Surveying forecasts of AI’s economic effects

Generative AI has generated an enormous amount of hype. Even if some of the more extreme predictions do not come to pass, we think it has the potential to substantially lift economic growth in the medium term.

It is likely to do so by automating routine tasks, optimising operations and enabling the more efficient use of resources – all of which are likely to enhance economic productivity. That’s important because ageing populations worldwide point to decreasing per capita output from labour.

The extent and timing of AI’s economic impact remain the subject of debate. Adoption of AI has been rapid – in the 24 months since the launch of ChatGPT, some 40 per cent of Americans are using it, which is double the adoption rate of the Internet over a similar period following its introduction. Nearly a quarter used it at least once in the week before the survey was conducted and more than one in 10 use it daily at work.¹

And in certain niches, AI’s economic impact is already being felt. The online education company Chegg lost 99 per cent of its value as students began turning to ChatGPT for help with their homework.² But so far, while individuals have embraced generative AI, companies outside of the tech industry have been more cautious.

Productivity debate

This, however, hasn’t prevented experts from making predictions about its transformative effects.

Several academics and consultancies have tried to forecast AI’s impact on economic productivity (see Table above). Overall, they anticipate significant economic gains – with one notable exception. Daron Acemoglu, the most recent Nobel laureate in economics, estimated a mere 0.07 per cent annual boos tto US productivity over the coming 10 years.³ By contrast, rival academics Aghion and Bunel⁴ forecast 0.68 per cent annual productivity gains for the US over the same period, using the same methodology. Industry estimates for productivity gains range from 0.5 per cent to 3.4 per cent per year, with most clustering around the 0.7 per cent to 1.3 per cent range.

We believe the most likely outcome is a slow or even negative impact on productivity over the next few years before the beneficial effects of AI start to snowball.

People will first need to discover AI’s full capabilities, learn how to use it and integrate it into various functions before we can expect productivity effects to flow. This follows the dynamic of all major innovations – from steam to silicon. The key question is how quickly these effects will start to flow, as the adoption of new technology can be a slow process.

Meanwhile, the capital spending on energy and silicon chips needed to run AI systems will prove a drag on overall economic productivity in the near term, with AI operators even talking about owning their own nuclear power plants. In a recent article, Bertin Martens of Tilburg University argued that “without significant productivity gains, the current investment cost trajectory is unsustainable.”⁵ He estimates “that 3 per cent annual productivity growth across advanced economies would be required to sustain AI model cost extrapolations to 2030.”

One additional near-term cost of this capital investment could be higher inflation, especially if productivity gains lag significantly. Longer term, though, the increase in productivity should be disinflationary if, as we expect, wages don’t rise as quickly.

Labour market frictions

Generative AI productivity gains will come with an additional short-term cost: labour market disruption.

The transport sector stands out as one of the most obvious sectors where workers are at risk. Fully autonomous vehicles are already operating in several cities and AI is likely to make this widespread. In the US alone, there are some 4 million driving-related jobs, of which 3.1 million are truck drivers, according to the Occupational Outlook Handbook. While the replacement of professional drivers by AI is a negative for labour, saving time for commuters is a positive. Those using autonomous vehicles will gain, on average, an extra hour each day to spend on more productive or pleasant activities instead of concentrating on the road.⁶

It’s not entirely clear where the most gains will be. Some early indications suggest that AI can help bring lower-skilled workers up to the average. But by the same token, AI might replace them entirely. Higher-skilled workers in some industries will be able to offload their more routine tasks to AI, allowing them to focus on where they truly add value.

Industrious AI

Industries with the highest investment in AI – healthcare, pharmaceuticals, biotechnology, IT infrastructure and hosting, media, social platforms and marketing – are likely to see the greatest productivity gains and the most significant impact on their workforces.

McKinsey surveyed industry experts for their estimates of the impact of generative AI on expected revenues by industry.⁷ High tech is seen benefiting the most, with a 4.8 per cent to 9.3 per cent increase in revenues, followed by pharmaceuticals and medical products at 2.6 per cent to 4.5 per cent, healthcare at 1.8 per cent to 3.2 per cent, and media and entertainment at 1.8 per cent to 3.1 per cent.

The differences in forecasts of likely productivity gains is largely down to different estimates of what fraction of tasks in an economy can profitably be performed by AI. The majority of forecasters use the same data source, O*NET. This database contains information on 800 different occupations, with data on average income, industry affiliation, required skill levels on 35 skill constructs and detailed task descriptions.

Estimates of task-fractions that can be automated profitably range from 6 per cent to 80 per cent. Acemoglu, whose forecasts lie at the lower end of that range, does not factor in that AI may create new products, services and jobs or shifts in labour concentrations to reflect changes in industry focus. We think higher estimates are more likely to be accurate than Acemoglu’s.