Could artificial intelligence really boost labour productivity?

Promising signs of labour market boosts

Artificial intelligence has long been theorised as a cure for the West’s ailing productivity growth. As the McKinsey Global Institute has shown, workplace productivity growth has been stagnant for about 40 years. But the discourse on the productivity effects of AI has been almost entirely speculative. Until recently, evidence of large-scale exposure to AI was absent from data.

In the last 10 years, scholars such as Daron Acemoglu and David Autor have shown how computerisation has led to an intra-firm re-allocation of skills and corresponding wage premiums via skill-biased technological change. Acemoglu and Pascual Restrepo showed that the equilibrium impact of industrial robots between 1990 and 2007 in the US reduced wages and employment in the local labour market.

However, the extent of these effects, as well as the question of whether they boosted productivity at the industry level, is a matter of debate. Thomas Phillippon, a French economist, has suggested that some of these productivity gains led to organisational splits between management and worker wages.

Another hypothesis is that general-purpose technologies – such as the steam engine or electricity – need time, investment and complementary innovations to convert an accumulation of intangible capital, like AI technology, into a productivity enhancer. This means that we should expect a stagnation period before a productivity boom.

Other possible explanations may be that we are not measuring productivity in the right way, as macroeconomic measures don’t usually account for intangible capital. It could also be that, because only a few firms are currently capturing most of the benefits from AI, this is not going to be reflected by macroeconomic measures.

Large language models are making an impact

Within generative AI, however, large language models are starting to have positive effects on productivity growth. These include faster and higher-quality creative output and professional writing, and extend to writing code and creating better advertisements. It should be noted that not all LLMs improved performance, and that some studies showed a larger boost for AI systems that allowed users to contribute with their own skills and knowledge.

LLMs seem to be spurring productivity growth for white-collar jobs. Several field studies showed that LLMs can improve the speed and quality of text-based businesses, such as customer service centres and consultancy. GenAI can also be employed to provide feedback to improve employees’ efficiency. AI-based performance-tracking systems and coaches have been shown to have a positive impact on employees’ performance.

Author and academic Erik Brynjolfsson used data from over 5,000 agents and found that AI assistance improves productivity by 14% in white-collar occupations. This especially benefitted novice and low-skilled workers with minimal impact on experienced and highly skilled workers. Brynjolfsson found that GenAI disseminates tacit knowledge, improves customer sentiment, reduces managerial intervention requests and enhances employee retention.

And yet, there are plenty of reasons to remain sceptical of AI’s productivity-boosting effects.

Too early to tell

First, it’s not clear how generalisable the results from the preliminary studies are. Most of the reports show an increase in the productivity of low- and medium-skilled workers, in line with trends from previous technological shocks. Brynjolfsson suggests that this may be due to the LLM encoding implicit expert knowledge from the most experienced and skilled workers.

Another study found an opposite effect of LLMs, with AI coaches enhancing the performance of high-skilled entrepreneurs and decreasing that of low-skilled entrepreneurs. We are only seeing the beginnings of AI use cases, so it’s hard to use preliminary evidence of generative AI effects as a barometer for broader impacts.

Second, some scholars have suggested that the measures of increased productivity correspond to a displacement of the workforce (automation) and realignment of intra-firm level returns, rather than an enhancement of the workers’ capacities (augmentation). Acemoglu has shown that some recorded increases in the productivity of US manufacturing industries corresponded to a decline in employment that was faster than the increase in output.

The issue of displacement effects is the third point. Acemoglu and Restrepo argue that displacement effects caused by automation (the substitution of labour with capital), together with a lack of new jobs for the displaced workers (or ‘reinstatement effect’) is one of the causes of the current productivity stagnation. This fits with evidence suggesting that automation may be increasing wage inequality while causing marginal gains in productivity. They suggest that this may be due to a lack of investment and low expected returns for innovation, which could generate reinstatement effects. This may lead actors towards marginal development and result in slower production growth.

Finally, the employment of ‘so-so technologies’ – technologies that don’t increase productivity compared to labour – could also explain the current stagnation. It is possible that AI will behave similarly to these early digitalisation technologies in spurring such effects.

So, what to make of existing evidence? While scholarship points to promising signs that AI will boost productivity, as has been predicted for years, it remains too early to tell exactly what this will look like and the technology is changing rapidly. However, we may be looking at a structural transformation of the modern economy, via productivity improvements, akin to the initial emergence of computers.

A challenge for policy-makers will be to navigate the potentially disequalising effects of AI, spurred through labour displacement. Such shocks can have disastrous economic and social effects, while also making productivity growth harder to sustain. We are once again witnessing a race between technology and public sector adjustment.

Julian Jacobs is Senior Economist, Digital Monetary Institute, OMFIF and Francesco Tasin is Research Assistant at the University of Oxford’s Future Impact Group.

Join OMFIF at the ‘Artificial intelligence as financial regulation technology’ roundtable on 5 March.

Join Today

Connect with our membership team

Scroll to Top