Iceland's Employment Policy 2035: A Systematic Blind Spot
An open, computationally verified audit of AI-driven workforce risk
Iceland's government published a ten-year employment strategy that never mentions unemployment. This open research project proves it — with data anyone can verify.
“An employment policy that does not mention unemployment is not an employment policy.
The Challenge
Iceland's Employment Policy 2035 was published without any risk assessment on AI-driven workforce disruption — despite every major international institution (IMF, OECD, WEF, ILO) having published such assessments, and despite Iceland's own ICT sector already showing measurable displacement signals. The challenge: demonstrate this blind spot rigorously enough that it cannot be dismissed as opinion. Every claim had to be computationally verified, every source linked, and the strongest counterarguments addressed before publication.
The Approach
Seven parallel research sectors, an adversarial Oracle agent, automated PDF text extraction, and a two-system verification pipeline (Manus 1.6 + Claude Code). All raw data, scripts, and verification reports published in an open-source repository. The approach draws on Responsibility Fog and Cognitive Debt from Beyond Fragmentation — applied here not to AI systems but to the governance system that produced the policy.
Outcomes
The Core Finding
Automated text extraction of Iceland's Employment Policy 2035 — reproducible by anyone in minutes — finds that AI (gervigreind) is mentioned three times, always as opportunity. The words "unemployment" (atvinnuleysi), "automation" (sjálfvirknivæðing), and "retraining" (endurmenntun) appear zero times each. A ten-year employment strategy, published April 2026, does not mention unemployment once.
Reproduce this yourself:
pip install requests && brew install poppler python3 scripts/02_verify_textanalysis.py
The ICT Trifecta
Iceland's ICT sector (NACE J, the European industry classification for information and communications) shows a four-signal pattern consistent with productivity displacement. Using conservative 2022 baselines:
- Revenue (J62 computer services): +28.1% (2022→2024)
- Employment: −1,300 jobs (peak 2023 to 2025, −11.4%)
- Vacancy rate: 5.4% → 0.3% (−94%)
- Labour productivity: +10.5% (highest of any sector in Iceland, 2025)
If this were a post-COVID correction, revenue would also have fallen. It grew 28%. The vacancy collapse from 5.4% to 0.3% eliminates that alternative: firms that need workers advertise for them. These firms did not. No Icelandic institution had assembled these three signals into a single finding before this analysis.
The Knowledge-Occupation Collapse
"This is not a forecast. This is a record of what has already occurred."
Between 2024 and 2025, Iceland lost 9,300 knowledge-intensive jobs — managers, professionals, associate professionals, and clerical staff (ISCO groups 1–4, −6.9%) — while gaining 6,900 physical and service jobs (ISCO groups 5–9, +7.4%). The economy did not contract; its composition shifted. The policy prescribes more STEM investment for a knowledge sector that is already contracting. The prescription directly contradicts the data.
Iceland Is Flying Blind
Iceland is the only Nordic country excluded from every key European AI and labour measurement instrument:
- Eurostat isoc_eb_ai: Denmark 42.0%, Finland 37.8%, Sweden 35.0%, Norway 28.9% — Iceland: not surveyed
- EWCS 2024 (Eurofound worker conditions survey): 35 countries, Iceland excluded
- Linked administrative data: Denmark has DREAM+IDA, Sweden LISA, Norway FD-Trygd, Finland Findata — Iceland has none
- Union AI agreements: Denmark (Hilfr2), Norway (LO-NHO), Sweden (Teknikföretagen-IF Metall) — Iceland: none
Additionally, Ríkisendurskoðun (National Audit Office) found a 17% discrepancy between Iceland's two official unemployment measures — Hagstofa (5.2%) vs. Vinnumálastofnun (3.3%) for the same month.
The Institutional Silence
When Iceland's government published its AI Action Plan for consultation — projecting 130,000 workers (55%) "greatly affected" by AI — not a single major labour union submitted a response in the 14-day window. ASÍ (Confederation of Labour), VR, Efling, BSRB, SFR, and KÍ were all absent. Iceland's largest union congress (ASÍ 46th, October 2024) passed zero resolutions on AI.
A Varða 2026 survey of 25,000 Icelandic workers found: 23% already affected by technology change, 36% fear job loss, 45% feel undertrained, 44% report increased work pressure (56% among immigrant workers).
Theoretical Framework
This analysis applies two concepts from Beyond Fragmentation (MA thesis, University of Akureyri, 2026) to a live policy document:
- Responsibility Fog: Systematic diffusion of accountability — institutions that should have responded to the AI consultation did not; no single actor bears responsibility for the collective silence.
- Cognitive Debt: A strategy built on international estimates never validated domestically, governing through borrowed statistics rather than observed data. The government's own "130,000 affected workers" projection uses a framework it has not validated for Icelandic conditions.
Verification
All five key findings were independently verified by two AI systems (Manus 1.6 and Claude Code). No finding was contradicted. Three precision corrections were applied. All verification reports are in the repository.
Eight Formal Data Requests
Eight formal information requests under Iceland's Freedom of Information Act (Upplýsingalög nr. 140/2012) have been prepared but not yet submitted to: Tryggingastofnun, Vinnumálastofnun, Hagstofa Íslands (×2), Sjúkratryggingar Íslands, Ríkisendurskoðun, Forsætisráðuneytið, and the Ministry of Higher Education, Industry and Innovation (Háskóla-, iðnaðar- og nýsköpunarráðuneytið). This page will be updated once the requests have been filed. Responses will be added to the repository as they arrive.
Technology Stack
Resources
Lessons Learned
- Computational verification is not just quality control — it is the argument itself. A script that downloads the government's own PDF and counts zero occurrences of "unemployment" is more persuasive than any rhetorical claim.
- Presenting both a conservative and a broad statistical baseline is stronger than the more dramatic number alone. Acknowledging the COVID-bubble concern preemptively closes the strongest counterargument.
- The strongest finding (9,300 knowledge jobs already lost) is not a forecast. Emphasising what has already happened is more powerful than projecting what might.
- Open-source evidence is harder to dismiss than a closed report. A repository with 150,000 rows of public data and reproducible scripts changes the terms of the debate.
- Verification by two independent systems identified three precision corrections before publication — including a timeframe inconsistency that critics would have used to discredit the entire analysis.