Working hours: the productivity converged, the time didn't

Status: drafted 2026-08-13 Reachability swept first: annual-working-hours-per-worker and labor-productivity-per-hour-pennworldtable both fetch clean (130 entities, 1870–2023). share-of-employment-part-time and part-time-employment-share both 404 — which is why the composition caveat below stays an open question rather than becoming a fourth panel. Vintage checked first: Penn World Table 2025 release via OWID (lastUpdated 2025-08-05, nextUpdate 2026-09-13), data through 2023. Cross-checked against the fresher upstream — OECD's own 2023 figures are US 1,799 / Germany 1,343 against PWT's 1,789 / 1,335, a 0.6% difference, and both give a gap of ~455 hours. The two sources agree; the story is decadal, so a one-year lag behind OECD is immaterial.

Segment: The rich world stopped shortening the working year — but only in some places

• Source theme(s): Batch 42 #11. Cross-refs: Batch 5 #20 (cohort hours at 30), Batch 5 #22 (employer tenure), Batch 18 #28 (labour share of GDP).
• Spine role: main story
• The framing people carry: Two of them. That technology was supposed to buy us leisure and didn't — Keynes promised fifteen-hour weeks and we got email at midnight. And that Americans work hard while Europeans take August off, which is a lifestyle preference Europe pays for in lower incomes.
• Hypothesis (from the bank): "A long secular decline flattening from roughly the 1980s in the US while continuing in continental Europe, producing a gap of several hundred annual hours; diary-measured hours lower than survey-recalled hours for long-hours workers…; the distribution polarising."
• Verdict: CONFIRMED on the headline, to the decade and the magnitude — with two sub-claims untestable on this data and therefore cut. The US flattening dates to almost exactly 1980; continental Europe kept falling; the gap is 454 hours. The diary-vs-recall and distribution-polarisation clauses need ATUS/MTUS microdata we have no connector for, and neither reaches the episode. The finding the bank did not make is the better story: productivity per hour converged completely while hours diverged — so the US–Germany gap in output per worker is now, essentially entirely, time.

The numbers (every row sourced — no exceptions)

Hours: OWID grapher annual-working-hours-per-worker — Feenstra et al., Penn World Table (2025) and Huberman & Minns (2005), accessed 2026-08-13. Tier A. Productivity: OWID grapher labor-productivity-per-hour-pennworldtable — GDP per hour worked, PPP-adjusted, constant 2021 international $, same PWT release, accessed 2026-08-13. Tier A.

The texture (what narrative flattens)

• The lead finding: the gap is time, not talent. In 2023 a German worker produced $82.50 an hour against an American's $83.53 — a difference of 1.2%, and it has been under 5% since 1994. Yet output per worker differs by 26%, because the American works 454 more hours. Almost the entire US–Germany income-per-worker gap is the length of the working year. The familiar reading — that Europe is poorer because it is less productive — is the wrong way round on this measure: Europe is poorer per head partly because it converted its productivity catch-up into time rather than into income. Same growth, different currency.
• The direction reversed inside living memory. In 1950 Germans worked 405 hours more than Americans. The lines crossed in 1973. By 2023 Germans work 454 fewer. That is a swing of about 860 hours in seventy years, and it happened while German output per hour went from a quarter of the American level to parity.
• "The fall stopped" is too gentle for the US. American hours did not plateau after 1980; they rose for two decades — 1,840 in 1980 to 1,888 in 2000 — before drifting back to 1,789. Net movement over 43 years: −52 hours, about one hour a week. Over the same span Germany shed 421, France 289, Australia 174, the UK 131.
• The century of decline was real and enormous, and it is the baseline the flattening is measured against. US hours fell 42% between 1870 and 2023; Germany's 59%. Roughly two-thirds of the entire US decline had already happened by 1950.
• The reputational inversions. Japan — the country whose language supplied karoshi, death from overwork — has worked fewer hours per worker than the United States every year since 1998, and is now 135 hours below it. South Korea, still the long-hours case at 1,910, has fallen 37% from its 1969 peak of 3,040: the largest reduction in the dataset, by a country nobody associates with shorter days.
• Keynes was directionally right and wrong about the destination. The fifteen-hour week did not arrive, but a German full-year worker at 1,335 hours is at about 27.8 hours a week across 48 weeks. The prediction failed by less in Europe than the folklore suggests, and by far more in the US.

Datoid artifact

• Nearest existing artifact is the Labor Market dashboard, and it does not overlap: that one is US-only, FRED-sourced, and about participation, openings, quits and U-6 since 2015. This is global, PWT-sourced, and runs to
• Tier A with zero new connector work — both graphers are on the existing OWID connector via @entity addressing. Eight countries × two metrics = 16 series, all annual, two unit types (count for hours, currency for output per hour) which land on the chart's two axes automatically.
• Artifact: the working-hours dashboard, 4 panels:
  1. The century of decline, 1870→2023 — US, Germany, France, UK, Australia.
  2. Where it stopped, 1970→2023 — the divergence, plus Japan and South Korea.
  3. Output per hour, 1950→2023 — the convergence.
  4. Two-axis: US and German hours against US and German output per hour — the punchline on one surface.

Story justification (Stage 1 rule 5)

• Does the argument have a reveal that needs sequencing? YES → build.
• Because: the punchline is a second dataset, and it only lands if it is off-screen while the trap is set. The sequence is: hours fell everywhere for a century → then they stopped, but only in some places → so the obvious inference is that Europeans chose leisure and pay for it in output → reveal productivity per hour, which is identical → the inference was wrong, and the income gap is the schedule. A dashboard panel can state that; showing it from the first frame throws the argument away. This is structurally the same case as maternal mortality, and StoryStep.activeSeriesSlugs (added for exactly this) makes it cheap to build now.
• Slug: /stories/working-hours. Deferred within this session if time-boxed — the brief + dashboard + episode are the cadence deliverable and do not depend on it; the story is recorded as owed, not as done.

Open questions / where I'm guessing

These must NOT become spoken assertions.

1. The denominator is "per worker", and part-time work is the biggest confound in the whole sheet. A country can cut average hours by having more people work part-time without any full-timer's week changing. Germany and the Netherlands both have high part-time shares, and the Netherlands' 1,439 hours is substantially a part-time artefact. We could not test this — both OWID part-time slugs 404. So: do not say "Germans work a 28-hour week", and do not attribute the fall to shorter full-time weeks. Say "hours per worker", and say that some of the gap is more people working part-time rather than everyone working less. This is the single most likely way to mislead a listener, and it should be stated in the episode, not just in this sheet.
2. Output per worker is my multiplication, not a published series. Hours × productivity, both PWT, so it is internally consistent — but it is derived, and it is not identical to PWT's own GDP-per-worker variable. Say "roughly" and "about a quarter", never $149,427.
3. PPP comparisons carry real uncertainty. "Constant 2021 international dollars" is a construct; a 1.2% gap between the US and Germany is well inside the error bar of any PPP exercise. The honest claim is "indistinguishable", not "Germany is 98.8% as productive" — and certainly not that one has overtaken the other.
4. Pre-1950 figures are historical reconstruction (Huberman & Minns), not measurement. The 1870 numbers are directionally solid and precisely soft. Round them and say "around three thousand hours".
5. Two of the bank's three sub-claims are untested here, not refuted. Diary-vs-recall discrepancy and distribution polarisation both need ATUS / MTUS microdata. They should not appear in the episode at all — including as "some research suggests". If they are wanted later, that is a separate Stage 1.
6. Causation is entirely absent from this data. Working-time regulation, union density, vacation mandates, tax wedges and industry mix are all plausible drivers of the divergence, and none of them is in this dataset. The quantities stand alone; "Europe legislated shorter hours" is an interpretation and must be flagged as one.
7. Ireland and Norway are excluded on purpose and should not be quoted from this dataset even if asked: Ireland's $158/hour is the well-known multinational-profit-shifting distortion of Irish GDP, and Norway's $130 is petroleum rents. Both would be nonsense as productivity comparisons.
8. The 2023 endpoint is post-COVID and still settling. Hours in several countries have not returned to their 2019 path, and the 2020–21 points are pandemic artefacts. Read the trend, not the last point.