Source notes for Reassessing Recursive Displacement During the Infrastructure Boom
Preview for Tyler Maddox • 14 October 2026

These notes support research commentary and a proposed framework amendment. They distinguish reported observations, modeled exposure, projections and hypotheses. Source text was inspected in the research conducted on 13–14 October 2026; underlying microdata and corporate accounts were not independently reproduced. This is a selective synthesis, not a systematic review. Bracketed numbers correspond to the article.

[1] Daron Acemoglu and Pascual Restrepo, Automation and New Tasks, Journal of Economic Perspectives 33(2), Spring 2019.
https://www.aeaweb.org/articles?id=10.1257%2Fjep.33.2.3
The publisher abstract supports the conceptual distinction between displacement and new-task reinstatement. Its historical US analysis predates the present generative-AI deployment cycle. It does not establish current failure of reinstatement.

[2] Tyler Maddox, The Theory of Recursive Displacement, public V1.0.1, February 20, 2026.
https://recursive.institute/theory/
Public version inspected October 13. Part III, Loop 2, distinguishes labor-budget reallocation from productive task substitution; Part VII discusses physical limits and financial vulnerability. These are existing elements of Maddox’s framework. The amendment proposed here makes their intervening links and tests more explicit; it does not claim to originate those distinctions. The public version is not asserted to be identical to repository drafts.

[3] International Energy Agency, Key Questions on Energy and AI, April 16, 2026.
https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
Inspected October 13 and 14. Global energy-system analysis and projections, including a 2026 project pipeline and horizons through 2030/2035. Datacenter electricity totals include non-AI uses. Physical constraints and sensitivity to financing support separating announced investment, construction, commissioned capacity and productive use. Neither projected electricity demand nor construction starts provide a labor-loss measure.

[4] Ed Zitron, The AI Demand Bubble, August 4, 2026.
https://www.wheresyoured.at/the-ai-demand-bubble/
Inspected October 13. Primary source for Zitron’s criticism, not for audited corporate financial performance. Concentration claims rely partly on analyst estimates with different AI/cloud definitions and fiscal/calendar periods; the underlying analyst notes were not independently reproduced. The article attributes the argument and does not import its concentration estimates as established facts.

[5] Ed Zitron, Annualized Enron Rate, October 13, 2026.
https://www.wheresyoured.at/annualized-enron-rate/
Inspected October 13. Primary source for the critic’s argument; underlying private-company financial claims rely partly on reporting and non-public information. The observation about maximum allowances versus average serving losses is this article’s methodological qualification, not a concession attributed to Zitron. Establishing losses requires actual consumption and cost data. Annualization is not inherently fraudulent.

[6] Amazon, OpenAI and Amazon announce strategic partnership, February 2026.
https://www.aboutamazon.com/news/aws/amazon-open-ai-strategic-partnership-investment
Inspected October 13. Corporate announcement of investment alongside a compute relationship. An announcement establishes stated commitments, not full disbursement or commissioned productive capacity. Investor/supplier/customer links warrant examination but do not establish absent external demand or misconduct.

[7] Microsoft, FY2026 Q4 earnings call, July 29, 2026; quarter ended June 30, 2026.
https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4
Inspected October 13. The $41 billion figure is quarterly capex; it is distinct from $35.8 billion in cash property/equipment payments. The call describes capacity constraints and wider customer demand, and explains guidance changes associated with lease classification without a change in underlying investment expectations. These are management disclosures across a mixed AI/non-AI business, not a standalone AI return-on-capital calculation.

[8] Office for National Statistics, Business insights and impact on the UK economy, October 8, 2026; BICS Wave 165.
https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/bulletins/businessinsightsandimpactontheukeconomy/8october2026
Questionnaire:
https://www.ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/conditionsanddiseases/articles/businessimpactofcovid19surveyquestions/21september2026to4october2026
Inspected October 13 and 14. Fieldwork September 21–October 4, 2026. The headcount denominator is AI-using businesses with at least ten employees in sampled private-sector industries. Six percent reported reduced overall headcount attributed to AI. The question asks what impact AI technologies have had and gives no explicit retrospective interval. Business shares are not worker shares or the magnitude of losses. The voluntary survey has coverage exclusions, including the public sector and finance/insurance, and geographically varying industry coverage. Self-reported attribution does not independently identify causation or measure forgone hiring.

[9] Anders Humlum and Emilie Vestergaard, Still Waters, Rapid Currents, March 13, 2026 working-paper revision; authors’ explanation October 6, 2026.
https://www.rfberlin.com/wp-content/uploads/2026/03/26078.pdf
https://www.brookings.edu/articles/still-waters-rapid-currents-early-labor-market-transformation-under-generative-ai/
Inspected October 13. Denmark: 25,000 workers in eleven exposed occupations, with administrative records through December 2024. The authors report null average earnings and recorded-hours effects and greater movement by adopters toward higher-paying, AI-relevant occupations. These findings cannot be relabeled as October 2026 outcomes, generalized to all workers, or treated as proof of durable causal earnings gains. The analysis was not replicated here.

[10] Stanford Digital Economy Lab and ADP Research, Canaries Dashboard, updated September 23, 2026, observations through August 2026; associated Canaries in the Coal Mine research revision, August 12, 2026, observations through June 2026.
https://digitaleconomy.stanford.edu/project/indicators/canaries-dashboard/
https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
Dashboard inspected October 13 and 14; revised research inspected October 14. Employment comparisons concern age and occupational-exposure groups within occupation-matched workers at balanced ADP firms. Exposure is not observed firm adoption. The revised research emphasizes reduced hiring and describes its indicators as descriptive rather than causal estimates. Education adjustments attenuate findings, and pretrends and weaker national-survey patterns matter. The sample is not nationally representative and excludes firm entry/exit. Dashboard and paper have different observation cutoffs; neither supplies a national AI-caused job-loss count.

[11] International Labour Organization, Disruption without dividend, Working Paper 166, March 17, 2026.
https://www.ilo.org/publications/disruption-without-dividend-how-digital-divide-and-task-differences-split
Inspected October 13. Modeled analysis covers 135 countries, with detailed task-survey comparisons for 46; data vintages differ. It supports attention to task composition and digital access across countries. Exposure estimates are not an October 2026 employment-loss census, and US task scores cannot simply be multiplied by other countries’ employment totals.

[12] US Census Bureau, The Microstructure of AI Diffusion, April 2026, Table 4, printed page 33.
https://www2.census.gov/library/working-papers/2026/adrm/ces/CES-WP-26-25.pdf
Inspected October 14. BTOS supplement fieldwork November 17, 2025–February 8, 2026; reference periods November–January, with employment effects over the preceding six months. Population: US private non-agricultural employer firms. The employment-effect denominator is AI-using firms, firm-weighted: 2.0% report a decrease, 2.3% an increase and 95.7% no change attributed to AI. This does not say actual employment was unchanged. Responses provide neither magnitudes nor gross hiring flows, and do not independently identify AI causation. The underlying AI-use question broadened in November 2025, limiting comparison with earlier production-only adoption series.

[13] Nine Raters, One Index, revision 4, August 6, 2026.
https://arxiv.org/html/2507.22748v4
Inspected October 14. UK study using British task-survey evidence, labor-force and online-vacancy data. Exposure is based on LLM-rated task–occupation cells weighted by worker task importance; raters produce materially different exposure levels. The pooled new-posting association adjusted for teleworkability is +0.094 log points per exposure interquartile range, p=0.21. It is not statistically distinguishable from zero and does not establish a zero effect. New-posting data are available through June 2026, but estimation excludes June because 30% of cells are suppressed, ending in May/partial Q2. Digital-work confounding and pretrends prevent a causal interpretation; adverts are not realized hires. Earlier revision summaries should not substitute for this version.

Interpretive boundary

The article’s proposed sequence from automation through reduced learning opportunities to weaker later reinstatement remains a hypothesis. None of these sources measures that complete causal sequence. The research agenda identifies evidence that could strengthen or weaken it; it does not manufacture national displacement totals, country risk scores or thresholds for irreversibility.