Pascual Restrepo says AGI will not automate most jobs because they are not worth replacing.
JAKARTA – The conventional fear that artificial general intelligence (AGI) will take over all human jobs has now been challenged by new research.
Pascual Restrepo, Associate Professor of Economics at Yale University who previously conducted research with Nobel laureate Daron Acemoglu, argues that most human jobs will not be automated in the AGI era.
In a working paper published by the National Bureau of Economic Research (NBER) titled “We Won’t Be Missed: Work and Growth in the AGI World”, Restrepo concludes that AGI does not make human skills obsolete but instead reassesses their value.
The main reason is not that artificial intelligence is incapable, but that most human jobs are considered insufficiently important or not worth the computational resources required to replace them.
As quoted by finance.yahoo.com (04/04/2026), Restrepo divides jobs in an AI‑driven economy into two categories.
The first is bottleneck work, which is essential for economic growth, such as energy production, infrastructure maintenance, scientific advancement, national security, reducing existential risks, surviving asteroid threats, and mastering fusion energy. These high‑level tasks will be fully automated by computation.
The second is supplementary work, such as arts, crafts, customer support, hospitality, design, academic research, and even professional economics. Because the economy can continue expanding without these roles, AI will ignore them due to the high cost of replicating human social elements.
In the long term, total computational resources in the economy are projected to reach 10⁵⁴ floating‑point operations per second (flops), making the combined computing power of all human brains, only around 10¹⁸ flops, economically insignificant.
Although humans are expected to retain supplementary jobs, Restrepo warns that surviving automation is not the same as benefiting from economic growth. In a post‑AGI world, wages will detach from Gross Domestic Product (GDP), and labour’s share of GDP will shrink towards zero as most income flows to entities that own computational resources.
This concern over wealth distribution aligns with BlackRock CEO Larry Fink’s annual letter, which warned that AI will concentrate wealth in the hands of a few.
Fink noted that the top 1 percent of US households now hold more wealth than the bottom 90 percent combined, while 40 percent of Americans have little or no exposure to capital markets.
Fink suggested solutions such as tokenisation and expanding pension investment, while Restrepo proposed universal income or treating computation as a public resource similar to land.
Regarding the transition to such a future, Restrepo identifies two modes: compute‑binding transitions, where adoption occurs gradually, and algorithm‑binding transitions, which leap suddenly and trigger extreme inequality. The second phase is currently unfolding in the United States in 2026, where technical workers are experiencing a surge in wage premiums.
Data from the AI recruitment platform Skillit shows that data‑centre construction workers now earn an average of US$81,800 per year, 32 percent higher than typical projects. Some electricians are earning as much as US$260,000 annually, with electrical work accounting for 45 to 70 percent of total data‑centre construction costs.
The United States is projected to need 300,000 new electricians in the next decade to replace the 200,000 expected to retire.
In aggregate, Restrepo argues that total labour income in a post‑AGI world will still be higher than previous baselines, meaning the population will not become poorer.
However, the title of his paper carries a sharp existential message. If today’s economy would collapse if half the workforce did not show up, in an AGI world the disappearance of human workers would not be missed and would not hinder economic activity.
Ultimately, for most workers, the question is not whether AI will take their jobs, but the realisation that their jobs may never have been important enough to replace. (DH/LM)
as a preferred source on Google