The Enablement–Regulation Axis

Sources#

• Meeting the Coming Wave: The Emerging Politics of AI and Work across 33 Parliaments

Summary#

Every other page in this domain measures what AI does to work — exposure, headcount, task composition, who gets displaced. This page holds the political supply side: what elected legislators actually say about it, at the scale of a whole corpus of parliamentary speech, and what shape the resulting conflict has.

The finding is a negative one about the literature's central expectation. Comparative political economy treats technological change as a shock that creates losers and then a politics of compensating them — social insurance, redistribution, retraining. In 1,514,950 substantive speeches across 33 national parliaments (January 2023 – April 2026), compensation accounts for 2.3 percent of broad response-frame mentions. Enablement and investment take 55.2 percent, regulation and restriction 21.8 percent, training 20.6 percent. The authors' term for what replaces the compensation axis is the enablement–regulation axis: contestation over whether public authority should accelerate and invest in AI's adoption or condition and constrain it — argued out before the distributive consequences are settled, which is the part the compensation literature is not built to observe.

Evidence note. empirical — a coded speech corpus with a published human-validation battery, which is the strongest tier this genre reaches. Two qualifications travel with every number below. (i) Speech is not policy: the paper's own first limitation is that "parliamentary speech does not directly measure enacted policy," and it makes no claim about legislation passed. Everything here is discourse composition. (ii) "The coming wave," the politics of adoption, and the enablement–regulation axis are the authors' framing, not measurements — the measurement is the frame distribution; the claim that this constitutes a new and durable cleavage is interpretation, and the paper says so ("the emerging structure may change rapidly").

The funnel, and which denominator each number sits on#

The corpus is assembled from official parliamentary sources — APIs, XML feeds, Hansard-style transcripts, PDF publications — normalized to one row per substantive intervention, lower house or unicameral chamber only, procedural and chair turns excluded. Getting the denominators right matters because three different ones are in play:

Quantity	Count
Substantive speeches (all parliaments, 2023–Apr 2026)	1,514,950
AI candidate speeches (multilingual dictionary retrieval)	19,411
AI-work relevant speeches (LLM classification)	5,317
— of which directly relevant	3,076
Country cases	33 national parliaments
Country-party units	936
Human-validation observations	728

The 1.5M figure is the retrieval denominator, not the analyzed sample: dictionary retrieval deliberately over-recalls (19,411 candidates, "including many AI speeches unrelated to work"), and the LLM classifier then cuts to 5,317. Every frame share on this page is computed over those 5,317 (or over the mentions they generate), not over 1.5M. The 33 countries are named in prose, not buried in a figure: Argentina, Australia, Austria, Belgium, Brazil, Canada, Chile, Croatia, Czechia, Denmark, Estonia, Finland, France, Germany, India, Ireland, Italy, Japan, Kenya, Latvia, the Netherlands, New Zealand, Norway, Poland, Singapore, Slovenia, South Africa, Spain, Sweden, Switzerland, Taiwan, the United Kingdom, the United States. It is explicitly not a probability sample — inclusion required recoverable speech-level official records.

Party families are identified for 94.7% of AI-work speeches; 85.5% fall inside the seven families analyzed (radical right, conservative, liberal, Christian democratic, social democratic, green, radical left). Parties that could not be mapped confidently were excluded rather than forced.

The validation, and the one stage that has none#

This is an LLM-coded corpus, so the classifier is the instrument — the concern Usage-Telemetry Classifier Validation exists to price. Unlike the usage-telemetry programs, this one publishes its numbers, and they are good:

• Coder: gemini-3.1-pro-preview, zero-shot, temperature 0, JSON structured output, accessed 6–7 July 2026. The annotator sees original-language speech text and the codebook, and is blind to party family, government status, and the hypotheses — the same anchoring discipline Usage-Telemetry Classifier Validation finds missing elsewhere.
• Primary benchmark (538 speeches, lead human coder): AI-work relevance at 97.8% accuracy, Cohen's κ = 0.947, 96.9% precision, 95.7% recall. Among jointly-relevant speeches, diagnostic coding micro-F1 0.893, response coding micro-F1 0.878.
• Adversarial stress test (190 speeches, second coder, enriched for boundary cases and rare responses): relevance 85.8% accuracy, κ = 0.679, micro-F1 0.894. Among the 114 jointly-relevant speeches, direct/indirect 96.5% (κ = 0.916), diagnosis micro-F1 0.891, response micro-F1 0.828.

The honest reading of the pair: κ = 0.947 against one in-house coder, κ = 0.679 against a second coder on the hard cases. The paper is explicit that the stress test "is not intended to estimate corpus-wide accuracy," so it is not a corrected headline — but it is the closest thing here to the blind-vs-anchored gap that Usage-Telemetry Classifier Validation shows separates an approval rate from an accuracy rate, and the response frames (0.828) degrade more than the diagnostic ones (0.891) exactly where the rare category lives.

The unvalidated stage is retrieval. Precision is measured at classification; recall is measured within the candidate set. Nothing in the validation design estimates how many AI-work speeches the multilingual dictionaries never surfaced — a false negative at stage one cannot be recovered at stage two, and the paper says so as a design rationale rather than as an error bar. Since the dictionaries cover "artificial intelligence, automation, and related algorithmic technologies," a speech about, say, works-council rights over automated scheduling that never names a technology is invisible by construction, and that omission would fall disproportionately on the regulation frame.

The distribution, and why it reconstructs exactly#

The headline response mix, over all broad response-frame mentions (primary and secondary tags counted):

Response frame	Share	Mentions
Enablement and investment	55.2%	3,022
Regulation and restriction	21.8%	1,194
Training	20.6%	1,129
Compensation	2.3%	125

The mention counts come from Figure 11's printed topic counts and are not stated in prose; they sum to 5,470 and reproduce all four published shares to the decimal (3,022/5,470 = 55.25%, 1,194/5,470 = 21.83%, 1,129/5,470 = 20.64%, 125/5,470 = 2.29%). That arithmetic closure is the reason this page cites them: the figure and the prose are the same measurement seen twice.

Compensation's absolute size is the part the percentage hides. 125 mentions, across the entire 33-parliament corpus, split 80 income-support / 45 reduced-working-time. The definition is deliberately broad — income protection, social protection, redistribution, reduced working time, basic income — and it still finds almost nothing. The paper's substantive point is sharper than the rarity: what compensation politics exists is also thin, "largely restating conventional welfare-state instruments rather than developing AI-specific remedies such as robot taxes or productivity-sharing schemes." The instruments the abundance case assumes will be legislated have, in three years of speech across 33 legislatures, 125 mentions.

The ordering survives two robustness cuts: a direct-only sample (compensation "rises somewhat but remains a small minority response") and dropping Singapore and Taiwan, the two highest-salience cases.

Emergence: fast, small, and very uneven#

Rising salience, low absolute level:

• AI attention (AI candidate speeches / substantive speeches): 1.06% in 2023 → 1.70% in 2026.
• Labor within AI (AI-work / AI candidate speeches): 18.0% → 33.8%. Parliaments are not just talking about AI more, they are increasingly attaching it to work.
• AI-work salience (AI-work / substantive speeches) stays well below one percent even in 2026. Per-year AI-work speech counts printed on Figure 1 — 813 / 1,502 / 2,251 / 751 — sum to exactly 5,317, so the trend and the corpus total reconcile. The 2026 value covers January through April only, which is what makes 751 look like a drop and is not one.

Cross-country dispersion is an order of magnitude wider than the trend. Figure 2's printed data labels: Taiwan 4.24%, Singapore 4.22% — the two outliers, both with explicit national AI-workforce strategies — then India 1.70%, Germany 1.30%, Norway 0.99%, Italy 0.84%, Finland 0.82%, Austria 0.71%, down through Spain 0.55%, Japan 0.53%, Sweden 0.52%, the UK 0.50%, and the United States at 0.35%, below Australia and Canada. The US supplies the paper's opening anecdote (the Sanders 100-million-jobs report, the Sanders–Ocasio-Cortez data-center moratorium bill) and one of the least AI-work-salient legislatures in the sample, with the most opportunity-tilted diagnostic profile of the 28 countries that clear the Figure 3 thresholds. Salience and framing are independent.

The single dimension#

Party families do not differ much in whether they raise AI and work. Predicted AI-work salience ranges from ≈0.25% (radical left) to ≈0.60% (liberal) with heavily overlapping 95% intervals — "no family clearly owns the issue in terms of attention." They differ enormously in what they say once they do.

Predicted shares from grouped-binomial multilevel models (party-family and year fixed effects, country and country-party random intercepts). Coding is multi-label, so rows need not sum to 100%:

Party family	Threat	Opportunity	Regulation	Enablement
Radical left	≈89%	≈31%	≈67%	≈25%
Green	≈66%	≈53%	≈45%	≈42%
Social democratic	≈57%	≈60%	≈36%	≈56%
Christian democratic	≈37%	≈81%	≈27%	≈68%
Liberal	≈31%	≈86%	≈17%	≈73%
Conservative	≈32%	≈87%	≈16%	≈77%
Radical right	≈39%	≈79%	≈16%	≈76%

These are point estimates read off plotted margins, so they carry a read-off error of a point or two; they are cross-checked arithmetically against Figure 9, which plots the same families on (threat − opportunity, regulation − enablement) difference coordinates and lands on +57/+43 for the radical left, +13/+3.5 green, −3/−20 social democratic, −44/−40 Christian democratic, −55/−56 liberal, −55/−60 conservative, −40/−60 radical right. The prose ranges corroborate the ends ("roughly nine in ten" radical-left threat, "fewer than one in three" radical-left opportunity, "roughly 80–86 percent" opportunity for the right bloc, threat "around one-third"). The result that matters is not any cell but that both differences order the families identically — one dimension, not two.

Three specific readings:

• The radical left is the only consistent contestatory force: the only family combining an overwhelming threat diagnosis with a predominantly governance-oriented response. And even there, regulation (≈67%) far outruns compensation (≈19%, on the widest interval of any left estimate) — the contestation is over deployment, not over transfers.
• Greens split evenly between regulation (≈45%) and enablement (≈42%), which the paper attributes to an ecological-modernization current that makes selective enablement available to them in a way it is not to the radical left.
• Social democracy is the adoption-oriented left. Threat and opportunity at roughly equal rates, and an aggregate response still on the enablement side (≈56% vs ≈36%). The paper reads this as the familiar social-democratic posture toward structural economic change — accept the transformation, manage the transition, reserve regulation for the worst consequences — with the consequence that "the strongest parliamentary challenge" on AI belongs to the radical left rather than to the historically dominant party of organized labor.

Diagnosis structures response, and that is where the conflict actually sits. Splitting by primary diagnosis (Figure 6, mutually exclusive primary frames), threat-diagnosing speeches lean heavily to regulation and opportunity-diagnosing speeches overwhelmingly to enablement, within every family — the social-democratic middle is not a moderate position but two substantial streams averaged. The n's are the useful part: opportunity-diagnosing speeches outnumber threat-diagnosing ones roughly 3:1 among conservatives (1,121 vs 337) and liberals (733 vs 199), run 744 vs 407 for social democrats, and invert only for the radical left (54 vs 89) and tie for greens (77 vs 77). Which consequences get made visible is most of the fight.

The radical-right null#

The cleanest negative result in the paper, and the one with an established literature to contradict. Workers exposed to automation risk are repeatedly found to shift toward radical-right parties (Anelli et al. 2021; Gallego & Kurer 2022; Im et al. 2019; Knotz 2025). If party supply followed electorate exposure, the radical right would be the entrepreneur of AI-related labor anxiety.

It is not. Radical-right parties sit with the mainstream right on both axes — opportunity ≈79%, enablement ≈76%, regulation ≈16%, threat only modestly above the conservative and liberal level (≈39% vs ≈32%/≈31%). Figure 9 places them next to liberals and conservatives in the adoption-oriented corner. The authors' explanation is repertoire, not exposure: national competitiveness, technological sovereignty and the danger of falling behind are frames the radical right already owns, while institutionalized workplace regulation and worker voice fit its repertoire badly. The electoral link to automation losers does not translate into a party-level position.

This is an association, not a test — the design cannot rule out that the exposure→vote link runs through frames the coding scheme does not capture (immigration, national decline, elite betrayal), which is precisely the channel the cited literature proposes.

The institutional gradient#

Government status adds a second, smaller gradient, estimated with binomial GLMs carrying government status plus party-family, year and country fixed effects. Government-minus-opposition margins, in percentage points, with 95% intervals that exclude zero except where noted:

• Opportunity diagnosis ≈ +15 pp; threat diagnosis ≈ −15 pp
• Enablement and investment ≈ +11 pp; regulation and restriction ≈ −8 pp
• Labor within AI ≈ +4 pp
• Training ≈ −1 pp, interval spanning zero — the portable frame is portable across institutional position too
• AI attention and raw AI-work salience: ≈ 0, indistinguishable

So governing parties are not more interested in AI and work; they are more likely to construct it as a governable opportunity once they raise it. The paper reads this descriptively and refuses the causal claim — confidence-and-supply parties and unattributable ministerial speeches are dropped, and "the analysis does not identify the causal effect of entering office." Note the direction relative to party family: the government effect (≈15 pp on diagnosis) is real but small against the family spread (≈89% to ≈31% threat), so ideology remains the organizing force, as E5 predicted it would.

Inside the frames: where regulation is actually pointed#

The exploratory layer — k-means over embedded English summaries of the coded speeches, human-labelled from TF-IDF terms and centroids, stable across seeds (mean adjusted Rand 0.889–0.928). The paper is emphatic that these are "interpretive aids, not additional validated outcome categories," and they enter no model. Read as composition, not as measurement:

Frame	Leading topics (share within frame, mentions)
Threat	Job displacement and income insecurity 53.1% (937); job quality and workplace control 19.5% (345); falling behind in the AI economy 14.8% (262); skills-transition and education gaps 12.5% (221)
Opportunity	Competitiveness and innovation 37.8% (1,531); productivity, job creation, labor-shortage relief 34.9% (1,413); AI infrastructure and energy 15.1% (612); skills and talent formation 12.2% (494)
Regulation / restriction	Copyright, creative-sector and media protection 36.1% (431); labor-market safeguards against displacement 27.2% (325); innovation-compatible AI regulation 23.7% (283); algorithmic management, platform-work safeguards and human oversight 13.0% (155)
Enablement / investment	Growth and national investment 39.1% (1,181); adoption support, productivity and public-sector capacity 31.3% (947); AI infrastructure and industrial strategy 17.2% (519); SME and business adoption support 12.4% (375)
Training	Workforce reskilling and employee training 45.6% (515); AI talent and research capacity 28.9% (326); schools and future skills 25.5% (288)
Compensation	Income support and safety nets 64.0% (80); reduced working time 36.0% (45)

Three things fall out of this table that the frame-level headline conceals:

• Regulation is currently more about copyright than about work. The largest single component of the regulation-and-restriction frame is intellectual property and creative labor — "a legacy concern … that generative AI has revived rather than invented" — and the topic closest to the surveillance-and-control concerns raised on the threat side is the smallest, at 155 mentions across 33 parliaments. The threat frame's second-largest topic is job quality and workplace control (345), so the diagnosis is roughly twice as loud as the remedy attached to it.
• Enablement is a public-investment frame at least as much as a business-support one. Growth/national investment plus public-sector capacity are seven in ten enablement mentions; SME and firm-level adoption support is 12.4%. This is the paper's explanation for why social democrats and greens register substantial enablement despite threat-leaning diagnoses — the frame is ideologically available to the left, because it is about what the state builds.
• Training's portability is why it is flat across families. Reskilling, national talent capacity and school curricula serve adoption, adaptation and long-run human capital respectively; a party can invoke it whatever it wants from the axis, which is exactly what makes it useless as a positional signal.

What this can and cannot support#

Stated in the paper and worth carrying intact:

• Speech is not enacted policy. Policy formation "also depends on coalition bargaining, veto institutions, bureaucracies, courts, unions, employers, and organized interests." No legislative outcome is measured anywhere in the study.
• Associational throughout. Party family, government status: "the models map associations rather than causal effects of ideology or government participation."
• The sample is availability-determined, and "differences in parliamentary institutions and transcript practices warrant caution in interpreting national rankings" — which bears directly on the Taiwan/Singapore outliers and on the US's low placement.
• An unusually early window. 2023–Apr 2026 is chosen to catch conflict before labor-market effects stabilize, which is also why the structure "may change rapidly": the paper names visible displacement strengthening compensation politics, deepening left contestation, or opening radical-right mobilization as three live futures.
• The representation gap is raised and explicitly not established. Demand-side work finds citizen support for slowing technological change and work-preserving intervention; that support has little parliamentary counterpart here. The paper refuses the inference — demand and supply studies "observe different populations, countries, and periods" — and offers only that public unease "is outpacing the positions parties have so far built around it." Treat the gap as a hypothesis with one side measured.

Connections#

• Procedural Value in AI Decisions — the two halves of one representation gap, measured independently and pointing the same way. Wang, Sturgis & de Kadt price what US job seekers want from an AI hiring decision: human decision authority +0.272 in choice probability, an appeal they can invoke +0.156, an opt-out +0.129, a system-level bias audit only +0.068. This page measures what legislatures supply, and the topic containing those instruments — algorithmic management, platform-work safeguards and human oversight — is the smallest component of the smallest-but-one response frame, 155 mentions of 5,470. The demand side ranks individually-invocable rights at the top; the supply side spends its regulation attention on copyright (431). Neither study can establish the gap alone (different populations, countries, periods, and one is stated preference), but the pair is as close as the vault gets to both ends of it
• Human-AI Accountability Redesign — the same regulatory content, one institutional level down, and the pairing shows which level is actually doing the work. That page's German co-determination section documents an accountability mechanism that is statutory, workforce-held and veto-strength (Works Constitution Act §87(1) no. 6) — and firms evading it by offshoring HR functions and buying capabilities that do not trigger it. This page shows what the parliamentary layer above it is doing meanwhile: algorithmic management and human oversight at 13.0% of a 21.8% frame. Where the workplace-governance content exists at all it is largely already-enacted national institutions being exercised, not new legislative demand being built — which makes the evasion channels that page documents harder to close from the top
• Post-Scarcity Macroeconomics — the political-feasibility check on the abundance transition. Musk's argument makes universal transfers non-inflationary in the steady state and concedes he cannot describe the path; the interviewer's question is how a polity gets there. This is the supply side of that path measured: universal basic income, redistribution and income protection together are 2.3% of response mentions, 125 in absolute terms, and what exists "largely restat[es] conventional welfare-state instruments." Whatever the deflation mechanism licenses economically, the constituency that would legislate it has not formed in three years of speech across 33 parliaments — and the paper's own explanation is the one the transition argument needs to answer: stable constituencies of AI losers may not exist at scale yet, so the politics is running upstream of the disruption instead
• The Tragedy of the Cognitive Commons — the public-policy rung of Lovett's Ostrom ladder, priced. That framework proposes training subsidies, credits for firms preserving entry-level pipelines, tiered credentials and AI-free competence assessment, explicitly as investment rather than prohibition. The training frame is 20.6% of mentions and — uniquely — invoked at similar rates across every party family, so the rung Lovett reaches for is the politically cheapest one available and the least positionally informative. The rung his argument actually needs is different: a credit for preserving junior work is a restriction on labor-replacing adoption, which lives in the 21.8% regulation frame, is concentrated in the radical left and greens, and inside that frame is outweighed three to one by copyright
• Usage-Telemetry Classifier Validation — a counter-example on validation practice, from outside the usage-economics genre. Where the AEI, ATLAS and the Codex study run taxonomy classifiers whose exact accuracy went unpublished (and in ATLAS's case is 22.6% at O*NET task level), this program publishes κ = 0.947 on relevance, micro-F1 0.893/0.878 on frames, and then commissions a second coder for an adversarial stress test that drops relevance to κ = 0.679 — reporting both. The discipline that page asks for (blind annotator, published κ, a hard-case ceiling check) is achievable on an LLM-coded corpus. It also inherits that page's structural warning one stage earlier: the retrieval dictionary has no measured recall, so the validated classifier sits on an unvalidated candidate set
• Telemetry vs. Survey Measurement — a ninth aperture: the elite-discourse text corpus. Not telemetry (nobody's behavior is logged), not a survey (nobody is asked), not a conjoint (nothing is randomized) — it is a near-census of what a specific, consequential population said on the public record, with the one property none of the other instruments has: the speaker had an institutional incentive to be on record. It buys latency and completeness on the supply side of policy and pays for it in construct — a speech measures positioning, not preference, not behavior, and certainly not enactment. Its definitional-sensitivity exposure is the retrieval dictionary rather than a threshold: what counts as an "AI and work" speech is a boundary the pipeline draws and never varies

Open Questions#

• The classifier is validated at stage two and unvalidated at stage one: precision and recall are both computed inside the 19,411 dictionary-retrieved candidates, so no number bounds how many AI-work speeches the multilingual dictionaries never surfaced. Does a recall audit — LLM-coding a random sample of the ~1.5M non-retrieved speeches, or a second retrieval pass with an orthogonal (embedding-based) recall stage — change the response mix, and specifically does it raise regulation, whose workplace-governance content is most likely to be phrased without naming a technology?
• Parliamentary speech and enacted policy are measured on different objects, and this design measures only the first. Does the enablement–regulation axis predict legislative output — do the countries and periods whose speech tilts toward regulation actually pass more workplace-AI statute than the enablement-tilted ones, net of party composition? Falsifiable against any coded AI-policy output dataset covering the same 33 countries over 2023–2026.
• Compensation's absence is currently explained by the absence of stable AI-loser constituencies, which is a prediction: if measurable displacement arrives, compensation's 2.3% share should rise and the axis should rotate back toward the compensation politics the literature expected. Trigger: a country in this sample with a documented AI-attributable employment decline, re-coded on the same pipeline.

Sources#

• Meeting the Coming Wave: The Emerging Politics of AI and Work across 33 Parliaments — Juliana Chueri (Department of Political Science and Public Administration, Vrije Universiteit Amsterdam) & Petter Törnberg (Institute for Logic, Language and Computation, University of Amsterdam), Meeting the Coming Wave: The Emerging Politics of AI and Work across 33 Parliaments, arXiv 2609.02296, 2026-09-02 (32pp, 3 tables, 11 figures; empirical). §2 the four-response issue space and the five expectations E1–E5; §3 corpus construction, the 33-country list, party-family harmonization (936 country-party units, 94.7%/85.5% coverage), the two-stage retrieval-then-classification design, the gemini-3.1-pro-preview annotation protocol, the two human-validation designs, the exploratory clustering procedure, and the grouped-binomial multilevel and GLM specifications; §4.1 the salience trends and country dispersion; §4.2 diagnostic margins; §4.3 the 55.2/21.8/20.6/2.3 distribution and the enablement–regulation axis; §4.4 government–opposition; §4.5 the exploratory topic composition; §5 limitations and the unestablished representation gap. Data and code promised in an open repository "upon publication"; no competing-interests or funding statement appears in the preprint — undisclosed rather than declared absent. Törnberg is also the author of two of the methods citations the annotation protocol rests on (Törnberg 2024, 2025), which is self-citation on method, not on result.
• Parse notes — all three tables reconciled against pdftotext -layout; one repaired cell confirmed. The raw carries exactly one > [!note] repair block, on Table 2, where docling dropped the leading 33 from the "Country cases" cell (33 national parliaments → national parliaments). The restoration is correct: pdftotext -layout p.13 reads Country cases 33 national parliaments. Tables 1 (response frames × political questions) and 3 (illustrative claims) carry no repair block and were therefore treated as unrepaired and checked cell-by-cell; both are byte-equivalent to the reference parse, including the hyphenation artifacts from wrapped labels (Regulation and re- striction, enablement invest- ment). The ingest warn on table-collapse was a false positive on the 2023-April 2026 date-range cell. One known en-dash/OCR corruption in the body text: the §4.1 heading renders as ## Al attention and labor salience rise over time (Al for AI); nothing is cited from it. Nothing on this page is cited from Table 3 — it is illustrative paraphrase, not data.
• Figures viewed and reconciled — and this paper's load-bearing numbers live in figures, so the provenance of each is recorded. All 11 images in were viewed; figure↔image order is 1:1 (image_000000 → Figure 1 … image_000010 → Figure 11). Figure 11 is the exception that needs no chart-reading: its topic labels, counts and percentages sit in the PDF text layer and were recovered with pdftotext -layout, and its counts reproduce every published share to the decimal (threat 937/345/262/221 of 1,765; opportunity 1,531/1,413/612/494 of 4,050; regulation 431/325/283/155 of 1,194; enablement 1,181/947/519/375 of 3,022; training 515/326/288 of 1,129; compensation 80/45 of 125) and the four headline response shares (3,022+1,194+1,129+125 = 5,470 mentions → 55.25/21.83/20.64/2.29%), plus the prose's "just 125 mentions." Figure 7 prints 21.8/55.2/20.6/2.3 as data labels matching the prose exactly. Figure 1 prints per-year n's whose AI-work series (813/1,502/2,251/751) sums to exactly 5,317; its AI-candidate series was not cited because the rendered 2024 label reads ambiguously (5,680 vs 5,880) and only 5,680 reconciles to the published 19,411 — an OCR-grade ambiguity caught by arithmetic, and the reason no per-year candidate count appears on this page. Figure 2a's country values are printed data labels read from the image. Figures 4, 5, 8 and 10 are dot-and-interval plots with no printed values; every ≈ estimate on this page is read from plotted points, and all of Figures 5 and 8 were cross-validated against Figure 9, which plots the same families on difference coordinates — every one of the seven families' (threat − opportunity, regulation − enablement) pairs matches the read-off margins within ~1 pp. Figure 6's per-family n's (threat 89/77/407/54/199/337/56; opportunity 54/77/744/215/733/1,121/162) are printed labels; its stacked-bar proportions are not cited. Figure 3 carries 28 country rows with printed n's — 28 of the 33 parliaments clear its ≥25-diagnoses/≥20-responses threshold — and only its qualitative ordering, which the prose states independently, is used.