AI will replace about 5% of what humans do within 10 years. Daron Acemoglu, a Nobel-winning economist at MIT, made that estimate. About two years on, in an essay for The Humanist Review, a magazine from Microsoft AI, he says it has held up reasonably well.
An October 2024 Pew Research Center survey he cites found 52% of American workers worried about AI’s future impact at work. He calls that worry misplaced and says the fear pushes people to build and use AI the wrong way.
So far, AI does not show up in productivity statistics, he says, and studies he mentions found little productivity gain at most companies adopting it. For the whole economy, he forecasts roughly 1.5% added to GDP over 10 years. AI models have improved faster than he expected, yet few easy-to-adopt applications have changed how firms produce goods and services.
Part of the 5% estimate rests on how fast earlier technologies such as electricity spread. New York and London had central power stations in the early 1880s, yet as late as the 1920s only about half of factories and homes used electricity. He expects AI to spread even more slowly, because companies must reorganize around it. Firms must decide which tasks go to AI and which stay with workers, while specialists learn to work with it and managers set goals for it.
Reliability adds to the delay, he says. Automating a whole job means handling every task in it, and AI is not yet reliable or versatile enough to do that. Even 99% accuracy falls short, because the last 1% is the hard part.
Where AI falls short on a task, he argues, it has to work alongside the people who do it. He wants developers to build AI that makes people better at their jobs, which he calls pro-worker AI. Existing models already make pro-worker AI possible, he says, and he expects it to beat attempts at full automation on productivity.
One obstacle to pro-worker AI is the tax code. Research he co-authored in 2020 found the US taxes labor at an effective rate above 28.5%, against about 5% on money invested in equipment and software. The difference works like a subsidy for automation, he explains, and discourages firms from hiring and training people.
Acemoglu calls the 5% figure a rough guess and says nobody knows how fast AI models will advance.
