Atlas of the Present Atlas · v0.1

As of 28 October 2026

Dossier 11

Work and Education

v2.1-draft · unchecked

1. As of

Date
Version
v2.1-draft
Author / model
Atlas generator
Reviewer
unchecked (Josef)

2. In one sentence

The education gap in generative AI compared with the internet (Watanabe; arXiv:2610.36181, submitted 2026-10-26, E2, abstract only): using public Eurostat survey data (ages 16–74), each country's 2025 generative AI use is matched to the internet year with similar overall use (within 5 percentage points); in the 16 of 35 countries that could be matched, the gap between high (ISCED 5–8) and low (ISCED 0–2) education was smaller in percentage points in 15 of 16 countries (median −27.6 points, p = 0.0005) and as a log ratio in 13 of 16 (median −0.625, p = 0.021) — the ratio result is fragile, and these are unadjusted contrasts in a non-random set of countries, not an education effect. The delegation divide (Espinal Maya, Jiménez Londoño; Universidad Nacional de Colombia/EAFIT arXiv:2610.30752, E2): two Claude models rate all 18,796 O*NET tasks on a preregistered rubric, weighted by employment in 146 countries — 35.2% of working time in 2026 is said to be at least partly delegable to AI agents, 2.5% fully; 48.7% in high-income vs 22.5% in low-income countries, because physical work dominates there; correlation with existing exposure indices r = 0.93 — a model judgement of technical feasibility, a ceiling rather than adoption, US task mix for every country, Haiku alone would give 51.4%. Better work, weaker result (Fu, Luo, Ramasubbu, Shen, Wang, Wang arXiv:2610.25603, E2): after an AI voice agent took over the cold leads in a Chinese insurance call centre, the share of hot leads reaching humans on the first three calls of the day rose by 23.4 percentage points — yet their conversion there fell by 4.5 pp, they handed leads back 8.1 pp more often and their calls were ~13% shorter; on hot leads overall 4.4% → 3.7% (descriptive). One firm, working paper. Metric ≠ teaching (Dominguez Figaredo & Fernández De la Cruz arXiv:2610.12125, E2): Gemini 3.1 Pro Preview audited on 2,000 synthetic learner cases, weakest cases corrected via LoRA on Qwen3-32B — Pedagogical Adaptivity Index +0.05 → +0.42, but 31 trained raters (graduate students, blind) 4.46 → 3.03 (p < 0.00001; controls 4.39 → 4.51); the corrected model repeats the same option across all five activities in 100% of cases, also out of sample (200 cases); human rating in-sample only. O*NET-BENCH (Lyu & Thompson arXiv:2610.02492, E1): 33 existing LLM-judge configurations against 4,501 test ratings from 45,796 worker ratings — ordering mostly above chance (25 configurations tie-aware ≥0.60; TF-IDF baseline nearly level), but the estimated share of acceptable AI responses ranges 3.0–97.9% vs workers 61.1%; calibrated scores explain ≤8.5% of worker variance — ranking ≠ acceptance rate ≠ occupational aggregate. EngiWorld (Gao et al. arXiv:2609.37686, E1): 1,301 expert-curated tasks; 6 domains (CAD/CAE/CAM/BIM/EDA/3D); 26 software platforms; GUI+CLI; artifact-centric verifiers; eval subset 300×7 frontier models; best EngiScore 44.3 (Claude Opus 5); GPT-5.6 Sol 38.0; multi-software only 3.6% (6/168); 300≠1,301; VM/license-specific. Medly micro-RCT (Harrison et al., E1): GCSE ITT g=0.33; Bastani Base −17%; Noy Science tasks≠jobs; PISA ≠ gen-AI RCT; UNESCO still unopened.

Established now · E1 / E2

3. What works today

Education gap in generative AI use in 2025 smaller than the internet's at a similar use level — in 15 of 16 European countries in percentage points (Watanabe, abstract)

  1. E2

    Watanabe (arXiv:2610.36181v1, submitted 2026-10-26, cs.IR/cs.CY; only the abstract was opened, 2026-10-28) asks whether the education gap in generative AI use in 2025 is smaller or larger than the internet's gap in the same country in the year when overall internet use was at a similar level. The data are public Eurostat survey data on ICT usage by individuals aged 16 to 74. Each country's 2025 rate of generative AI use in the last three months is matched to the internet year whose overall last-three-months use was closest, within 5 percentage points, and the gap between people with high (ISCED 5 to 8) and low (ISCED 0 to 2) completed education is measured as a log ratio and as a percentage-point difference. Sixteen of 35 countries could be matched; most of the others were early internet adopters. In the matched countries the generative AI gap was smaller in 13 of 16 countries on the log ratio (median change −0.625, 95% bootstrap interval −0.969 to −0.291, two-sided sign test p = 0.021) and in 15 of 16 on the percentage-point gap (median change −27.6 points, interval −32.2 to −16.8, p = 0.0005). The point-gap result keeps its sign in every generative-AI-versus-internet sensitivity analysis with 8 or more countries, including comparisons restricted on both sides to ages 25 to 54 and 55 to 74, but its size is sensitive to matching that takes education composition into account (about −10 to −11 points in 7 and 8 countries).

    Watanabe GenAI vs internet education gap. Is the education gap in generative AI use smaller than the internet's was at a similar level of use? Evidence from 16 European countries. https://arxiv.org/abs/2610.36181. As of arXiv:2610.36181v1 (submitted 2026-10-26; cs.IR, cs.CY; per comment 16 pp., 2 figures, code, data and outputs as ancillary files); abstract via export.arxiv.org API only, PDF and ancillary files not opened; affiliation not stated in the abstract record. Checked 2026-10-28. Type: Paper (arXiv; abstract only; secondary analysis of public Eurostat ICT-usage survey data, self-report).

Agent Delegability Index: 35.2% of working time partly, 2.5% fully delegable to AI agents — 48.7% rich vs 22.5% poor countries (Espinal Maya, Jiménez Londoño)

  1. E2

    Espinal Maya (Universidad Nacional de Colombia), Jiménez Londoño (Universidad EAFIT; arXiv:2610.30752v1, submitted 2026-10-22, 32 pp., PDF opened 2026-10-23): Claude Sonnet 5.5 (primary) and Claude Haiku 5.5 rate all 18,796 tasks in O*NET 30.2 on a frozen, preregistered rubric — no role / augmentation / partial / full delegation to an agent — plus four barriers (physical, accountability, trust, access); carried to 433 ISCO-08 unit groups and weighted by ILOSTAT employment in 146 countries (2,336 million workers). Result (consensus = lower level): 35.2% of working time at least partly, 2.5% fully delegable; Sonnet alone 35.3%, Haiku alone 51.4%. By income 48.7 / 35.0 / 28.3 / 22.5% (high to low), slope 7.23 percentage points per log point of GDP per capita (95% interval 6.52–7.83); with ESCO skills instead of O*NET 4.66. Top Singapore 64.2%, Luxembourg 63.5%, UK 54.9%; US 51.9%, India 27.8%; bottom Burundi 14.7%.

    Espinal Maya & Jiménez Londoño Agent Delegability Index. The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries. https://arxiv.org/abs/2610.30752. As of arXiv:2610.30752v1 (submitted 2026-10-22; econ.GN; 32 pp.); PDF research/pdfs/arXiv2610.30752.pdf; index, prompts and MCP server github.com/Cespial/delegation-divide not opened. Checked 2026-10-23. Type: Paper (arXiv; PDF opened; preregistered model-rated index, 146 countries).
  2. E2

    Same source, occupations and cross-checks: managers 81.8%, clerical support 78.3% (by far the largest share of full delegation), professionals 69.9%, elementary occupations 12.3%; the physical barrier flags 72–94% of task weight in manual and service groups — the divide comes from the composition of employment, every country receives the same occupational index. Full delegation peaks for tasks of about 11.8 minutes (the minutes themselves a model estimate). Correlation with exposure per Eloundou et al. r = 0.93, per Felten et al. 0.89, with observed Claude usage by occupation 0.69; the preregistered assumption that middle-income countries use AI less than their potential predicts is not supported (r = 0.86 across 101 countries, p = 0.270). Model agreement κ = 0.77; blind human rating of 300 tasks κ = 0.85.

    Espinal Maya & Jiménez Londoño Agent Delegability Index. The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries. https://arxiv.org/abs/2610.30752. As of arXiv:2610.30752v1 (submitted 2026-10-22; econ.GN; 32 pp.); PDF research/pdfs/arXiv2610.30752.pdf; index, prompts and MCP server github.com/Cespial/delegation-divide not opened. Checked 2026-10-23. Type: Paper (arXiv; PDF opened; preregistered model-rated index, 146 countries).

AI voice agent takes the cold leads: early in the day humans get 23.4 pp more hot leads — and close 4.5 pp less often there (Fu et al.)

  1. E2

    Fu, Luo, Ramasubbu, Shen, Wang, Wang (Georgia State, Renmin, Pittsburgh, Waterdrop Research Institute, PolyU, Sydney; arXiv:2610.25603v1, submitted 2026-10-19, working paper, PDF opened 2026-10-20): a licensed Chinese online health-insurance broker (name withheld under NDA, ~10,000 outbound calls per workday) changed its routing on 28 August 2024 — cold leads (dropped out before the payment page) go by default to a scripted AI voice agent, hot leads (dropped out on the payment page) to humans with priority; pay, training, scripts, evaluation and hiring stayed unchanged according to the firm. Regression discontinuity in time over ±60 days, 409,622 human and 157,468 AI calls, about 13,000 agent-days. On the first three calls of the day the probability of a hot lead rose by 23.4 percentage points (later towards zero); exactly there human conversion fell by 4.5 pp (next three −2.5 pp, middle positions unchanged, end of shift +3.4 pp), leads were handed back to the shared queue 8.1 pp more often, calls were ~13% shorter.

    Fu et al. parallel AI voice agent. When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition. https://arxiv.org/abs/2610.25603. As of arXiv:2610.25603v1 (submitted 2026-10-19; econ.GN; working paper, version September 2026); PDF research/pdfs/arXiv2610.25603.pdf. Checked 2026-10-20. Type: Working paper (arXiv; PDF opened; field data from one firm).
  2. E2

    Same source, Table 6 (descriptive accounting, not an effect estimate): conversion on hot leads 4.4% → 3.7% (−15.9%), across all human calls 3.2% → 2.9%; human calls −31.8%, AI calls +351%. Humans lose 50.5 hot-lead conversions per day, the AI recovers 7.7 on cold leads (~7 : 1). The metric “conversions per 50 human calls” still rises by 6.3% — only because the denominator falls faster. Model decomposition of hot-lead yield: −17.0% = better ordering +0.58% versus changed behaviour −17.58%. The day-level RDiT average effects on conversion and duration are small and not statistically distinguishable from zero (Table C1).

    Fu et al. parallel AI voice agent. When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition. https://arxiv.org/abs/2610.25603. As of arXiv:2610.25603v1 (submitted 2026-10-19; econ.GN; working paper, version September 2026); PDF research/pdfs/arXiv2610.25603.pdf. Checked 2026-10-20. Type: Working paper (arXiv; PDF opened; field data from one firm).

AI tutor: the metric rises, the teaching falls (Dominguez Figaredo & Fernández De la Cruz)

  1. E2

    Dominguez Figaredo & Fernández De la Cruz (UNED Madrid; arXiv:2610.12125v1, submitted 2026-10-08; PDF opened 2026-10-16): the Pedagogical Adaptivity Index (PAI, −1 to +1) scores five decision points per activity (content complexity, cognitive task, support, engagement, disciplinary approach) one by one against fixed case conditions. Audited: Gemini 3.1 Pro Preview on 2,000 synthetic learner cases; the weakest 589 cases corrected via LoRA on the open proxy Qwen3-32B; 31 trained raters (graduate students) blind, scale 1–5.

    Dominguez Figaredo & Fernández De la Cruz PAI. The learner who does not learn: when optimizing a pedagogical metric degrades LLM tutoring. https://arxiv.org/abs/2610.12125. As of arXiv:2610.12125v1 (submitted 2026-10-08); PDF research/pdfs/arXiv2610.12125.pdf. Checked 2026-10-16. Type: Paper (arXiv; PDF opened; preprint).
  2. E2

    Before: expert ratings unrelated to the PAI (ρ = −0.15, p = 0.39); lowest 30 vs highest 8 PAI cases 4.46 vs 4.39 (p = 0.53). After correction, same 30 cases: PAI +0.05 → +0.42, rating 4.46 → 3.03 (Wilcoxon p < 0.00001, all six items); unmodified controls 4.39 → 4.51 (p = 0.38). The corrected model picks the same option five times in 100% of cases (at LoRA rank 1, 4, 8, 16) and likewise on 200 unseen cases; on cognitive task all 200 learners got the same option, sub-score 0.24 → 0.04 — the overall PAI still rose.

    Dominguez Figaredo & Fernández De la Cruz PAI. The learner who does not learn: when optimizing a pedagogical metric degrades LLM tutoring. https://arxiv.org/abs/2610.12125. As of arXiv:2610.12125v1 (submitted 2026-10-08); PDF research/pdfs/arXiv2610.12125.pdf. Checked 2026-10-16. Type: Paper (arXiv; PDF opened; preprint).

O*NET-BENCH: AI judges order right, scale wrong (Lyu & Thompson)

  1. E1

    Lyu & Thompson (MIT; arXiv:2610.02492v1, PDF opened 2026-10-05): O*NET-BENCH — audit suite built from an existing corpus of 45,796 ratings of AI responses to O*NET workplace tasks by workers with self-reported occupational experience (Mertens et al. 2026; no new ratings). 33 existing LLM-judge configurations from six model families plus six train-fitted controls on 4,501 test ratings (task- and worker-disjoint splits).

    Lyu & Thompson O*NET-BENCH. Right Order, Wrong Scale: Auditing LLM Judges for Occupational AI Measurement. https://arxiv.org/abs/2610.02492. As of arXiv:2610.02492v1 (2026-10-01); PDF research/pdfs/LyuThompson_ONETBench_arXiv2610.02492.pdf. Checked 2026-10-05. Type: Paper (arXiv; PDF opened; preprint under review).
  2. E1

    Ordering: all 33 above chance; 25 reach tie-aware pair accuracy ≥0.60 (median 0.631, best 0.658); a response-only TF-IDF baseline 0.643. Scale: all 33 mean-bias intervals exclude zero (median −1.18 points on the 9-point scale). Workers rate 61.1% of test responses acceptable (≥7); judges 3.0% (Claude Opus 4.7) to 97.9% (Llama-3.3-70B, near-constant), median 43.9%. Worker majority acceptable in 68.7% of 307 occupations; Opus in 0.3% (disagreement 68.4%). Calibration removes mean bias, but scores explain ≤8.5% of individual worker-rating variance.

    Lyu & Thompson O*NET-BENCH. Right Order, Wrong Scale: Auditing LLM Judges for Occupational AI Measurement. https://arxiv.org/abs/2610.02492. As of arXiv:2610.02492v1 (2026-10-01); PDF research/pdfs/LyuThompson_ONETBench_arXiv2610.02492.pdf. Checked 2026-10-05. Type: Paper (arXiv; PDF opened; preprint under review).

EngiWorld: frontier agents in professional engineering software (Gao et al.)

  1. E1

    Gao, Qu, Lei et al. (Tsinghua / Zhiman / multi; arXiv:2609.37686v1, PDF opened 2026-09-30): EngiWorld — 1,301 expert-curated tasks; 6 domains (CAD/CAE/CAM/BIM/EDA/3D visualization); 26 software platforms; GUI+CLI; artifact-centric domain verifiers (geometry/physics/rule compliance).

    Gao et al. EngiWorld. EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments?. https://arxiv.org/abs/2609.37686. As of arXiv:2609.37686v1 (2026-09-29); PDF research/pdfs/EngiWorld_arXiv2609.37686.pdf. Checked 2026-09-30. Type: Paper (arXiv; PDF opened).
  2. E1

    Eval subset 300 tasks × 7 frontier models; best EngiScore 44.3 (Claude Opus 5); GPT-5.6 Sol 38.0 (Table 2). Multi-software: only 3.6% of attempts succeed (6/168 across models). Zero-shot; Windows/Ubuntu VM snapshots.

    Gao et al. EngiWorld. EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments?. https://arxiv.org/abs/2609.37686. As of arXiv:2609.37686v1 (2026-09-29); PDF research/pdfs/EngiWorld_arXiv2609.37686.pdf. Checked 2026-09-30. Type: Paper (arXiv; PDF opened).

Writing, code, customer service: tasks, not jobs

  1. E1

    Noy & Zhang MIT WP 2023 (source 1, still E1 for WP figures): 444 college-educated professionals, two writing tasks, GPT-3.5. Time −10 minutes / 37% vs. control mean 27 min. Grades +0.45 SD. 68% submit ChatGPT first output unedited. Inequality within the task compresses. Limits: short self-contained tasks; control leakage 10–20%; employment unobserved. Science figures (n=453 / −40% / +18%) now source 14 — not inferred from the WP.

    Noy & Zhang. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf. As of MIT WP 2023-03-02. Checked 2026-08-28. Type: Working paper.
  2. E1

    Noy & Zhang Science 381:187–192 (2023-07-13; DOI 10.1126/science.adh2586; author PDF opened 2026-09-11): preregistered online RCT, n=453 college-educated professionals, occupation-specific writing tasks, ChatGPT (GPT-3.5 era). Abstract: average time −40%; output quality +18%. Body: time −11 minutes (0.75 SD) vs. control mean 27 min (p<0.001); grades +0.45 SD (p<0.001). Almost everyone submitted lightly edited or unedited ChatGPT output; 33% without editing, 53% editing (editors only ~3.3 min active).

    Noy & Zhang Science. Experimental evidence on the productivity effects of generative artificial intelligence. https://doi.org/10.1126/science.adh2586. As of Science 381:187–192 (2023-07-13); author PDF research/pdfs/Noy_Zhang_Science_adh2586.pdf. Checked 2026-09-11. Type: Paper (Science; author PDF opened).
  3. E1

    Noy & Zhang Science follow-up: treatment ~2× as likely to report job use at 2 weeks (34% vs 18%); ~1.6× at 2 months (42% vs 27%). Control leakage ~10–20% ChatGPT use. Employment/wages unobserved. WP figures 444 / −37% / 68% unedited remain warrants of source 1 — do not restate as Science warrant.

    Noy & Zhang Science. Experimental evidence on the productivity effects of generative artificial intelligence. https://doi.org/10.1126/science.adh2586. As of Science 381:187–192 (2023-07-13); author PDF research/pdfs/Noy_Zhang_Science_adh2586.pdf. Checked 2026-09-11. Type: Paper (Science; author PDF opened).
  4. E1

    Peng et al.: RCT May–Jun 2022, 95 Upwork developers, 35 completers per arm. HTTP server in JavaScript. Completers: 71.17 vs. 160.89 min → −55.8% time. Success rate +7 pp, n.s. Code quality unmeasured. n small; authors Microsoft/GitHub. Extrapolation to US computer/math occupations in the discussion is speculation, not fact.

    Peng et al. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. https://arxiv.org/html/2302.06590. As of arXiv:2302.06590, experiment May–Jun 2022. Checked 2026-08-28. Type: Paper.
  5. E1

    Brynjolfsson, Li, Raymond NBER WP, rev. Nov 2023: 5,179 customer-support agents, one Fortune-500 software firm. Resolutions/hour +0.30 / +13.8%. Novices +34%; top skill ~0, quality partly −. Tenure <1 month +46%; >1 year ~0. Adherence ~38%. No evidence on wages, headcount, skill mix. QJE 2025 (15%, 5,172) not opened.

    Brynjolfsson et al. Generative AI at Work. https://www.nber.org/system/files/working_papers/w31161/w31161.pdf. As of NBER WP 31161, revised Nov 2023. Checked 2026-08-28. Type: Working paper.

Education: GPT Tutor design vs. unguarded GPT-4 (one field RCT)

  1. E1

    Bastani et al. PNAS 2025 (author PDF opened): field RCT at one Turkish high school, grades 9–11, mathematics, GPT-4-0613, four 90-min sessions; arms Control / GPT Base / GPT Tutor. Practice with tool: GPT Base +48%, GPT Tutor +127% vs control (Table 1: +0.137 and +0.361 on control mean 0.284; 2,848 observations). Subsequent unassisted exam: GPT Base −17% (−0.054; control mean 0.321); GPT Tutor ≈ control (−0.004, n.s.) — harm largely mitigated, no positive learning effect. Main sample n=839 (honors/non-survey excluded). Pre-reg: aspredicted.org/4DL_Q3J.

    Bastani et al. Generative AI without guardrails can harm learning: Evidence from high school mathematics. https://hamsabastani.github.io/education_llm.pdf. As of PNAS 122(26) e2422633122, published online 2025-06-25; DOI 10.1073/pnas.2422633122; experiment Fall 2023–2024. Checked 2026-08-30. Type: Field RCT (PNAS; author PDF opened).
  2. E1

    Bastani §4.1: on 57 practice problems (10 “What is the answer?” queries each), GPT Base is correct only 51% of the time on average (42% logical, 8% arithmetic errors). Authors’ mechanism: students often use GPT Base as a “crutch” (copy answers); logical-error spillover onto the exam not significant — not a universal hallucination warrant.

    Bastani et al. Generative AI without guardrails can harm learning: Evidence from high school mathematics. https://hamsabastani.github.io/education_llm.pdf. As of PNAS 122(26) e2422633122, published online 2025-06-25; DOI 10.1073/pnas.2422633122; experiment Fall 2023–2024. Checked 2026-08-30. Type: Field RCT (PNAS; author PDF opened).

PISA 2022: OECD means and 2018→2022 trend (measurement, not an AI RCT)

  1. E1

    OECD PISA 2022 Results Volume I (DOI 10.1787/53f23881-en; PDF opened 2026-09-14): data from 81 countries/economies; nearly 700,000 students. OECD mean scores 2022: mathematics 472, reading 476, science 485 (Executive Summary). E1 for official OECD measurement — not for AI causal claims.

    OECD PISA 2022 Vol. I. PISA 2022 Results (Volume I): The State of Learning and Equity in Education. https://doi.org/10.1787/53f23881-en. As of OECD Publishing, Paris, 2023; PDF research/pdfs/OECD_PISA2022_Vol1.pdf. Checked 2026-09-14. Type: International assessment report (OECD; PDF opened).
  2. E1

    PISA 2022 trends (Executive Summary): OECD mean mathematics down almost 15 points vs 2018, reading about 10 points; science remained stable. SG foreword: maths decline three times greater than any previous consecutive change; about one in four 15-year-olds a low performer in maths, reading and science on average across OECD. About 69% at least basically proficient in maths; about three out of four basic proficiency in reading and science.

    OECD PISA 2022 Vol. I. PISA 2022 Results (Volume I): The State of Learning and Equity in Education. https://doi.org/10.1787/53f23881-en. As of OECD Publishing, Paris, 2023; PDF research/pdfs/OECD_PISA2022_Vol1.pdf. Checked 2026-09-14. Type: International assessment report (OECD; PDF opened).

Retraining: meta of 56 US RCTs (Roodman & Massenkoff)

  1. E1

    Roodman & Massenkoff (arXiv:2609.07011; Anthropic Economic Index; PDF opened 2026-09-19): evidence review including a meta-analysis of 56 randomized US job-training trials since 1973 (146 impact estimates) — per authors the most exhaustive meta in this niche.

    Roodman & Massenkoff. An evidence review of worker retraining. https://arxiv.org/abs/2609.07011. As of arXiv:2609.07011 (published 2026-08-12; Anthropic Economic Index); PDF research/pdfs/Roodman_Massenkoff_arxiv2609.07011.pdf. Checked 2026-09-19. Type: Evidence review / meta-analysis (arXiv; PDF opened).
  2. E1

    Average impacts positive but modest: employment +1.7 percentage points in years 3–5 (control baseline ~63%); annual pre-tax earnings ~+$800 per person offered training. Average cost ~$13,598 per treatment-group member; programs roughly break even under debatable persistence assumptions; government recoups ~¾ of cost via taxes/transfers.

    Roodman & Massenkoff. An evidence review of worker retraining. https://arxiv.org/abs/2609.07011. As of arXiv:2609.07011 (published 2026-08-12; Anthropic Economic Index); PDF research/pdfs/Roodman_Massenkoff_arxiv2609.07011.pdf. Checked 2026-09-19. Type: Evidence review / meta-analysis (arXiv; PDF opened).
  3. E1

    Large federal efforts (JTPA, Workforce Investment Act, Job Corps) small or null. Sector programs raise earnings roughly ~10× as much (~$5–10k/year) but filter out >80% of applicants and are hard to replicate; not built for years-long reskilling of highly skilled professionals.

    Roodman & Massenkoff. An evidence review of worker retraining. https://arxiv.org/abs/2609.07011. As of arXiv:2609.07011 (published 2026-08-12; Anthropic Economic Index); PDF research/pdfs/Roodman_Massenkoff_arxiv2609.07011.pdf. Checked 2026-09-19. Type: Evidence review / meta-analysis (arXiv; PDF opened).

Jagged frontier: tasks inside vs. outside (BCG, GPT-4 2023)

  1. E1

    Dell’Acqua et al., Organization Science 37(2):403–423 (March–April 2026; online 11 March 2026). Preregistered lab-in-the-field, 758 BCG consultants (~7% of the individual-contributor workforce), GPT-4 as of end of April 2023. Inside the frontier (n=385, 18 realistic product tasks): AI arms complete 12.2% more tasks (control ~82%, GPT+overview ~93%, GPT only ~91%) and are 25.1% faster on average (body: overview 22.5% / only 27.6% on the first 17 questions). Human quality, control mean 4.38: overview +1.48 / +33.9%, only +1.31 / +29.9%; text: quality ~+32% on average.

    Dell’Acqua et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. https://pubsonline.informs.org/doi/pdf/10.1287/orsc.2025.21838. As of Organization Science 37(2):403–423, March–April 2026; published online 2026-03-11; DOI 10.1287/orsc.2025.21838; GPT-4 end of April 2023; first WP 2023-09-18 / HBS 24-013. Checked 2026-08-30. Type: Paper (field RCT, journal).
  2. E1

    Outside the frontier (n=373, one strategy case with spreadsheet + interviews): control correct ~84.5%; AI arms 60% and 70.6% — body: −19 percentage points averaging the AI arms (overview −24.5 pp, only −13.9 pp). Abstract says “19% less likely”; Atlas follows the body (percentage points). Coherence/persuasiveness still rises (overview +25.1% on the coherence scale — not the same 25.1% as inside speed).

    Dell’Acqua et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. https://pubsonline.informs.org/doi/pdf/10.1287/orsc.2025.21838. As of Organization Science 37(2):403–423, March–April 2026; published online 2026-03-11; DOI 10.1287/orsc.2025.21838; GPT-4 end of April 2023; first WP 2023-09-18 / HBS 24-013. Checked 2026-08-30. Type: Paper (field RCT, journal).

Automation substitutes and complements; aggregates still invisible

  1. E1

    Autor 2015: automation substitutes routine tasks and complements non-routine. Employment-to-population in the 20th century did not collapse. US agriculture 41% (1900) → 2% (2000) — occupations vanish, work does not. ATMs 100k→400k (1995–2010), bank tellers 500k→~550k (1980–2010). Polarization 1979–2012: middle-skill 60% → 46%. Forecast “middle-skill jobs will persist” = Autor conjecture, E3.

    Autor 2015. Why Are There Still So Many Jobs? The History and Future of Workplace Automation. https://economics.mit.edu/sites/default/files/publications/why%20are%20there%20still%20jobs%202014.pdf. As of JEP 29(3), 2015. Checked 2026-08-28. Type: Review/essay (JEP).
  2. E1

    Acemoglu, Autor, Hazell, Restrepo JOLE 2022: Burning Glass 2010–Oct 2018, pre-ChatGPT. AI vacancies rise in exposed establishments; non-AI hiring there falls (Felten −13.8%). Aggregate employment/wages in exposed occupations/industries not detectable. NAICS 51/54 excluded. Vacancies ≠ employment stock.

    Acemoglu et al. Artificial Intelligence and Jobs: Evidence from Online Vacancies. https://economics.mit.edu/sites/default/files/publications/AI%20and%20Jobs%20-%20Evidence%20from%20Online%20Vacancies.pdf. As of JOLE 40(S1), 2022. Checked 2026-08-28. Type: Paper.
  3. E1

    OECD Employment Outlook 2023 Ch. 3: “little evidence of significant negative employment effects due to AI.” High-skill occupations most AI-exposed and with relative employment gains. Occupations at highest general automation risk = 27% of employment (≠ AI displacement). Evidence without generative AI; adoption low.

    OECD EO 2023. OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market. https://www.oecd.org/content/dam/oecd/en/publications/reports/2023/07/oecd-employment-outlook-2023_904bcef3/08785bba-en.pdf. As of 2023-07-11. Checked 2026-08-28. Type: International report.

Adoption 2026 real, not universal; the macro turning point is demography

  1. E1

    US Census BTOS, 14 Dec 2025 – 3 May 2026: firm AI use 17–20% nationally; expected next 6 months 20–23%. ≥250 employees: 37%; 100–249: 32%; <4 employees <20%. Information 39.7%, finance 33.9% vs. national 19.8% (3 May 2026); retail ~14%. Firm-count, not employment-weighted (CES WP not opened).

    Census BTOS. Large Firms With at Least 20 Employees Biggest AI Users. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html. As of 2026-05-26 (BTOS 14 Dec 2025 – 3 May 2026). Checked 2026-08-28. Type: Official statistics story.
  2. E1

    OECD Employment Outlook 2025: working-age population (20–64) in the OECD aggregate stops growing in 2025 and starts to fall. Old-age dependency 2060 52%, almost 3× 1980. Title/focus: demographic crunch / labour shortage — not AI mass unemployment. E1 for demography as an OECD statement; E3 as “AI has no 2025 employment effects” — that is not what it says.

    OECD EO 2025. OECD Employment Outlook 2025: Can We Get Through the Demographic Crunch?. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/07/oecd-employment-outlook-2025_5345f034/194a947b-en.pdf. As of 2025. Checked 2026-08-28. Type: International report.

Medly micro-RCT: GCSE science AI tutoring (Harrison)

  1. E1

    Harrison, Khowaja, Dobson, Uwimpuhwe, Higgins (What Worked Education / Durham; arXiv:2609.14789v1, PDF opened 2026-09-23): four-week multisite individual randomisation of Medly (AI tutor for constructed written responses) vs business-as-usual self-directed revision in GCSE Biology/Chemistry/Physics (England); Year 9–10.

    Harrison et al. Medly. Evaluating AI Tutoring at the Speed of Innovation: Practitioner-Led Micro-Randomised Trials of an AI Tutoring Platform in GCSE Science. https://arxiv.org/abs/2609.14789. As of arXiv:2609.14789v1 (2026-09-13); PDF research/pdfs/AI_Tutoring_GCSE_arXiv2609.14789.pdf. Checked 2026-09-23. Type: Paper (arXiv; PDF opened; micro-RCT).
  2. E1

    Primary ITT: 929 pre-test / 644 post-test (attrition 30.7%); adjusted post-test 13.95 vs 11.39; difference coefficient +2.56 marks (95% CI 1.39–3.73) → Hedges’ g = 0.33 (95% CI 0.18–0.48). By subject: Physics g=0.31; Chemistry g=0.32; Biology g=0.52 (interaction not evident).

    Harrison et al. Medly. Evaluating AI Tutoring at the Speed of Innovation: Practitioner-Led Micro-Randomised Trials of an AI Tutoring Platform in GCSE Science. https://arxiv.org/abs/2609.14789. As of arXiv:2609.14789v1 (2026-09-13); PDF research/pdfs/AI_Tutoring_GCSE_arXiv2609.14789.pdf. Checked 2026-09-23. Type: Paper (arXiv; PDF opened; micro-RCT).
  3. E1

    No clear heterogeneity by Pupil Premium status (interaction +0.57 marks, CI overlaps 0). Engagement correlates exploratorily (~+0.18 marks/question) but is post-randomisation — not a causal complier effect. Authors: preliminary given attrition, curriculum-aligned outcomes, low process-eval return (6/39).

    Harrison et al. Medly. Evaluating AI Tutoring at the Speed of Innovation: Practitioner-Led Micro-Randomised Trials of an AI Tutoring Platform in GCSE Science. https://arxiv.org/abs/2609.14789. As of arXiv:2609.14789v1 (2026-09-13); PDF research/pdfs/AI_Tutoring_GCSE_arXiv2609.14789.pdf. Checked 2026-09-23. Type: Paper (arXiv; PDF opened; micro-RCT).

Claimed · E3

4. What is claimed, not shown

Generative AI is spreading more evenly across education levels than the internet once did (reading of the research question; the author himself words it cautiously)

  1. E3

    Watanabe: the starting point is the observation that new technologies are usually adopted first by people with more formal education; for the matched countries the results suggest that this pattern is weaker for generative AI in 2025 than for the internet at a comparable level of use. The abstract itself limits the claim: these are unadjusted education contrasts in a non-random set of countries, not an education effect, and the results do not generalise to the EU as a whole. Whether the smaller gap is due to the technology, the timing, the expansion of education since the internet years or the kind of use is not shown by this.

    Watanabe GenAI vs internet education gap. Is the education gap in generative AI use smaller than the internet's was at a similar level of use? Evidence from 16 European countries. https://arxiv.org/abs/2610.36181. As of arXiv:2610.36181v1 (submitted 2026-10-26; cs.IR, cs.CY; per comment 16 pp., 2 figures, code, data and outputs as ancillary files); abstract via export.arxiv.org API only, PDF and ancillary files not opened; affiliation not stated in the abstract record. Checked 2026-10-28. Type: Paper (arXiv; abstract only; secondary analysis of public Eurostat ICT-usage survey data, self-report).

Delegation rather than exposure as the measure for the agent era (author reading)

  1. E3

    Espinal Maya & Jiménez Londoño: exposure was the right construct for chatbots; for agents what counts is whether a task can be handed over, and the index supplies levels and barriers for the whole world. To forecast employment effects it must be combined with adoption, prices and institutions — on its own it supports none. A conceptual reading; their own finding of r = 0.93 with exposure shows the ranking of occupations barely changes.

    Espinal Maya & Jiménez Londoño Agent Delegability Index. The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries. https://arxiv.org/abs/2610.30752. As of arXiv:2610.30752v1 (submitted 2026-10-22; econ.GN; 32 pp.); PDF research/pdfs/arXiv2610.30752.pdf; index, prompts and MCP server github.com/Cespial/delegation-divide not opened. Checked 2026-10-23. Type: Paper (arXiv; PDF opened; preregistered model-rated index, 146 countries).

Better expected leads lower persistence with the current one — AI routing is a behavioural intervention (author reading)

  1. E3

    Fu et al.: by the logic of sequential search a better pool of future leads raises the “reservation value”, so workers give up on the current lead sooner — most strongly early in the day, when many chances still lie ahead. Deploying AI in parallel with humans is therefore not behaviourally neutral; in the calibrated model a smoother rhythm of releasing hot leads would raise conversions per call by ~13.5% (95% bootstrap 10.5–16.6%). Mechanism inferred, not measured directly; model calculation, not tested.

    Fu et al. parallel AI voice agent. When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition. https://arxiv.org/abs/2610.25603. As of arXiv:2610.25603v1 (submitted 2026-10-19; econ.GN; working paper, version September 2026); PDF research/pdfs/arXiv2610.25603.pdf. Checked 2026-10-20. Type: Working paper (arXiv; PDF opened; field data from one firm).

Measurement validity ≠ optimisation validity (author reading)

  1. E3

    Author reading: any metric that scores decisions one by one and averages them is maximised by repeating the single best decision; benchmarks that assess or train AI tutors should score trajectories, not isolated moves. Weight analysis: the correction sits in the output projection, the model memorised training strings. A design principle, not a measured learning effect in students.

    Dominguez Figaredo & Fernández De la Cruz PAI. The learner who does not learn: when optimizing a pedagogical metric degrades LLM tutoring. https://arxiv.org/abs/2610.12125. As of arXiv:2610.12125v1 (submitted 2026-10-08); PDF research/pdfs/arXiv2610.12125.pdf. Checked 2026-10-16. Type: Paper (arXiv; PDF opened; preprint).

Exposure scores and scenarios, not displacement 2026

  1. E2

    Frey & Osborne 2013: 47% of US employment (BLS 2010) in high-risk (p>0.7). Paper’s own qualification: “potentially automatable over some unspecified number of years, perhaps a decade or two.” E2 as a 2013 occupation score; E4 as observed displacement 2013–2026 (contradicts OECD 2023, Acemoglu 2022, Census adoption).

    Frey & Osborne. The Future of Employment: How Susceptible are Jobs to Computerisation?. https://oms-www.files.svdcdn.com/production/downloads/academic/The_Future_of_Employment.pdf. As of Oxford Martin WP 2013-09-17. Checked 2026-08-28. Type: Working paper.
  2. E2

    Eloundou et al. 2023: ~80% of the US workforce with ≥10% of tasks affected (β); ~19% with >50%. LLM alone ~15% of tasks “significantly faster”; with software 47–56%. “We do not make predictions about the development or adoption timeline.” Exposure without an augmentation/displacement distinction. OpenAI authors; rubric subjective.

    Eloundou et al. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. https://arxiv.org/html/2303.10130. As of arXiv:2303.10130 (rev. Aug 2023). Checked 2026-08-28. Type: Paper (OpenAI).
  3. E2

    IAB research report 23/2025: QuBe scenario, assumptions. GDP +0.8 pp/year, cumulative 4.5 trillion €. Employment after 15 years similar to the reference. Turnover ~1.6 million. IT +110,000 / business services −120,000. Without new business models “clearly more negative.” Explicitly: not for creating accurate future answers. E3 as a world forecast.

    IAB FB 23/2025. Künstliche Intelligenz: Potenzielle Effekte für den deutschen Arbeitsmarkt. https://doku.iab.de/forschungsbericht/2025/fb2325.pdf. As of IAB-Forschungsbericht 23/2025. Checked 2026-08-28. Type: Scenario report.
  4. E2

    OECD EO 2023 survey (finance+manufacturing, 7 countries, 2022): 63% of AI-using workers report more enjoyment; 60% worried (infographic). Two sectors, not all occupations.

    OECD EO 2023. OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market. https://www.oecd.org/content/dam/oecd/en/publications/reports/2023/07/oecd-employment-outlook-2023_904bcef3/08785bba-en.pdf. As of 2023-07-11. Checked 2026-08-28. Type: International report.
  5. E2

    Roodman/Massenkoff: retraining is the most popular policy among experts and the public against AI labour-market disruption — authors doubt current programs can meet a crash moment (extrapolation into an uncertain AI world).

    Roodman & Massenkoff. An evidence review of worker retraining. https://arxiv.org/abs/2609.07011. As of arXiv:2609.07011 (published 2026-08-12; Anthropic Economic Index); PDF research/pdfs/Roodman_Massenkoff_arxiv2609.07011.pdf. Checked 2026-09-19. Type: Evidence review / meta-analysis (arXiv; PDF opened).

Constrained · Limit

5. Bottleneck and limit

Education gap GenAI vs internet: self-report, 16 of 35 countries, ratio result fragile, size depends on matching, abstract only

  1. E2

    Watanabe 2026: the basis is a survey (Eurostat ICT survey, self-reported “use in the last three months”), not measured use, intensity, purpose, or any learning or work outcome. Only 16 of 35 countries could be matched; most of the others were early internet adopters, so the set is non-random and per the abstract does not generalise to the EU as a whole. The log-ratio result is fragile: it fails the decision rule when countries matched at their first internet observation are removed (9 of 11, p = 0.065). The percentage-point result keeps its sign but shrinks from −27.6 to about −10 to −11 points when matching takes education composition into account (7 and 8 countries). Unadjusted contrasts, not a causal education effect. The analysis follows an unpublished plan whose timing cannot be verified and is explicitly not presented as pre-registered. Preprint without peer review, single author; only the abstract via the arXiv API was read (no PDF, figures, country list, country-level gap sizes or affiliation in the record); code, data and outputs (ancillary files) not checked. Do not net against the delegability index (source 22, model-rated ceiling) or Census firm use (source 9).

    Watanabe GenAI vs internet education gap. Is the education gap in generative AI use smaller than the internet's was at a similar level of use? Evidence from 16 European countries. https://arxiv.org/abs/2610.36181. As of arXiv:2610.36181v1 (submitted 2026-10-26; cs.IR, cs.CY; per comment 16 pp., 2 figures, code, data and outputs as ancillary files); abstract via export.arxiv.org API only, PDF and ancillary files not opened; affiliation not stated in the abstract record. Checked 2026-10-28. Type: Paper (arXiv; abstract only; secondary analysis of public Eurostat ICT-usage survey data, self-report).

Delegability index: models from one developer rate, US task mix for every country, ceiling ≠ adoption

  1. E2

    Espinal Maya & Jiménez Londoño 2026: technical delegability with the technology of October 2026, not what firms do, regulators allow or workers accept — per the authors a ceiling; O*NET describes US work, the task mix is carried unchanged to every country (agriculture only 193 task statements; with ESCO the income slope drops from 7.23 to 4.66); both raters and the usage data come from the same developer (Anthropic), circularity possible; the human check is weaker than registered — 300 instead of 1,500 tasks, led by a coauthor with students, no human–human agreement measurable (abstract gives κ = 0.85, introduction α = 0.88 — two measures of the same check); no temperature control; an “agent” framing in the prompt raised the mean level in the pilot from 1.33 to 1.64. Preprint; data repository not opened. Not to be equated with field measurements (call centre, source 21) or exposure (Eloundou, source 5).

    Espinal Maya & Jiménez Londoño Agent Delegability Index. The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries. https://arxiv.org/abs/2610.30752. As of arXiv:2610.30752v1 (submitted 2026-10-22; econ.GN; 32 pp.); PDF research/pdfs/arXiv2610.30752.pdf; index, prompts and MCP server github.com/Cespial/delegation-divide not opened. Checked 2026-10-23. Type: Paper (arXiv; PDF opened; preregistered model-rated index, 146 countries).

AI voice agent in a call centre: one firm, one product, 2024 — per-human metric ≠ performance

  1. E2

    Fu et al. 2026: one anonymised insurance broker in China, one routing change, ±60 days in 2024; call duration is only a proxy for effort; the mechanism is inferred from the within-day pattern; day-level average effects are not significant; no wage, employment or customer-satisfaction data; one co-author is from an industry research institute; working paper without peer review. Not transferable to call centres in general or to generative-AI assistance (Brynjolfsson, source 3) — here the AI does not assist, it works in parallel.

    Fu et al. parallel AI voice agent. When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition. https://arxiv.org/abs/2610.25603. As of arXiv:2610.25603v1 (submitted 2026-10-19; econ.GN; working paper, version September 2026); PDF research/pdfs/arXiv2610.25603.pdf. Checked 2026-10-20. Type: Working paper (arXiv; PDF opened; field data from one firm).

Tutor-metric study: synthetic cases, in-sample rating, no learners

  1. E2

    Human rating in-sample only (the 30 cases were in the corrective training set); nobody rated the unmodified proxy; corrected texts were also shorter and generic (11.9 vs 16.7 words), so the rating drop cannot be cleanly split between monotony and text; raters were graduate students, not in-service teachers; case narratives synthetic; no real learners, no learning outcome measured. No warrant against AI tutoring in general — a warrant against training on such per-move metrics.

    Dominguez Figaredo & Fernández De la Cruz PAI. The learner who does not learn: when optimizing a pedagogical metric degrades LLM tutoring. https://arxiv.org/abs/2610.12125. As of arXiv:2610.12125v1 (submitted 2026-10-08); PDF research/pdfs/arXiv2610.12125.pdf. Checked 2026-10-16. Type: Paper (arXiv; PDF opened; preprint).

O*NET-BENCH: worker ratings are the target, not the truth

  1. E1

    Corpus: English text outputs for US occupational tasks (screened for text suitability), self-reported experience, not employment-weighted; one rating per response — whether worker, judge or both are wrong cannot be determined. Test-split analyses exploratory; only the pointwise→listwise reversal in one fine-tuned lineage is prespecified. No warrant for displacement, job-exposure figures or other modalities — but a warning sign for any AI-and-work measurement that turns judge scores into acceptance rates.

    Lyu & Thompson O*NET-BENCH. Right Order, Wrong Scale: Auditing LLM Judges for Occupational AI Measurement. https://arxiv.org/abs/2610.02492. As of arXiv:2610.02492v1 (2026-10-01); PDF research/pdfs/LyuThompson_ONETBench_arXiv2610.02492.pdf. Checked 2026-10-05. Type: Paper (arXiv; PDF opened; preprint under review).

EngiWorld: 300-subset ≠ full 1,301 / industry VM

  1. E1

    Eval on 300-subset, not full 1,301; Windows/Ubuntu VMs with specific industrial licenses/env; zero-shot — no warrant for live CAD/CAE production or multi-tool office workflows outside the bench.

    Gao et al. EngiWorld. EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments?. https://arxiv.org/abs/2609.37686. As of arXiv:2609.37686v1 (2026-09-29); PDF research/pdfs/EngiWorld_arXiv2609.37686.pdf. Checked 2026-09-30. Type: Paper (arXiv; PDF opened).

What has not shifted

  1. E1

    Medly g=0.33 / curriculum-aligned GCSE post-test ≠ standardised national exam; 30.7% attrition and active digital comparator (not “no support”) limit external validity; micro-RCT ≠ definitive education effectiveness; Bastani Base −17% remains a separate harm warrant.

    Harrison et al. Medly. Evaluating AI Tutoring at the Speed of Innovation: Practitioner-Led Micro-Randomised Trials of an AI Tutoring Platform in GCSE Science. https://arxiv.org/abs/2609.14789. As of arXiv:2609.14789v1 (2026-09-13); PDF research/pdfs/AI_Tutoring_GCSE_arXiv2609.14789.pdf. Checked 2026-09-23. Type: Paper (arXiv; PDF opened; micro-RCT).
  2. E1

    Pattern through BTOS 2026: writing/seeing/planning speeds executable tasks in RCT/field; education shows the Bastani pattern (practice ↑, unaided exam Base ↓); Dell’Acqua jagged frontier. Employment statistics and firm adoption remain far behind Frey-47% and Eloundou-80%.

    What has moved, what has not. Opened measurements only.
    MeasurementMovedNot moved
    Noy/Zhang WPTime, grade, draft→edit (444 / −37% / 68%)Employment, wages
    Noy/Zhang ScienceTime −40% / quality +18%; follow-up job useEmployment, wages
    Peng CopilotJS-task timeCode quality, headcount
    BrynjolfssonRPH, novice onboardingFirm headcount, wages
    Bastani PNAS 2025Exam −17% Base; Tutor ≈ controlUniversal education, other subjects, long run
    Dell’Acqua OrgSci 2026Inside: completion +12.2%, speed ~25%, quality ~+32%; outside: correctness −19 pp, coherence +Employment, education, 2026 models, other firms
    Acemoglu 2022AI vacancies, non-AI hiring −Occup./industry employment
    OECD 2023Exposure scores, surveyAggregate employment levels
    Census BTOS 2026Firm adoption 17–20%“Every other job”
    OECD 2025Working-age population turnsAI mass unemployment
    OECD PISA 2022 Vol. IOECD means 472/476/485; trend −15/−10/stableGen-AI productivity, AI harms/helps education
    Roodman/Massenkoff meta+1.7 pp employment; ~+$800; sector ~10×AI-retraining guarantee; white-collar crash response
    Noy & Zhang. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf. As of MIT WP 2023-03-02. Checked 2026-08-28. Type: Working paper.Noy & Zhang Science. Experimental evidence on the productivity effects of generative artificial intelligence. https://doi.org/10.1126/science.adh2586. As of Science 381:187–192 (2023-07-13); author PDF research/pdfs/Noy_Zhang_Science_adh2586.pdf. Checked 2026-09-11. Type: Paper (Science; author PDF opened).Peng et al. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. https://arxiv.org/html/2302.06590. As of arXiv:2302.06590, experiment May–Jun 2022. Checked 2026-08-28. Type: Paper.Brynjolfsson et al. Generative AI at Work. https://www.nber.org/system/files/working_papers/w31161/w31161.pdf. As of NBER WP 31161, revised Nov 2023. Checked 2026-08-28. Type: Working paper.Bastani et al. Generative AI without guardrails can harm learning: Evidence from high school mathematics. https://hamsabastani.github.io/education_llm.pdf. As of PNAS 122(26) e2422633122, published online 2025-06-25; DOI 10.1073/pnas.2422633122; experiment Fall 2023–2024. Checked 2026-08-30. Type: Field RCT (PNAS; author PDF opened).Dell’Acqua et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. https://pubsonline.informs.org/doi/pdf/10.1287/orsc.2025.21838. As of Organization Science 37(2):403–423, March–April 2026; published online 2026-03-11; DOI 10.1287/orsc.2025.21838; GPT-4 end of April 2023; first WP 2023-09-18 / HBS 24-013. Checked 2026-08-30. Type: Paper (field RCT, journal).Acemoglu et al. Artificial Intelligence and Jobs: Evidence from Online Vacancies. https://economics.mit.edu/sites/default/files/publications/AI%20and%20Jobs%20-%20Evidence%20from%20Online%20Vacancies.pdf. As of JOLE 40(S1), 2022. Checked 2026-08-28. Type: Paper.OECD EO 2023. OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market. https://www.oecd.org/content/dam/oecd/en/publications/reports/2023/07/oecd-employment-outlook-2023_904bcef3/08785bba-en.pdf. As of 2023-07-11. Checked 2026-08-28. Type: International report.Census BTOS. Large Firms With at Least 20 Employees Biggest AI Users. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html. As of 2026-05-26 (BTOS 14 Dec 2025 – 3 May 2026). Checked 2026-08-28. Type: Official statistics story.OECD EO 2025. OECD Employment Outlook 2025: Can We Get Through the Demographic Crunch?. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/07/oecd-employment-outlook-2025_5345f034/194a947b-en.pdf. As of 2025. Checked 2026-08-28. Type: International report.
  3. E2

    Eloundou: human α only 14% of occupation tasks directly; ζ 46% needs software co-invention. Only 3% of US workers have >50% of tasks exposed without extra software. Frey: 2013+13 years = 2026, and OECD/Census/Acemoglu show no 47% event.

    Eloundou et al. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. https://arxiv.org/html/2303.10130. As of arXiv:2303.10130 (rev. Aug 2023). Checked 2026-08-28. Type: Paper (OpenAI).Frey & Osborne. The Future of Employment: How Susceptible are Jobs to Computerisation?. https://oms-www.files.svdcdn.com/production/downloads/academic/The_Future_of_Employment.pdf. As of Oxford Martin WP 2013-09-17. Checked 2026-08-28. Type: Working paper.Autor 2015. Why Are There Still So Many Jobs? The History and Future of Workplace Automation. https://economics.mit.edu/sites/default/files/publications/why%20are%20there%20still%20jobs%202014.pdf. As of JEP 29(3), 2015. Checked 2026-08-28. Type: Review/essay (JEP).
  4. E1

    Productivity gains compress task inequality: Noy low-baseline benefits more; 68% unedited can cut labour demand (authors, E3). Brynjolfsson: novices +34%, experts 0 — reversal of skill-biased IT (Autor 2015).

    Noy & Zhang. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf. As of MIT WP 2023-03-02. Checked 2026-08-28. Type: Working paper.Brynjolfsson et al. Generative AI at Work. https://www.nber.org/system/files/working_papers/w31161/w31161.pdf. As of NBER WP 31161, revised Nov 2023. Checked 2026-08-28. Type: Working paper.
  5. E1

    Limits (authors §5): one school, Turkey, mathematics only, GPT-4 (Fall 2023), short-term same-session exams — not employment, not long-run learning, not other subjects, not 2026 SOTA tutors. No warrant for universal “AI harms education” or “AI helps education.” Productivity RCTs (Noy/Peng) measure output with the tool, not unaided learning. PISA 2022 Vol. I now source 15 (E1 measurement) — PISA ≠ gen-AI RCT; UNESCO still unopened.

    Bastani et al. Generative AI without guardrails can harm learning: Evidence from high school mathematics. https://hamsabastani.github.io/education_llm.pdf. As of PNAS 122(26) e2422633122, published online 2025-06-25; DOI 10.1073/pnas.2422633122; experiment Fall 2023–2024. Checked 2026-08-30. Type: Field RCT (PNAS; author PDF opened).
  6. E1

    Dell’Acqua does not measure jobs, wages, headcount, occupations, or learning. The finding is task quality/speed/correctness in a BCG lab-in-the-field with GPT-4 (April 2023). One firm; early-career consultants; one outside task; subjective grades. The frontier is, per the paper, not fixed; 2026 models are unmeasured. “Writing/planning always better” is contradicted by the outside arm in this setting.

    Dell’Acqua et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. https://pubsonline.informs.org/doi/pdf/10.1287/orsc.2025.21838. As of Organization Science 37(2):403–423, March–April 2026; published online 2026-03-11; DOI 10.1287/orsc.2025.21838; GPT-4 end of April 2023; first WP 2023-09-18 / HBS 24-013. Checked 2026-08-30. Type: Paper (field RCT, journal).
  7. E0

    Not in the dossier: BLS EP 2024–34 (bot block); Cui Copilot field; QJE-15% (Brynjolfsson); ifo 27% DE (not opened). Noy Science is source 14; PISA Vol. I source 15; Roodman/Massenkoff source 16.

  8. E1

    PISA 2022 measures 15-year-olds’ literacy in reading/maths/science — not a generative-AI productivity or learning RCT. No warrant for “AI harms education” or “AI helps education.” Bastani remains the opened education RCT.

    OECD PISA 2022 Vol. I. PISA 2022 Results (Volume I): The State of Learning and Equity in Education. https://doi.org/10.1787/53f23881-en. As of OECD Publishing, Paris, 2023; PDF research/pdfs/OECD_PISA2022_Vol1.pdf. Checked 2026-09-14. Type: International assessment report (OECD; PDF opened).
  9. E1

    Historical US job-training meta (56 RCTs) ≠ guarantee that AI-era retraining works. Sector programs ~10× with >80% filtering and replication problems — no warrant for a scalable AI crash response. Study populations (often low-income/youth) ≠ typical white-collar AI exposure.

    Roodman & Massenkoff. An evidence review of worker retraining. https://arxiv.org/abs/2609.07011. As of arXiv:2609.07011 (published 2026-08-12; Anthropic Economic Index); PDF research/pdfs/Roodman_Massenkoff_arxiv2609.07011.pdf. Checked 2026-09-19. Type: Evidence review / meta-analysis (arXiv; PDF opened).

6. Actors and incentives

Who measures tasks, who forecasts, who measures adoption

  1. E1

    Noy/Zhang (MIT WP + Science 381), Peng (Microsoft/GitHub/MIT), Brynjolfsson/Li/Raymond: RCT and one firm set. Human-in-the-loop by design.

    Noy & Zhang. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf. As of MIT WP 2023-03-02. Checked 2026-08-28. Type: Working paper.Noy & Zhang Science. Experimental evidence on the productivity effects of generative artificial intelligence. https://doi.org/10.1126/science.adh2586. As of Science 381:187–192 (2023-07-13); author PDF research/pdfs/Noy_Zhang_Science_adh2586.pdf. Checked 2026-09-11. Type: Paper (Science; author PDF opened).Peng et al. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. https://arxiv.org/html/2302.06590. As of arXiv:2302.06590, experiment May–Jun 2022. Checked 2026-08-28. Type: Paper.Brynjolfsson et al. Generative AI at Work. https://www.nber.org/system/files/working_papers/w31161/w31161.pdf. As of NBER WP 31161, revised Nov 2023. Checked 2026-08-28. Type: Working paper.
  2. E1

    Bastani et al.: education field RCT (high-school mathematics, GPT-4). Dell’Acqua/BCG/HBS: task RCT (jagged frontier, GPT-4 2023).

    Bastani et al. Generative AI without guardrails can harm learning: Evidence from high school mathematics. https://hamsabastani.github.io/education_llm.pdf. As of PNAS 122(26) e2422633122, published online 2025-06-25; DOI 10.1073/pnas.2422633122; experiment Fall 2023–2024. Checked 2026-08-30. Type: Field RCT (PNAS; author PDF opened).Dell’Acqua et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. https://pubsonline.informs.org/doi/pdf/10.1287/orsc.2025.21838. As of Organization Science 37(2):403–423, March–April 2026; published online 2026-03-11; DOI 10.1287/orsc.2025.21838; GPT-4 end of April 2023; first WP 2023-09-18 / HBS 24-013. Checked 2026-08-30. Type: Paper (field RCT, journal).
  3. E1

    OECD and US Census measure adoption and aggregates; the 2025 headline is labour shortage, not a jobless future. OECD PISA 2022 Vol. I (source 15) measures student performance, not AI.

    OECD EO 2023. OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market. https://www.oecd.org/content/dam/oecd/en/publications/reports/2023/07/oecd-employment-outlook-2023_904bcef3/08785bba-en.pdf. As of 2023-07-11. Checked 2026-08-28. Type: International report.OECD EO 2025. OECD Employment Outlook 2025: Can We Get Through the Demographic Crunch?. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/07/oecd-employment-outlook-2025_5345f034/194a947b-en.pdf. As of 2025. Checked 2026-08-28. Type: International report.Census BTOS. Large Firms With at Least 20 Employees Biggest AI Users. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html. As of 2026-05-26 (BTOS 14 Dec 2025 – 3 May 2026). Checked 2026-08-28. Type: Official statistics story.
  4. E2

    Eloundou (OpenAI), Frey/Osborne, IAB/BIBB/GWS: exposure and scenarios, not 2026 employment actuals.

    Eloundou et al. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. https://arxiv.org/html/2303.10130. As of arXiv:2303.10130 (rev. Aug 2023). Checked 2026-08-28. Type: Paper (OpenAI).Frey & Osborne. The Future of Employment: How Susceptible are Jobs to Computerisation?. https://oms-www.files.svdcdn.com/production/downloads/academic/The_Future_of_Employment.pdf. As of Oxford Martin WP 2013-09-17. Checked 2026-08-28. Type: Working paper.IAB FB 23/2025. Künstliche Intelligenz: Potenzielle Effekte für den deutschen Arbeitsmarkt. https://doku.iab.de/forschungsbericht/2025/fb2325.pdf. As of IAB-Forschungsbericht 23/2025. Checked 2026-08-28. Type: Scenario report.

7. State of the dispute

47% at risk vs. no aggregate effects vs. 17–20% adoption

  1. E4

    Frey 47% “at risk” (2013, technical feasibility) vs. OECD 2023 “no signs of slowing labour demand (yet)” vs. Census 2026 adoption 17–20% vs. Acemoglu establishment hiring − under exposure, aggregate 0. Autor 2015 cites exactly this Frey passage as overstated ML optimism.

    Frey & Osborne. The Future of Employment: How Susceptible are Jobs to Computerisation?. https://oms-www.files.svdcdn.com/production/downloads/academic/The_Future_of_Employment.pdf. As of Oxford Martin WP 2013-09-17. Checked 2026-08-28. Type: Working paper.OECD EO 2023. OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market. https://www.oecd.org/content/dam/oecd/en/publications/reports/2023/07/oecd-employment-outlook-2023_904bcef3/08785bba-en.pdf. As of 2023-07-11. Checked 2026-08-28. Type: International report.Acemoglu et al. Artificial Intelligence and Jobs: Evidence from Online Vacancies. https://economics.mit.edu/sites/default/files/publications/AI%20and%20Jobs%20-%20Evidence%20from%20Online%20Vacancies.pdf. As of JOLE 40(S1), 2022. Checked 2026-08-28. Type: Paper.Census BTOS. Large Firms With at Least 20 Employees Biggest AI Users. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html. As of 2026-05-26 (BTOS 14 Dec 2025 – 3 May 2026). Checked 2026-08-28. Type: Official statistics story.Autor 2015. Why Are There Still So Many Jobs? The History and Future of Workplace Automation. https://economics.mit.edu/sites/default/files/publications/why%20are%20there%20still%20jobs%202014.pdf. As of JEP 29(3), 2015. Checked 2026-08-28. Type: Review/essay (JEP).

8. Open questions

  1. Brynjolfsson QJE 2025 full text.
  2. BLS Employment Projections 2024–34 (bot block).
  3. Census CES-WP-26-25 employment-weighted adoption.
  4. Cui et al. Copilot field RCTs.
  5. IAB short report on substitutability 2024 + DiWaBe 2.0.
  6. PISA 2022 Vol. I opened 2026-09-14 (source 15, E1). UNESCO / further OECD Education volumes still unopened.

9. Changes

  • v2.1-draft2026-10-28: Watanabe education gap GenAI vs internet arXiv:2610.36181 (source 23, E2, abstract only) — Eurostat ICT survey ages 16–74, matching within 5 pp, 16 of 35 countries; ISCED 5–8 vs 0–2; percentage-point gap smaller in 15 of 16 (median −27.6, interval −32.2 to −16.8, p = 0.0005), log ratio in 13 of 16 (median −0.625, 95% bootstrap −0.969 to −0.291, p = 0.021); with education composition about −10 to −11 points; ratio fragile (9 of 11, p = 0.065); not pre-registered, unadjusted, not an education effect; preprint, abstract warrant.
  • v2.0-draft2026-10-23: Espinal Maya & Jiménez Londoño Agent Delegability Index arXiv:2610.30752 (source 22, E2, PDF) — Claude Sonnet/Haiku 5.5 rate 18,796 O*NET tasks, 146 countries; world 35.2% partial / 2.5% full (Haiku alone 51.4%); 48.7% high vs 22.5% low income, slope 7.23 (ESCO 4.66); peak at ~11.8 min; r = 0.93 with Eloundou; adoption gap not supported; human κ = 0.85 on 300 tasks.
  • v1.9-draft2026-10-20: Fu et al. parallel AI voice agent arXiv:2610.25603 (source 21, E2, PDF) — RDiT ±60 days, 409,622 + 157,468 calls; first three calls: hot leads +23.4 pp, conversion −4.5 pp, hand-back +8.1 pp, duration ~−13%; hot leads 4.4% → 3.7%; per 50 human calls +6.3% only via a smaller denominator; decomposition +0.58% / −17.58%; one firm, working paper.
  • v1.7-draft2026-10-05: O*NET-BENCH Lyu & Thompson arXiv:2610.02492 (source 19, E1, PDF) — 33 judges / 45,796 worker ratings / 4,501 test; acceptance 3.0–97.9% vs workers 61.1%; 25 configs pair ≥0.60, TF-IDF 0.643; calibrated ≤8.5% variance; target ≠ truth.
  • v1.6-draft2026-09-30: EngiWorld Gao et al. arXiv:2609.37686 (source 18, E1, PDF) — 1,301 tasks / 6 domains / 26 platforms; EngiScore best 44.3 (Opus 5); multi-software 3.6% (6/168); 300≠1,301.
  • v1.5-draft2026-09-23: Medly micro-RCT Harrison et al. arXiv:2609.14789 (source 17, E1, PDF) — GCSE science ITT g=0.33 (929→644; attrition 30.7%); Physics/Chem/Bio 0.31/0.32/0.52; preliminary; narrows education-RCT gap.
  • v1.4-draft2026-09-19: Roodman & Massenkoff arXiv:2609.07011 (source 16, E1, PDF) — meta 56 US RCTs; +1.7 pp employment / ~+$800; cost ~$13.6k; sector ~10× with >80% filter; meta ≠ AI-retraining guarantee.
  • v1.3-draft2026-09-14: OECD PISA 2022 Results Vol. I (source 15, E1, PDF) — 81 systems / ~700k students; OECD means 472/476/485; 2018→2022 ≈ −15 maths / −10 reading / science stable. PISA ≠ gen-AI RCT.
  • v1.2-draft2026-09-11: Noy & Zhang Science 381 (source 14, DOI 10.1126/science.adh2586, author PDF) opened — n=453, time −40% / quality +18%, body −11 min / +0.45 SD; follow-up job use 2×/1.6×. WP (source 1) remains E1 for WP figures 444/−37%/68%.
  • v1.1-draft2026-08-30: Bastani PNAS education RCT (source 12) and Dell’Acqua OrgSci jagged frontier (source 13) opened — education no longer E0; jagged frontier established. PISA/UNESCO still unopened.
  • v1.0-draftFirst version from the 2026-08-28 verification log. Education remains a gap (no opened education RCT).

10. Sources

No. Source As of Checked Grade
1 Noy, S.; Zhang, W.. Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. https://economics.mit.edu/sites/default/files/inline-files/Noy_Zhang_1.pdf. Type: Working paper. E1
2 Peng, S.; Kalliamvakou, E.; Cihon, P.; Demirer, M.. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. https://arxiv.org/html/2302.06590. Type: Paper. E1
3 Brynjolfsson, E.; Li, D.; Raymond, L. R.. Generative AI at Work. https://www.nber.org/system/files/working_papers/w31161/w31161.pdf. Type: Working paper. E1
4 Autor, D. H.. Why Are There Still So Many Jobs? The History and Future of Workplace Automation. https://economics.mit.edu/sites/default/files/publications/why%20are%20there%20still%20jobs%202014.pdf. Type: Review/essay (JEP). E1
5 Eloundou, T.; Manning, S.; Mishkin, P.; Rock, D.. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. https://arxiv.org/html/2303.10130. Type: Paper (OpenAI). E2
6 Acemoglu, D.; Autor, D.; Hazell, J.; Restrepo, P.. Artificial Intelligence and Jobs: Evidence from Online Vacancies. https://economics.mit.edu/sites/default/files/publications/AI%20and%20Jobs%20-%20Evidence%20from%20Online%20Vacancies.pdf. Type: Paper. E1
7 OECD. OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market. https://www.oecd.org/content/dam/oecd/en/publications/reports/2023/07/oecd-employment-outlook-2023_904bcef3/08785bba-en.pdf. Type: International report. E1
8 OECD. OECD Employment Outlook 2025: Can We Get Through the Demographic Crunch?. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/07/oecd-employment-outlook-2025_5345f034/194a947b-en.pdf. Type: International report. E1
9 Grundy, A.; Breaux, C.; Khatiwoda, D.. Large Firms With at Least 20 Employees Biggest AI Users. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html. Type: Official statistics story. E1
10 Frey, C. B.; Osborne, M. A.. The Future of Employment: How Susceptible are Jobs to Computerisation?. https://oms-www.files.svdcdn.com/production/downloads/academic/The_Future_of_Employment.pdf. Type: Working paper. E2
11 Zika, G.; Hassemer, T.-M.; Hummel, M.; Krebs, B.; Maier, T.; Mönnig, A.; Schneemann, C.; Weber, E.; Zenk, J.. Künstliche Intelligenz: Potenzielle Effekte für den deutschen Arbeitsmarkt. https://doku.iab.de/forschungsbericht/2025/fb2325.pdf. Type: Scenario report. E2
12 Bastani, H.; Bastani, O.; Sungu, A.; Ge, H.; Kabakcı, Ö.; Mariman, R.. Generative AI without guardrails can harm learning: Evidence from high school mathematics. https://hamsabastani.github.io/education_llm.pdf. Type: Field RCT (PNAS; author PDF opened). E1
13 Dell’Acqua, F.; McFowland III, E.; Mollick, E.; Lifshitz, H.; Kellogg, K. C.; Rajendran, S.; Krayer, L.; Candelon, F.; Lakhani, K. R.. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. https://pubsonline.informs.org/doi/pdf/10.1287/orsc.2025.21838. Type: Paper (field RCT, journal). E1
14 Noy, S.; Zhang, W.. Experimental evidence on the productivity effects of generative artificial intelligence. https://doi.org/10.1126/science.adh2586. Type: Paper (Science; author PDF opened). E1
15 OECD. PISA 2022 Results (Volume I): The State of Learning and Equity in Education. https://doi.org/10.1787/53f23881-en. Type: International assessment report (OECD; PDF opened). E1
16 Roodman, D.; Massenkoff, M.. An evidence review of worker retraining. https://arxiv.org/abs/2609.07011. Type: Evidence review / meta-analysis (arXiv; PDF opened). E1
17 Harrison, W.; Khowaja, R.; Dobson, E.; Uwimpuhwe, G.; Higgins, S. (What Worked Education; Durham University). Evaluating AI Tutoring at the Speed of Innovation: Practitioner-Led Micro-Randomised Trials of an AI Tutoring Platform in GCSE Science. https://arxiv.org/abs/2609.14789. Type: Paper (arXiv; PDF opened; micro-RCT). E1
18 Gao, H.; Qu, H.; Lei, Y.; et al. (Tsinghua University; Zhiman Inc.; multi-institution). EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments?. https://arxiv.org/abs/2609.37686. Type: Paper (arXiv; PDF opened). E1
19 Lyu, H.; Thompson, N. (Massachusetts Institute of Technology). Right Order, Wrong Scale: Auditing LLM Judges for Occupational AI Measurement. https://arxiv.org/abs/2610.02492. Type: Paper (arXiv; PDF opened; preprint under review). E1
20 Dominguez Figaredo, D.; Fernández De la Cruz, R. (UNED, Madrid). The learner who does not learn: when optimizing a pedagogical metric degrades LLM tutoring. https://arxiv.org/abs/2610.12125. Type: Paper (arXiv; PDF opened; preprint). E2
21 Fu, X. (Georgia State University); Luo, S. (Renmin University of China); Ramasubbu, N. (University of Pittsburgh); Shen, P. (Waterdrop Research Institute); Wang, X. (Hong Kong Polytechnic University); Wang, Y. (University of Sydney). When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition. https://arxiv.org/abs/2610.25603. Type: Working paper (arXiv; PDF opened; field data from one firm). E2
22 Espinal Maya, C. (Universidad Nacional de Colombia, Facultad de Minas); Jiménez Londoño, S. (Universidad EAFIT). The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries. https://arxiv.org/abs/2610.30752. Type: Paper (arXiv; PDF opened; preregistered model-rated index, 146 countries). E2
23 Watanabe, K.. Is the education gap in generative AI use smaller than the internet's was at a similar level of use? Evidence from 16 European countries. https://arxiv.org/abs/2610.36181. Type: Paper (arXiv; abstract only; secondary analysis of public Eurostat ICT-usage survey data, self-report). E2

11. Uncertainty log

Overall uncertainty of this entry, bound to the verification log of 2026-08-28 plus openings 12+13 (2026-08-30), Noy Science source 14 (2026-09-11), PISA Vol. I source 15 (2026-09-14) and Roodman/Massenkoff source 16 (2026-09-19) plus Medly micro-RCT source 17 (2026-09-23) plus EngiWorld source 18 (2026-09-30) plus O*NET-BENCH source 19 (2026-10-05) plus pedagogical metric source 20 (2026-10-16) plus AI voice agent in a call centre source 21 (2026-10-20) plus Agent Delegability Index source 22 (2026-10-23) plus education gap GenAI vs internet source 23 (2026-10-28, abstract only). 23 openings (22 full text, 1 abstract only). PISA ≠ gen-AI RCT; EngiWorld 300≠1,301; meta ≠ AI-retraining guarantee. Not used as warrant: QJE snippets, BLS bot-block, WEF, McKinsey, LinkedIn Work Trend. Keep WP figures and Science figures separate.

  • Established (layer 1): education gap GenAI 2025 vs internet at similar use: smaller in 15 of 16 countries (median −27.6 pp) and 13 of 16 (log ratio −0.625) (23, E2, abstract); delegability index world 35.2% partial / 2.5% full, 48.7% high vs 22.5% low income (22, E2, model-rated); parallel AI voice agent: hot leads early +23.4 pp, conversion −4.5 pp, hot leads 4.4 → 3.7% (21, E2); Noy Science −40% time / +18% quality (n=453); Noy WP −37% / +0.45 SD / 68% unedited; Peng −55.8%; Brynjolfsson +14% RPH, novices +34%; Bastani exam −17% Base / Tutor ≈ control; Dell’Acqua inside +12.2% completion / outside −19 pp; PISA 2022 OECD means 472/476/485 and trend −15/−10/stable; Roodman meta +1.7 pp / ~$800 / sector ~10×; O*NET-BENCH judges 3.0–97.9% vs workers 61.1% acceptable (19); EngiWorld EngiScore 44.3 / multi 3.6% (18); Medly GCSE ITT g=0.33 (17); Autor task framework; Acemoglu aggregate 0; OECD 2023 little evidence; Census 17–20%; OECD 2025 demography.
  • Claimed (layer 2): GenAI more evenly spread across education groups than the internet once was, matched countries only (23, E3); delegation rather than exposure for agents (22, E3); reservation-value mechanism, AI routing as a behavioural intervention (21, E3); Frey 47% high-risk; Eloundou 80% / 19% exposure; IAB scenario net jobs ≈ 0 under assumptions.
  • Constrained (layer 3): education gap self-report, 16 of 35 countries non-random, ratio fragile (9 of 11, p = 0.065), with education composition only −10 to −11 pp, not an education effect, abstract only (23); delegability index ceiling ≠ adoption, one developer rates, US task mix (22); call-centre field study one firm/2024, per-human metric ≠ performance (21); exposure ≠ displacement; productivity RCTs ≠ learning; jagged frontier (one firm, GPT-4 2023); PISA ≠ gen-AI RCT (UNESCO still unopened); O*NET-BENCH ranking ≠ acceptance rate, worker rating = target ≠ truth (19); EngiWorld 300-subset/VM/license ≠ live engineering; Medly curriculum-aligned ≠ national exam / 30.7% attrition; retraining meta ≠ AI guarantee; Frey 47% as a 2026 event E4.

Not opened (not a warrant)

  • Brynjolfsson QJE 2025 — NBER WP 5,179 / 14% is the warrant.
  • BLS EP 2024–34 / MLR / TED bot block.
  • Census CES-WP-26-25 HTTP 500; HTOS Aug 2026 JS-empty.
  • Cui Copilot field; Autor New Frontiers; Arntz/Gregory/Zierahn; DiWaBe 2.0; IAB-KB 2024.
  • WEF Future of Jobs; McKinsey/Goldman; UNESCO / further OECD Education volumes (PISA Vol. I opened as source 15).