Why we need to be realistic about generative AI’s economic impact

The astonishing pace of technological progress over recent decades has failed to lift growth rates in advanced economies such as the United States. During the pandemic, many rushed to declare the accelerated use of digital services a turning point. But as we wrote then (and since), the exuberant growth impacts were unlikely to materialize – and they didn’t.

Understanding technology’s past disappointments helps clarify future potential. One reason for the meek showing is technology is only the fuel. Productivity growth also needs a spark to ignite effective technological adoption.

Persistently tight labour markets can be that spark because firms are first nudged, then forced, to replace labour with technology when they can’t hire. The sustained labour market slack in the 2010s meant firms did not have to reinvent their processes, but the re-emergence of tightness since COVID may provide that spark.

An obstacle to higher productivity growth has been the lack of technology that can comprehensively replace labour, particularly in labour-intensive services. What automation achieved in manufacturing has no technological equivalent in services, which depend on non-standardized, reciprocal human interaction.

Generative artificial intelligence (AI) promises to change some of that. But to realistically gauge its plausible impact we must take a closer look at the mechanisms that translate technology to broad productivity growth.

Far too often productivity growth is framed as technological marvels delivering product innovation. Though important, big productivity shifts come from at-scale cost reduction more than new or better products. It is the deflationary nature of technology that is macroeconomically powerful.

We have used the story of the humble taxi to highlight the misguided focus on snazzy apps instead of the gritty world of costs. Uber, Lyft, and Grab may be the epitome of what drives society forward, literally and figuratively, but where is the productivity growth? As every first-year economics student knows, productivity growth is about improving the ratio of inputs to outputs.

Apps have not radically changed that: labour and capital inputs – in this case, the driver and the car – are unchanged; the matching of drivers and riders is somewhat better; but higher prices tell us there has been no productivity transformation – if there was, prices would be falling.

Why? Firms that can replace labour with technology will lower prices to take market share from higher-cost competitors. When that process rolls through sectors, the macroeconomy experiences strong productivity growth.

In the case of transportation, the big shift will arrive if and when algorithms and sensors replace drivers. It will not arrive because the app indicates the driver’s location or because you settle payment silently.

It has been tech’s inability to spark that technology-cost-price effect that left its growth impact wanting. Now generative AI’s potential to remove cost from the service economy through replacing non-linear interactions – from call centres to marketing to research and design – makes impact more likely.

And it has usefully focused the conversation on cost reduction, which is what will deliver macro benefits because it kicks off the technology-cost-price effect.

The evidence of this deflationary effect is strong. Consider the experience of goods versus services over the last 30 years. In durable goods, automation, along with outsourcing, led to a steep fall in labour inputs and prices.

In the three decades to 2020 the price index of durable goods fell 35%, and it rose by only 15% for all goods. In services labour intensity has declined modestly or not at all. Therefore prices typically rose, often steeply. In transport, prices rose 79%, in education 348% and for services overall 120%.

If cost reduction and falling prices is how technology delivers significant productivity growth, then consumers will be the winners. Lower prices will boost real incomes that can be spent elsewhere.

Consider that food once took a significant share of people’s wallet, but as prices fell – via mechanization and, later, fertilizers – income was freed up to spend on household goods and services, such as tourism. This is how tech drives aggregate growth – and how dystopian predictions of mass unemployment have not come to pass because new spending also creates new jobs.

For firms, this means that the productivity cascade – tech-cost-price-income – is a threat as much as opportunity. Firms that can lead down the cost curve, maintain relative advantage, lower prices and capture market share will be winners – at the expense of those who can’t.

While generative AI will create new corporate titans, or reinvigorate existing ones, some sectors may see AI puts industry profits at risk for all firms.

This happens when labour-saving technology is so widely accessible that all firms employ it easily. A price war and dwindling profits ensue. Productivity gains in auto manufacturing, shipping, or airlines haven’t left booming industry-wide profits, but rather low prices, fierce competition and modest profits.

For firms, therefore, the strategic implications of generative AI and other tech is as much defensive (cut costs to remain viable) as it is offensive (cut costs to gain advantage).

Generative AI is a critical piece of a technological mosaic – including sensors, 5G, robotics, biotechnology and more – that can drive productivity growth, but by how much? Exuberance typically accompanies remarkable innovations and some recent estimates have argued that US productivity growth could shift up by more than 300 basis points (bps).

That’s too exuberant. Though it’s tempting to use bottom-up case studies to make macroeconomic predictions, such estimates remain exercises in assumption and extrapolation. There are unknowable hurdles such as regulatory friction and societal acceptance that will stretch timelines and limit impact.

The estimate of 300bps implies the US economy would more than double its widely accepted trend growth rate, from around 2% to 5%. It is the same kind of thinking that predicted that greater digital penetration during the pandemic would lift productivity growth by 100bps or more, which didn’t happen.

Past productivity boosts provide clues about plausible impact. Is the current wave of technology likely to be bigger, smaller, or similar to the information and communication technology (ICT) boom that delivered a productivity boost between the mid-1990s and mid-2000s? That was the last time a tight labour market intersected with the availability and enthusiasm for a new wave of technology, accelerating growth by about 100bps for around 10 years.

As in the 1990s, the tight US labour market can help accelerate technological adoption. Many of the new innovations are promising, but will take time to build and be usefully adopted. In addition, regulatory and other frictions must be overcome. We think this points to a more modest impact – an acceleration closer to 50bps than 150bps, and certainly not 300bps.

Generative AI is full of credible promise, but does it have the same “general purpose” application in the economy that the internet had? Perhaps one day it will, but likely with long lags. Consider the information and communications technology boom gathered pace (and critical mass) over a 30-year period (from the late 1960s) before its macroeconomic impact was visible. Even if today’s lags are shorter – a credible proposition in the digital age – they have not disappeared.

But keep in mind that in macroeconomics, small numbers mean big things. If generative AI and the confluence of new tech can deliver a 50bps growth bump it would add an extra $8 trillion to US gross domestic product over a decade (roughly twice the size of Germany’s output in 2021), which is nothing to scoff at.