As part of your AI Literacy homework I have collected 9 recent (2026), serious (?) sources on the ‘future of work’ and graded them in terms of ‘sophistication of theory’ ‘strength of empirical evidence’ and ‘practical utility’. I then synthesise them into their most interesting bullets.
Well, I say ‘I have…’ but of course I mean ‘AI has…’. My overall ambition is to fire myself and jam with Barney…
Sources
1. Stanford HAI AI Index 2026, Chapter 4 (Economy) — 13 Apr 2026 Aggregated economy-wide AI labour data; ~20% fall in entry-level developer employment since 2024. Theory 2 / Evidence 5 / Practical 3
2. ILO Research Brief: Workers’ exposure to AI — what indicators tell us, and what they don’t — 17 Apr 2026 Methodological critique of AI-exposure measures. Theory 4 / Evidence 2 / Practical 3
3. World Bank (with ILO): “Disruption without dividend?” — 17 Mar 2026 135-country GenAI exposure, background paper for World Development Report 2026. Theory 3 / Evidence 4 / Practical 4
4. IMF Note: Global Economic and Financial Implications of AI — Apr 2026 Scenario-based macro and labour assessment. Theory 4 / Evidence 3 / Practical 4
5. NBER Working Paper 34836 — Yotzov, Barrero, Bloom, Davis et al., “Firm Data on AI” — Feb 2026 (revised Mar 2026) 6,000 executive survey across US, UK, Germany, Australia; 69% firms use AI, but 9 in 10 report no employment/productivity impact yet. Theory 3 / Evidence 4 / Practical 3
6. Brookings: “AI growth acceleration versus distributional fairness” — 5 May 2026 Policy-economics synthesis on whether AI growth gains will be broadly distributed. Theory 4 / Evidence 3 / Practical 4
7. OECD: Building an AI-ready Public Workforce — 19 Jan 2026 Public-administration AI-readiness playbook. (More recent OECD pieces exist — Apr 2026 “Anticipating Skill Needs” and Feb 2026 neurodivergent-learners report — but this is the most AI/workforce-central.) Theory 2 / Evidence 3 / Practical 5
8. WEF: Four Futures for Jobs — AI and Talent in 2030 — 14 Jan 2026 Four AI-and-talent scenarios drawn from chief-strategy-officer dialogues. Theory 4 / Evidence 2 / Practical 4
9. Anthropic Economic Index report: Learning Curves — 24 Mar 2026 Real Claude usage data from ~1M conversations; introduces learning-by-doing framework; ~49% of jobs now have at least a quarter of tasks performed using Claude. (Caveat: I am Claude. Conflict of interest in recommending Anthropic’s own research.) Theory 4 / Evidence 4 / Practical 3
Synthesis
Top 5 theoretical concepts / lenses
- Adaptive capacity × exposure — Two-dimensional framework: exposure is half the picture. Financial security, skill transferability, geographic mobility, and age determine who can actually transition. Brookings / Manning & Aguirre
- Task-level concentration > mean exposure — Two jobs with identical mean exposure can diverge: the one with exposure concentrated in fewer tasks suffers less, because workers can reallocate effort to unaffected tasks. NBER Hampole et al.
- Exposure ≠ outcome — Exposure measures capture what AI could do under static task lists; they do not predict displacement, since adoption depends on profitability, workflow change, demand shifts, and institutional friction. ILO brief
- Augmentation / automation continuum + learning curves — Five interaction types (directive, feedback loop, task iteration, validation, learning) replace the binary. Experienced users move toward iterative augmentation, not automation. Anthropic
- Disruption-without-dividend asymmetry — In developing economies, disruption can materialise faster than productivity gains because digital infrastructure and task composition gate the upside. ILO–World Bank
Top 5 links to value from AI
- Software development — 35% of Claude.ai conversations; coding now migrating from chat interfaces to programmatic / agentic workflows. Anthropic
- Customer service / support — Documented 14–34% productivity gains, largest for novice and low-skill workers. Brookings synthesis (Brynjolfsson et al.)
- Public-sector administrative processing — Finland’s Kela (social-security agency) saves ~38 years of full-time equivalent work per year on document classification alone. OECD
- Consumer surplus from free generative AI tools — US consumer surplus ~$172B annually by early 2026, up from $112B; median per-user value tripled in a year. Stanford HAI
- Wage premium for new AI-adjacent skills — Job postings with one new skill pay ~3% more; four or more new skills pay up to 15% more (UK) / 8.5% (US). IMF
Top 5 practical recommendations for people
- Iterate, don’t one-shot — Experienced users show ~10% higher success rates and use AI as thinking partner; learning-by-doing pays measurably. Anthropic
- Build complementary, not substitutable, skills — Cognitive, creative, and technical skills that combine with AI rather than compete with it. IMF
- For early-career workers: route around exposed entry points — Employment for 22–25-year-old software developers has fallen ~20% since 2024; entry pipelines are where damage concentrates. Stanford HAI
- Build adaptive capacity, not just skills — Savings, geographic mobility, and skill transferability matter as much as retraining when transition becomes necessary. Brookings
- Track where your function is migrating — Coding migrated from chat to agentic API within months; watch the migration pattern in your domain before it surprises you. Anthropic
Top 5 practical recommendations for companies
- Redesign workflows before training people — AI-investing firms also alter internal hierarchies and skill mixes; training bolted onto unchanged workflows teaches unusable skills. Brookings
- Build in-house AI capability rather than outsource — Reduces dependency, prevents information asymmetries in procurement, aligns AI with institutional needs. OECD
- Plan for entry-level pipeline disruption — Junior roles handle the work AI does best; cannibalising them now eats your pipeline of senior talent in 5–10 years. Stanford HAI
- Stress-test against multiple AI futures, not one forecast — Four divergent AI-and-talent scenarios for 2030 should drive workforce planning, not a single point estimate. WEF
- Discipline attribution: avoid AI-washing layoffs — 9 in 10 executives report no realised employment/productivity impact yet but expect large forward effects; conflating macro and AI-driven cuts erodes credibility. NBER W34836
Top 5 risks
- Skill-biased technical change widens within-country inequality — Gains concentrate among workers who can leverage AI; others see displacement or wage compression. IMF
- Entry-level pipeline collapse — Largest measured contraction so far is in 22–25-year-old white-collar roles; risks generational talent gap. Stanford HAI
- Disruption-without-dividend in developing economies — Disruption may materialise before productivity gains; risks widening gap between high- and low-income countries. ILO–World Bank
- 6.1 million US workers face both high exposure and low adaptive capacity — Concentrated in clerical and administrative roles; 86% are women. Brookings
- Standards capture and infrastructure concentration — Foundation Model Transparency Index fell from 58 to 40 in a year; the most capable models disclose the least. Stanford HAI