | Provincial education / human resources departments. ⚠️ You must also check whether the post categories are the same as the previous year's — when the bases differ, do not compare percentages |
One last piece of methodological discipline, for you and for me: in Section 5.3 this essay wrote down a pre-registration rule — if more complete program-instance data are obtained in future, only when all four conditions hold simultaneously ("a single year's change exceeding ±3×," "two consecutive years in the same direction," "using the official annual basis and reporting the whole basket of related programs to rule out relabeling," and "reporting an incidence rate with the number of institutions not yet offering the program as the denominator") may we say that "the expansion of program instances has indeed turned" — and even then only as auxiliary description, never as a basis for career advice. Write the rule first, look at the data after: this is the only constraint this essay imposes on itself, and it is also the standard you can apply to any recommendation about fields of study: was its decision rule set before it saw the data, or after?
occupation.xlsx Table 1.10 / Table 1.2; the BLS official evaluation pages Occupational Projections Evaluation: 2006–2016, 2012–2022, 2014–2024; BLS methodology pages (separations.htm, replacements.htm); the July 2005 external evaluation article and the February 2025 AI case study in the Monthly Labor Review (both are BLS's own publications, not peer reviewed); the 2014–24 and 2016–26 historical tables. Most BLS primary documents were obtained via raw web-archive snapshots.The load-bearing claims in this essay went through 180 votes of three-vote adversarial verification and 24 seats of dual-seat auditing of single-source empirics, with 10 core figures withdrawn or downgraded; faithful retelling is not the same as factual truth, which is why evidence grades are attached item by item. Related research on this site: Is There Still a Place for Junior Engineers in the AI Era? (the other side of entry-level jobs), An Evidence Check on "95% of Pilots Fail" (the definitional ladder for reading survey numbers), Learning Science: What Actually Works (how to study once you are in university), and Is This Round of AI Capex 1999? (what kind of macro cycle you will graduate into).