# PROJECT NOTES — filed 31 August 2026 **Project:** *Nobody to Blame: Attribution, the Hiring Margin, and the Coming Politics of Artificial Intelligence* — Aaron Long, draft 4.0. **Provenance:** the drafting and verification work recorded here was carried out 30 July – 1 August 2026. This is the standing reference file for the project: what was built, what is true and how well established it is, what the paper will be attacked on and what the answers are, what changed across four drafts, and what to do next. **Read this file before touching the paper again.** Several of the numbers below were wrong in earlier drafts and were corrected in a direction that is not obvious. The error log at §10 exists so the same mistakes are not made twice. --- ## 1. Where the project stands | Deliverable | Status | |---|---| | `nobody_to_blame_v4_core.pdf` | 26pp. The argument. | | `nobody_to_blame_v4_full.pdf` | 30pp. Adds §9 historical record, §10 ownership and the ε limit case. | | `build_paper_v4.py` | Single build script, both editions. `build(False)` core, `build(True)` full. | | `attribution.py` → `attribution_result.json` | The model: attribution structure, hiring margin, pipeline, RCT reframe, 5 predictions, 8 indicators. | | `attribution_figures.py` | Figures 1–3. | | `data.py` | Every numeric series, each with a `source` string and a `status` tag. | | `limit_case.py` | The ε derivation. | | `README.md` | Public-facing summary. | | `ATTACK_NOTES.md` | Hostile-referee file, now with a draft-4 section. | | `NOTES_2026-08-31.md` | This file. | Superseded but retained: `third_disembedding_v1_draft.pdf`, `..._v2_draft.pdf` (Polanyi and Marx at length), `slowness_condition_v3_{core,full}.pdf` (scenario taxonomy, Locke, limit case). **Draft 3's central claim is retracted** — see §3. Figures in the paper: (1) hiring margin, (2) attribution 2×2, (3) pipeline equilibria, [full only: (4) tool phases, (5) ownership + Locke], (6/4) METR time horizon, (7/5) electrification. --- ## 2. The thesis, stated precisely > Economic displacement always produces a political response. What determines the > **form** of that response is not how large the displacement is but how it can be > **attributed**. The operative variable is **agency plus out-group membership**. > > - Agency + out-group → demands for **protection** (block the outsider). > - Agency + in-group → demands for **compensation** (make them share). > - No agent → **diffuse disaffection with no policy vehicle**. > > Artificial intelligence is the first major displacement in modern economic history > with no foreigner attached to it. The claim is **form, not magnitude**. That distinction is the whole contribution, and it is what reconciles two literatures that appear to disagree: one finds trade displacement driving political realignment, the other finds robot displacement doing it at least as strongly. Both are right. They were measuring magnitude; the disagreement is about form, which neither was designed to see. I have not found this reconciliation stated anywhere. Two authors get close and are credited in the text: **Rodrik** ("we do not see populists campaign against technology or automation"; trade is "a convenient scapegoat, since politicians can point to identifiable foreigners") and **Mutz** (8:1 ratio of trade to automation newspaper coverage; trade attribution raises the belief that "the jobs can come back" — a cleaner mechanism than mine, and adopted). ### Two mechanisms carry it forward 1. **The hiring margin.** Adjustment runs through hiring, not separations. Nobody is fired; the junior requisition never opens. Consequence: the cost falls on people who do not yet have a job — young, unorganised, dispersed, represented by no institution, and invisible to the statistic everyone watches. 2. **The pipeline.** Junior work is how senior competence gets produced. It was free to the firm because it was a by-product of output the firm wanted anyway. AI supplies that output directly, so training acquires a positive marginal cost for the first time — in a market that underprovides it. Training equilibria come in pairs, so the failure mode is a tipping point, invisible for a decade, then a step. --- ## 3. The retraction (draft 3 → draft 4) Draft 3 argued that **no protective countermovement followed the post-1980 displacement**, reading the halving of US union density as the absence of an institutional response — 46 years of measured harm, no reaction, the Polanyian mechanism simply failed. Formally: λ ≈ 0. **That was wrong.** 2016–2026 is a countermovement by every criterion Polanyi states — protective, piecemeal, non-doctrinal, not centrally coordinated, arising from those most immediately affected: - Brexit; two Trump administrations - Section 301 tariffs from 2018, expanded 2025; **effective US tariff rate 11.8 % in April 2026**, the highest since the early 1940s - Sustained immigration restriction; net migration possibly negative for the first time in 50+ years - CHIPS Act, Inflation Reduction Act, the general return of industrial policy Union density was simply the wrong instrument for detecting it. **Two further premises under draft 3 also fail, and both are worth stating plainly because both are widely believed:** **(a) Automation did *not* destroy more US manufacturing employment than trade.** Robots: **360,000–670,000** jobs, 1990–2007 (Acemoglu & Restrepo's own aggregate). China shock: **~2.0–2.4 million**. Trade is **three to seven times larger** on the best available estimates. The circulating "88 % was automation" figure is Hicks & Devaraj (Ball State, 2015) — a productivity **residual**, not an estimate, and its own two components sum to 101.4 %. Houseman's critique is decisive on measurement: computers are 10–13 % of manufacturing value added, and excluding them, 1997–2007 manufacturing real GDP growth falls from 3.6 %/yr to 1.2 %; manufacturing ex-computers had **5 % lower** real value added in 2011 than in 2000. Fort, Pierce and Schott's verdict on separating the two is the honest one: *"extraordinarily difficult."* **(b) Automation did *not* escape political consequence.** Anelli, Colantone & Stanig: 14 countries, 1993–2016, **both shocks in one regression** — a one-SD robot shock raises the radical-right vote by **1.8 pp on a 5.7 % base**, while the China shock is positive but small and insignificant. Frey, Berger & Chen: 95 % lower robot exposure flips Michigan, Pennsylvania and Wisconsin to Clinton in 2016 — and note that the only "this flipped the Rust Belt" counterfactual in the literature is about *robots*, not imports; Autor and co-authors decline to run the trade version. **Why the retraction is load-bearing rather than embarrassing.** Both broken premises pointed the same way: the paper had assumed technology escapes blame. It does not. Once that is given up, the surviving question is not *whether* displacement produces politics but *what shape* the politics takes — which is the draft-4 thesis. The error produced the result. --- ## 4. Fact ledger Verification tags: **V** = verified against a primary source; **V2** = verified twice under different framings; **U** = unverified, not used in text. Everything below that is used in the paper is V or V2. ### 4.1 The attribution experiment — the paper's foundation Di Tella, R. & Rodrik, D. (2020), *Economic Journal* 130(628): 1008–1030. **V** Identical vignette — a 900-worker garment plant closes — with the **cause** randomised across arms. Demand for import protection: | Arm | Protection demand | |---|---| | control (no cause given) | 0.09 | | technology shock | 0.13 | | outsourcing to France | 0.23 | | outsourcing to Cambodia | **0.29** | Two further findings, and they matter more than the headline: - **The sign flips on the other margin.** Trade attribution *reduces* demand for compensating the losers by **6–8 points**; technology and bad-management attribution *raise* it. Attribution redirects the response, it does not merely scale it. - **Bad management identifies the variable.** Fully agentic, fully domestic → **zero** protectionism and the **largest** transfer demand of all. So the operative variable is not attributability. It is agency **plus out-group membership**. *What this does not license:* stated preferences over a hypothetical closure. The mapping to enacted policy is an assumption, and it is the joint the whole theory turns on. Said in the text, in a caption, and in the disclosure. ### 4.2 The hiring margin | Fact | Value | Tag | |---|---|---| | JOLTS hires rate, May 2026 | **3.3 %** (vs 4.3 % Jan 2022) | V | | JOLTS quits rate, May 2026 | **1.9 %** (vs 3.0 % peak Nov 2021) | V | | JOLTS layoffs & discharges | **1.1 %** — 0.2 pp above the all-time low of 0.9 % | V | | BLS unemployed new entrants, Jun 2019 → Jun 2026 | **509,000 → 772,000 (+52 %)** | V | | …as a share of total unemployment | **8.6 % → 10.9 %** | V | | NY Fed recent-graduate underemployment | **41.5 %** | V | | U-4 and U-5 | flat year on year | V | **Firm-level, and the strongest single piece of evidence in the paper:** Hosseini Maasoum & Lichtinger, 65 million résumés across 280,000+ firms — junior employment falls in AI-adopting firms while senior employment is unchanged, *"driven primarily by slower hiring rather than increased separations."* The mechanism in the authors' own words. **The correction that matters:** an earlier version claimed the entry-level adjustment was *invisible to official statistics*. Wrong. It is visible — in a series nobody looks at. What is genuinely invisible is **underemployment**: U-6 counts involuntary part-time work, not skill mismatch. A philosophy graduate working forty hours as a barista is fully employed in U-3 through U-6 alike. If AI closes the graduate-entry rung without reducing aggregate labour demand, the observable consequence is not unemployment at all — it is a cohort working, earning, counted as employed, and not accumulating the human capital the job title used to come bundled with. ### 4.3 The pipeline - **The unbundling argument** (the paper's own contribution, and narrow): junior training was free to the firm because it was a by-product of wanted output — the associate's document review, the resident's rounds, the graduate's tickets. AI supplies that output directly. Training acquires a positive marginal cost **for the first time**, and must be paid deliberately by a firm that does not capture its full return because trained workers leave. A textbook underprovision problem *newly created by a technology*, not an old one made worse. - **The price of mentorship:** Emanuel, Harrington & Pallais — across 1,055 engineers, senior engineers write **0.76 fewer programs per month** while mentoring. That cost was always there. What changed is that the offsetting benefit can now be bought elsewhere. Bundled, it was never priced. Unbundled, it is a line item, and line items get cut. - **Why a cliff, not a slope:** Afrouzi, Blanco, Drenik & Hurst (2026) formalise entry tasks as *"the curriculum through which workers accumulate human capital"* and prove equilibria always exist in **pairs** — high-learning and low-learning — with cheaper technology raising learning in the first and driving it to zero in the second. Acemoglu & Pischke (1998) had the same structure for general training under frictions. This is **not** post-2008 cohort scarring. Scarring is a slope that closes. A training equilibrium is a tipping point between two stable states, and the transition is invisible while the existing stock of seniors is still working. - **The forecasting problem in one sentence:** the firm that stops training sees no cost in its own accounts within any horizon over which its managers are evaluated. ### 4.4 The RCT reframe — the objection that looks fatal and is not Every well-identified trial finds AI gains are **junior**-biased: - Cui et al., 3 RCTs, 4,867 developers: juniors **+21–40 %**, seniors **+7–16 %** - Brynjolfsson, Li & Raymond: novices **+34 %**, experienced ≈ **0** If AI helps juniors most, how can it displace them? > **The mechanism does not require seniors to become more productive. It requires > SENIOR + AI to substitute for SENIOR + JUNIOR.** That condition holds precisely when > AI is most productive at junior-type tasks — which is exactly what the trials find. > Read correctly, the junior-biased gradient is evidence *for* the mechanism. **Citation hazard, circulating now:** METR's February 2026 follow-up found **−18 %** for the original cohort (CI −38 % to +9 %) and −4 % for new recruits, and the team is redesigning the study because selection broke it. A widely shared 2026 summary reports this as a "+18 % speedup" — a **sign error on a headline result**. The only sources claiming senior-biased gains are vendor telemetry with no stated methodology. ### 4.5 Rates METR time-horizon series (the one capability measure on a *time* scale rather than a benchmark score). Fitted doubling times: **197 days** full sample, **131 days** since 2023, **89 days** since 2024. Tag: V for TH1.1 rows, U for pre-2023 rows. Four caveats, all from METR themselves, all in the paper: 1. Intervals wide — upper bounds run **1.66× to 2.28×** the point estimates. 2. Only **5 of 31** tasks over eight hours have measured human baselines. 3. The TH1.0 → TH1.1 instrument revision moved the trend by ~20 %; old models fell 35–57 %, new models rose 11–55 %. **Part of the apparent acceleration is a composition artefact.** 4. Software and ML engineering only. Nothing licenses extrapolation to legal, medical, managerial or physical work. **But the rate that matters is deployment, not capability.** Electrification: first central generating stations 1881; 5 % of factory mechanical drive electrified in 1899, 50 % by 1919, 80 % by 1929; US manufacturing TFP growth 1.5 %/yr (1899–1914) → 5.1 %/yr (1919–29). ~40 years, and the binding constraint was reorganisation — unit drive, factory layout, the whole management system — not the electricity. Three lags separate capability from economic effect: **invention, diffusion, reorganisation**, and only the first is on the METR curve. Software diffuses faster than capital equipment, so 40 years is an upper bound; the reorganisation lag has no particular reason to have shortened, because it is bounded by how fast organisations and the people in them can change. **This is why the paper does not forecast.** The argument is built to be informative whatever the rate. What depends on the rate is only *when* the indicators move. ### 4.6 The feedback loop, and the caveat that cuts against it **Claim:** protection and immigration restriction raise the price of labour relative to capital, and labour scarcity induces automation. A countermovement aimed at the wrong target does not merely fail — it accelerates the thing it is reacting to. **Evidence, clean:** Clemens, Lewis & Postel (*AER* 2018) on the 1964 Bracero termination — excluding ~half a million Mexican farm workers produced **no detectable wage gain** for domestic workers and immediate adoption of the mechanical tomato harvester. Acemoglu & Restrepo: demographic ageing explains ~**40 %** of cross-country variation in robot adoption. **Caveat, stated rather than buried:** the 2024 Section 301 escalation put 25–50 % duties on semiconductors, batteries and solar — tariffs on **capital goods**, which raise the price of automation rather than lowering it. Flaaen & Pierce found machinery and computers among the sectors most damaged by the 2018 round and put the net effect on manufacturing employment at **−2.7 %**: input costs and retaliation outweighed the protection. So the general mechanism is well evidenced; its application to the actual 2018–2026 tariff schedule is **ambiguous in sign** — even as that schedule is unambiguously negative for the employment it was enacted to protect. ### 4.7 The historical record — three questions, three answers The single most useful table in the project, because nearly all confusion in this area is answering one question while being asked another. | Question | Answer | What it does not license | |---|---|---| | Do displaced **occupations** recover? | **No. Essentially never.** Hand-loom weavers 240,000 (1820) → 10,000 (1860). UK coal 1,191,000 (1920) → 360 (2022). US coal 863,000 (1923) → 41,600 (2020) *while output rose to an all-time peak in 2008*. US agriculture 41 % of employment (1900) → 1.9 % (2000). | Almost nothing on its own. Selection on the dependent variable: occupations that were not destroyed are not in the sample. | | Do the **specific displaced people** recover? | **Mostly not, within their own working lives — roughly 1 in 7.** China-shock literature: 86 % of net job loss absorbed by a falling employment rate rather than reallocation, effects still measurable in 2019. Displaced workers lose ~1.4 years of prior earnings in expansions, 2.8 in recessions. | **This is the level the public argument is actually about, it is where the record is most pessimistic, and it is the level least often cited.** | | Does **labour in aggregate** recover? | **Always, so far. Roughly one for one.** No episode reduced aggregate employment over the medium run. Bessen, 317 occupations × 243 industries, 1980–2013: net effect of computer use on total employment **−0.07 %/yr**. Composition, not level. | **Every recovery ran through generational replacement over 50–100 years, not retraining.** Aggregate recovery is compatible with total individual ruin, and historically has been. | **Tool phases** (years from commercial introduction to displacement, where a datable transition exists): container shipping 10 · tractor 14 · power loom 20 · industrial robots 30 · ATM 35 · spreadsheet 40. Median 25. AI dated from ChatGPT: 3.7 years. The datings are **interpretive** and the sample is selected on transitions that happened. The usable point is only that the tool phase is real, has lasted roughly a decade to four, and **has never been announced as ending in advance**. ### 4.8 Distribution (full edition, §10) - Top 1 % of US households hold **50.2 %** of corporate equities and mutual fund shares; 90th–99th a further **37.2 %** (top decile ≈ **87 %**); bottom half ≈ **1 %**. Federal Reserve Distributional Financial Accounts, Q1 2026. - Therefore: whatever share of AI's gains accrues to equity accrues roughly half to the top 1 % and roughly seven-eighths to the top decile. **This requires no contested economics — only a share register.** A weaker AI produces a smaller surplus distributed in the same proportions. **The proportion is the settled part.** - **"Only capital survives" is wrong.** Capital is *reproducible*, and nothing reproducible stays scarce. In the limit the returns accrue to whatever cannot be built: land, energy, spectrum, compute at a binding physical constraint, regulatory position, data with a legal moat. Terminal share of the non-produced factor, by ε (calibrated at γ = 0.05, the natural-resource share Trammell & Korinek use, as the reproducible composite grows ten-thousand-fold): | ε | terminal share | reading | |---|---|---| | 0.5 (complements) | **96.5 %** | rentier terminus | | 1.0 (Cobb–Douglas) | 5.0 % | shares constant by construction; nothing happens | | 1.5 (substitutes) | 0.65 % | the constraint is engineered around | Note the trap in the ε = 1 row: γ = 0.05 is the figure Trammell & Korinek use to argue resources will not bind — but under their Cobb–Douglas nesting that share is **constant by assumption**. The functional form forbids the result they are testing for. For a genuinely non-produced factor the standard assumption is ε < 1. **Three orders of magnitude, and ε is not identified.** That is why the paper does not forecast a distributional terminus. What it does establish is the *direction of the correction*: the intuition points at the wrong asset, and the real candidate for terminal concentration is a claim about land, energy and law. --- ## 5. The defence — attacks, and the answers Ordered by how much damage they do. The full hostile-referee version is in `ATTACK_NOTES.md`; this is the short form to have in hand. **A1. "You built a theory of politics on one survey experiment with hypothetical vignettes, then used it to read a decade of real elections."** *This lands.* The defence is real but not yet sufficient: (i) the distinction is stated as falsifiable predictions P3 and P4 with resolution dates in Nov 2026 and Nov 2028; (ii) the compensation-margin **sign flip** is not something a magnitude story predicts, so it is a genuine discriminating observation, not a re-description. Until P3 or P4 resolves, the central claim is a well-motivated conjecture and the paper says so in §4's caption and in the disclosure. **Do not defend this harder than that.** **A2. "The entry-level evidence is not an AI result."** *Correct, and conceded in a box in §7.* Emanuel, Harrington & Pallais attribute **64 %** of the rise in young-graduate unemployment (2017–19 → 2022–24) to **remote work**, through the *same* lost-mentorship mechanism. And the recent-graduate/overall unemployment crossover is **February 2019** — pre-pandemic, pre-AI. The defensible claim is: the entry rung is under strain, and the mechanism is mentorship. Anyone citing this as an AI result is overreading it, and the paper's own claims box says so. **A3. "The pipeline has no case behind it."** *Correct, and it is the largest weakness in the paper.* No documented instance of a training pipeline collapsing **while demand held up**, with clean numbers and a decade-lagged consequence. UK apprenticeships halved 1979–95 but the industries died first — causation runs the wrong way. What the argument actually rests on: a theorem about equilibrium multiplicity (solid), a measured mentorship cost of 0.76 programs/month (solid, one firm, one occupation), and a claimed regime change in whether training is free (**unmeasured — the paper's own contribution**). Indicator 6 would settle it and no time series exists. **A4. "The junior-posting evidence is contested and you resolved it by assertion."** Partly fair. Indeed: entry postings **−7.5 %** y/y against senior **+15 %**. Two Lightcast-based studies: junior and senior demand moving broadly in parallel. The paper's resolution — *the bin is stable; its contents are not* — is supported by PwC (entry roles demanding senior skills **+35 %**, non-seniorised entry roles **−10 %**) and Indeed (**30 %** of entry-level applications now from people with 10+ years' experience). That is suggestive, not dispositive. **Indicator 4 states the test; nobody has run it.** Measuring on title will miss the effect entirely — measure on task text. **A5. "The RCTs refute you."** Answered fully at §4.4 above. This is the objection the paper is best positioned against, and the reframe should be stated early when presenting. **A6. "Your feedback loop is contradicted by the actual tariff schedule."** Conceded in §6 rather than buried. See §4.6. The concession costs the section its punchline and buys the paper its credibility. **A7. "Five sections of philosophy in a paper about JOLTS."** Reduced sharply from draft 3 — Locke survives only as full-edition §10 and two sentences in §12.4. The answer: the limit case *exactly reverses* Locke's explicit empirical premise (labour supplies 9/10 or 99/100 of value; "nature and the earth furnished only the almost worthless materials", §43), and the contractarian exclusion problem is the philosophical shape of the §12.4 political worry — contractarian justifications have an exclusion problem exactly when a group's cooperation stops being needed. Thin, but not decorative. --- ## 6. Weaknesses, ranked (this is the order in the disclosure) 1. **The pipeline argument (§8)** — theoretically well-founded, historically unevidenced. 2. **Attributing entry-level weakness to AI** — remote work takes 64 %, and the crossover predates any plausible AI effect. 3. **ε (§10, full edition)** — unidentified; the rentier case depends entirely on it. 4. **AI exposure measures throughout** — vendor telemetry, not representative instruments. 5. **Extending a survey-experimental result to enacted policy** — the joint the whole theory turns on, and the experiment was not designed to bear it. --- ## 7. Five predictions | | Claim | Falsified if | |---|---|---| | **P1** | Displacement shows up in hiring rates and new-entrant unemployment **before** the headline unemployment rate — and possibly without ever showing up there. | Unemployment rises materially while the hires rate is stable. | | **P2** | The junior-to-senior hiring ratio falls in AI-exposed occupations, visible in **task content** of entry-labelled postings before it is visible in their count. | Task-text seniority is flat while counts fall — an ordinary demand story. | | **P3** | Political response takes the form of **compensation** demands rather than protection, because there is no foreigner to exclude — unless a target is manufactured. | AI displacement produces trade or immigration restriction aimed at AI specifically; or produces no policy demand at all. | | **P4** | The AI countermovement does **not** inherit the trade countermovement's coalition, because it does not hit the same places. | AI-exposed districts move as manufacturing-exposed districts did after 2000. | | **P5** | Attempts are made to give AI displacement a **legal name**, because a compensation demand needs a cause to attach to. | Such attempts stop appearing. *Partially confirmed already.* | **The two I would bet against myself on:** P3 fails if a target is successfully manufactured, and there is no strong reason to think it will not be. P4 fails if county-level AI exposure is a poor proxy for who actually loses work, which is entirely possible given how those measures are built. --- ## 8. Eight indicators, with thresholds and dates | # | Indicator | Reads now | Fires / Dies | When we'd know | |---|---|---|---|---| | 1 | Unemployed new entrants (BLS) | 772k, +52 % since 2019; 10.9 % of unemployment vs 8.6 % | **Fires:** share keeps rising while headline is flat/falling · **Dies:** reverts to 8–9 % while adoption continues | monthly; 12–18 months | | 2 | Recent-graduate underemployment (NY Fed) | 41.5 %; no BLS measure captures it | **Fires:** rises while U-3 and U-6 stable · **Dies:** falls to pre-2019 | monthly; 12–18 months | | 3 | JOLTS hires rate | 3.3 % (May 2026) vs 4.3 % (Jan 2022), layoffs 1.1 % | **Fires:** hires depressed with layoffs at record lows · **Dies:** hires recover with no rise in layoffs | monthly; 12–18 months | | 4 | **Junior:senior posting ratio, on task text not title** | **Contested** — see A4 | **Fires:** task-text seniority of entry roles keeps rising · **Dies:** entry roles retain entry-level task content | months — analysis, not data, is the constraint | | 5 | **Within-occupation return to experience, by AI exposure** | **Not estimated** | **Fires:** gradient compresses in exposed occupations only, concentrated in first decade of tenure · **Dies:** no differential compression | **a day's work on CPS or ACS** | | 6 | Training / mentorship spend per junior hire | **No time series exists** | **Fires:** training spend falls while senior productivity rises · **Dies:** training holds up in AI-adopting firms | needs a study nobody has run | | 7 | Whether AI acquires a **legal name** as a cause of layoffs | Great American AI Act discussion draft (June 2026) would amend the WARN Act to require employers to state whether AI caused a mass layoff, and study AI adjustment assistance modelled on TAA | **Fires:** it passes, or any jurisdiction mandates AI attribution in layoff notices · **Dies:** it dies and nothing replaces it — displacement stays nameless, compensation channel stays closed | legislative calendar | | 8 | Political geography of AI exposure | **62 of the 100 most AI-exposed US counties voted Democratic in 2024** — inverting the China-shock and robot-shock geography | **Fires:** AI-exposed districts shift measurably 2026→2028 · **Dies:** no realignment by 2028 | Nov 2026, decisively Nov 2028 | > **Stated kill condition: if four or more of the eight have not moved in the predicted > direction by the end of 2028, the paper is wrong and should be discarded rather than > patched.** --- ## 9. What follows, for the next five to ten years (§12) 1. **The politics gets angrier and less coherent at the same time.** Disaffection arrives on schedule; a compensation demand needs a cause to attach to and AI supplies none. Predicted signature: high and rising disaffection with no stable policy vehicle — volatile support, rapid switching, anti-incumbency without a programme. Roughly what 2024–2026 has looked like across the OECD. 2. **Someone will manufacture a target.** Three candidates already visible: *legislative* (the WARN Act amendment — a device for creating an attribution where none exists); *corporate* (the frontier labs are a small, named, wealthy, geographically concentrated set of firms — the most available agent in the picture); *foreign* (reframing AI as a Chinese technology race converts an agentless domestic cause into an out-group one). **The third is the one to watch — it is the only route by which AI displacement produces classical protectionism.** 3. **The damage lands on entrants, and lands late.** A cost borne by the politically weakest group, arriving on a lag longer than an electoral cycle, attributed to nothing in particular. Close to a worst case for institutional response. Note the structural trap: unions represent incumbents, and displaced-worker programmes are triggered by **separations** — exactly the event that is not happening. 4. **The compensation question arrives before the automation question.** Long before any aggregate employment effect is measurable, the distributional question is live, because the surplus accrues to equity and equity ownership is already what it is. The likely form is not UBI-for-technological-unemployment but ordinary distributional conflict over an unusually concentrated gain: taxation of returns, sectoral bargaining, adjustment assistance, licensure and professional-scope fights. --- ## 10. Error log — corrections made, and their directions Three independent verification passes with different framings. **Pass 2 found 24 items pass 1 missed, including one figure pass 1 had corrected in the *wrong direction*.** That is the argument for doing it more than once. Anyone reusing a number below should still check it. **Substantive reversals (draft 3 → 4):** - λ ≈ 0 since 1980 → **retracted**. The countermovement arrived; it was aimed elsewhere. - "88 % of manufacturing job loss was automation" → a **residual**, not an estimate; components sum to 101.4 %; trade is 3–7× larger. - "Technology escapes political blame" → **false**; reframed as form-not-magnitude. - "The entry-level adjustment is invisible to official statistics" → **false**; it is visible in BLS new entrants. Underemployment is the invisible margin. **Citation-level hazards, all of which are circulating in wrong form:** - Acemoglu & Restrepo automation TFP: **3.4 %** in the published *Econometrica* version, 3.8 % only in the working paper. *Pass 1 corrected this in the wrong direction.* - Brynjolfsson, Li & Raymond published figures: **15 % / 30 %**, not the widely quoted 14 % / 34 %. - METR February 2026 follow-up: **−18 %**, circulating as "+18 %". A sign error on a headline result. - Polanyi passage is **p. 75**, not p. 71. - Hejeebu & McCloskey wrote **"relies on"**, not "presupposes". - Locke's two fractions are **both in §40**, with different denominators — not in separate sections as the secondary literature often implies. - Paine's premise is the earth in its "natural **uncultivated** state"; the most-linked online transcription corrupts this to "cultivated", reversing his meaning. - Boyer quote is from the **EH.Net encyclopedia**, not the source usually credited. - Blaug quote was truncated in a way that changed its force. - "Union **coverage**" should read "union **membership**". - US union density series are **not splice-compatible** across the 1970s denominator change. Two incompatible series are carried separately in `data.py` and must not be joined. **Build and code errors worth remembering:** - ReportLab Times-Roman has no glyph for İ, ṁ, ℓ, ⟹, 𝒲, 𝔼, ᾱ or combining macron — they render as filled black boxes. Probe with `pdfmetrics.getFont("Times-Roman") .stringWidth(ch, 10)` before using any non-ASCII character. Dotted derivatives were replaced with Leibniz notation; 𝒲→*H*, 𝔼→*E*, ℓ→*l*. - The toy-model parameter sweep was dominated by sampling noise because each parameter value drew fresh noise. Fixed by passing `noise` into `make_world()` and using **common random numbers** across 40 replications. - A blanket string replacement in `scenarios.py` (`"eps "` → `"ε "`) matched inside "keeps a", producing "keε a". Never do blanket replacements on short substrings. - Draft 4 build: `figure()` and `table()` return different types — `table()` returns a single flowable and must be wrapped, `ext([table(...)])`. - Figure 7 in draft 4 was originally `fig09_three_lags.png`, which contains a "no response at scale" row asserting exactly what §3 now retracts. **Swapped for `fig07_electrification.png`.** If reinstating the three-lags figure, regenerate it first. --- ## 11. What to do next, in priority order 1. **Run indicator 5.** Within-occupation return to experience by AI exposure, CPS or ACS, roughly a day's work. If the gradient compresses differentially in exposed occupations and concentrates in the first decade of tenure, the expertise-rent mechanism has an identified test rather than one field experiment in one occupation. **This is the single highest-value action available and it is cheap.** 2. **Re-run the posting analyses on task text rather than titles.** Indicator 4. The data exists; the analysis does not. 3. **Search harder for one historical pipeline collapse with demand intact.** If it does not exist, say the mechanism is *novel* rather than analogous — a stronger claim and a more exposed one. 4. **Wait for November 2026 and November 2028.** P4 resolves on real votes. 5. Optional, skipped for context reasons in the draft-4 session: a third research stream on tariffs-as-automation-subsidy and recursive improvement rates. Key pieces are already captured in dossier 11 (Bracero; Acemoglu–Restrepo ageing; the Section 301 capital-goods caveat). --- ## 12. Reproduction ``` python3 attribution.py # model + indicators -> attribution_result.json python3 attribution_figures.py # figures 1-3 python3 generate_figures.py # historical figures python3 scenario_figures.py # scenario / ownership figures python3 build_paper_v4.py # -> nobody_to_blame_v4_{core,full}.pdf ``` Simulations use **fixed seeds** (`np.random.default_rng(20260731)`) and **common random numbers** across parameter sweeps. Every series in `data.py` carries a source string and a verification tag; anything tagged **U** is not used in the text. Primary philosophical texts are quoted to stated editions **with section numbers**, because section numbers are stable across editions and page numbers are not. --- ## 13. One paragraph, if that is all there is time for The public asks whether AI will take the jobs and watches the unemployment rate. Both are wrong. A frozen labour market adjusts on the cheapest margin, which is the requisition that never opens — hires down a full point since 2022, layoffs near an all-time low, junior employment falling in AI-adopting firms while senior employment does not, people seeking a first job up 52 % in seven years, and none of it in the headline rate. The aggregate always recovers and the specific displaced people mostly do not, within their own working lives; every recovery in the record ran through generational replacement, not retraining. What is genuinely new is narrower: junior work was how senior competence got produced, and it was free because it was a by-product of wanted output — AI supplies that output directly, so training gets a positive marginal cost in a market that underprovides it, and training equilibria come in pairs, so the failure is a tipping point a decade away rather than a slope. And the politics: displacement produces political response whatever causes it, but **attribution determines the form** — an outside agent produces demands to block, an inside agent produces demands to share, and no agent produces disaffection with nowhere to go. **AI is the first major displacement in modern economic history with no foreigner attached to it. The anger will arrive on schedule. What it asks for is being decided right now, by whoever succeeds in giving the thing a name.** --- ## 14. Draft 4.1 addendum (research pass of 1 August 2026) Executed next-steps §11.1–.3. Four additions, all folded into the paper: 1. **PATCO 1981–92 — the reverse experiment (new §8.1).** No forward case of a pipeline collapsing with demand intact exists (still true), but the reverse experiment measures the pipeline's **time constant**: ~10,400 of 17,000 controllers fired 5 Aug 1981 (4,669 left); demand intact; pipeline run at maximum throughput (400,000+ tested 1981–92, ~50,000/yr peak, 25,277 to the Academy); recovery declared complete **mid-1992** (11 yrs), headcount ~14,400 only by **mid-1995** (14 yrs). Use "11–14 years, both definitions stated." *Citation hazard:* popular accounts say 11,345/11,359 fired; the FAA's own report (Broach, DOT/FAA/AM-98/23, 1998) says ~10,400 of 17,000. Cite the FAA report. PATCO bounds **repair time, not collapse probability** — do not overclaim. 2. **NY WARN AI disclosure read ZERO (indicator 7 partially fired).** NY added AI-attribution disclosure to state WARN in March 2025; first year: **160+ notices, zero cited AI/automation** (Hunton, May 2026). This is *predicted* by the hiring-margin mechanism: WARN triggers on separations, the margin that isn't moving. Indicator 7's fires/dies conditions rewritten accordingly — a WARN-type instrument reading materially nonzero would now *hurt* P1. Caveats: statute not amended, key phrase undefined, checkbox not audit. 3. **Dallas Fed (Davis, 24 Feb 2026) — indicator 5's between-occupation cousin.** Experience premium (median 40%) positively correlated with AI exposure; wage effect of exposure: −0.28pp (zero-premium), −0.05pp (median), **+0.2pp** (90th-pct premium); exposed-sector employment −1 %, disproportionately under-25. Direction: seniority protects. **Do not cite as indicator 5 firing** — the within-occupation compression test is still unrun; `indicator5_experience_gradient.py` (new) is the runnable spec (IPUMS CPS ASEC, Felten AIOE exposure, spline × exposed × post, occupation FE, 2019–22 placebo). 4. **Named counter-evidence added to §7 box.** EPI (Fast & Gould, 7 May 2026): young noncollege unemployment rose similarly (to 7.1 %) → nothing college-specific; credential-erosion story (attainment 18.0 % → 31.6 %). NY Fed (Audoly, Guerin & Topa, 14 May 2026): junior/senior postings parallel in high-exposure occupations; exposed-occupation decline **began before ChatGPT**. Reply on record: the mechanism doesn't predict college-specificity. Data: May 2026 JOLTS re-confirmed 3.3 / 1.9 / 1.1 (release 30 Jun 2026); June JOLTS not yet out at time of pass. New series in `data.py`: PATCO_RECOVERY, NY_WARN_AI, DALLAS_FED_EXPERIENCE, ENTRY_LEVEL_COUNTERARGS (all V). Paper is now **Draft 4.1** (both editions rebuilt); references added: Broach 1998; Davis 2026; Fast & Gould 2026; Audoly, Guerin & Topa 2026; NYS DOL WARN disclosure. ### 14.1 Indicator 5 execution status (1 Aug 2026, second pass) - **Sandbox cannot download microdata** (Census, NBER, GitHub raw all unreachable; package registries only). The pipeline is therefore fully built and TESTED instead: `indicator5_experience_gradient.py` now loads NBER MORG .dta files, merges Felten AIOE via the Census-occ→SOC crosswalk, absorbs occupation FE by weighted within-demeaning, and computes occupation-clustered SEs in pure numpy (statsmodels unavailable). **Synthetic self-test passed**: recovers a known −0.02 first-decade compression within ±0.01, with correct nulls on later segments. - **One manual step remains**: 4 files, ~60 MB, listed with URLs in `INDICATOR5_DOWNLOADS.md`. Drop them in the project folder and say "run indicator 5." - **PwC June 2026 barometer verified at source** (indicator 4 row updated): one billion postings / 27 countries / 2.4M US entry-level jobs; AI-exposed entry roles **7×** more likely to demand senior skills; seniorised entry roles +35% since 2019 vs conventional −10%; AI wage premium 62% (from 57%). Release 15 Jun 2026. - MORG hazards logged in the script docstring: earnwke topcode ~2,884.61 (drop, and say so); occupation-code vintage check (occ18/docc03/occ2012); AIOE preferred over vendor telemetry for comparability with Davis (Dallas Fed).