Sourced, ranked, and honest about the gaps

Evidence

The European evidence base behind the draft, with sources, and a dead-citations list: numbers that circulate in this debate that we checked and will not use, including some that would have helped us.

On this page
  1. 1. Ownership is narrow
  2. 2. The gains concentrate at the platform
  3. 3. What the evidence does not show
  4. 4. What happens to pooled claims
  5. 5. What we will not cite
  6. 6. Measurement traps
  7. 7. The long horizon

A draft law owes its readers three things: the evidence for its premise, the evidence against it, and the numbers it refuses to use. This page carries all three. Every figure was verified at source, not relayed from commentary; assembled 18 August 2026, extended 19 August 2026, corrections welcome by pull request.

The Regulation's premise is stated in its own recitals and tested by its own Article 14: hyper-automated production decouples output from labour at the level of the firm, the gains concentrate in whoever owns the platform, and the ownership of productive capital in Europe is too narrow for those gains to reach citizens by themselves. Each section below maps to a part of that premise, and section 3 records what the European evidence does not show, because the instrument is designed around that honesty.

1. Ownership is narrow

The condition the instrument answers: Europeans own homes and pension promises, not productive capital.

  • In the euro area, 60 % of households own their home and 11 % own shares in a company. Belgium 67 against 13. Italy 75 against 4. Greece 68 against 1. ECB Household Finance and Consumption Survey, wave 5, 2023 reference year, published June 2026. Same file, same population, same survey.
  • The richest tenth of euro-area households own 83 % of directly held shares; the poorest half own 2 %. Italy: 93 against 0,5. ECB Distributional Wealth Accounts, 2025-Q4. Caveat to state: every DWA observation is an ECB estimate.
  • The bottom half of euro-area households hold 63 % of their assets as housing and 3 % as equity. Belgium: 76 % housing, 2 % equity. ECB DWA, 2025-Q4. The premise as one row of a table.
  • The typical Dutch household is worth EUR 135 500, or EUR 23 000 once the house is taken away. CBS 83834NED, 1 January 2024, tax registers, not a survey. State the pension exclusion first.
  • Belgian households have accrued EUR 1 740 bn of pension claims and own EUR 136 bn of it. About 92 % is a promise from future taxpayers. OECD/Eurostat Table 29, 2021. Netherlands 49 % funded; France 0,0 %.
  • Danish households hold pension assets worth 206 % of national output; German households hold 6 %. OECD, 2024. Caveat: the German figure excludes Direktzusage book reserves.

2. The gains concentrate at the platform

Why Article 3 designates what it designates: where production is automated, the price that holds is the licence price, and the volume that falls is the hours.

  • European cloud companies tripled their revenue in seven years and their share of their own home market still fell from 29 % to 15 %, because the market grew sixfold. Amazon, Microsoft and Google take 70 % of a EUR 61 bn European market; SAP and Deutsche Telekom hold about 2 % each. Synergy Research Group, 24 July 2025. Completed revenue history, not a forecast.
  • France 2025: services firms -1,8 %, technology consulting -2,5 %, software and cloud +8,2 %, in the trade body's own words "surtout de la hausse des tarifs", with budget increases "absorbée par les évolutions des plateformes cloud". Numeum/PAC, December 2025.
  • Germany 2026: IT services +3,1 %, software +9,9 %, cloud software +21,9 %, AI platforms +75,8 %. Bitkom/IDC, August 2026.
  • Germany produces a third more software and IT services in real terms than five years ago, with the ICT vacancy rate down from 6,3 % to 2,5 %. Production volume index 99,2 (2021-Q1) to 134,0 (2026-Q1). Eurostat, verified against the API.
  • In Germany and France the whole job market is back at pre-pandemic levels while software-developer postings sit at half. Germany 50,6 against 104,5; France 51,1 against 100,0 (February 2020 = 100). Indeed Hiring Lab, 7 August 2026. The control group is built in.
  • Hays Germany: fees +3 % while volumes fell 9 %. Hays plc H1 and Q4 FY26, audited and listed. Belgium: project-sourcing revenue +0,4 % while billable hours fell 3,4 % (Federgon). Netherlands: the seconded rate rose 3,3 % to EUR 71,32 while seconded FTE fell 9,3 % (VvDN, self-reported trade body).

3. What the evidence does not show

The instrument is built around three facts that cut against the loudest versions of the automation story, and it does not need those versions.

  • There is no European labour-share collapse. The EU27 labour share moved about one point in thirty years; Germany's is at a series high; the Dutch series (81,4 % in 1995 to 70,6 % in 2025) has risen three years running. Anyone resting a law on a collapsing labour share loses the argument in Europe, so this law does not.
  • Aggregate displacement is not yet demonstrable. The EU27 employment rate is at a record 76,3 %; ICT specialists rose from 3,5 % to 5,0 % of employment; compensation of employees was 48,09 % of EU27 GDP in 2025, above 2015, 2019 and 2022. Danish register data on ~25 000 workers finds precise nulls two years after ChatGPT (Humlum and Vestergaard); the German federal government formally finds "keine Hinweise" of junior displacement (Bundestag Drucksache 21/3722, 19 January 2026). Same technology, opposite choices: Publicis added roughly 5 800 staff in the year WPP shed 9 389.
  • The mechanism has not diffused far. Only 19,95 % of EU firms with ten or more employees used any AI in 2025. The serious claim is a forecast about a mechanism, not a report about the past.

This is why the Regulation claims nothing in the past tense. Designation under Article 3 requires demonstrated decoupling at the level of the firm, not an economy-wide story; the warrant under Article 5 takes nothing until gains actually crystallise; and Article 14(3) carries the falsification condition on its face: if automation spreads without the decoupling and concentration the premise asserts, the Commission must report that the premise is unsupported and propose amendment or repeal. A law founded on an empirical claim should carry its own test.

4. What happens to pooled claims

Why Chapter VI of the Regulation looks the way it does: the history of pooled public assets in Europe is a history of raids, and the history of individual claims is better.

  • Poland, one line of law: 51,5 % of the money in every citizen's pension account cancelled on a named day. Act of 6 December 2013, Dz.U. 2013 poz. 1717, Article 23(1). Article 23(2) took the Treasury bonds first, so the state could retire its own debt. The raid is not a hypothesis; it is a statute, and Article 12 of this Regulation is drafted against its enumerated verbs.
  • Spain's pension reserve fund fell 97 %, from EUR 66,8 bn (2011) to EUR 2,1 bn, cumulative withdrawals EUR 80,3 bn. The law protecting it was passed in 2023, after it was empty.
  • Ireland's National Pensions Reserve Fund stands, in real terms, at about 99 % of its 2007 position eighteen years on, because it was liquidated into the bank rescue. Hungary's -75,9 % is the other benchmark. These are the strongest cases against this Regulation's own mechanism, and they are why the Reserve's protections are operative articles rather than promises.
  • Estonia proves a claim is necessary but not sufficient. When the second pillar went voluntary in January 2021, 149 083 people took out EUR 1,32 bn in a single month; leavers averaged age 41 with EUR 8 468, earning slightly below the national average. That is why the entitlement in Article 10 is incapable of surrender: a claim that can be bought out of a citizen's hands will be, cheapest first.
  • A Danish worker's frozen 1978 pay rise of DKK 4 368 is worth DKK 119 506 tax-free today, about 27 times over 46 years, in an account with the worker's name on it. Lønmodtagernes Dyrtidsfond, at 31 December 2025. Deferred income converted into owned capital: the closest existing instrument to what this Regulation builds.
  • Two-thirds of Norway's fund was never oil: NOK 15 210 bn of NOK 22 683 bn is compound return. The answer to "we have no oil" is that Norway largely does not either, any more; it has a rule, and Annex II is that rule made law.
  • Across Europe's development and reserve funds, citizens hold no individual claim at all: Invest-NL, the Nationaal Groeifonds, COFIDES, SEPI, Spain's FRSS, Portugal's BPF and FEFSS, Greece's Growth Fund, Finland's Solidium, Bpifrance, Italy's CDP, Slovenia's SDH, Poland's PFR, Ireland's FIF, ICNF and ISIF. No account, no unit, no inheritable right. Individual claims exist only in pension vehicles, which is the gap Article 10 closes.

5. What we will not cite

Numbers that circulate in this debate, checked and retired, including some that would have helped us.

  • "The OECD says only 9 % of jobs are automatable." Superseded by Nedelkoska and Quintini (WP 202, 2018): 14 % above 70 % probability, another 32 % between 50 and 70.
  • "Bruegel found AI exposure correlated with employment growth." Misattributed (it is OECD WP 265, and equivocal). Bruegel's own European result is negative: one extra robot per thousand workers cuts the employment rate 0,16 to 0,20 points.
  • European cloud share "27 % to 13 %". Wrong; it is 29 % to 15 %.
  • Klarna as "700 people replaced by AI". A hiring freeze plus 15 to 20 % annual attrition; "700 agents" was a chat-volume calculation.
  • "SAP cut 10 000 jobs for AI." Headcount rose in both years.
  • The 160 000 telecom job losses as an AI story. The sector's own assessment: "this dramatic shrinkage owes very little to AI"; the same operators shed over 380 000 between 2014 and 2022.
  • Job-board causation claims (entry-level postings "down 32 % since ChatGPT" and similar): windows containing the entire rate-tightening cycle, asserted as AI effects by parties selling recruitment products.

6. Measurement traps

  • Three defensible labour-share measures differ by up to 19 points. Italy 2024: unadjusted 39,4 %, adjusted 58,3 %. Never quote one without naming it.
  • WID, HFCS and DWA figures must never share a table: different units and methods. The Dutch top-10 % share is 44,8, 52,0 or 56,1 depending on which you pick.
  • Every major European household-wealth statistic excludes pensions, which inverts the ranking and makes the Netherlands and Denmark look asset-poor.
  • Norway's wage share is a gas-price denominator effect; do not use it. Germany has a hard series break at 1991; never splice.
  • Eurostat ICT vacancy data stop at 2025-Q4 while price and employment series run into 2026; the legs are not aligned in time.

7. The long horizon

The Reserve is an instrument built for a horizon politics rarely rewards, so this page states plainly what fifty years do, and what they do not.

  • Alaska has paid every resident a dividend for 44 consecutive years. The Permanent Fund holds more than USD 86 bn for about 740 000 residents, roughly USD 116 000 of fund capital per resident; a resident collecting every dividend from 1982 to 2016 received USD 37 027 in total. Alaska Permanent Fund Corporation; Alaska Department of Revenue, dividend table. No legislature has touched the principal in half a century, because every citizen had a reason to defend it.
  • Alaska also shows the failure mode. Since 2016 the dividend has been set by annual political bargain; the statutory formula remains in law and is simply not followed, and the 2025 dividend of USD 1 000 is the smallest in the programme's history once adjusted for inflation. A distribution rule in ordinary statute is capturable by a budget fight. That is why the rule here is Annex II of the Regulation itself, why the entitlement is an operative article, and why the Commission reports on the formula rather than sets it.
  • The AI-value forecasts stay out of the premise. PwC puts artificial intelligence at USD 15,7 tn of global product by 2030; Goldman Sachs at about USD 7 tn over a decade; McKinsey at USD 2,6 to 4,4 tn a year. These are forecasts of a mechanism, with error bars measured in trillions, and section 5's discipline applies: this Regulation does not rest on them. They appear in one place only, as labelled scenario inputs to the public simulator, beside a sceptic setting that assumes almost nothing.
  • What the arithmetic of Annex II does with fifty years. Run cautiously (EUR 150 bn of covered revenue at designated firms, Norway-class real returns), the simulator's central curve reaches a stake of roughly EUR 1 200 of Reserve capital per citizen and a dividend of roughly EUR 47 a year by year fifty; run at the forecasts above, roughly EUR 10 000 and EUR 370, in constant 2026 euros: a fund on the scale Norway's rule built. Both curves spend their first decade indistinguishable from zero. That is the design: Annex II preserves capital before it distributes, and the stake compounds while the payout waits. If neither curve materialises, Article 14(3) obliges the Commission to report that the premise failed and to propose amendment or repeal.

Working notes for the three research strands sit in this directory; they carry the full sourcing trail. Dispute a number by opening an issue or a pull request: a figure that does not survive contact with its source comes off this page and goes into section 5.