The Economics of AI: Why Do Economists Disagree on the Gain?

In June 2023 McKinsey estimated that generative AI could add between $2.6 and $4.4 trillion a year to the global economy, and set that against the United Kingdom's 2021 gross domestic product of $3.1 trillion. About a year later Daron Acemoglu, a professor of economics at the Massachusetts Institute of Technology, published a paper concluding that the rise in total factor productivity would not exceed about 0.66% over ten years.

Which is closer to the truth? Put that way, the question misleads, because the two figures answer different questions.

How does each side build its number?

Acemoglu starts from a simple logic: if firms' gains come from cost savings in particular tasks, the effect on the whole economy can be estimated from the share of tasks affected and the average saving in each. He then adds a caution. Early evidence comes from tasks that are easy to learn; if AI spreads to harder tasks the saving shrinks, and the estimate falls below 0.53%.

McKinsey, for its part, counts the possible use cases, 63 of them, and values them against the structure of the global economy in 2022. It does not count what entirely new products and services might create, and it assumes that affected workers move into other activities, which is what allows the productivity gains to materialise. It is an estimate of what is possible, not a measurement of what has happened. Nor can the two figures be set side by side without care: McKinsey itself estimates that generative AI could raise labour productivity growth by 0.1 to 0.6 percentage points a year through 2040, a different measure from the total factor productivity Acemoglu analyses, over a different horizon.

What does the field say?

The clearest evidence is a field experiment studied by Erik Brynjolfsson and his team. A generative assistant was introduced in stages to 5,179 customer-support agents. The number of issues resolved per hour rose by about 14% on average, by about 34% for novices and less experienced agents, and only slightly for experienced ones.

The lesson cuts two ways. The gains are real in specific tasks, and they tend to lift the weaker performers more than the stronger. But the leap from one call centre to a whole economy is one the study itself does not guarantee.

Costs fall, but not to zero

It is often said that the cost of an extra unit of AI tends to zero. The data are more precise. According to the Stanford AI Index 2025, the cost of running a model at GPT-3.5 level (about 64.8% on the MMLU benchmark) fell from $20 per million tokens in November 2022 to $0.07 in October 2024, a drop of more than 280-fold.

That is a steep fall, but it is not zero, and it sits beside heavy spending: US private investment in AI reached $109 billion in 2024, and 78% of organisations in the survey said they use AI, up from 55% a year earlier. Adoption is fast, and the economists' question is when it will show up in aggregate productivity.

Who gains, and where do we stand?

The International Monetary Fund estimates that about 40% of jobs worldwide are exposed to AI: about 60% in advanced economies, 40% in emerging markets and 26% in low-income countries. Exposure is not loss. About half of the exposed jobs in advanced economies may benefit from working alongside the technology, while demand for the other half may fall.

Kristalina Georgieva, the Fund's Managing Director, judges that inequality will probably widen in most scenarios. She calls for social safety nets and retraining, and for emerging economies to start with digital infrastructure and skills. In his paper Acemoglu goes further, arguing that AI may widen the gap between capital income and labour income, and that some of the tasks it creates may be of little or negative social value. A gain in aggregate does not necessarily mean a gain for everyone.

In Saudi Arabia, on 12 May 2025 the launch of HUMAIN, owned by the Public Investment Fund, was announced; its aims are set out in the article "AI Diplomacy: Whoever Writes the Rules Helps Shape the Game". The bet is that infrastructure and models will turn into value inside sectors. Assessing the economic effect will need data that is published and reviewed over the coming years. For more on Arabic models, see the article "Arabic Artificial Intelligence: Do We Make or Consume? Sovereignty and Arabic Language Models".

Four questions before believing any number

• Is it a theoretical possibility or a field measurement?
• Which tasks does it cover, and are they easy to learn or hard?
• Over what time horizon?
• Who reaps the fruit, and who bears the cost?

This is knowledge to aid understanding, not investment advice; anyone wishing to commit money should consult a licensed professional and observe the rules of Sharia. Taking the means, through learning and scrutinising figures, does not replace reliance on God, just as reliance on God does not replace taking the means.