AI Employment Effects

Confidence 0.95 · 56 sources · last confirmed 2026-07-01

The empirical record of how AI is affecting employment levels, hiring, wages, and occupational composition in real labor markets. As of late 2025, the wiki’s evidence converges on a specific pattern: early-career workers in highly AI-exposed occupations are seeing relative employment declines, concentrated in automation rather than augmentation uses, with adjustments visible in headcount more than wages.

This page is distinct from enterprise-ai-adoption (which is about how organizations use AI) and from automation-vs-augmentation (which is about how AI is deployed task-by-task). The focus here is the labor-market consequences.

Working definition

“AI employment effects” = the measurable changes in employment levels, hiring rates, wages, and occupational composition attributable (in whole or in part) to the deployment of generative AI and AI agents in workplaces. The literature distinguishes:

Direct displacement — workers losing jobs because AI now does the task
Reduced hiring — firms hiring fewer new workers because existing workers are more productive (or AI substitutes for them)
Compositional shift — jobs not lost in aggregate, but the mix of work changes (tasks redistributed across roles)
Wage effects — pay differentials between AI-exposed and non-exposed workers
Skill-biased effects — differential outcomes by age, education, experience

Key claims

The 2024–25 inflection in entry-level employment (Brynjolfsson et al. 2025)

Using ADP payroll data (~25M U.S. workers, monthly Jan 2021–July 2025):

Early-career workers (ages 22–25) in the most AI-exposed occupations: ~13% relative decline in employment since late 2022, after controlling for firm-level shocks.
Software developers ages 22–25: employment declined nearly 20% from peak in late 2022 to July 2025.
Updated job-postings signal (Thompson 2026 (NYT The Daily)): Thompson cites a more recent Brynjolfsson job-postings analysis showing software-developer hires down 16% “in the last year or so” — i.e. extending the 2022-onset effect into 2025-26 at the hiring (not just employment-stock) layer. Thompson’s framing: “if that’s happening when the AI coding tools are really just going from a crawl to a walk to a run, what might happen when they’re sprinting?” — primary-source ingest of the underlying job-postings analysis is an open target.
22–25 year-olds in highest AI-exposure quintiles: -6% absolute from late 2022 to July 2025.
35–49 year-olds in same occupations: +6% to +9%.
22–25 year-olds in lower exposure quintiles: +6% to +13% (no clear ordering by exposure).
Less-exposed occupations (stock clerks, health aides): no age divergence pattern.

The pattern emerged specifically around late 2022, coincident with ChatGPT’s release. Did not appear during the COVID-19 unemployment spike — so it’s not a generic “young workers struggle in downturns” effect.

Automation vs. augmentation matters more than exposure level (Brynjolfsson et al. 2025 Fact 3)

Using the Anthropic Economic Index classification of AI uses as automation vs. augmentation:

Occupations with highest automation share: declining employment for youngest workers.
Occupations with highest augmentation share: NO similar pattern — most-augmentative quintile shows fastest growth.

This is empirical evidence for a distinction that has been mostly conceptual in other sources (Anand-Wu’s 2×2 framework, Cisco’s “give agency with human oversight”). See automation-vs-augmentation.

Wage stickiness: employment moves before wages (Brynjolfsson et al. 2025 Fact 5)

Annual salary trends show little difference by age or AI-exposure quintile.
“AI may have larger effects on employment than on wages, at least initially.”
Implication: labor markets adjust through hiring/headcount before pay — consistent with a long literature on wage rigidity.

The codified-vs-tacit knowledge mechanism

Brynjolfsson et al. 2025 propose a clean theoretical mechanism:

AI replaces codified knowledge (the “book-learning” of formal education) more than tacit knowledge (the idiosyncratic tips and tricks accumulated with experience). Young workers supply relatively more codified than tacit knowledge, so they face greater task replacement.

This maps directly to Anand-Wu’s “type of knowledge” axis (explicit data vs. tacit knowledge) — a strong cross-source convergence.

Workforce expectations vs. realized outcomes

MIT CISR McKinsey survey:

31% of orgs expect little change in workforce size over 3 years.
43% expect decreases (8% by >20%, 14% by 11–20%, 21% by 3–10%).
23% expect increases.
46% expect >20% of workforce to need reskilling.

Brynjolfsson et al. 2025 empirical correlate: those expected decreases are showing up in the data, disproportionately at the entry level, not uniformly across all roles. The McKinsey expectation that AI may not shrink overall headcount (only 8% expect >20% decrease) is consistent with the Brynjolfsson finding that overall employment continues to grow.

Aggregate productivity estimate (Anthropic Economic Index, 4th report)

Anthropic’s earlier estimate that widespread AI adoption could add +1.8 pp/yr to U.S. labor productivity growth is revised in the fourth report once task-level reliability is accounted for:

+1.2 pp/yr for tasks completed on Claude.ai
+1.0 pp/yr for tasks completed on the 1P API (more challenging task mix)

Even +1 pp/yr would return U.S. productivity growth to late-1990s / early-2000s rates. The estimate does not account for further model improvement or for AI use becoming more sophisticated.

Effective AI coverage at the occupation level (Anthropic Economic Index, 4th report)

The fourth Anthropic report introduces effective AI coverage — the share of a worker’s time-weighted duties that Claude could successfully perform. Outliers when this is plotted against simple task coverage:

Effective coverage exceeds task coverage: data entry keyers, radiologists, medical transcriptionists.
Effective coverage is below task coverage: teachers, software developers.

Pooled across all reports through Nov 2025: 49% of jobs have Claude used for at least 25% of their tasks (up from 36% in Jan 2025), but the success-weighted picture is more uneven.

Observed exposure: the methodological primary ( Anthropic, March 2026)

This Anthropic report is the methodological backbone for the wiki’s “observed exposure” claims (and the exact study AWS presents on stage). It introduces a displacement-risk measure that combines theoretical LLM capability (Eloundou et al. 2023 task β-ratings) with real-world usage (the Anthropic Economic Index), weighting automated and work-related uses more heavily than augmentative ones — the weighting being what turns “exposure” into “displacement risk.”

Headline results (US, Current Population Survey + O*NET + AEI usage):

Capability ≫ adoption. Tasks fully feasible for an LLM (β=1) are 68% of observed Claude usage vs 3% for β=0 — but actual coverage remains “a fraction of what’s feasible.” Most-exposed occupations: Computer Programmers (75% coverage), Customer Service Representatives, Data Entry Keyers (67%), financial analysts; 30% of workers have zero coverage (cooks, bartenders, lifeguards…).
Exposure ↔ weaker projected growth. Employment-weighted regression: +10pp coverage → −0.6pp in the BLS 2024–2034 growth projection (slight but real; the Eloundou measure alone shows no such correlation — the usage-weighting adds the signal).
Who’s exposed: most-exposed workers are more likely female (+16pp), more educated (graduate degrees 17.4% vs 4.5%, ~4×), and earn 47% more — inverting the “low-skill-first” narrative and corroborating the page’s existing skill-biased findings.
No unemployment effect yet; young-hiring slowdown. Difference-in-differences (top-quartile vs zero-exposure) since ChatGPT is indistinguishable from zero (detectable threshold ~1pp). But the CPS job-start panel shows a ~14% drop in the job-finding rate for ages 22–25 into exposed occupations vs 2022 (barely significant; none for >25) — independently echoing Brynjolfsson et al. on a different dataset.

The report’s stance is methodologically disciplined: establish the measure before effects emerge and revisit periodically, because past labor forecasts over-predicted (offshorability) and AI is “more like the internet or China-trade than COVID” — gradual and confounded. It chooses unemployment as the priority outcome (most directly captures harm). This is the wiki’s citable primary for the capability-exceeds-adoption and entry-level-felt-first threads that the AEI reports, Allen/Brovich, and Giles otherwise carry second-hand.

Task-content shift and deskilling (Anthropic Economic Index, 4th report)

A finding distinct from displacement and from wage effects: AI-covered tasks skew toward higher-education content (Claude-covered: 14.4 years average; economy-wide: 13.2 years). As a first-order effect, removing AI-covered tasks from a job deskills the remaining task mix on average — shifting it toward lower-education content while the job itself persists. Most-affected named occupations: technical writers, travel agents, teachers. See ai-deskilling for the dedicated concept page.

Capability as a leading indicator: GDPval (GDPval, OpenAI, Oct 2025)

A measurement-stance distinct from everything above. Adoption, usage, and GDP-attribution metrics are lagging indicators — the invention-to-permeation lag for electricity, computers, etc. runs years to decades. GDPval proposes the alternative: directly measure model capability on real expert deliverables as a leading indicator of economic relevance, visible before adoption shows up in employment data. Across 44 occupations in the top-9 US-GDP sectors, the best model (Claude Opus 4.1) produced deliverables graded wins-or-ties vs human experts on 47.6% of gold-subset tasks (Oct 2025), improving roughly linearly.

The load-bearing caveat for this page: capability is not displacement. A 47.6% win rate is a ceiling on what a model can produce one-shot under ideal context, not a measured labor-market outcome — GDPval tasks are precisely-specified, one-shot, and non-interactive (the paper’s own limitation). This is the supply-of-capability mirror of Narayanan’s capability-reliability gap and of the micro-productivity-trap: high task-level capability has repeatedly failed to convert directly into firm-level or labor-market change. GDPval also names the most-exposed occupations its design surfaces (knowledge work with digital deliverables — developers, lawyers, and similar), echoed by Argenti (HBR), who cites GDPval and extrapolates the trend to ~80% by mid-2026.

The supply-side answer: durable skills (Globerson et al. 2026)

Brynjolfsson et al. (2025) showed early-career employment declines concentrated in automation uses but did not operationalise which skills remain valuable for the workers who keep their jobs (or for those entering the labor market). The 2026 durable-skills literature — anchored by Globerson et al. (Google Research) — provides the supply-side measurement: collaboration, creativity, critical thinking are the canonical examples, and the Vantage / Executive LLM platform makes them measurable at scale (large-N validation; Pearson 0.88 with human experts on creativity assessment).

Together with ai-deskilling, this gives the wiki a complete labor-skill carve:

What gets hollowed out: codified knowledge and procedural tasks (deskilling).
What gets retained: open-ended, socially/contextually situated skills (durable skills).
What it implies for the entry-level pipeline: Canaries showed the squeeze; durable-skills shows what training/curriculum/hiring should target if firms want to build their next generation of workers around what AI cannot substitute.

Productivity gains alongside employment declines

The wiki’s primary source for the customer-support productivity finding is now Brynjolfsson, Li & Raymond (2025) QJE — “Generative AI at Work”:

+15% productivity in resolutions per hour (with full year-month + agent + agent-tenure FE: +15.2%).
30% gain for less-experienced / lower-skilled workers with quality also improving.
Most-experienced / highest-skilled workers: small speed gains AND small DECLINE in quality — the equalizing effect is not strictly Pareto-improving.
AI accelerates the experience curve ~3×: treated 2-month-tenured agents perform like untreated 6+-month-tenured agents.
Effects are durable — workers maintain higher productivity even during AI outages, suggesting the AI is teaching skills rather than just providing real-time scaffolding.

Reconciling productivity gains with the Canaries paper employment declines: AI raises individual productivity (especially for low-skill workers in a role) while reducing the number of workers needed in that role. Whether net employment goes up or down depends on demand elasticity — how much demand expands when costs fall. The customer-support equalizing effect at task level does not imply pro-employment outcomes at occupation level.

The two-paper Brynjolfsson arc is the wiki’s clearest illustration of the task-level vs. occupation-level paradox: same author, same lab, same methodological rigor, two complementary findings that together are more informative than either alone.

Skill-biased technological change: high-tenure users gain more (Anthropic Economic Index, 5th report)

The fifth Anthropic Economic Index report (Feb 2026 sample) introduces an explicit skill-biased technological change framing — a counterpart to the customer-support equalizing effect, but pointing the other way at the frontier of usage:

High-tenure Claude users (those with longer histories of Claude.ai use) report ~3–4 percentage points higher task-success than lower-tenure users, after controlling for task type, model selection, and conversation complexity. The gap survives controls — it is not just about high-tenure users picking easier tasks.
The mechanism is identified as learning by doing: skilled users learn what Claude is good at, structure prompts to reach the model’s strengths, choose the right model tier (Haiku/Sonnet/Opus) per task, and recover from failure faster.
Model selection itself becomes a skill. The 5th report shows users select Opus differentially for higher-value tasks: +1.48 percentage-points higher Opus share per +$10/hour task value on Claude.ai; +2.79 pp per +$10 on the 1P API (about twice as steep). High-value-task selection of premium models is a learned behavior visible in usage logs.

Reconciling with Brynjolfsson, Li & Raymond (2025)‘s within-role equalizing effect: the two findings are not contradictory but operate on different populations:

Setting	Effect direction	Population
Customer support augmentation (BLR 2025)	Equalizing — low-skill +30%, high-skill ~0%	Within-role workers given a fixed AI tool
Open-ended Claude usage (AEI 5, 2026)	Skill-biased — high-tenure +3–4pp success	Self-selected users across all tasks/models

The reconciliation: constrained, single-tool augmentation equalizes; open-ended, multi-tool deployment rewards expertise. As enterprise deployments mature from single-purpose copilots to general-purpose Claude (or multi-model) access, the 5th-report pattern likely dominates over the customer-support pattern. This matters for ai-deskilling and for durable-skills in two ways: AI-literacy is becoming a learnable skill with measurable returns, and access to premium models becomes a productivity differentiator alongside underlying competence.

Hopes outpacing fears: the worker-attitude reversal (Ransbotham et al. 2024)

A counterweight finding to the displacement narrative — and the wiki’s strongest signal that current workers’ subjective expectations have improved, not worsened, since 2017. From the MIT SMR × BCG 8th-annual survey appendix:

Attitude (% agree/strongly agree)	2017	2024
Hope AI will assist with some of my tasks in 5 years	70%	84%
Fear AI will assume some of my tasks in 5 years	31%	20%

Hope rose by 14 pp; fear declined by 11 pp. The report’s interpretation: experience with generative AI may be showing people what these models can — and cannot — do well. The displacement panic that surged with ChatGPT’s release in late 2022 has not been borne out at the worker-attitude level — at least not within the MIT SMR × BCG sample (3,467 respondents, 21+ industries, 136 countries).

Reconciliation with the Canaries paper’s 13% relative decline finding: the two are about different populations. Brynjolfsson measures future hiring of 22–25-year-olds in AI-exposed occupations; Ransbotham measures current incumbents’ expectations about themselves. Both can be true: incumbents feel safer than ever, while new entrants find the door narrower than ever. The compositional shift is generational, not subjective.

The expert-as-evaluator: the durable role under augmentation (Kiron & Schrage 2026)

A specific claim about which roles retain durable value under sustained AI augmentation: “The expert as evaluator is not a transitional role.” The mechanism:

AI compressed implementation time dramatically; it did not compress the formation of expertise.
Without prior expertise, only two moves exist for an AI output: accept or reject.
With expertise, a third move opens up: stay in the encounter and learn.

Operational implication for hiring and reskilling:

Don’t deploy AI first in domains where expertise is shallow — the org loses the evaluator capability that makes compounding possible.
Preserve evaluation expertise as a deliberate capability. Domain experts repositioned as evaluators rather than producers.
The Polanyi tacit-dimension breach makes codified knowledge work more substitutable, but the evaluation role becomes more valuable, not less.

This adds a second durable-role argument to the wiki, alongside the durable-skills supply-side measurement of collaboration / creativity / critical thinking. Where Globerson et al. measure individual durable skills without reference to AI deployment, Kiron-Schrage make the explicit deployment-context claim: the durable role is the expert as evaluator of AI outputs, in domains where the human has prior deep expertise.

Engineer identity-threat as named phenomenon (Forsgren & Macvean 2026)

The wiki’s first source naming identity threat as a labelled phenomenon affecting software engineers under AI augmentation — surfaced from DORA research and addressed directly on stage at Google I/O 2026:

“Our DORA research highlights that many developers are currently feeling a very real identity threat. Historically, we derived immense intrinsic value just from the sheer craft of writing code. We need to shift how we measure our worth. We have to lean into the broader value of our work, which comes from solving real human problems and real business problems.”

This is structurally distinct from the AI Index 2026’s -20% entry-level employment finding — that’s an outcome-side labour-market signal; identity-threat is the content-side role-evolution signal among engineers who remain employed. The two are complementary: employment can be falling for the youngest cohort while the role-content shifts for the remaining workforce, and both surface from the same underlying mechanism (AI taking execution).

The talk’s reframing names the durable role under the role-content shift: engineers become value translators — “taking user needs and business needs to precise requirements”. The named exemplar: “an AI might be able to find and refactor inefficient loops. But a Value Translator first asks, ‘improve performance for whom? Under what conditions?‘” This is the engineering-discipline-specific articulation of the expert-as-evaluator durable role Kiron-Schrage name above — same mechanism (humans keep judgment about what to evaluate against), different vocabulary (value-translation rather than evaluation, but operationally the same load-bearing capability).

Convergent with Masad’s “only two jobs left” role-evolution thesis from the founder-vantage; convergent with Thompson’s 75-developer first-person framing of “hard technical skills are easier to automate than talking to customers and figuring out what should we be building”. The talk also names the role-evolution-across-disciplines observation: “software engineering isn’t the only discipline getting more T-shaped — we are seeing this exact same evolution in our product managers, our UX researchers, and more”, generalising the value-translator role beyond engineering to the broader knowledge-work population.

Software developers ages 22–25: -20% from 2024 (AI Index 2026)

A 2026 update extending the Brynjolfsson Canaries finding to a more recent timeframe and via a different data path:

U.S. software developers ages 22–25: employment fell nearly 20% from 2024, even as headcount for older developers continued to grow.
Productivity gains from AI are 14–26% in customer support and software development — clustered in the same fields where entry-level employment is declining.
AI agent deployment remains in single digits across nearly all business functions, suggesting the employment effect is not driven by full agent replacement but by augmentation reducing demand for entry-level labor.

Crowdsourcing labor markets and HAI substitution (Boussioux et al. 2024)

A field study comparing human crowd (HC, n=125 global Freelancer.com solvers) vs human-AI (HAI, single prompt engineer + GPT-4) on a circular-economy ideation challenge finds:

HC retains a novelty advantage (especially at the top decile of solutions).
HAI delivers higher strategic viability, environmental value, financial value, and overall quality.
Cost: HAI ~$27 vs HC ~$2,555 (≈94× cost reduction); time: ~5.5 hours vs ~2,520 human-hours.

This is direct evidence on substitution dynamics in crowdsourcing labor markets (Freelancer.com–style platforms): HAI may displace HC for value/quality-oriented tasks while HC retains a comparative advantage on extreme novelty. The paper does not predict full crowd displacement; the authors frame the patterns as complementary, with HAI augmenting early-innovation phases.

Equalizing effect within elite knowledge workers (Dell’Acqua et al. 2026)

A randomized field experiment with 758 BCG consultants finds:

Bottom-half-skill AI users (within an already elite professional sample) gain +31% on quality scores.
Top-half-skill users gain +11% — meaningfully more than zero.

Both groups gain. The pattern echoes the within-occupation equalizing effect documented in customer support, extended to high-end knowledge work, with the additional finding that top performers are not zero-gain in this domain.

Subjective coherence persists when correctness fails (Dell’Acqua et al. 2026)

Outside-the-frontier AI users produce more confident, more coherent recommendations even when their answers are wrong. This decouples quality of presentation from quality of substance — a labor-market signal that may mask poor performance from supervisors and clients in the short run.

Convergence in communication patterns

A specific finding from Brynjolfsson, Li & Raymond 2025 worth flagging: AI access produces convergence — low-skill agents begin communicating more like high-skill agents. The mechanism appears to be that the AI, fine-tuned on top performers’ conversations, propagates their communication patterns to less-experienced workers.

This adds a skill-leveling dimension to AI’s labor effects — beyond raw productivity numbers, AI may be reducing within-role skill differentiation. Whether that’s good (faster onboarding, lower training costs, better customer experience) or worrying (homogenization, loss of original problem-solving approaches, training-data degradation as top performers reduce their original contributions) is genuinely contested.

Role-vindication: designers under AI (Spiegel 2026)

A complementary signal from the operator side. Lenny frames the designer-PM-engineer triad under AI as a three-way standoff (each role thinking it’s the future and doesn’t need the other two — a framing he attributes to a prior Marc Andreessen episode). Spiegel rejects the framing but takes the underlying observation seriously: “designers feel vindicated in a lot of ways. A lot of designers had parents who were saying, ‘Why aren’t you studying computer science? What are you going to do with this skill set, drawing things?’ … And I think today, a lot of our designers are now shipping code, which is extraordinary.”

The substantive claim — the role most-augmented by AI is the one whose career legitimacy was historically contested — is an operator-grade counterweight to the wiki’s predominantly displacement-flavoured framing. Snap’s operational instantiation: bottom-up adoption (not mandated), AI-driven code review catching “close to 10,000 bugs” auto-detected, agents debugging shake-to-report submissions, near-term forecast of agents implementing fixes (not just suggesting them). The wiki’s first billion-user-scale operator-narrated guardrail-stack that lets non-engineer roles ship code safely — a structural condition for the role-vindication claim to hold at production scale rather than at toy scale.

Position relative to the displacement literature: not contradictory. Brynjolfsson et al.’s entry-level-decline finding is about new entrants to occupations highly automatable by current LLMs; Spiegel’s vindication finding is about senior contributors in adjacent crafts that AI augments rather than replaces. The wiki holds both as role-specific empirical claims at different career stages, not as a contradiction.

The junior-hiring crisis as named phenomenon + the hourglass-organization counter-prescription ( AWS London Exec Forum, 21 May 2026)

Jonathan Allen at the AWS London Executive Forum 2026 anchors the entry-level employment pattern as “the junior-hiring crisis” at AWS-Executive-in-Residence vantage, with four datapoints (~23:36–24:00):

73% entry-hiring level collapse in European tech — per Ravio’s 2025 study.
7% of new grads entering big tech, down from 15% — “halving in roughly one cycle.”
7.7% junior headcount decline; 9% senior employment growth — the within-firm compositional shift that lines up with the Brynjolfsson Canary entry-level-employment finding at a different vantage.
Fastest-growing US top job title: AI engineer; 1.3M new AI roles created; 67K open software-engineer roles vs ~52K layoffs — a job-family change, not aggregate destruction.

Allen pairs the empirical anchor with the hourglass-organization prescription — the operating-model counter to the “seniors 20×-ing their productivity, cutting back juniors” default ratio. The argument (~25:42–25:54): “if you stop training the juniors, where on earth do your seniors come from in 5 to 10?” — and Allen names Matt Garman (AWS CEO) as the leadership channel endorsing protection of the junior learning path. This is the wiki’s first vendor-CEO-endorsed prescriptive counter to the seniors-only AI-team default, structurally compatible with the Brynjolfsson Canary’s automation-share interpretation: the within-firm compositional shift toward seniors is observed by both, but Allen names it as changeable by leadership choice rather than as fixed market structure. See also the ai-deskilling page for Allen’s “protect that junior learning path” framing as active anti-deskilling discipline.

The MGI shift-not-elimination framing + 72% shared-skills finding ( MGI 2025)

[[2025-11-25-yee-mgi-agents-robots-and-us-skill-partnerships|MGI’s Agents, Robots, and Us]] (November 2025) reframes the AI-employment question at structural-modeling altitude. The headline framing:

“Today’s technologies could theoretically automate more than half of current US work hours. This reflects how profoundly work may change, but it is not a forecast of job losses. Adoption will take time. As it unfolds, some roles will shrink, others grow or shift, while new ones emerge — with work increasingly centered on collaboration between humans and intelligent machines.”

Four substantive contributions to this concept:

(a) Explicit avoidance of the AI-will-eliminate-jobs frame. MGI’s “Framing the jobs debate” sidebar (Ch.1, pp.12–13) names four open questions — how close are AI agents and robots to matching all economically relevant human capabilities? will an AI-centric economy create enough jobs? how might the composition of work change? will we adapt fast enough? — and explicitly states “whether AI proves to create or reduce net jobs depends on how effectively it is used to build new industries and markets… a question beyond the scope of our analysis.” The report’s contribution is measuring skill change, not predicting job losses — the wiki’s most rigorously hedged structural source on the AI-labour question.

(b) The 72% shared-skills finding. “More than 70 percent of the skills sought by employers today are used in both automatable and non-automatable work.” The implication is that most skills remain relevant — what changes is how and where they’re used. This is the skill-level decoupling from the occupation-level archetype distribution: an accountant (agent-centric occupation) and a nurse (people-centric occupation) share most of the same skills (communication, problem-solving, detail orientation) — they just apply them in different mixes of automatable / non-automatable activities.

(c) The Skill Change Index (SCI) as a quantitative complement to Brynjolfsson’s empirical signals. Brynjolfsson Canaries supplies the current empirical signal that automation displaces while augmentation doesn’t. MGI’s SCI supplies the prospective skill-by-skill decomposition: highly exposed (top quartile) skills decline in demand, middle-quartile skills evolve, low-exposure skills endure. The two frameworks are descriptive + prospective: Brynjolfsson measures what’s already happening; MGI forecasts which skill-categories will change most by 2030.

(d) AI fluency demand growing faster than any other skill. AI fluency rose 6.8× in two years (1.0M → 7.0M US employees in occupations where job postings call for AI-fluency skills, 2023→2025). Technical AI skills (Govern AI + Develop AI, 55 skills) grew 1.6× (2.1M → 3.3M). The growth in AI-fluency demand is faster than for any other skill in US job postings — including faster than the prior demand surge for cloud, cybersecurity, and other digital skills. 75% of AI-skill demand is concentrated in 3 occupation groups (Computer & mathematical, Management, Business & financial operations); 9 occupation groups (~40% of workforce) have near-zero AI-skill demand — the AI-skill labour market is currently bifurcated.

The MGI framing converges with Brynjolfsson’s empirical evidence: both reject the AI-will-eliminate-jobs framing in favour of AI-will-reshape-which-skills-are-applied-where. The pair forms the wiki’s strongest descriptive-empirical + structural-prospective anchor on the AI labour question.

The lump-of-labor fallacy counter-frame ( Lenny’s Podcast May 2026)

Evans (May 2026) supplies the wiki’s clearest independent-analyst-altitude reframing of the displacement debate as a textbook case of the lump-of-labor fallacy:

“Every time we have a new technology, it automates away a bunch of jobs and then that automation — whether it’s price elasticity and the enablement of the fact that they became automated — unlocks a bunch of new jobs. And you can always see the job that’s going to go away, and you don’t know the new job because it doesn’t exist yet. And it’s like something that sounds dumb anyway — like railway engineer? what’s a railway? Why would that be a thing? Who would care, who would want to go that fast?”

The historical pattern Evans names: typesetters, telephone operators, typists — all crap jobs seen retrospectively; the new jobs that replaced them are better because GDP keeps going up. The analytical move is to invert the doomer framing: instead of asking “what fraction of current jobs will AI eliminate”, ask “what is the new job that doesn’t exist yet”. The answer is unknowable in advance — but the lump-of-labor frame requires that the answer be zero, which the historical record contradicts at every prior automation wave (1800 → present).

Three corollaries Evans draws:

(a) AI labs themselves keep hiring. “Even just looking at the most advanced AI companies — Anthropic, OpenAI — everyone’s just increasing headcount. Like the companies you would think would be least likely to add humans are adding many, many humans.” The companies most exposed to their own technology are not shrinking headcount; they are scaling it. Same observation surfaces in Anthropic engineering-org (Boris Cherny’s practitioner-altitude account) and convergent with the McKinsey team-shape pieces (Moon et al., Sternfels) which note professional-services capacity is being added not subtracted.

(b) The Excel-and-bankers / accountants-and-spreadsheets analogy. “Young people won’t believe this but before Excel, junior investment bankers worked really long hours and now thanks to Excel, Goldman’s associates all work at lunchtime on Fridays. Like — well, why is that not what happened?” The number of US accountants has risen continuously through 20th-century waves of adding-machines → punch-cards → mainframes → ERP → cloud → spreadsheets and continues to rise into the 21st century. Same pattern with software engineering: “Before IDEs and libraries and operating systems, developers had to write all the code. Now if you write an iPhone app, 90% of the code is written for you by Apple … so we’ve got like a tenth as many engineers now? Well, no.” The Jevons / price-elasticity mechanism converts cost reductions into demand expansion.

(c) The 5-to-10-year sales-cycle speed limit. “Typical big-company enterprise software sales cycle is 18 months if you’re lucky. So no, people aren’t just going to tear out SAP and replace it with XYZ. Maybe in three, five, ten years yes that whole estate will look radically different and all those jobs will have changed — but it will take three, four, five, ten years and it will take time sector by sector.” Evans calls people who assert “every big company is going to buy ChatGPT tomorrow and then in two weeks time they’ll fire all their staff” “morons”. Convergent with enterprise-ai-adoption’s 18-month-enterprise-sales-cycle observation: the speed of displacement is bounded by the speed of procurement, integration, and workflow redesign — not by the speed of model capability improvements.

The framing does not contradict Brynjolfsson’s Canaries (Evans accepts AI is real, big, and displacing); it disputes the aggregate-employment-collapse reading by pointing to the historical record and to the labs’ own behaviour. Sits in productive tension with the more catastrophist framings ([[2026-04-25-masad-replit-ceo-only-two-jobs-left|Masad’s only two jobs left]] is the wiki’s clearest contemporary maximalist version), with Evans explicitly rejecting the X% of senior partner’s work is automatable framing as “horseshit” — see automation-vs-augmentation §19 for the task-vs-job alternative.

The April-2026 layoff anchor + AI-as-#1-cited-reason monthly attribution ( WP Intelligence May 2026)

Giles (May 2026) supplies the wiki’s first monthly-attribution datapoint for AI-as-cited-reason-for-layoffs at executive-readership altitude:

April 2026 US layoffs: 83,387 announced job cuts (+38% from March 60,620); third-highest month since 2009 per outplacement firm Challenger, Gray & Christmas.
“AI is the number one reason cited for job cuts in both March and April of this year” — Challenger, Gray & Christmas data. “Other reasons included concerns about the economic outlook and company closures.”
Single-firm anchors: Block (Square’s parent; Jack Dorsey CEO) February-2026 announcement of a ~40% workforce reduction (over 10,000 → less than 6,000) citing “intelligence tools” including agents. Meta (Zuckerberg) 8,000 job cuts the same month, with the 2026 capex outlook raised to $125–145B from $115–135B as the named justification.

The Giles framing puts this alongside the NACE survey of 183 employers (late 2025): “only 14 percent of respondents had considered replacing entry-level roles with AI. Most companies cutting jobs cited the uncertain economic outlook and related budget cuts.” The wiki’s read: AI-as-#1-cited-reason is consistent with AI as a partial contributor rather than the sole driver — the macro-economic-uncertainty + budget-cut framing is the load-bearing companion explanation in the NACE survey-data, and the gap between the two readings is where the empirical noise lives.

The IBM counter-trend single anchor: tripling US entry-level hiring in 2026 announced February. Nickle LaMoreaux (IBM CHRO): “while reducing entry-level hiring saves money in the short term, it can lead to capability gaps later that have to be filled by expensive external hires who need time to adapt to a new corporate culture.” + Agi Garaba (UiPath CPO): “If you stop bringing in young workers, you are ultimately eliminating your growth engine.” The IBM/UiPath counter-trend pairs with the Evans AI-labs-themselves-keep-hiring observation as twin reasons-to-doubt the AI-will-collapse-aggregate-entry-level-hiring default reading.

The Carl Benedikt Frey no-automobile-industry counter-precedent ( WP Intelligence May 2026)

The wiki’s first named-academic-altitude empirical counter to the every-prior-automation-wave-created-new-jobs historical-induction argument:

“It’s entirely plausible that AI will create new kinds of businesses,” said Carl Benedikt Frey, associate professor of AI and work at Oxford University. “But it’s hard to see it creating something like the automobile industry that [generated] many new jobs.” — quoted in Giles 2026.

Frey is the named voice on the historical-induction-may-not-hold-here side, against the Evans typesetters-and-telephone-operators / lump-of-labor counter-frame. The two framings are not strictly contradictory — Frey allows that AI could create new businesses; he specifically doubts that the scale of new-job-creation will match prior platform shifts. Useful as the third pole in the wiki’s displacement debate triangulation: Brynjolfsson Canaries (empirical signal of early-displacement happening now), Evans (historical-induction argument for new-jobs-will-emerge), Frey (academic-altitude historical-induction-may-not-extrapolate-to-this-wave caveat).

The Peron do-more-with-the-same healthcare-domain reframe ( MIT SMR May 2026)

Peron (May 2026) adds a structural-supply-shortage-domain worked example that explains why the displacement-by-default framing does not fit cleanly in healthcare:

“A decade ago people were saying we’re never going to have radiologists again because the machines are going to do everything. That narrative has really not played out at all.”

The mechanism Peron names: in a structurally-undersupplied domain (radiology globally, but also clinical practice broadly — “we have deserts, areas where people don’t have any access to care” exist in the US, Europe, Asia, not just low-income settings), AI changes which work the clinician does without reducing the number of clinicians needed. The binding economic constraint is supply, not labor cost. “As we incorporate technology we’ll be able to do more with the same not with less.”

This sharpens the wiki’s reading of the AI-employment-effects debate at a domain-conditional layer: where supply is structurally short (healthcare clinicians, certain skilled-trade professions, eldercare), AI augments without displacing; where supply is at-or-above demand (mass-market customer-support, entry-level white-collar processing, language-only-tasks marketing), AI displacement is more likely. The Giles April-2026-layoff data sits closer to the supply-at-or-above-demand pole; the Peron healthcare-radiology case sits at the supply-short pole. A single occupation-level will-AI-displace-this-role? prediction without naming the supply-side condition is structurally incomplete.

The Linux Foundation not-a-jobs-crisis survey anchor + the European entry-level corroboration (LF Global + LF Europe 2026)

The two 2026 Linux Foundation State of Tech Talent reports supply the wiki’s first workforce-survey-side evidence on the AI-and-jobs question (n=400 global / 157 European; Marco Gerosa & Adrienn Lawson), and they cut both ways on the debate above:

Aggregate: AI is not eating IT jobs. Net hiring effect +26% (2025), +31% (2026) globally; +27%/+17% in Europe. Only the largest organisations report a negative net effect (−4% global; −15% Europe for 20,000+-employee firms for 2027). Demand for AI-specific roles is +64% (Europe). This is the survey-side complement to the lump-of-labor counter-frame and to the aggregate-employment-still-growing reading of Brynjolfsson, Chandar & Chen.
Compositional: the entry-level tier is where the pressure concentrates. The Europe report finds a −3% contraction in entry-level technical roles (vs +14% in the rest of the world), warning of a future mid-to-senior shortage if the junior pipeline is not maintained. This is the wiki’s first independent survey corroboration of the Brynjolfsson-canaries entry-level-decline finding (which used ADP payroll data) — two different instruments converging on the same compositional twist, now with a regional qualifier (sharper in Europe). It directly substantiates the debate’s central nuance: aggregate IT employment can rise while entry-level employment falls.
Mechanism: it’s an operationalisation/skills constraint, not a headcount cut. The reports frame the gap as capability, not supply (“not a jobs crisis, but a skills crisis”) — roles are being redefined and expanded, not eliminated. The displacement that exists is concentrated (largest enterprises; junior tier), consistent with the wiki’s domain-/tier-conditional reading (see the Peron supply-shortage reframe above): a single occupation-level will-AI-displace-this? prediction is incomplete without naming the firm-size and seniority conditions. The upskilling-as-pipeline-investment response lives in durable-skills.

A caveat the wiki carries on both reports: the publisher (Linux Foundation Education) sells the training/certifications the reports recommend, so the prescription is interest-aligned even where the employment data corroborates independent sources — hence confidence held at 0.75 on the sources, with the entry-level finding’s weight resting on its agreement with Brynjolfsson.

The developer-tooling-vendor displacement-but-net-growth framing ( Microsoft Dec 2025)

The GitHub/Microsoft Agentic DevOps keynote states the wiki’s debate from inside the developer-tooling vendor — and notably, the presenters foreground their own stake (“I like to stay employed… I want to stay employed too”). The framing: displacement is real and continuous (Randell’s 1980s typing-pool analogy — a job that simply no longer exists), but it is set against the World Economic Forum projection of 78 million net new jobs by 2030, with the prescription being collaborative use (“it’s about using these agents… to work collaboratively with the agents to get stuff done. So you’re not going to go away”). This is a vendor-practitioner restatement of the lump-of-labor counter-frame above and the jobs-redefined-not-eliminated reading of the Linux Foundation surveys — calibrated by the same honest caveat the wiki applies elsewhere: “this is the data we have today. In five years, it could all be different.” The corollary the keynote ties to jobs is skill-shift, not headcount — the durable move is from writing every line yourself to reviewing, steering, and orchestrating agents (see durable-skills and agentic-engineering).

Performative oversight + reshaping not unemployment: the June-2026 leadership read (MIT SMR + AWS)

Two June-2026 sources add a leadership-altitude reading of AI’s labour effects:

The skeptical-CIO worry: humans pushed into hollow oversight roles ( CIO Symposium). Thomas Davenport’s worry is a labour-quality claim, not a headcount one: agents work so much faster that human review becomes “performative” and “cursory” — people “pestered to approve things rapidly” and “not going to want to be auditors of what AI is doing.” The displacement risk he names is of meaningful human work being replaced by rubber-stamp auditing nobody wants. Max Chan’s counter-prescription is the optimistic version of the same dynamic: humans “replace themselves at the lower level” (training the AI to do their old work) while moving up to “top-line and bottom-line considerations” — explicit job-laddering-up as the augmentation path, contingent on the org actually creating the higher-level roles.
The empirical spine restated: reshaping, felt at entry level ( AWS Sydney). Brovich presents the Anthropic “March study” (Massenkoff & McCrory) as the talk’s ground truth: observed AI exposure ≪ theoretical (computer/math ~33% of what’s possible; office/admin theoretical ~90%, observed a fraction); no systematic unemployment rise since ChatGPT; junior hiring slowed ~14%; and — counter to the old narrative — the most-exposed workers are “disproportionately older, more educated, and better paid.” His summary (“not mass unemployment… a reshaping, and entry-level is where the reshaping is happening first”) is a vendor-altitude restatement of the Brynjolfsson Canaries entry-level finding, paired with the hourglass-organization prescription already on the Allen page.

Reshape ≫ replace: the BCG six-segment model ( BCG, April 2026)

BCG’s microeconomic role-model (Revelio 1,500-role taxonomy + O*NET) gives the wiki its most structured role-level taxonomy of AI’s labor impact, and — crucially — a method-independent corroboration of the Anthropic observed-exposure report: two entirely different methods (BCG demand-modeling vs Anthropic usage-data) land on the same headline — reshape ≫ replace, slow substitution, entry-level first, not a near-term mass-unemployment story.

The headline split: 50–55% of US jobs reshaped within 2–3 years (same role, radically new expectations); 10–15% vulnerable to elimination over 4–5 years (full substitution is slower than augmentation). 43% of jobs are ≥40% automatable; the other 57% are physical/interaction-heavy.

The six AI Labor Disruption Segments (with the load-bearing insight that automation potential alone doesn’t predict job loss — demand expandability does; see automation-vs-augmentation):

Segment	Logic	Share
Amplified	augment + expandable demand	5%
Rebalanced	augment + bounded demand	14%
Divergent	substitute + expandable demand (entry-level exposed)	12%
Substituted	substitute + bounded demand (net losses)	12%
Enabled	AI embedded, structure unchanged, skill bar rises	23%
Limited-Exposure	low automation feasibility	34%

Four transformation side-effects that “headline job numbers can mask”: (1) upskilling/redeployment speed is the binding constraint; (2) entry-level hiring shrinks then is redefined toward supervising-AI work, with AI fluency becoming a complement to tenure (sometimes advantaging AI-fluent juniors); (3) skill thresholds rise (durable roles need higher credentials/seniority → barriers to entry — see durable-skills); (4) cognitive load intensifies. These sharpen the wiki’s entry-level-decline thread (Brynjolfsson, Massenkoff-McCrory) with a mechanism: the rungs aren’t just fewer, they move up.
The CEO warning worth keeping verbatim: “Those who cut their workforce beyond AI’s ability to replace it will see productivity drop, institutional knowledge disappear, and critical talent walk away.” — the labor-side statement of the micro-productivity-trap.

Occupation boundaries blur, domain expertise persists: the agentic-coding read ( Anthropic, June 2026)

The AEI’s agentic-coding report gives the wiki a usage-side, session-level read on the same occupation-blurring the observed-exposure report (same authors) measured at the occupation level — and adds a sharp claim about which human input keeps its value.

Coding-occupation boundaries are dissolving in practice. Across ~400,000 Claude Code sessions, in code-producing work every one of the ten largest occupations reaches verified success within ~7 points of software engineers (≈34% vs 29% for non-software professions; the gap has neither widened nor narrowed over seven months). “Coding agents are making a coding background less relevant to successful programming” — a direct, usage-grounded counterpart to the exposure report’s finding that the most-exposed occupations are no longer the low-skill ones.
Returns to domain expertise persist and may sharpen. Success rises with task-specific expertise (verified success 15% novice → 28–33% intermediate+), and experts extract far more from each instruction (~12 actions / 3,200 words per prompt vs ~5 / 600 for novices). The labor signal: agentic tools absorb implementation-heavy work while rewarding command of the problem domain — a complement to durable-skills and a tension with pure ai-deskilling (handled on those pages).
A leading indicator to watch. Anthropic flags that if returns to expertise begin to fall, it would signal models supplying the judgment users currently bring (gains broadening beyond domain experts); and if non-software occupations’ success keeps rising, software production may be becoming ordinary work in every field. Both would reshape who benefits from agentic work — the kind of compositional shift this page tracks.

The professions-will-lobby + new-jobs-and-solo-founders read (Mollick & Chou, June 2026)

Two June-2026 interviews add a political-economy and an entrepreneurship angle the page’s empirical anchors don’t cover:

A Bit of Optimism — the professions will fight, and the fights determine outcomes. Mollick’s reading of prior industrial revolutions: the benefits got spread “not because the technology made everything great alone” but because labour fought capital (unionisation). Applied to AI: doctors and lawyers have guilds, licensure, and political power (much of Congress is lawyers) and will pass laws requiring a human sign-off “even if they’re worse than AI” — whereas coders “do not have the protection that doctors or lawyers or actors” have. He also restates the jobs-we-can’t-imagine point (80% of today’s jobs didn’t exist 20 years ago → 80% of jobs in 20 years are unimaginable now) and warns of the prompt-engineering-degree fallacy (a skill that vanished in months). Adds a guild/regulation-mediated mechanism to the page’s mostly-empirical displacement story.
YC Lightcone — the abundance / more-small-business counter-read. Chou’s “AI white pill” against the doomer scenario: AI is “the great equalizer” that lets the deep-but-narrow builder also do go-to-market, so “entrepreneurship might become way more important” and there will be “way more small business in the future.” The age-of-the-40-year-old-solo-founder thesis is the optimistic compositional shift — not net job loss but a redistribution toward many tiny, experience-led firms (one person + agent “clones”). A useful counterweight to the page’s entry-level-decline findings: the displaced-mid-career worker as a prospective founder, not only a casualty.

Per the Lifecycle rules these two interview/popularisation sources do not lift the concept’s confidence; they widen the page’s coverage to the political-economy and entrepreneurship angles alongside the empirical displacement panel.

A third June-2026 CEO voice, Sridhar Ramaswamy (Snowflake), holds both the anxious and the growth read at once: he is “terrified about the future employment” of his two software-engineer sons (26 and 24) and pushes technical staff to be “at the cutting edge … protect your livelihood,” yet his own firm’s response is redeploy-not-cut (the demo team moved to other roles; “real opportunities for people I haven’t been able to fill before”) and grow output, not shrink headcount. The “uber programmers 50–100×” claim is the productivity-per-worker correlate of the uneven-adoption split he also names (“there are people … struggling to adopt”). A CEO-altitude statement of the augmentation-with-reinvestment beats displacement posture — with explicit acknowledgement that the transition is uneven across workers.

Nadella (Microsoft, June 2026) supplies a platform-CEO articulation of the social-permission dimension of AI employment effects — and an unusually self-critical one about the industry’s rhetoric. His analogy: in the early 1980s nobody predicted “four billion typists” would become knowledge workers creating artifacts; the future of work under “20,000 employees and 20 million agents in a loop” is similarly not yet conceptually understood. But the binding constraint he names is trust, not capability: the sector lost social permission by saying “all economic opportunity will go away for knowledge workers… white-collar jobs are gone” while being excited to build it — “why would anyone want you to be successful?” (he cites a college-commencement speaker being booed for an AI-displacement line). The repair he prescribes is to stop arguing the lump-of-labour fallacy abstractly and instead name “the new jobs, the wages of the new jobs, what I can train myself for” — concretely, locally, with proof points. This converges with Argenti’s and the Giles/WP role-redesign material on what the new work is, but adds the macro claim that demonstrating tangible benefit is itself the gating variable for whether AI’s employment transition is socially permitted to proceed.

The offshoring vantage: India’s IT-services sector + the human-in-the-loop training-labor economy ( Kaushik, June 2026)

The FT documentary adds the wiki’s first field-journalism, Global-South vantage on AI’s labor effects — and it sits at a structurally informative point in the offshore chain: the same IT-services work the wiki’s US/European empirical sources measure (Brynjolfsson Canaries, AI Index 2026) is the work India built a ~$330–340bn export economy around over 25+ years.

The displacement vantage is the receiving end of the offshore chain. India’s IT/ITeS export — the country’s largest single export and a primary dollar-income source — was built on “under-promise, over-deliver” at “the right price point” (“if you are paying $1… I can get it done for 20 cents”). The film’s framing: AI is “the toughest challenge India’s IT services sector has faced in their existence,” because the “high-volume, repetitive tasks” with “a margin of error that might be somewhat acceptable” are “a prime candidate for AI substitutability.” This is the same substitutability mechanism the page documents for US entry-level coding/CS roles, viewed from the vendor economy that absorbed that work in the first wave of offshoring — a recursive twist the empirical US-payroll sources can’t see.
A new labor layer the page hadn’t named: human-in-the-loop training labor at population scale. The film documents the data-annotation and RLHF labor that trains models built elsewhere — “India is probably the second largest AI workforce in the world today”; “almost all the products that are there globally, at some point, will be trained by somebody in India.” The newest frontier is egocentric data for humanoid robots: textile workers in Karur wearing GoPros / Meta glasses 6–8 hours/day to capture folding, sweeping, dish-washing for robot training (“San Francisco offices don’t train the robot… it’s people sitting in small-town India”). This is labor that AI demand creates — “the booming AI industry needs more and more humans in the loop” — and it complicates the displacement-only reading: the same technology that erodes the IT-services tier spins up a lower-value annotation tier beneath it. (Whether that constitutes climbing or descending the value chain is the film’s central open question — flagged as a candidate concept page: human-in-the-loop data-labor / “AI factory” economy.)
The lump-of-labor counter-frame appears here too — “while certain kinds of tech jobs may go out, there will be certain other newer kinds of tech jobs which will get created,” and “a distinct possibility that the overall number of jobs in the tech sector in India will actually go up.” This is the same Evans-style historical-induction argument, contested within the film by the extraction-skeptics (see responsible-ai). The binding constraint stated plainly: “India has talent at very large scale. But if you don’t have enough jobs for these people, what is the value of the scale? It’s nothing.”

This is qualitative, journalistic evidence, not a measured labor-market study — it does not lift the page’s confidence. Its value is widening the page’s geography (offshore-vendor economies) and surfacing the training-labor-creation counter-current to displacement.

The Netherlands sectoral vantage: occupational-exposure scoring for IT + business services ( Rabobank, June 2026)

The wiki’s first Dutch national / sectoral empirical source applies the occupational-exposure method-family to the Netherlands: RaboResearch (June 2026) maps Gmyrek et al. (2025, ILO Working Paper 140) task-level GenAI-automatability scores onto the CBS Nederlandse Beroepsclassificatie, binning every task into hoog / potentieel / laag / niet-mogelijk. It is a method cousin of Anthropic (both convert task ratings into occupational exposure) and corroborates the Brynjolfsson Canaries / BCG reshape ≫ replace compositional finding from a fresh geography and instrument.

Sector exposure: 86% of IT tasks (Informatie & communicatie: 49% hoog + 37% potentieel) and 64% of business-services tasks (33% + 31%) have (hoog) automatiseringspotentieel — among the most-exposed sectors in the Dutch economy, alongside financial services; agriculture, horeca and construction are least exposed.
The compositional-shift mechanism, sector by sector: accountancy, advocatuur and consultancy all show the same pattern — routine execution (boekkeeping, contract analysis, jurisprudence search, data-analysis/reporting) automates while value concentrates in complex judgment, advice and client work, and demand for junior/instap roles falls. In IT, senior developers move toward quality assurance, architecture and “IT governance” (defining the context agents operate in) while routine programming, testing and documentation are absorbed — the executing→advisory shift this page tracks, restated at the Dutch-sector level.
Per-occupation ranking (Gmyrek-derived): most-exposed are boekhoudkundig medewerkers (~100% hoog), secretaresses, administratief medewerkers, callcentermedewerkers, databank-/netwerkspecialisten; least-exposed are schoonmakers and beveiligingspersoneel — the same routine-cognitive-high / physical-low gradient documented elsewhere on this page.
NL ahead of EU average, not the leader (CBS enterprise survey, firms >10 employees): Dutch IT and business-services firms use AI above the EU mean in every business function but below the EU leader in every function — a national-adoption datapoint mirrored on enterprise-ai-adoption.
Potential ≫ realisation is stated explicitly: the scores measure potential when the technology is further developed and implemented, and RaboResearch flags that experimental productivity gains may not survive into daily practice — the bridge to the micro-productivity-trap (see that page for the study’s task-level-vs-firm-level productivity table).

This is a research-note-altitude sectoral study (bank research arm; figures sourced from CBS/UWV/ILO), not a peer-reviewed large-N panel — it widens the page’s geography to the Netherlands and its method coverage to a second occupational-exposure instrument, without lifting confidence past the ceiling already set by the US empirical anchors.

AI washing: the attribution confound gets its own page ( BBC, June 2026)

The wiki’s attribution-skepticism thread — “Are these declines really AI?” (Debates, below), the [[2026-01-09-baron-signals-for-2026|O’Reilly “the decisions of people deploying it”]] framing, the Giles “AI #1 cited reason” vs. NACE 14% gap — now has a dedicated concept page, ai-washing, anchored by labor economist Kathryn Anne Edwards (BBC, June 2026). Her contribution is the mechanism beneath the over-attribution: AI is not treated neutrally by the stock market, so “we pivoted to AI” is a more shareholder-friendly story for layoffs than “we overhired” — a valuation premium that incentivises citing AI for cuts a firm would have made anyway (analogised to greenwashing, reinforced by downturn peer pressure). This is the corporate-incentive explanation for why the attribution noise on this page runs predominantly toward over-crediting AI; it converges with Sam Altman’s executive-altitude AI washing counter-framing (via Everitt).

Edwards also adds a counter-data point to the entry-level-decline narrative: Indeed software-development job postings rose from far below overall postings in early 2024 to ~4× higher by mid-2026, outperforming overall jobs over the prior year — a journalist-surfaced counterweight to the AI Index / Canaries software-dev figures, in the same direction as the Evans lump-of-labor and Linux Foundation “not a jobs crisis” poles. It is narrative critique, not a new measurement (single chart, single platform) — weight accordingly. Full treatment, including the measurement-impossibility thesis (“we may never know… inconclusive forever,” the jobs-lost-to-computers-since-1955 analogy), lives on ai-washing.

AI reshapes hiring at the job-definition layer, not selection ( Stanford GSB, June 2026)

A narrow but distinct mechanism, surfaced as an aside in Glenn Carroll’s organizational-culture talk (he explicitly defers on AI but relays the finding). Citing Isabel Fernández-Mateo (London Business School), Carroll’s claim is that the common assumption — “AI is really just going to affect who gets hired and how” (i.e. selection) — is wrong about where the change lands:

“The real change is going to occur not with the selection but with the way the jobs get defined and the way people care about those jobs in the first place… how you attract the candidates and how you shape the applicant pool.”

Two consequences worth tracking against the page’s selection-focused displacement evidence:

The locus of AI’s hiring effect is upstream of selection — job design, candidate attraction, and applicant-pool composition — not (only) the resume-screen/matching step that most “AI in hiring” discussion fixates on.
Applicant-pool homogenisation. Candidates increasingly use AI to produce application materials, which are becoming “more and more homogeneous,” so “the hiring process has become very, very difficult” even though AI lets firms process far more applicants. This is a demand-side analogue to the communication-convergence finding (AI making low-skill workers communicate like high-skill ones): AI-mediated applications converge, eroding the signal hiring relies on.

This is qualitative, second-hand testimony, not a measured study — it does not move the page’s confidence. Its value is naming a mechanism (hiring-process effects concentrate on job-definition + pool-shaping, and AI-written applications homogenise) the page’s displacement/exposure literature does not otherwise cover. Carroll’s broader thesis — culture as the strategic-renewal microfoundation that governs how empowered employees feel to act — also bears on ai-knowledge-hiding (whether workers disclose their AI use).

”We won’t run out of jobs unless we run out of ideas”: the onlyness/agency read (Wood, WorkLab, June 2026)

Aneesh Raman (LinkedIn’s Chief Economic Opportunity Officer) adds a named voice to the page’s lump-of-labor counter-frame cluster (Evans; the WEF 78-million-net-new-jobs projection): “we’re not going to run out of jobs unless we run out of ideas” — credited in the interview to Jensen Huang (Nvidia CEO). The argument for why ideas won’t run out: most economic growth to date has optimised for “consumer convenience and enterprise production” under industrial-age math, leaving “any number of arenas, including societally beneficial arenas, from health to climate” under-explored — and AI “lowers the barrier to expertise and knowledge” as well as “the barriers to entrepreneurialism and innovation.”

Two contributions distinct from the page’s empirical anchors:

Identity de-anchored from job title (“onlyness”). Raman’s framing — professional identity should be a unique, non-linear combination of curiosities and capabilities rather than a job title inherited from assembly-line-era org charts — is the individual-level complement to the page’s occupation-level displacement data. It converges independently with Schoening’s “cultivating agency matters more than job titles” framing (near-identical title, independent interview) and with Argenti’s mindset-not-skillset inversion — three sources now converge on identity/mindset shift as the individual-level response, distinct from (and upstream of) any specific skills inventory.
Labor-market opacity + Global-South equity framing. Raman names the pre-AI labor market as already “one of the least efficient, least dynamic, least transparent markets humans have ever created” — most hiring runs on guesswork and pedigree signals, a baseline dysfunction independent of AI. His equity argument: unlike prior general-purpose technologies, which diffused “top down, over years” before reaching workers, AI is unfolding “bottoms up, middle out, overnight” — a structural opportunity if paired with (a) radical transparency about what’s happening and (b) broad adoption, especially in the Global South, where Microsoft president Brad Smith’s electricity-access cautionary tale (large parts of the world still lack electricity, and were correspondingly locked out of the growth it enabled) is invoked as the risk case. This is a popular-press restatement of the social-permission thread Nadella names elsewhere on this page, applied at global rather than domestic-workforce scale.

Per Lifecycle rules this is a single-source, promotional-book interview — it does not lift the page’s confidence past its 0.95 ceiling. Its value is a named addition to the lump-of-labor cluster (Jensen Huang) and an individual-identity vantage the page’s occupation-level data doesn’t otherwise carry.

Debates and supersession

The “equalizing effect” vs. employment displacement. As above — both can be true simultaneously, but the popular reading of “AI helps low-skill workers most” has been over-extended to imply pro-employment outcomes. The Brynjolfsson 2025 paper sharpens that.
Are these declines really AI? Brynjolfsson et al. are explicit that other factors may be in play. They’ve ruled out: COVID-era effects, tech-sector contraction, remote-work / outsourcing pressure, firm-level shocks, and education declines from COVID school closures. But they cannot rule out all confounds. Treat as strong correlational evidence, not definitive causal proof.
Will the equalizing-effect hold at scale? Robust in early studies (customer support, consulting, software). Open question: as AI tools mature, do high-skill workers eventually catch up by leveraging more sophisticated workflows?
Replacement vs. augmentation in the long run. AI Index 2025 notes that the share of orgs predicting workforce reductions has declined YoY — business leaders are becoming less convinced AI will shrink workforces. Yet Brynjolfsson 2025 shows entry-level employment is declining in automation-exposed occupations. Resolution: aggregate workforce expectations remain stable while compositional shifts disadvantage entry-level workers.
Geographic and platform effects. ADP data is U.S.-only; somewhat overrepresents Northeast and manufacturing/services. Whether the Brynjolfsson pattern generalizes to other countries is an open question.

automation-vs-augmentation — the conceptual cut driving the empirical pattern
ai-deskilling — task-composition shift within retained jobs (distinct mechanism from displacement)
enterprise-ai-adoption — the organizational decision-making side
generative-ai — the technology driving the displacement
ai-agents — concentrated in the automation quadrant where employment is declining
responsible-ai — the labor disruption is an under-attended RAI concern
ai-washing — the attribution-confound layer: how much of the AI-cited layoff signal is real vs. shareholder-facing narrative

The “AI is not taking jobs” attribution claim ( O’Reilly, January 2026)

The O’Reilly Radar editorial entry to the wiki adds a quotable attribution claim that complements the empirical record: “AI is not taking jobs: The decisions of people deploying it are” — Tim O’Reilly (via [[2026-01-09-baron-signals-for-2026|Baron’s annual Signals for 2026 outlook]]). The wiki absorbs this as the deployment-decision-attribution framing — useful as a rhetorical counter to “AI as automatic-displacement” narratives, but not as a claim that AI lacks labor-market effects (the empirical record on Brynjolfsson’s canary and the AI Index 2026’s −20% software-dev employment from 2024 are both load-bearing). The wiki’s reading: the labor-market effects are real and measurable; the attribution to deployment decisions rather than to the technology itself is a useful framing for thinking about agency and policy levers.

Mixed-signal labor-market data from Loukides’s monthly digests (Apr–May 2026)

The April 2026 Radar Trends digest adds two trade-press datapoints worth tracking:

Software-development employment −20% from 2024 — direct match to the AI Index 2026 headline figure.
Product-manager roles at decade highs; engineering demand recovering. The PM/engineering split is the wiki’s first observation that AI’s labor-market effect is role-asymmetric within tech — not a uniform tech-sector decline. Track for future ingest of any large-N study that decomposes by role.
“Copilot correlates with reduced management and collaborative time.” First wiki signal on AI’s labor-market effect at the team-shape level — not just headcount but what work the remaining people do. Adjacent to ai-deskilling (the skill-formation side) but distinct: the collaboration-erosion claim is about how teams work, not what individuals know how to do. Worth a future synthesis-pair with the February digest’s Kent Beck reframe (“AI augments junior developers, accelerates learning cycles”) — Beck’s claim and the Copilot-collaboration finding are not strictly contradictory but they cut in opposite directions on whether AI grows or shrinks the team-coordination surface.

AI-Wiki

Explorer

ai-employment-effects