Sources · Checked weekly · Last checked November 15, 2026
Which sources does Applied Frontier rely on, and why?
Direct answer
This page lists the published research Applied Frontier relies on. None of it is our own data. We start with what AI labs, payroll processors and multilateral agencies publish, because those sources report on the labor market as it is now. The core sources are OpenAI’s GDPval, the Anthropic Economic Index, Stanford’s analysis of ADP payroll data, Microsoft Research’s occupational study, the IMF and ILO exposure estimates, and the African Union’s Continental AI Strategy. Each source is checked every week, and changes are recorded on this page.
Week of November 15, 2026
What changed this week?
- Think tanks and multilateral research. Added fourteen publications on AI and jobs in developing countries, from the Center for Global Development, Brookings, the Carnegie Endowment, Bruegel, the Tony Blair Institute, IDOS, ODI Global, the World Bank, the ILO and the Asian Development Bank.
- Stanford Digital Economy Lab. The August 2026 update of “Canaries in the Coal Mine” extends the payroll data to June 2026. Workers aged 22–25 in the most AI-exposed occupations are now about 19% below where they would be had they kept pace with peers in less exposed occupations, up from 15% a year earlier. The authors describe this as a descriptive pattern, not a causal estimate. The table below is updated.
- Anthropic Economic Index. The June 2026 report adds survey results. Respondents in high-income countries report that AI can do about 10 percentage points less of their current tasks than respondents in lower-income countries, and concern about job loss is higher in lower-income countries. Anthropic notes that the survey is not representative.
- All sources. Links to the original publication added for all twelve sources.
Every source on this page is checked each week for new editions and new data. Each week’s changes stay here, newest first.
Why start with industry sources?
Because they are the most current. The labs publish what their models can do and what people use them for; payroll processors publish who is being hired; the IMF, ILO and World Bank publish exposure by country. Peer-reviewed research on AI and labor is careful but often lags the labor market, so we cite it after these sources, where it adds something.
Every finding below is summarized as its source reports it, with the year and a link. These are other organizations’ findings and estimates, and several are forecasts. Check the original before using a figure in a policy paper.
What do the sources say?
| Source | Year | Finding | How we use it |
|---|---|---|---|
| OpenAI, GDPval | 2025 | Frontier models matched or beat industry experts on roughly half of real work tasks across 44 occupations | Analysis, coding and writing are commoditized |
| Stanford Digital Economy Lab, ADP payroll data (“Canaries in the Coal Mine”) | 2025, updated August 2026 | Workers aged 22–25 in the most AI-exposed occupations are about 19% below where they would be had they kept pace with peers in less exposed occupations, as of June 2026 | The entry-level white-collar job is disappearing |
| Dario Amodei, Anthropic (Axios interview) | 2025 | AI could eliminate half of entry-level white-collar jobs and push unemployment to 10–20% within one to five years | The scale and speed of the shift |
| Anthropic Economic Index | 2025–26 | Usage concentrated in software, writing and analysis; the share of work fully automated rather than assisted is rising; per-capita usage tracks GDP per capita | What work is delegated first, and where |
| Microsoft Research, “Working with AI” | 2025 | Ranks occupations by AI applicability from Copilot usage; writing, translation, sales and customer service highest, physical trades lowest | The stop-training list |
| Microsoft, Work Trend Index (“The Frontier Firm”) | 2025 | Firms reorganizing around human-agent teams led by people who direct and verify agents | The outcome-owner and workflow-builder job families |
| OpenAI, Jobs Platform and Certifications | 2025 | A lab building its own hiring signal and pledging to certify 10 million Americans by 2030 | The degree signal is being replaced from outside the university |
| IMF, “Gen-AI: Artificial Intelligence and the Future of Work” | 2024 | About 40% of global jobs exposed; 60% in advanced economies, 40% in emerging, 26% in low-income | The developing world is less exposed but less ready |
| ILO, “Generative AI and Jobs” | 2023, 2025 | Automation risk is low in low-income countries but augmentation potential is large if infrastructure and skills exist | The opportunity in the developing world |
| African Union, Continental Artificial Intelligence Strategy | 2024 | Skills, data, compute and governance as continental priorities | Alignment with Africa’s own agenda |
| Dario Amodei, “Machines of Loving Grace” | 2024 | AI could lift growth in developing countries far above historical norms | The leapfrog argument |
| Goldman Sachs, “The Potentially Large Effects of AI on Economic Growth” | 2023 | 300 million full-time jobs exposed globally | The baseline exposure figure |
What are think tanks and multilateral researchers finding about AI and jobs in developing countries?
That developing countries are less exposed to job loss than rich countries, that the jobs gap is older and larger than AI, and that the binding constraints are connectivity, devices and delivery. The Center for Global Development, Brookings, the Carnegie Endowment, Bruegel, the Tony Blair Institute, IDOS and ODI Global are the policy institutes we follow most closely. Each finding is summarized as its source reports it.
| Source | Year | Finding | How we use it |
|---|---|---|---|
| Center for Global Development, “The Developing World’s Jobs Crisis Was Here Before AI” | 2026 | Developing countries need about 800 million more jobs over the next decade than they are on track to produce. Of 1.2 billion young people reaching working age, about 420 million can expect a job. The gap existed before AI. | The scale of the jobs problem in every country report |
| Center for Global Development, “SMS Charges May Be Inhibiting the Use of AI for 1 Billion People” | 2025 | About 1.7 billion mobile phones lack active internet service, serving roughly a billion people. For a text-based AI service, about 98% of the cost per message would be the SMS charge. | AI tools and training must work on the devices people have |
| Center for Global Development, “(How) Do the Poorest Use AI?” | 2026 | A smaller share of people in low- and middle-income countries use AI than in high-income countries, and many of the poorest have only basic phones. | Readiness is measured by use at work, not by access |
| Center for Global Development, “Cutting Through the Noise: My AI Priors” | 2026 | Markets will leave public-interest gaps in low- and middle-income countries, such as tools for some 500 million smallholder farmers. Adoption depends on delivery, internet access and behavior change. | Why frontline AI roles need public funding |
| Brookings, “Why Africa should sequence, not rush into AI” | 2026 | Internet penetration in Africa is 38%, against 68% globally. Nearly 12 million people enter the job market each year, and about 3 million formal jobs are created. | The starting point for the Africa playbook |
| Carnegie Endowment for International Peace, “The Other AI Race: An Export Promotion Strategy for the Global South” | 2025 | Recommends concentrating AI infrastructure support on strategic hubs such as Brazil, India, Indonesia, Kenya and Nigeria, and pairing data centers with public-sector applications. | Where enablement hubs are likely to form first |
| Bruegel, “Artificial intelligence, jobs and growth in developing economies” | 2026 | Hosted a discussion of the World Development Report 2026 in Europe, noting that low measured exposure to AI in developing economies may reflect the mix of jobs, not protection from change. | Low exposure is not a reason to wait |
| Tony Blair Institute for Global Change, “Delivering AI Impact: A Leadership Agenda for Turning Technology into Public Value” | 2026 | Low- and middle-income countries should treat AI as public infrastructure, and have fewer legacy systems to replace. | Enablement hubs as public institutions |
| Tony Blair Institute and Singapore’s Institute for Adult Learning | 2025 | Nearly 80% of Southeast Asian workers are in the informal economy, and six of ten ASEAN economies have national AI strategies. | The Southeast Asia playbook |
| German Institute of Development and Sustainability (IDOS), “Reforming the development sector for agentic AI (and whatever comes next)” | 2026 | Development agencies are adopting AI faster than they can govern it, and most pilots stay exploratory. Proposes shared capacity to test AI tools and data readiness assessments before AI-assisted programs. | Why pilots publish cost, method and outcomes |
| ODI Global, “Why AI can hold back Africa’s industrialisation and what to do about it” | — | Generative AI could add $61 billion to $103 billion in annual value in Africa, with early gains in services. Proposes a continental micro-credentialing scheme for technical skills. | Credentials based on demonstrated work |
| World Bank, World Development Report 2026 | 2026 | Generative AI puts 4.5% of jobs in low- and middle-income countries at risk of automation, against 14.2% in high-income countries. AI could meaningfully raise productivity in 16.2% of jobs in developing economies. | The exposure baseline for regions and countries |
| ILO and World Bank, “Disruption without dividend? How the digital divide and task differences split GenAI’s global impact” | 2026 | In a study of 135 countries, developing economies are less exposed, but disruption may arrive faster than productivity gains because many workers who could benefit lack reliable internet. | Connectivity comes before productivity |
| Asian Development Bank, “Workers’ Exposure to Artificial Intelligence Across Development Stages” | 2026 | Lower AI exposure in low-income countries comes mainly from differences in the tasks workers do within the same occupation. Use of digital tools and human capital explain most of the variation. | Train for tasks, not job titles |
Where is Applied Frontier’s own evidence?
In the country reports and in targeted pilots. The country reports for Kenya, India, Bangladesh, Indonesia and Brazil use Applied Frontier’s dataset on what it costs to create a professional, and set out a benchmark for the cost of training for an AI-economy role. Targeted pilots measure cost per job, time to placement and retention for one job family in one market. Both are published with the method and the sources.
Read the country reports → How a targeted pilot works →Questions about the sources
- Which think tanks does Applied Frontier follow on AI and jobs in developing countries?
- The Center for Global Development, Brookings, the Carnegie Endowment for International Peace, Bruegel, the Tony Blair Institute for Global Change, the German Institute of Development and Sustainability (IDOS) and ODI Global, alongside research from the World Bank, the ILO and the Asian Development Bank. Their findings are summarized on the Sources page with links to the originals.
- Does Applied Frontier use academic research?
- Yes, after the industry sources. Peer-reviewed work on AI and labor is careful but often lags the labor market. The labs, payroll processors and multilateral agencies publish what is happening now, so we start there and cite academic work where it adds something.
- Are these figures Applied Frontier’s own?
- No. Each is another organization’s finding or estimate, reported as the source states it, with the year and a link to the original. Applied Frontier’s own evidence, from pilots and cost-per-job research, will be published separately as it is produced.
- How often is this page updated?
- Every week. Each source is checked for new editions and new data, and what changed is recorded at the top of the page, newest first.