Will AI make it easier to get the productivity growth we want?

Artificial intelligence, or AI, is not new. The term was coined by John McCarthy during the legendary Dartmouth Conference in 1956, where some of the brightest minds in computing science gathered to discuss ‘thinking machines.’ Since then, the field has experienced cycles of enthusiasm and decline.

Early optimism eventually led to an 'AI winter', a period of cuts in investment, following overhyped expectations that fail to deliver, and the term AI fell out of favour.

However, in the past decade, fuelled by increased computing power and vast amounts of data, the field has decisively shifted from early rules-based systems to more flexible bottom-up machine learning, like large language models (LLMs), culminating in OpenAI's release of ChatGPT and a new cycle of hype. Yet despite renewed excitement, generative AI’s impact on the economy remains uncertain.

True, numerous studies have shown that current AI capabilities can already deliver significant productivity gains in areas like customer service, writing and coding, and work done by mid-level professionals.

Such findings raise hopes for reversing the decline in global productivity growth, which has been slowing for more than a decade. But these applications generate efficiency gains by mainly automating tasks already performed by humans rather than creating new tasks and products.

Here we outline three key factors that will influence AI's contribution to productivity growth, and the crucial role of policy in shaping them.

While automation can provide a short-term productivity boost, it’s a one-time gain. If all progress since 1800 had been limited to automation, we would have efficient agriculture and cheap textiles, but little more.

The size and timing of the productivity effects of generative AI are uncertain. Daron Acemoglu estimates that without AI contributing to scientific discovery or new product creation, automation with state-of-the-art AI might increase US productivity by just 0.07% annually over the next decade. Meanwhile, others, like Erik Brynjolfsson, are more optimistic about AI’s potential to boost labour productivity through augmentation and innovation.

Yet in the end, AI's trajectory and productivity contribution will be influenced by policy choices. For instance, tax policies have inefficiently favoured automation in some countries by effectively taxing equipment, hardware and software that are classified as capital at lower rates than labour.

To the extent that automation is driven by tax incentives rather than for efficiency gains, it can encourage indiscriminate substitution of labour and fail to improve productivity. Careful consideration should be given to rebalancing the mix of depreciation allowances, credits and other tax incentives that create such a bias.

Adjusting this imbalance could incentivize companies to leverage AI for augmentation rather than automation, prioritizing capital savings over labor replacement. As a result, AI-driven productivity gains would increasingly enhance tasks already done by machines, such as robot training, self-driving cars and improving credit card fraud detection. Reversing the trend decline in capital income taxes seen in recent decades can boost productivity and make those gains widely shared at the same time.

Secondly, AI has the potential to revolutionize scientific discovery, and generate the technological progress that boosts future productivity. Already, AI-predicted protein-folding provides new insights in biomedical applications and AI-assisted discoveries of new drugs are helping with pharmaceutical research and development. But acceleration of innovation in a broad range of fields is not a given.

Personal computers and the internet were once hailed as ultimate research tools, enabling unprecedented access to information, aiding statistical analysis and fostering global collaboration. However, despite an initial productivity boom around the turn of the millennium, computers failed to unlock significant breakthroughs in research.

Maintaining Moore's Law – the observation that the number of transistors on a silicon computer chip would double every two years – now requires 18 times more researchers than in the early 1970s, reflecting a broader trend of diminishing returns across the US economy.

One explanation is that start-ups, which traditionally drive innovation, face increasing barriers to entry. Larger companies have scale in both data and computing power that could lead to excessive concentration of AI development. But studies also document a substantial rise in the concentration of patenting among superstar firms, and this defensive patenting helps protect their technological lead.

There was a time when start-ups eagerly entered industries with high profitability. However, by the late 1990s, profits no longer predicted the entry of new firms. Whether due to defensive patenting or lobbying for restrictive regulation by incumbents, they are finding it harder to enter the market.

This decline in dynamism is particularly concerning, given that start-ups play an outsized role in bringing new research to market. And start-ups are also more likely to launch products that create new industries and new tasks for labour. Revitalizing competition to spur innovation is thus a key task facing policy-makers going forward.

Third, since translating scientific breakthroughs into marketable products can take many years, or even decades, productivity growth in the next decade will largely depend on how quickly AI is adopted.

On a positive note, research indicates that the time lag between when different countries adopt new technologies has decreased significantly over the centuries.

For example, while steamships took 131 years to reach Indonesia after being introduced in England, the personal computer arrived in Vietnam just 11 years after its invention in the US. The trouble is that the degree to which these technologies spread across the population after their initial adoption has declined over time.

The question is, how can policy-makers boost the intensity of AI diffusion and adoption around the world? This will require complementary investments.

First of all, foundational AI preparedness will be vital, requiring investments in digital infrastructure and digital data and skills, particularly in emerging market and developing economies. In fact, when it comes to upskilling and reskilling, AI can be part of the solution by offering improved training options for workers—provided the necessary infrastructure is in place.

Secondly, pre-AI studies of digital adoption suggest that this broad diffusion pattern is not guaranteed, and divergence is possible or even likely. For AI to achieve full economic impact over time, it must be accessible throughout the economy, to companies large and small, in rural as well as urban areas.

Finally, effective AI adoption will require appropriate regulations to manage AI risks (without harming innovation by disadvantaging startups), and creating effective guardrails around AI to bolster confidence in its use.

AI holds the potential to boost productivity and lift growth prospects. A comprehensive policy approach is needed to harness its full benefits and broaden the gains for all.

The authors, Carl-Benedikt Frey, Era Dabla-Norris, Rob Hornby and Laura D’Andrea Tyson, are members of the Global Future Council on the Future of Growth.