Seniority-Biased AI Adoption: The Junior Share at Adopting Firms

Sources#

Summary#

The vault's entry-level evidence has come from two directions that never met. Exposure designs (Canaries, Indeed's postings) find young or entry-level workers in AI-exposed occupations falling behind, but they cannot see which firms adopted. Adoption designs (Ramp × Revelio) see firms, and found intensive adopters growing their entry-level headcount fastest. This page holds the first adoption design in the vault that is instrumented, international, and split by seniority: Chandar & Klein Teeselink, How Does AI Change Labor Demand? Evidence from 41 Countries (Stanford / King's College London, 2026-09-20, empirical).

Its answer is that adoption changes which workers an affiliate employs more than how many:

Junior share −1.9pp at treated affiliates by March 2026 against matched controls (juniors are 57.1% of baseline positions, so this is 3.3% of the baseline share).
Mostly because seniors grow. Senior employment is +6.7% (p<.01). Junior employment is −2.5% and not significant. Total employment is +3.3%, significant only at 10%. For the average matched treated affiliate (≈255 positions, 145 of them junior), that means about 8 more seniors and 4 fewer juniors than without adoption.
Seniors move toward exposed work and juniors move away from it. The top-minus-bottom exposure-quartile share rises +0.9pp among seniors (p<.05) and falls −0.6pp among juniors (n.s.).
It is widespread. The junior share is negative in 23 of 31 countries with a usable first stage, significant in seven, with a random-effects mean of −1.1pp.

In the paper's model, AI is labor-saving for juniors and labor-expanding for seniors in exposed occupations, with a modest scale effect on top.

Evidence note. empirical: a working paper (not peer-reviewed) with a formal model, a 2SLS event study, and two uninstrumented designs as checks. Figures below come from the prose, and the Table 3 and Table 4 values were checked against pdftotext -layout. Four limits travel with every number. (1) Population. The instrumented estimates cover foreign affiliates of multinationals (100,813 in the estimation sample; 26,102 treated and 27,697 controls matched), which are larger and more integrated than the average firm. The workforce data is LinkedIn-derived Revelio, which covers 67% of official employment in the US and 63% in the UK, but 24% in Germany and Brazil, 7.9% in India and 2.1% in Japan. It over-represents managers (24% of positions against 5% of official employment in the median country). (2) Adoption is a hiring signal. A company counts as an adopter if it posted at least one LLM-verified generative-AI job ad after November 2022. The median adopter posted three, and AI jobs are 0.60% of adopters' postings. Firms that use AI without advertising for it count as controls, which the authors argue biases estimates toward zero. (3) Treatment is dated at ChatGPT's launch for every adopter, not at the first AI ad. (4) Authorship overlap. Chandar co-authors the Canaries series, so this is not an independent confirmation of the junior-decline thesis by an unaffiliated team. The design and data are independent of Canaries (Revelio, not ADP; firm adoption, not occupational exposure).

The design, briefly#

Adoption measure. A keyword screen (ChatGPT, Claude, Copilot, LLM, LangChain, RAG, and similar terms) flags more than 2 million candidate postings. Llama 3.1 8B then keeps the ones where the job genuinely involves generative AI, which leaves 1.36M postings, 99% of them after ChatGPT. Adoption is set at the company level, and only 30% of treated affiliates advertised an AI job in their own country. The measure follows Hosseini Maasoum & Lichtinger (2025), a US postings-based adoption study the paper cites as finding junior declines; the vault does not hold that study.
Matching. Each treated foreign affiliate is matched to up to five control affiliates in the same country and 12-sector industry, on log employment, average Eloundou exposure, and junior share, all as of October 2022. Controls must come from corporate groups with no adopter anywhere.
Instrument. The instrument is the leave-out adoption rate among other multinational parents headquartered in the same metro as the affiliate's parent. Those peers sit in a different country from the affiliate, and the design needs at least nine of them. A 1-SD rise in peer adoption raises the probability that the affiliate's company adopts by 16pp, with an effective first-stage F = 104.2. The first stage is positive in all eleven parent sectors and all three size terciles.
Estimation. The paper runs long differences from October 2022, month by month, with matched-set fixed effects. That makes the specification saturated, so each coefficient is a positively weighted LATE for compliers: affiliates pushed into adoption by their parent's peers.

Where it shows up#

Within occupations, not between them. Occupation shares barely move: 21 of 22 groups change by less than 0.5pp. The exception is computer & mathematical, whose share rises +0.8pp and survives a Bonferroni correction. The junior share within occupations falls in 16 of 22 groups. The five significant declines are legal −4.9pp, computer & mathematical −3.3, business & financial operations −3.2, sales −2.6, and management −1.0. All five are in the upper half of the exposure ranking. Only computer & mathematical and business & financial survive Bonferroni. In the chart (Figure 5; a chart reading, not a printed value), healthcare practitioners and architecture & engineering have within-occupation junior-share intervals that span zero. That ordering matches the high- and low-vulnerability professions predicted in The Tragedy of the Cognitive Commons. It is only suggestive: this is a headcount outcome, and the legal estimate does not survive the multiple-comparison correction.

Exposure is the gate. Above-median-exposure affiliates show a junior share of −3.4pp, senior employment of +10.4%, and junior employment of −6.5%. Below the median, the coefficients are near zero. Technology affiliates show −3.9pp, +14.6% senior and +7.8% total. Outside tech the effect is smaller but present: −0.9pp, +2.9% senior, and flat total employment. By size, the junior-share effect is the same on both halves (−1.8 to −1.9pp), and growth concentrates in smaller affiliates. Grouped by the occupation of the company's first AI job ad, the junior-share effect is negative in all ten estimable categories and significant in six. So the effect is not just a by-product of hiring AI engineers.

Across countries (31 of 41 have first-stage F ≥ 10):

Significant junior-share declines appear in the US, UK, Spain, Poland, Brazil, Mexico and Saudi Arabia. Heterogeneity is moderate (I² = 42%).
Senior employment rises +3.4% on the meta-analytic mean (marginal).
No country shows a significant fall in total employment.
Country correlations. Junior losses run deeper in richer economies (r = −0.37) and more digitized ones (r = −0.29 for digital readiness), and shallower where labor law is more rigid (r = +0.20). Only one of these correlations is significant.
The developing-economy claim rests on Brazil, Mexico, Saudi Arabia, and imprecise negatives in Egypt, Colombia and Indonesia. India and Japan are not among the 31: the instrument is too weak there (read from Figure 6's country list). The paper ranks India first by AI share of postings (a software/IT posting mix, not national adoption), even though population-use telemetry puts it near the bottom.

Robustness, and the one place it wobbles#

The design holds. The junior-share decline survives every instrument-validity check:

Check	First-stage F	Junior share	Senior employment
Local-boom "bad" control added	116.6	−1.9pp	+6.9%
Country × sector peer instrument	20.1	−2.3pp	+9.8%
Out-of-sector peers only (imitation check)	100.9	−1.5pp	+4.6%
Drop any one of the top-10 metros	≥ 90	−1.5 (drop SF) to −2.0 (drop Paris)	+5.7 to +7.5%

Four more checks:

Without the instrument. Synthetic DiD gives −1.4pp on foreign affiliates and −1.3pp on all affiliates, domestic included.
Adoption intensity. At roughly the 90th-percentile adopter, the junior share falls −4.1pp and senior employment rises +16% (F = 40.1).
Placebo launch date (January 2022): the junior share moves by at most 0.2pp.
Parent-level adoption. Measuring adoption at the parent halves the effect (−1.1pp), and it shifts onto junior employment. So the affiliate's own company adoption carries more of the recomposition than headquarters centralization does.

The level of junior employment is the fragile part. Its sign depends on the design:

2SLS: −2.5% (n.s.)
Synthetic DiD: +1.5% on foreign affiliates, −1.1% on the full sample
With controls required to have posted in the treated company's AI-job occupation: −6.0%
Plain matched DiD: +7.0% (with total employment +10.8%). The authors discard this estimate because it shows clear pre-treatment divergence, meaning adopters were already on steeper growth paths.

The share is robust. Whether juniors are cut or merely outgrown is not settled.

Timing is the paper's own soft spot. Some effects appear within a year of ChatGPT, before many treated companies posted their first AI ad:

The cleanest cohorts respond in order. For the 2023 and 2024 cohorts (F = 227 and 75), the 2024 cohort moves later and catches up. By March 2026, the junior share is −1.6pp (2023) and −2.0pp (2024).
The response can precede the first ad. The 2024 cohort's junior share starts falling during 2023, before its first ad, at about two-thirds of the 2023 cohort's pace.
The 2025 cohort (F = 24) has pre-trends, and its junior share falls before any company in it had advertised.

The authors read that last cohort as unreliable rather than as evidence that the response precedes adoption. They concede they "cannot rule out that part of the response among the latest adopters reflects factors other than their own adoption."

The Census graduate study on AI-Exposed College Majors at Labor-Market Entry runs into the same puzzle from the worker side: effects on new graduates appear to precede broad firm adoption. It proposes a channel that needs no adoption at all. If AI-inflated grades stop signalling skill, employers hire fewer graduates before they use AI themselves.

Against Ramp: the same seniority coding, the opposite tilt#

Ramp × Revelio (Firm AI-Spend Intensity and Headcount Growth) is the vault's other adoption design, and it pointed the opposite way on composition. The earlier contradiction between Ramp and Indeed was reconciled on the page as different units (headcount stock against vacancy flow). That excuse is not available here, because both papers use Revelio's workforce records and define junior/entry-level as seniority levels 1–2.

	Ramp × Revelio (Kharazian, Simon & Stevens, Jun 2026)	Chandar & Klein Teeselink (Sep 2026)
Workforce data	Revelio, seniority 1–2 = entry	Revelio, seniority 1–2 = junior
Adoption measure	AI-vendor spend on Ramp cards/bills, graded by $/employee/month	≥1 LLM-verified generative-AI job ad
Population	21,559 US firms	~54K foreign affiliates of multinationals, 41 countries, US included
Identification	Staggered event study against not-yet-adopters	Matched sets + cross-border peer IV; synthetic DiD as a check
Entry/junior headcount	+12.0% (high intensity)	−2.5% (n.s.)
Entry/junior share	+1.15pp (high intensity); −0.52pp (low)	−1.9pp; −4.1pp at ~90th-percentile intensity
Total headcount	~+10% (high intensity)	+3.3% (p<.10)

How it resolves, and how far. The three remaining differences are the adoption measure, the population, and the identification.

Identification: a warning for Ramp, not a refutation. This paper's own uninstrumented matched design reproduces Ramp's direction and magnitude: junior +7.0%, total +10.8%. It fails the pre-trend test because adopters were already growing faster. Ramp's not-yet-adopter design is built against exactly that selection, and Ramp's own never-adopter comparison showed the same contamination.
Intensity: the gates point opposite ways. Ramp finds that the junior tilt appears only with intensity, while low-intensity adopters tilt senior. Here, higher adoption intensity makes the junior share fall further. The intensity measures differ (vendor spend against AI hiring), and nothing in either paper says which captures the "real" adoption that matters.
Weighting. Both sources are empirical. On composition, this paper is the better-identified source: it has an instrument, synthetic DiD agrees with it, and it reports a placebo and pre-trends. Ramp's +12.0% stands as a measured fact about US firms that pay AI vendors intensively. The broader claim that adopting firms tilt junior is now contested by a better-identified design, and it is not established.

What falls: Ramp's page proposed that the two findings could be reconciled "if adopting firms are absorbing reallocated juniors." That is contradicted on this instrument, because at postings-defined adopters juniors do not grow relative to matched non-adopters.

Adopters vs the economy: why the sign can flip#

The paper's model makes a point the vault's firm panels have been exposed to without stating it. An adopter-vs-non-adopter difference (what this paper, Ramp and OpenAI's enterprise study all estimate) and the general-equilibrium change in an occupation's employment are different objects:

The DiD measures a gap between firms. Adopters cut costs, lower prices and take market share. That scale effect raises their demand for every worker type, including types whose tasks AI displaces. So "adopters grow relative to non-adopters" can hold while aggregate employment falls.
The closed form (Appendix A3.3). Take a uniform productivity gain with AI capital. The GE employment change is "never positive: it is zero without AI capital and negative with it," because AI capital takes a slice of costs and lowers the labor share. The DiD, meanwhile, is positive whenever the productivity gain is large enough.
Simulation (Table A1). All four combinations of signs occur, including adopters employing fewer of an occupation than their twins while market employment in that occupation rises.

This gives the vault's cross-design comparisons a rule. Exposure designs and adoption designs are not two estimates of one number. An adoption DiD with a positive total-employment sign says nothing, by itself, about whether AI grows jobs in the economy. This is the formal version of the endogenous-adoption point on Task Saturation: Broad but Shallow AI Diffusion's micro–macro gap. The paper's scale-and-bias decomposition (Table 4, verified):

	Juniors	Seniors	All
Bias toward most-exposed quartile, α(Q4) − α(Q1)	−2.0% (n.s.)	+5.0% (p<.05)	—
Scale term (least-exposed quartile)	−1.8% (n.s.)	+4.3% (n.s.)	+2.5% (n.s.)

The senior-minus-junior scale gap is 6.1pp, significant at 5%. An OLS version of the same decomposition puts the all-worker scale term at 9.7% against 2.5% under 2SLS, which is the selection the instrument removes.

What it does not measure#

Wages. Treated and control affiliates share a labor market, so wage effects are differenced out. The authors list "whether senior wages rise and junior pay declines" as open. The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay has the only offer-side reading in the vault.
Developmental content. As with every instrument on The Tragedy of the Cognitive Commons, headcount by seniority says nothing about whether the junior roles that remain still build expertise. The authors cite the pipeline concern (Ide 2025) and stop there.
New work. Recomposition is measured across existing occupations. AI-created job types that do not map onto O*NET are invisible to it.
The long run. The data covers just over three years after ChatGPT. The authors call it "the initial reorganization," not an equilibrium.

Connections#

Firm AI-Spend Intensity and Headcount Growth — the contradiction set out above: the same Revelio seniority coding, and the opposite composition tilt at adopters. Ramp's +12.0% stands for its population; the general "adopters tilt junior" reading does not
The Tragedy of the Cognitive Commons — the first firm-level, multi-country evidence that adoption shrinks the junior share. The within-occupation ordering (legal, computing, business/finance declining; healthcare and engineering not) matches that page's predicted vulnerability ranking on the headcount margin, though not on validation capability
The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay — the same seniority tilt on a third instrument: Indeed's entry share of exposed postings fell 29% → 10%. Postings and headcount at adopters now agree on direction
Codified vs Tacit Knowledge Exposure — the age gradient under a different variable. Here, seniors shift toward the most-exposed occupations and juniors away, within the same firms and inside occupations, which an occupation-level schooling index cannot generate
Returns to Expertise in Agentic Coding — the labor-demand side of the expertise premium: AI is labor-expanding for exposed seniors (+5.0% bias) and labor-saving for exposed juniors
Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — the model's point that one exposure score (Eloundou β here) can imply higher demand for one seniority and lower for another in the same occupation, so exposure is not a verdict on a job
The Enterprise AI Adoption Gradient — a fourth adoption instrument next to vendor records, card spend and surveys: generative-AI job ads, with the same selection (larger, more exposed, better-paid firms adopt)
Task Saturation: Broad but Shallow AI Diffusion — the micro–macro gap there (experiments compress the expert premium, observational data widen it, and endogenous adoption reconciles them) gets its formal counterpart here: adopter DiDs and GE changes can differ in sign
AI-Exposed College Majors at Labor-Market Entry — where the juniors who are not hired end up. Census PSEO×LEHD records show graduates of the most-exposed majors losing 5pp of initial employment and 13% of initial earnings, about half of it from sorting into retail and food service. This page measures the share inside adopters; that page measures the outcome for everyone who applied
Stanford Digital Economy Lab — the lab's first firm-level design, closing the "cannot see firms" limitation of the ADP series

Open Questions#

Is the composition gap between Ramp and this paper caused by the adoption measure (vendor spend against AI job ads) or by the population (US firms against multinational affiliates)? Falsifiable: apply both adoption definitions to the same Revelio firm panel. If spend-defined adopters tilt junior and postings-defined adopters tilt senior on identical firms, the measure is the explanation.
Does the recomposition carry into wages, with the senior premium rising and junior pay falling at adopters? The design differences out wages by construction. Falsifiable: an employer–employee panel with pay (Nordic registers, or ADP with firm identifiers) cut by adopter status and seniority.
Does the junior-share gap at adopters keep widening past March 2026, or plateau as the paper's "initial reorganization"? Trigger: a later vintage of the same design, which the authors frame as reusable.

Sources#

How Does AI Change Labor Demand? Evidence from 41 Countries — Bharat Chandar (Stanford) & Bouke Klein Teeselink (King's College London / AI Objectives Institute), How Does AI Change Labor Demand? Evidence from 41 Countries, Stanford Digital Economy Lab working paper, 2026-09-20, 156pp, empirical. Cited: §1 (headline numbers), §2 (Proposition 1, labor-saving/expanding/scale decomposition), §3.3–3.5 (adoption measure, descriptives, representativeness), §4 (matching, instrument, long-difference 2SLS), §5.1 (Table 3, Figures 3–5), §5.2 (Figure 6 and country correlations; exposure/size/tech/first-AI-job splits), §5.3 (instrument-validity and additional robustness, timing cohorts), §5.4 (Table 4), §6, Appendix A3.3–A3.4 (closed-form case and simulated sign combinations, cited from prose only), Appendix A8 (coverage). Parse warning: docling mangled Table 2 (affiliate-level predictors of treatment). The junior-share and log-wage coefficients sit one row off their labels, and column 3's N and R² are welded into the "Outcome mean" cell. The table was not cited; its coefficients are quoted from the §3.4 prose only. Table 3 lost its standard-error row in the parse (a dropped-row failure no check flagged). Its point estimates and SEs were verified against pdftotext -layout. Table 4 parsed intact and matches the PDF. Figure 5 and Figure 6 statements marked as chart readings are approximate
A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment — Kharazian, Simon & Stevens (Ramp Economics Lab × Revelio Labs, June 2026, empirical). Cited only for the comparison row (entry-level +12.0%, share +1.15pp / −0.52pp, ~10% total, seniority 1–2, not-yet-adopter design); full treatment on Firm AI-Spend Intensity and Headcount Growth
Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors — Orr, Tucker & Warren (US Census Bureau CES-WP-26-56, September 2026, empirical). Cited only for the graduate-side timing puzzle and the signal-erosion hypothesis (§8); full treatment on AI-Exposed College Majors at Labor-Market Entry