Inflation, wages, public deficit: generative artificial intelligence (GAI), ChatGPT style, could, in theory, solve everything. The acceleration of productivity gains linked to the use of this technology would, in fact, have crucial economic and financial consequences. By reducing companies’ production costs, it would reduce inflation and interest rates. By increasing the efficiency of workers, it would increase their wages. By improving economic growth and therefore tax revenues, it would resolve a large part of our public deficits.
But will this miracle happen? The debate on the link between AGI and productivity has just taken a new turn with the publication, in mid-May, of an article entitled “The simple macroeconomics of AI”, written by one of the most respected economics professors renowned in the world, Daron Acemoglu, currently stationed at the Massachusetts Institute of Technology (MIT). This text had the effect of a cold shower for the techno-optimists – including the author of these lines. Let’s summarize the argument.
Acemoglu points out that AGI can generate productivity gains in two ways: by completely automating easy tasks or by increasing the productivity of humans for complex tasks. According to him, the total productivity gain would nevertheless be small in the end, of the order of 0.5% within ten years, with no impact on GDP growth. Why is this result so disappointing? Firstly, tasks that can easily be automated – verifying a person’s identity, summarizing a text, writing a press release, etc. – are in reality few in number. And single-task jobs, which can be completely automated – there has been a lot of talk in recent weeks about series or film dubbers – are even less widespread.
Second, the adoption of AGI by companies still seems slow, in particular because it entails complex and costly reorganizations. In fact, and in France in particular, the phenomenon seems faster than Acemoglu suggests; I will come back to this in a future column. Third, it is not known whether the release of working time permitted by the AGI will be reallocated to productive work. It can be for activities with low added value, or even for leisure time. Fourth, AGI intended for complex tasks – such as verifying the reliability of information – is trained on human data, itself heterogeneous and sometimes false. Its progress therefore tends to plateau.
From Patrick Sébastien to Baudelaire
Acemoglu’s article will reassure those who fear that the speed of adoption of AI will change work and jobs at a speed incompatible with the capacity of our societies to adapt, or that it will explode inequalities. However, it would be wrong to see this as a victory for the techno-pessimists. Already, AGI continues to progress rapidly, as evidenced by the dramatic performance differences between different versions of ChatGPT. Then, this analysis highlights an essential point: the economic and social impact of AGI will be the consequence of the uses that users and businesses have of it.
Asking ChatGPT to transform a Patrick Sébastien song into a Baudelaire poem is fun, spectacular, but destructive of productivity through wasted time and wasted energy. As for the production and propagation of fake news in this way, they can generate colossal costs for our societies. Conversely, if companies quickly switch to IAG to train their employees, increase their technical or linguistic skills, make them more productive in their current tasks and redeploy the time saved towards commercial action, customer service or research and development of new products, this revolution could generate massive productivity gains. Ultimately, it’s the trade-off between slow and thoughtless use on the one hand or fast and intelligent use on the other that will make the difference. The choice is collectively in our hands. Let’s be adults.
.