Login 

Share

Letter from America: AI and economic growth: Dreams versus theory

Daron Acemoglu questions optimism about the productivity gains from AI

Artificial intelligence (AI) has captured imaginations, with the promise of human-like machines and boundless productivity growth. Some of these predictions may have come closer to reality over the last year, with spellbinding advances in generative AI and large language models that produce text, information and images – and even Shakespearean sonnets – in response to user prompts. ChatGPT, originally released on November 30, 2022, soon became the fastest-spreading tech platform in history, with an estimated 100 million monthly users only two months after launch.

Generative AI will have implications for the macroeconomy, productivity, wages and inequality, but all of them are very hard to forecast. This has not stopped economists, business organizations and other pundits from trying to do so.

The macroeconomic impact of generative AI will be intertwined with the productivity gains it will create. Some experts believe that truly transformative effects, sometimes equated with “singularity” or artificial general intelligence (AGI), may be around the corner.

Other forecasters predict big (if somewhat more modest) effects on output. Goldman Sachs predicts a 7% increase in global GDP, equivalent to $7 trillion, and a 1.5 percentage point increase in the productivity growth rate over a ten-year period. Recent McKinsey Global Institute forecasts suggest that generative AI could offer a boost as large as 15–25% on top of the earlier estimates of productivity growth from increased work automation. They reckon that the total impact of AI and other automation technologies could produce a 3-4 percentage point rise in average annual productivity growth in advanced economies over the coming decade, with as much as 0.7% attributable to generative AI alone.

There are several sources of productivity gains. If generative AI produces a truly transformative breakthrough – the hope of believers in rapid progress towards AGI – then the pace of innovation can become much faster. New crystal structures discovered by the Google subsidiary DeepMind and recent neural-network enabled advances in protein folding suggest that AI may already be improving our capabilities for scientific discovery. Nevertheless, the focus of many AI discussions is on the gains from automation: generative AI or other AI modalities can now perform tasks (such as clerical functions, text summary, data classification, and pattern recognition) more cheaply, and even sometimes more accurately, than humans are able to do.

We can gain some insight about how large the productivity gains from automation are likely to be – and how workers’ wages may change – by considering a well-known formula in growth economics. Originally stated by Charles Hulten (and thus referred to as Hulten’s theorem), this formula estimates the aggregate economic impact of “total factor productivity” growth at the sectoral level. As with all results, this theorem has a few caveats; most importantly, the formal theorem assumes competitive product and factor markets and applies only to small changes. Nevertheless, the formula is useful even if we depart from these assumptions, as it provides a major component of the aggregate productivity effects from technological change under more general conditions.

In the context of artificial intelligence – or more generally of automation – it may be useful to think in terms of the tasks that workers perform across a range of occupations and industries, rather than sectors. Then Hulten’s theorem links aggregate output gains to productivity improvements (or cost savings) from AI at the task level.

Suppose, additionally, that the main effect of AI will come through automation. Then Hulten’s theorem implies:

              aggregate productivity effect of AI =

              fraction of tasks that are automated  average cost savings from automation

Average cost savings (or productivity gains at the task level) are, of course, difficult to estimate. But several careful studies offer some guidance, using randomized controlled trial methods to measure AI-driven productivity gains for narrowly defined tasks. Work by Shakked Noy and Whitney Zhang looks at the impact of access to ChatGPT on simple or mid-level professional writing tasks (such as summarizing documents or writing routine grants or marketing material). They estimate that access to the generative AI agent reduces time taken by about 40% on average, and also leads to better quality output.

Erik Brynjolfsson, Danielle Li and Lindsey Raymond study the impact of a generative AI conversational assistant on customer service output (number of cases successfully resolved per hour) and find a 14% improvement on average. Both studies estimate that the effects are concentrated among less experienced or less expert workers – an intriguing pattern which I will return to later.

The two papers together imply an approximate 27% labour cost savings from automation for a range of white-collar tasks. Combining this with a labour share of 60%, we  arrive at average cost savings of about 16%. This is a nontrivial productivity boost, but not out of line compared to earlier automation technologies. Industrial robots are also estimated to have reduced costs by about 30% for the tasks that they were deployed to perform.

What about the fraction of tasks to be automated? This may be even harder to ascertain. Based on work by Daniel Rock and collaborators, we can estimate that about 20% of current tasks in the US economy can be ultimately performed by large language models and related technologies, such as computer vision. Based on estimates for computer vision, an upper bound on how much of this can happen within 10 years is about 23%. If we take 20% × 23% = 4.6% as an upper bound for the next 10 years (an aggressive assumption), then Hulten’s theorem implies an aggregate productivity gain of no more than 16% × 4.6% = 0.7% (or an annual average productivity gain of less than 0.07%) from generative AI and related technologies by 2034 – much less than the optimistic forecasts of AI’s macroeconomic impact. The true number is likely to be lower, because some of the yet-to be automated tasks may be quite hard for generative AI, leading to lower cost savings.

This underscores a general point: it is difficult to get big productivity gains from automation. Automation substitutes somewhat cheaper capital and algorithms for tasks that labour used to do – but in many cases, labour used to perform those tasks reasonably well, so the room for further productivity improvement is often limited.

So where could bigger gains come from? My work with Pascual Restrepo suggests that we should look at new tasks, meaning new creative activities or complex functions for human workers. A large part of employment growth in the US comes from occupations that have added significant new job tasks or “job titles”.

Unfortunately, it is even more difficult to forecast how many new tasks generative AI will create and how much these new tasks will add to productivity. Although the potential for new task creation using AI is present, the current focus in the industry may be too much on automation, and potential gains from new tasks may not be realized.

Who will benefit from the possible AI-driven productivity gains? The fact that AI tools enabled workers with less experience and less expertise to score the largest productivity boost in the studies mentioned above – bringing them up-to-par with other participants who were already well-equipped for the tasks – may suggest that generative AI could play an equalizing role. This is what The Economist magazine concluded late last year, predicting that generative AI will contribute to a “blue-collar bonanza”.

Here, too, there are lessons from economic theory. At first, one might think that a 16% improvement in productivity for workers will translate into 16% higher wages. But this is not so. Simply improving the productivity of workers in the tasks that they already perform will tend to have much smaller effects on wages, because the increased output from those tasks will also tend to drive prices down. This is the reason why, in general, labour-augmenting improvements in technology do not much impact the labour share of income. It is also the reason any gains will be split with firms and capital owners.

In fact, the apparent larger productivity benefits for less expert workers mentioned above may not much help low-wage workers. If labour’s productivity increases in a set of tasks, such as writing, then many workers performing these tasks will be displaced. When that happens, those who had greater expertise in the tasks they were previously performing may both be the ones keeping their jobs and may also be better placed to translate their skills into occupations which have not been equally impacted by AI. Hence, for wages to increase – including the wages of lower-pay workers – new tasks are vital. Additionally, for low-paid workers to benefit from technological changes, institutional factors like collective bargaining, minimum wages and limits to employer monopsony power may turn out to be important as well.

The recent advances in generative AI technologies are impressive in many ways. Nevertheless, the predictions of strikingly large productivity gains within the next ten years and major equalizing effects do not seem as likely when examined from the perspective of basic economic theory.

Daron Acemoglu, 20 March 2024

References

Acemoglu, D. (2023). The Letter from America: Will AI remake or destroy education? RES Newsletter, October.

Acemoglu, D. and Autor, D. (2011). Skills, tasks and technologies: Implications for employment and earnings. In O. Ashenfelter and D. Card (eds.) Handbook of Labor Economics, Volume 4B, Elsevier, Amsterdam.

Acemoglu, D., Autor, D. and Johnson, S. (2023). Can we have pro-worker AI? Choosing a path of machines in service of minds. MIT Policy Memo, revised September 19, 2023.

Acemoglu, D. and Johnson, S. (2023). Power and progress: our thousand-year struggle over technology and prosperity. PublicAffairs, New York.

Acemoglu, D. and Restrepo, P. (2018). The race between man and machine: Implications of technology for growth, factor shares, and employment. American Economic Review, 108(6), 1488-1542.

Acemoglu, D. and Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3-30.

Acemoglu, D. and Restrepo, P. (2020). Robots and jobs: Evidence from US labor markets. Journal of Political Economy, 128(6), 2188-2244.

Acemoglu, D. and Restrepo, P. (2022). Tasks, automation, and the rise in U. S. wage inequality. Econometrica, 90, 1973-2016.

Autor, D., Chin, C., Salomons, A. M. and Seegmiller, B. (2022). New frontiers: The origins and content of new work, 1940-2018. NBER working paper no. 30389, August.

Baqaee, D. R. and Farhi, E. (2020). Productivity and misallocation in general equilibrium. Quarterly Journal of Economics, 135(1), 105-163.

Brynjolfsson, E., Li, D. and Raymond, L. R. (2023). Generative AI at work. NBER working paper no. 31161, April.

The Economist (2023). A new age of the worker will overturn conventional thinking. The Economist, 30 November 2023.

Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2023). GPTs are GPTs: An early look at the labor market impact potential of large language models. arXiv preprint, https://doi.org/10.48550/

Goldman Sachs: https://www.goldmansachs.com/intelligence/pages/generative-ai-could-raise-global-gdp-by-7-percent.html

Hulten, C. R. (1978). Growth accounting with intermediate inputs. Review of Economic Studies, 45(3), 511-518.

Jumper, J., Evans, R., Pritzel, A. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583-589.

Kurzweil R. (2005). The singularity is near: when humans transcend biology. Viking, New York.

McKinsey Global Institute report: The economic potential of generative AI: The next productivity frontier. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-AI-the-next-productivity-frontier#introduction

Noy, S. and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381,187-192.

Reuters: https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/

Rock, D. (2023). Natural questions about Artificial Intelligence. RES Newsletter, October.

Svanberg, M., Li, W., Fleming, M. Ghoering B. and Thompson, N. (2024). Beyond AI exposure. Which tasks are cost-effective to automate with computer vision? https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4700751

Time report: https://time.com/6340681/deepmind-gnome-ai-materials/

Trammell, P. and Korinek, A. (2023). Economic growth under transformative AI. NBER working paper no. 31815, October.