IMF and OECD Warn AI Gains Could Widen Gaps Between Economies

Why two international institutions are sounding an alarm

The International Monetary Fund and the Organisation for Economic Co-operation and Development are warning that artificial intelligence could boost economic output while widening differences between countries, firms and workers. Reuters reported their concerns on October 22. Wealthier economies are often better placed to deploy AI because they have dependable electricity, fast connectivity, computing resources and people trained to use the technology. A productivity gain that appears positive at the global level can still leave some communities behind. The central policy challenge is not whether AI has potential, but how its benefits and costs are distributed.

Growth projections are not guarantees

The IMF has discussed a possible boost to annual global growth of up to about half a percentage point, while the OECD has outlined a range of potential productivity improvements among its members. Such estimates are scenarios based on assumptions about adoption, investment and how firms reorganise work. They are not measured increases already delivered to every economy. A country may buy AI subscriptions without seeing much productivity improvement if workers lack useful data or if regulations and business processes prevent effective deployment. Policymakers should therefore separate a modelled opportunity from results demonstrated in company accounts and national statistics.

Infrastructure is the first dividing line

Running modern AI services requires more than an internet connection. Data centres need power, cooling, communications networks and capital. Businesses need cybersecurity, reliable records and the ability to connect new software to existing systems. Where electricity is expensive or unreliable, an AI application that appears cheap in a wealthy country may be difficult to deploy at scale. Public infrastructure investment can improve access, but governments must weigh it against schools, health systems and other priorities. The question is how to build capacity that supports many industries rather than subsidise a small number of fashionable projects.

Why smaller firms may lag

Large companies can hire specialists, negotiate computing contracts and run controlled experiments across departments. A small retailer, manufacturer or professional service firm may lack the time to evaluate tools, protect customer data and redesign processes. Even when software is available, the cost of training and checking outputs can be significant. That does not mean smaller businesses cannot benefit. Well-designed services can reduce administrative burdens, improve forecasting and help with translation or customer communication. The challenge is to make these benefits practical and safe without requiring every small business to build an AI department.

The particular risk for younger workers

Entry-level jobs often involve drafting, data collection, routine analysis and other tasks that AI can accelerate. Those tasks also provide the experience from which workers learn to exercise judgement. If employers automate the early steps without creating alternative training routes, new graduates may find it harder to develop professional skills. The risk is not simply that every junior position disappears; roles may change, and some new work will emerge. Employers need to consider apprenticeships, supervision and opportunities to review AI-generated material. Education policy should teach people how to verify outputs and understand domain knowledge, not only how to write prompts.

What exposure to AI actually means

The IMF has previously estimated that a substantial share of jobs worldwide may be exposed to AI. Exposure is not the same as elimination. Some occupations contain tasks that are easy to automate and others that require trust, physical presence or complex judgement. A teacher might use AI to prepare materials while remaining responsible for students; a finance worker might automate reconciliations while handling exceptions and accountability. Employment effects depend on demand, regulation, business choices and how quickly workers can move between roles. Headlines that equate task exposure with a fixed number of lost jobs exaggerate what the research establishes.

Policies that could narrow the gap

Countries can improve digital infrastructure, support workforce training, strengthen privacy safeguards and encourage competition among technology suppliers. Universities and vocational institutions can teach statistical reasoning and subject expertise alongside software skills. Governments may also need to help smaller enterprises test tools safely and evaluate whether promised savings actually occur. International cooperation matters because AI markets cross borders, while data-protection rules and access to computing resources differ widely. None of these interventions guarantees equal outcomes, but they can reduce the chance that adoption simply follows existing patterns of wealth and market power.

How to judge progress over time

Useful evidence would include measured productivity in different industries, adoption rates among small firms, changes in entry-level hiring and whether public services become more accessible. Researchers should compare benefits with energy use, security incidents and the costs of human review. Countries may also choose different approaches based on language, labour markets and existing infrastructure. The IMF and OECD warnings should not be read as a prediction that AI inevitably increases inequality. They are a call to design policies that make broad participation possible while tracking who actually gains from the technology.

Measurement will determine whether promises are realised

Companies often report that employees save time on individual tasks, but time saved does not automatically translate into higher national productivity. Staff may use the extra time for better service, more output or additional checks; some organisations may struggle to change workflows at all. Researchers need to compare similar firms and account for the cost of software, supervision and computing. The effects may also differ across languages and industries. A system trained mainly on widely available English material could be less useful for a small business working in a less represented language. Better measurement would reveal where public investment and training have the greatest value, and prevent governments from mistaking impressive demonstrations for broad economic transformation.

Access to high-quality local-language material is another dimension of inequality. A public-service chatbot that performs well in one language may give weaker answers in another, particularly for specialised legal or agricultural terminology. Testing across languages and communities should be part of procurement, not an afterthought. The same principle applies to disability access and the cost of connectivity. Broad access means systems are usable by people with different needs, not simply available on a website.