Economic Growth and the AI Revolution

Whither AI?

A few years ago, we were asked by the Royal Society of Edinburgh (RSE) to write a short piece on AI and productivity growth in the UK (short post here). At the time, we hedged and took a wait-and-see approach.

We’ve since been asked to return to the topic, this time by colleagues in Dubai, with a focus on how AI might play out in Gulf Cooperation Council countries. That gave us the chance not only to look at the GCC, but also revisit and update our earlier assessment of the UK.

Economic Growth and AI

In popular imagination, the AI Revolution begins with the release of ChatGPT in November 2022. In reality, it builds on a decade of advances in machine learning. A key moment was the 2017 paper “Attention is all you Need” by Google scientists that introduced the transformer architecture underpinning models such ChatGPT, Gemini, and Claude.

Since then the likely impact of AI on the economy has been widely debated. Estimates vary sharply. Daron Acemoglu suggests very modest effects, in the region of 0.06 percentage points of annual productivity growth over a decade. By contrast, Philippe Aghion and Simon Bunel argue that AI would lead to productivity growth in the region of 0.8 to 1.3 percentage points annually. The gap is large. Scenario exercises, such as those published by the Federal Reserve Bank of Dallas in June 2025 range from incremental change to extreme outcomes (singularity or extinction). The scenario kind of sums up where the current understanding of AI impact on productivity growth is: utopia, dystopia, or just meh.

AI Scenarios from the Federal Reserve Bank of Dallas

More speculative accounts have painted scenarios of rapid productivity growth alongside rising unemployment and inequality.

Can we see AI in the productivity statistics?

Robert Solow once remarked that we could see the computer age everywhere but in the productivity statistics. Is the same true for AI? Aggregate productivity in the US, as a leader in the adoption and innovation of large language models, is the best place to look. Recent data show Total Factor Productivity growth of 1.5 percent in 2024, falling to 0.8 percent in 2025. Over a longer horizon, TFP growth since 2019 was modestly higher (0.4 percentage points) than the period after the Global Financial Crisis (2007-2019), but still below the stronger performance of the early 2000s.

There is no clear break in the data that can be attributed to AI. Productivity has improved slightly, but not in a way that suggests a transformative technological shock. This may reflect lags in diffusion or measurement, as firm-level gains are not yet visible in aggregate statistics. Even so, the data point to incremental change rather than acceleration. Measurement error and timing issues remain important (see a recent work on the productivity puzzle), so it is still early.

Trend growth in US Private Nonfarm Total Factor Productivity Growth

Productivity Gains in the UK?

The UK has experienced a productivity slowdown (as highlighted by Eoin in an earlier post). Our initial approach to the question of AI and productivity was therefore to ask what impact it is likely to have in the UK context.

UK TFP growth

Unlike the US, productivity growth in the UK has continued to decline. The most recent TFP¹ estimates are negative: -1.7 in 2023 and -0.51 in 2024. Figures for 2025 are not yet available.

The UK’s TFP trajectory differs markedly from that of the US, declining at at average rate of 0.13 between 2019 and 2024. This compares with growth of 0.42 over 2010-2019 and around 1 percent over 2000-2009. There is no clear evidence of an AI-related effect in the UK data, and it remains uncertain whether productivity has reached its lowest point.

Productivity gains in the GCC?

If the productivity effects of AI are difficult to detect in the United States and absent in the UK, they are even harder to identify in GCC economies.

Economic growth across the GCC has slowed over time, reflected in declining GDP per capita growth across most member states. The 1990s and early 2000s saw periods of relatively strong performance, but growth since 2010 has been weaker, and, in some cases, negative. This pattern is consistent across countries including Kuwait, Oman, and the UAE.

Trend growth in GDP per capita

Evidence from sources such as the Conference Board and the Penn World Tables points to a common pattern: productivity growth has been weak. Economic expansion has been driven by capital accumulation (e.g., construction) and labour expansion rather than efficiency gains. Within the region, the UAE performs somewhat better, with modest positive TFP growth. Even here, however, productivity improvements remain limited and volatile.

These estimates should be treated with caution. Measuring productivity in resource-rich economies is difficult, and standard growth accounting frameworks are sensitive to assumptions about capital and output in the oil sector. Even so, the broad picture is clear: there is little sign – yet? – of a sustained, AI-driven productivity acceleration.

Trend Growth in Total Factor Productivity, 1990-2005

Recent geopolitical tensions have added further headwinds affecting trade, investment, as well as key sectors such as tourism and logistics. A recent UNDP report estimates that the cost of the Iran war could lead to decline in GDP in the region of 3.7 to 6 percent across the region. The latest World Bank forecasts downgraded estimates of growth in the region from 4 percent to 1.8 percent. The war has led to destruction of key infrastructure and has cut off vital supply chains. Qatar’s economy is forecast to contract by 5.7 percent. One area that has been affected is the tourism industry with losses estimated to be between $13 and $32 billion. The GCC economies have diversified in recent years and are no longer dependent solely on fossil fuels. One aspect of the diversification is the roll out of data centres and cloud computing in the region. However, these have become targets in the war.

Negative externalities and risks

Even where gains are possible, several risks could limit AI’s impact on productivity.

One concern is data quality. Models trained on open data can be distorted by deliberate manipulation (‘poison the well’). A recent example was the feeding a fake disease, Bixonimania, by researchers that was uploaded in preprints. These were then adopted uncritically by LLMs and relayed as though they were factual. This raises the risk that systems learn from unreliable or contaminated inputs.

AI is also affecting scientific publishing. The rapid increase in AI-generated content has led to a rise in low-quality or unreliable research, often described as AI-slop, with persistent issues of hallucination. The issue of slop has translated to ‘workslop’, leading to a disconnect between employees and employers. The latter have reduced workforces in the expectation of AI being more productive. This could perhaps explain why many companies have reported not seeing any profitable return on AI investment.

Geopolitical tensions are another constraint. Competition between the United States and China has intensified, particularly around semiconductor supply chains. Production remains highly concentrated, with a small number of firms dominating advanced chip manufacturing. One point of friction is Taiwan where one company produces almost all AI chips. This creates vulnerabilities that could slow the diffusion of AI technology or cause negative spillover effects on other sectors.

Labour market effects are also emerging. Entry level positions appear to be declining (as highlighted by a recent paper), though it is difficult to separate out the impact of AI from broader restructuring trends (not to mention economic wars!). There are also concerns about the longer-term effects, including the possibility of model collapse. The 2026 Standard AI index report highlights the shortage of real data owing to the contamination of data through the use of AI generated slop. A different kind of labour market effect is how job search works. It is very easy for people to use AI to generate sharp-looking CVs and job applications, and this in turn makes it more costly for firms to screen applicants because, e.g., they need to interview more or have applicants sit test. And applicants themselves need to send out more applications just to keep up - an example of what evolutionary biologists call the “Red Queen effect”. In Lewis Carroll’s Through the Looking Glass she famously says, “Now, here, you see, it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that!”

Perhaps most ominously is the potential for AI to facilitate crime. AI coding tools are great at detecting vulnerabilities in IT systems, so much so that Anthropic is holding back release of its latest AI model “Mythos” because it found so vulnerabilities in the systems of major companies, giving them time to patch. One suspects there are other active operators who are not so community-minded. The tools are also boon to small-scale cyberfraudsters, who can quickly and cheaply set up professional looking websites that promise to deliver products for a fee and instead disappear. We are in the early days of a new technological arms race between cybercriminals on the one hand, and firms and security agencies on the other. Technological advances usually generate overall economic growth, but “usually” and “overall” are doing some heavy lifting here. “Red Queen effects” galore.

Where to?

What does the future hold? There range of possibilities remains wide. Productivity growth could continue on an incremental path or shift toward more rapid acceleration. This is closer to Robert A. Heinlein’s “Where to?” scenarios, and in some respects resembles the framing used by the Federal Reserve Bank of Dallas.

Robert Heinlein’s Scenarios of Technological Progress

For now, however, the evidence points in one direction. Across both frontier and non-frontier economies, the pattern is similar. Productivity growth has picked up slightly in recent years, but not enough to suggest a transformative technological shock of the kind associated with past general-purpose technologies and the ‘Solow paradox’ seems to apply to AI. In short, the data still look more like the early stages of diffusion than a full productivity boom. There is a strong expectation that AI will eventually generate a substantial surge in productivity, but if so, we are not there yet.

1. Technically this is Multi-Factor Productivity (MFP) which is similar to TFP.